A Theory of Language and Information A Mathematical Approach

A Theory of Language and Information
A Theory of
Language and Information
A Mathematical Approach
ZELLIG HARRIS
CLARENDON PRESS • OXFORD

1991
\ 1 ‘r to
Oxford University Press, Walton Street, Oxford ox2 6dp

Oxford New York Toronto
Delhi Bombay Calcutta Madras Karachi
Petaling Jaya Singapore Hong Kong Tokyo
Nairobi Dar es Salaam Cape Town
Melbourne Auckland
and associated companies in
Berlin Ibadan
Oxford is a trade mark of Oxford University Press
Published in the United States

by Oxford University Press, New York
© Zellig Harris 1991
All rights reserved. No part of this publication may be reproduced,

stored in a retrieval system, or transmitted, in any form or by any means,
electronic, mechanical, photocopying, recording, or otherwise, without
the prior permission of Oxford University Press
British Library Cataloguing in Publication Data

Harris, Zellig S. (Zellig Sabbettai) 1909-
A theory of language and information: a mathematical
approach.
1. Mathematical linguistics
I. Title
410
ISBN 0-19-824224-7
Library of Congress Cataloging in Publication Data

Harris, Zellig Sabbettai, 1909-
A theory of language and information: a mathematical approach/
Zellig Harris.
Includes index.
1. Linguistics. 2. Language and languages. 3. Mathematics.
4. Information theory.
P121.H354 1990 415' ,01'51—dc20 90-39472
ISBN 0-19-824224-7
Set by Hope Services (Abingdon) Ltd.

Printed in Great Britain by
Biddles Ltd
Guildford & King’s Lynn
Preface
This book presents several interrelated matters: a formal theory of syntax;

an example of how something in the real world—the basic structure of talk
and writing—can be represented as a mathematical object; the intrinsic
relation between language and information, and the structure of information of
individual sciences; and, lastly, indications of how language is a self¬
organizing and evolving system.
The theory of syntax is stated in terms related to mathematical Information
Theory: as constraints on word combination, each later constraint being
defined on the resultants of a prior one. This structure not only permits a
finitary description of the unbounded set of sentences, but also admits
comparison of language with other notational systems, such as mathematics
on the one hand and music on the other.
Since frameworks arise not arbitrarily but from how one considers the
data, this book assesses first what methods are relevant to the data of
language (ch. 4), and then what kind of theory can be reached from those
methods. For reasons given in Chapter 4, one cannot rely on episodic
examples; it is necessary to survey the full range of data in a language, in a
way that establishes their regularities. Hence the analysis presented here
draws upon a companion volume (the writer’s Grammar of English on
Mathematical Principles, Wiley-Interscience, New York, 1982), which covers
in some detail the data of one language in terms of the present theory;
references to this grammar (indicated by G) are given throughout the
present book.
The analysis in this book thus arises from the results obtained by particular
methods of description and decomposition of the word combinations found
in a language. The intent is not only to understand language and information,
but also to see what the methods used here can show about how structure
comes to be.
One explanatory note may be needed here. In characterizing (‘explaining’)
the structure of word combinations, a method is used whereby one com¬
bination is derived from another (its source). In some cases the derivation
for a given form involves reconstructing a source which is not actually said in
the language. These sources, however, are never abstract models. They are
word combinations made in accordance with the actual grammar of the
language, and from which the given form would be derived by the existing
VI Preface
derivational relations of the language. That they are not actually said may be
due to their complexity or length; in many cases it would be impossible to
formulate a regular grammar of the language that would exclude them as
ungrammatical. Therefore the unsaid sources can be counted upon as
derivational sources for the given word combinations.
Since some of the chapters may be read independently of the others, data
relevant to several different issues have been repeated in the respective
different chapters. A reader who seeks only the more general formulations
can safely omit various sections, chiefly 1.1.1,3.2.2-3.2.5, 3.3.2.1-3, 3.4.2-
3.4.3, 5.3.2-5.3.4, and Chapters 7 and 8.
It may be in order to state what language-acquaintance underlies the
results presented here. The statements in this book have been based on
detailed grammatical formulations of several European and Semitic languages,
and on outlinings of the structures in a variety of American Indian languages
(especially Algonquian, Athapaskan, Siouan, Iroquoian) and Eskimo, and
also in some African languages (especially Swahili) and some Dravidian.
The statements have been checked against existing grammatical descriptions of
a sampling of languages: Basque; languages of ancient Asia Minor (where I
had the advantage of information from Anna Morpurgo Davies); Sumerian;
some languages of the following groups: Australian, Austronesian, Papuan
in New Guinea, Micronesian; and finally, Chinese and Korean. The variety
of structures in the languages surveyed is of course vast; but it was possible to
see that the dependence relation of operator on argument, on which rests
the syntactic structure presented here, is a possible underlying description
for them. There was no possibility of determining whether some of the other
constraints, defined here upon the basic dependence, hold for all the
languages.
The work here starts off from the distributional (combinatorial) methods
of Edward Sapir and of Leonard Bloomfield, to both of whom I am glad to
restate my scientific and personal debt. I am also indebted as always to
many enlightening and felicitious conversations with Henry Hiz, Henry
Hoenigswald, M. P. Schutzenberger, and Maurice Gross, and to very
valuable critiques and suggestions by Andre Lentin and Danuta Hiz.
Contents
List of Figures xii
I. INTRODUCTION
1. Overview 3
1.0. Introduction 3
1.1. A theory in terms of constraints 5
1.1.0. The major constraints 5
1.1.1. Example: The set-vs.-member connectives 9
1.1.2. Example: The passive construction 12
1.2. Mathematical properties 15
1.3. Science sublanguages 18
1.4. Information 21
1.5. The nature of language 24
2. Method 30
2.0. Introduction 30
2.1. The issue to be resolved: Departures from equiprobability 30
2.2. The relevant method: Least redundancy in description 31
2.3. Simplicity 36
2.4. Can meaning be utilized? 40
2.5. Properties of the data 43
2.5.0. Introduction 43
2.5.1. Unbounded language, finite speakers 43
2.5.2. Discrete elements 43
2.5.3. Linearity 44
2.5.4. Contiguity 44
2.5.5. Variety within a class 45
2.5.6. Synchronic single-language coexistence of forms 46
2.5.7. Diachronic change 47
2.6. From method to theory 48
viii Contents
II. THEORY OF SYNTAX
3. A Theory of Sentences 53
3.0. Constraints on equiprobability 53

3.1. Zero vs. non-zero likelihood 54
3.1.1. Dependence 54
3.1.2. Dependence on dependence 58
3.1.3. Partial order 59
3.1.4. Predication 60
3.2. Inequalities of likelihood 61
3.2.1. Likelihood constraints 62
3.2.2. Likelihood types 65
3.2.3. Reconstructed likelihoods 68
3.2.4. Likelihood restricted to zero 73
3.2.5. Word subsets 74
3.2.6. Frequency; low information 76
3.3. Reduction 79
3.3.1. The reduced shapes 80
3.3.2. Likelihood conditions 83
3.3.2.0. Introduction 83
3.3.2.1. Broad selection 85
3.3.2.2. High frequency 87
3.3.3. Required reductions 93
3.3.4. How reduction takes place 94
3.4. Linearizations 97
3.4.1. Normal linearization 98
3.4.2. Alternative linearizations 99
3.4.3. Permutation 100
3.4.4. Interruption: Modifiers 101
4. Language Structure: The System Created by the Constraints 105

4.1. The base set of sentences 106
4.2. Sentences derived from the base 109
4.3. Systemic properties following from the constraints 114
4.4. Properties following from additional conditions 117
5. Metalinguistic Apparatus within Language 123

5.1. The metalinguistic capacity 123
5.2. Metalinguistic sentences 124
Contents ix
5.3. Reference
128
5.3.0. Introduction
128
5.3.1. Repetition
128
5.3.2. Locating the antecedent 129
5.3.3. Sameness
133
5.3.4. Demonstratives: Reference to the external situation 133
5.3.5. Summary: Reference from cross-reference 134
5.4. Use/mention 135
5.5. Performatives 137
5.5.1. Isay
138
5.5.2. Imperative and interrogative 139
5.5.3. Adverbs on say. And, or, not, any, tense 141
6. On the Mathematics of Language 145

6.0. Why a mathematical approach? 145
6.1. Language properties having mathematical relevance 147
6.2. Increasing the amenability to mathematical description 151
6.3. Mathematical structures in language 153
III. GRAMMATICAL ANALYSIS
7. Analysis of Sentences 163

7.0. Classes and sequences 163
7.1. Phonemic distinctions 164
7.2. Phonemes and letters 166
7.2.0. Phonemes 166
7.2.1. The phonemic system 167
7.2.2. The phonemic alphabet 168
7.3. Morpheme and word boundaries 171
7.4. Words and morphemes; complementary alternants 175
7.5. Word classes 180
7.6. Word-class sequences: Continuous and contiguous strings 182
7.7. Sentence boundaries 185
7.8. Sentence structure 189
8. Analysis of the Set of Sentences 196

8.1. Sentence transformations 198
8.1.0. Word-class grammars and word-choices 198
8.1.1. The transformational method 199
8.1.2. Sentence acceptability grading 200
8.1.3. Transformational Analysis 207
8.1.4. Survey of English elementary transformations 210
8.1.5. Algebra of the set of sentences under transformations 214
X Contents
8.1.6. Structures in respect to transformations 219

8.1.6.0. Introduction 219
8.1.6.1. Unambiguous subsets of sentences and decompositions 219
8.1.6.2. The distinguishing partial sentence 224
8.1.6.3. Graphs of the set of sentences 225
8.1.6.4. Language-pair grammars 227
8.1.7. Transformations as a theory of syntax 228
8.2. Operator theory 231
8.2.0. Motivation 231
8.2.1. Operator-argument analysis 234
8.2.2. Adequacy of the analysis 238
8.3. Sentence semilattices 247
8.4. Individual language structures 248
IV. SUBSETS OF SENTENCES
9. Sentence Sequences 255
9.0. Motivation: constraints between sentences 255

9.1. Conjoined sentences 256
9.2. Metasentences 257
9.3. Discourse analysis 259
10. Sublanguages 272

10.1. Grammar-based sublanguages 273
10.2. The metalanguage 274
10.3. Subject-matter sublanguages 278
10.4. Science sublanguages 283
10.5. Formulaic languages of sciences 290
10.5.1. A distinct linguistic system 291
10.5.2. Formulas of information 294
10.5.3. Boundaries of fields 298
10.5.4. Information and argumentation 299
10.5.5. Notation 303
10.6. Mathematics as a linguistic system 304
10.7. Non-linguistic systems; music 308
V. INTERPRETATION
11. Information 321

11.1. How words carry meaning 322
11.2. Information in sentences: Predication 332
Contents xi
11.3. How language carries information 334,

11.3.1. Where the information is located 334
11.3.2. Successive stages of information-bearing constraints 335
11.3.3. Non-informational constraints 338
11.3.4. The meanings of word relations 341
11.4. Form and content 342
11.5. A theory of language or a theory of information? 345
11.6. Whence the power to‘mean’? 347
11.6.1. Co-occurrence and the perceived world 347
11.6.2. Language and the perceived world 350
11.7. The structure of information 354
12. The Nature and Development of Language 357

12.1. Overview of language structure 357
12.2. Structural considerations from language change 362
12.3. Formation and development of language 365
12.3.0. From descriptive order to time order 365
12.3.1. Pre-syntactic speech 367
12.3.2. Early development of sentences 368
12.3.3. Further development of syntax 371
12.3.4. The limits of language 374
12.3.5. From use to constraints: Institutionalization 377
12.4. The nature of language 382
12.4.1. The language system; language and thought 382
12.4.2. Systematizing factors 387
12.4.3. A self-organizing system 391
12.4.4. An evolving system 394
12.4.5. What is needed for language to exist? 402
Index 407
List of Figures
7.1. Word boundaries in a sentence 172

7.2. Morpheme boundaries in words 173
7.3. Cycling automaton for an English sentence 194
8.1. Semilattice of an English sentence 218
8.2. Semilattice of a Korean sentence 220
8.3.1. Decomposition lattice 1 of an ambiguous sentence 222
8.3.2. Decomposition lattice 2 of an ambiguous sentence 222
8.4. Decomposition 1 of an ambiguous sentence 223
8.5. Decomposition 2 of an ambiguous sentence 224
9.1. Analysis of a text 262-267
All figures are from Z. Harris, Mathematical Structures of Language, Wiley-

Interscience, New York, 1968.
I
Introduction
1
Overview
1.0. Introduction
This book presents in the first place a formal theory of syntax (chs. 3, 4), in
which a single relation on word-sequence gives to the set of these sequences
the structure of a mathematical object, and produces a base set of sentences
which carry all the information that is carried in the language. Within this set
of sentences there acts a set of reductions which change the shape of words
but not the information in the sentences, thereby producing the remaining
sentences of the language. Prior material to this theory is an analysis of how
words are constructed from sounds (7.1-4), and preliminary considerations
of how words combine into sentences (7.5-8, 8.1). Following upon the
establishment of sentences, we can go on to an analysis of how sentence sets,
especially sentence-sequences, are constrained in longer utterances and in
discourses (ch. 9), and finally in sublanguages (ch. 10). The relation of
syntactic structure to information and to the development of language is
then discussed in Chapters 11, 12.
The general structure of language consists then of: (1) the construction of
words from sounds, in a way that does not affect the structure of syntax;
(2) the construction of sentences from words, by a single partial-ordering
relation; (3) the transformation of sentence shapes, primarily by a process of
reduction. The independence of word-sounds from sentence construction
leaves room for various languages to have sound-syntax correlations:
particular sound similarities among words having the same sentential status.
This is the chief basis for additional grammatical constructions (beyond
those listed above), such as conjugations and declensions.
In the chain of operations, from individual sounds to complete discourses,
each operation is defined on the resultant of prior operations, and each word
that enters into the making of a sentence enters contiguously to the con¬
struction created thus far. Because of this, every construction (aside from
statable exceptions) can be analyzed in an orderly way in terms of these
operations as first principles.
It is thus not the case that language has no precise description and no
structural relation to meaning, but rather that these properties are reached
only via particular methods. These methods arise from a crucial property,
namely that language has no external metalanguage in which one could
4 Introduction
define its elements and their relations (2.1). Hence, the elements can be
established only by their redundancies in the set of utterances, i.e. by the
departures from equiprobability in the observable parts of speech and
writing. From this it follows that no matter what else a theory of language
does, it must account for the departures from equiprobability in language.
Furthermore, this accounting must not confuse the issue by adding to these
any additional redundancies arising from its own formulation. That is, we
need the least grammar sufficient to produce precisely the utterances of a
language. It then turns out—necessarily so, on information-theoretic grounds
(1.4)—that the elements and relations which are reached in a least-redundant
description have fixed informational values which together yield the informa¬
tion of the sentences in which they occur. In this approach, finding the
structure simultaneously exhibits the information (11.4).
To find the departures from equiprobability, we first show that not all
combinations of elements occur (3.0), or are equally probable, in the
utterances of a language (e.g. Go a the is not an English sentence), and then
show what characterizes the occurring combinations as against the non¬
occurring. Given the great number and the changeability of the sentences of
a language, the word combinations cannot be listed. However, they can be
characterized by listing constraints on combination, such as can be under¬
stood to preclude all the non-occurring combinations, leaving those which
are indeed found (3.0). It has been possible to find constraints that are
relatively stable, and are such that each contributes a fixed meaning to the
sentences formed by it. The initial and basic constraint shows that sentences
are produced by a partial order on words in respect to their ability to occur
with other words in the set of sentences. When the constraints can be
formulated in such a way that each one acts on the resultant of another, we
obtain not only an orderly description of the structure of the utterances, but
even a way of quantifying the effect of each constraint without counting
anything twice (12.4.5). These constraints are presented in 1.10 and in
Chapter 3.
This least set of constraints, which characterizes language, is found to
have mathematical properties that are relevant to their informational effect.
In particular, the initial constraint, in producing the set of possible sentences,
necessarily creates a mathematical object (1.2, ch. 6).
The same least-constraint method shows that sentence-sequences within
longer utterances can best be described as satisfying additional constraints,
of word repetition (ch. 9). This produces finally a double array of the
sentence-making partial order among the words (9.3).
It is further found that when the constraint-seeking methods used to
obtain the structure of a whole language are applied to the notes and articles
of a single sub-science, they yield not the grammar of the language, but a
somewhat different kind of grammar which has some of the properties of
mathematical notation, and which reflects the structure of information in the
Overview 5
given science. These methods provide a procedure for organizing the

specific information in each document (or conversation) of the sub-science
(1.3, ch. 10).
Because they can be formulated as constraints on combination, the
entities and processes of this syntactic theory have an intrinsic relation to
information, and suggest a theory of the information carried by language
(1.4, ch. 11). Furthermore, the constraints can be regarded as processes
which in creating the structure of language limit what comes to be sayable in
it. As a result, they give some indication of the development of the
construction of sentences, and of language as a whole (1.5, ch. 12).
There remain other aspects of language, such as its detailed historical
changes (and dialectal differences), its analogic processes, and its use as art
material. Some of these (e.g. analogy and change) can affect the structure,
though most of them are domesticated into the structure created by the
constraints. Others (e.g. social dialects and literary arts) involve the study of
language as institutionalizations of a set of human behaviors, which for
systemic reasons have to stay within the framework of the set of constraints
described here.
1.1. A theory in terms of constraints
1.1.0. The major constraints
The theory presented in Chapter 3 states that all occurrences of language are
word-sequences which satisfy certain combinatory constraints; furthermore,
for reasons related to mathematical Information Theory, these constraints
express and transmit information (1.4). The first constraint (3.1) is what
gives a word-sequence the capacity to express fixed semantic relations
among its words: it consists in a partial ordering on words (or morphemes,
7.4), whereby a word (or morpheme) in a sentence serves as ‘operator’ on
other words, called its ‘argument’, in that sentence. (In Sheep eat grass, we
take eat as operator on the argument pair sheep, grass.) The essential
generality of this partial-order relation is considered in 1.2. All the other
constraints act on the resultants of the first.
The second constraint (3.2) reflects the specific meaning in a sentence, by
recognizing that for each argument word certain words are more likely than
others to appear as operator on it. (Eat is more likely than drink and than
breathe as operator on the pair sheep, grass.) The meanings of words are
distinguished, and in part determined, by what words are their more likely
operators or arguments; and the meaning of a particular occurrence of a
word is determined by the ‘selection’ of what words are its operator or
arguments in that sentence (11.1).
6 Introduction
The third constraint (3.3) on word-sequence makes sentences more

compact: it results from reducing—e.g. to an affix or to zero—certain words
which have highest likelihood, and thus contribute little or no information,
in a particular operator-argument sequence. (Pronouning and zeroing of
repeated words under and reduces Sheep eat grass and sheep eat weeds to
Sheep eat grass and they eat weeds, and to Sheep eat grass and weeds.)
In addition, the partially ordered operator-argument words are put into
one or more linear forms (3.4), needed since speech is an event in the single
dimension of time. (Thus "eat operating on sheep, grass' becomes Sheep eat
grass.) These four steps create, out of the set of all possible permutations of
words in a language, precisely the set of possible (grammatical or potentially
grammatical) sentences of that language: the term ‘sentence’ here applies to
something which is not only roughly recognizable by speakers of a language,
but is also (7.7) delimited, by a stochastic process, as a segmentation of
utterances (i.e. of connected occurrences of speaking or writing).
In addition to these constraints that make a sentence, there are two
constraints on connected sentences. One of these demands interrelated
word repetition in conjoined sentences. The other constraint demands
repetition of the operator-argument relation (above) in interrelated sets of
sentences: in connected discourse it can create a double array of words,
beyond the linear form seen in sentences (ch. 9). And in specialized subject
matters such as sciences it creates distinguished sentence-type formulas out
of distinguished word subclasses; here the sentences are found to satisfy
special conditions which make them a formally definable sublanguage that
represents the special information in that science (ch. 10).
The theory presented here is not a model devised out of various con¬
siderations or assumptions about language, but a direct generalization of
results obtained from syntactic analysis of languages.1 The central result at
this point is that when the sentences of a language are reconstructed into
their unreduced form (by reversing the reductions that carry out the third
constraint), the unreduced form is found to satisfy the first (operator-
argument) constraint.
The discovery of such reconstruction, and of the form of unreduced
sentences, arose out of an attempt to find the least system (grammar)
sufficient to characterize, or produce, the sentences of a language. In
following this line of research, it was found that after sentences had been
described compactly in terms of component word-sequences (e.g. subject-
verb-object and various modifiers), the description of language could still be
tightened by showing that certain sentences contained in a regular way the
' That the picture of language presented here is similar in some respects or others to various
views of language (cf. n. 6) does not obviate the need for reaching the present theory
independently by detailed analyses of language. This is so partly because a more precise and
detailed theory is then obtained, but chiefly because we then have a responsible theory in direct
correspondence with the data of language rather than a model which is speculative no matter
how plausible.
Overview 7
same component word-sequences as certain other sentences which were

paraphrastic to them (i.e. which preserved the meanings of the first). There
are two cases: (a) Many sentences consisted simply of other sentences plus
additional (higher) words, with the meaning of the included (lower) sentence
being both preserved and added to: I know sheep eat grass, For sheep to eat
grass is quite probable; Sheep's eating grass occurs all the time, (b) Many
sentences consist of another (source) sentence with no addition but with a
change, in most cases a reduction (or transformation, 8.1), that leaves the
meaning of the source sentence unaltered: He's at the grocer’s has to be from
He's at the grocer's place; He reads all day is from He reads things all day (this
derivation avoids having both a transitive and an intransitive read)-, similarly
for the two component source sentences in Sheep eat grass and weeds above.
In (a), the sentences are derived from elementary base sentences (such as
Sheep eat grass) plus further operators. In (b), the sentences are derived
ultimately from base (i.e. unreduced) sentences by means of reductions or
other transformations.
In a varied sample of languages (cf. the Preface), it was found that in each
language there is a particular set of reductions (with possibly a few non-
reductional transformations) and particular conditions necessary for their
being carried out. Each reduction leaves a trace; in some cases two or more
different reductions on two or more different sentences produce by a
degeneracy the same word-sequence, which thereupon becomes a pluri-
source, ambiguous, sentence. Given a sentence, one can recognize from its
structural traces whether or not it has undergone reduction, and if so which
reduction, up to this ambiguity. One can then reconstruct the unreduced
source sentence from which it was derived. This derivability of sentences,
from intermediate source sentences and ultimately from unreduced source
sentences, was the first step toward the syntactic theory.
The next step was the finding that the unreduced sentences satisfy the first
constraint above (4.1). More precisely: for the great bulk of sentences, the
unreduced form is seen immediately to satisfy the operator-argument
constraint; for the remaining sentences it has proved possible to define
variants of the existing types and conditions of reduction, which would
produce these remaining sentences as reduced forms of unreduced (operator-
argument) sentences.
The first constraint, which creates sentences and is thus the essential one,
holds that it is possible to collect the words of a language into classes which
serve as operator and argument relatively to each other, such that in the set
of unreduced sentences no word of an operator class appears unless the
sentence also contains one or another word from each of its argument
classes. Thus the occurrence of a word in an operator class depends on an
occurrence of a word from its argument class. For example, sleep is an
operator on one ‘noun’-type argument, eat an operator on two: this produces
not only Sheep sleep but also Grass sleeps, and not only Sheep eat grass but
8 Introduction
also Sheep eat sheep. Grass eats sheep, etc.; but all of these are syntactically
possible sentences. This constraint suffices to create all possible unreduced
sentences.
The operator-argument constraint imposes a partial order on the words
of a language with respect to their defined ability to co-occur in sentences or
utterances: in this order eat is higher than sheep or grass, while sheep is
neither higher nor lower than grass; know, probable, occur are higher than
eat (in (a) above). Each sentence is a set of word-occurrences satisfying this
partial order (i.e. of word-occurrences within which the operator-argument
relation holds), possibly with (statable) reductions. The partial ordering
itself has a meaning: the operator is said about (i.e. predicated of) the
argument (3.14). Hence, given the meanings of the words (by real-world
referents, or by selection of operators and arguments), finding the operator-
argument relations among the words of a sentence yields its meaning
directly: that meaning is the hierarchy of predicatings among the meanings
of the words of the sentence. The operator hierarchy, i.e. the partial
ordering, is overt in the unreduced sentences; it and its meaning—though
not its overtness—is preserved under reduction.
The grammar of a language is then the specification of these word classes
in their partial ordering, and a specification of the reductions and linear¬
izations. This seems quite far from the traditional contents of grammar. But
the usual grammatical classes, such as verb and adjective, are obtainable
from the partial order. Some (e.g. adjectives) are by-products of reductions
on operators; others (e.g. verbal nouns) are secondary statuses such as that
of an operator {eat) appearing as argument (eating) of a further operator
(e.g. under occurs all the time, in Sheep’s eating grass occurs all the time).
And such syntactically local constructions as plural, tenses, conjugations,
and declensions are merely patternings of the form, likelihood, and syntactic
status of certain reductions. Thus the usual complex grammatical entities
and constructions, important as they may be for practical descriptions of a
language, are only incidental to (and derivable from) the universal and
simple form-and-meaning constraints of language structure. The traditional
constructions mentioned above can be brought out by seeking the similarities
and interrelations among the reductions in a language, and correlations
between the shape of words and their syntactic status.
In general, the constraints—including the vocabulary and the co-occurrence
likelihoods of the words—are discovered from the regularities seen in
sentences. Since each constraint acts on the resultant of previous constraints,
the successive constraints have monotonically descending domains; and
more importantly, at each stage of analysis the remaining constraints do not
alter what has been obtained thus far. Furthermore, each sentence is a
partially ordered structure of words and reductions. In its unreduced form,
the words are ordered by their operator-argument relations there—in
effect, by their ordered entry into the making of the sentence. The reductions
Overview 9
and transformations which create a reduced sentence are ordered by the

property that if they are carried out it is done as soon as the conditions for
them are satisfied in the making of the sentence. Given all the properties
noted above, the constraints make it possible to devise a strategy for
analyzing each sentence in a language, recognizing the reduction traces and
the operators and arguments, from the latest entries back to the ultimate
elements. Conversely, the constraints make it possible to synthesize every
sentence. For, when the words of each class are specified, with their ranges
of meaning or their likelihood in respect to their argument words, the
syntactic theory presented here produces actual sentences of the language,
with direct indication of their meaning. The syntax of a sentence indicates its
semantics because each constraint has a fixed meaning, and any difference in
meaning that a word may have in its various operator-argument statuses is
specifiable a priori.
In addition, this theory has certain useful properties. One is that while
consisting of computable forma! relations it has a behaviorally functional
explanation (12.3.5). Another is that the theory is explanatory for gram¬
matical detail: a great many special and seemingly arbitrary or irregular
word combinations are seen to result from a constellation of regular processes
in a way that explains both their form and their meaning.
1.1.1. Example: The set-vs. -member connectives
To give some impression of how analysis in tqrms of the theory explains

grammatical construction in detail, we consider two particularly complex
illustrations. One is the case of words, such as only, except, but, that can
occur in various unique environments all of which are created by zeroings
associated with the words everything, everyone. Only and the others are a
subset of the bi-sentential operators, which can occur as connectives between
any two (component) sentences. In this subset there is considerable likeli¬
hood for one of the component sentences to contain the word everything, or
everyone, or the like, while the other component sentence contains a word
for an individual or subset of that every. When the component sentences
have this high-expectancy set-vs.-member matching, the every is predictable
and hence zeroable. Under certain of these operators the occurrence of not
in one of the component sentences but not in both is also highly likely and
hence zeroable. It is these special zeroings, together with zeroings of
repeated material, that can make the environments of these operators
unique (G401).2
First, only appears in two environments, with different meanings: (1) The
meeting was fine, only John didn’t come (similar to saying except, but), which
2 For the reference G, see n. 4 below.

10 Introduction
conjoins any arbitrary two sentences; and the apparently independent (2)
Only John didn't come (with stressed only, similar to saying John alone
didn't come). However, the only of (2) turns out to be a case of the only of
(1), if we derive (2) from (3) Everyone (else) came, only John didn't come
(which is structurally of type (1) above). In (3) the whole first component
sentence is zeroable (reducible to zero) on grounds of low information in
respect to the set-vs.-member selection of the bi-sentential operator (con¬
junction) only. Note that the first component of (3) is identical with the
second component, except that everyone (else) with optional else replaces
the argument (John), and that the two verbs differ in negation alone (didn’t
come). When the two sentences which are arguments of only differ in just
this way the whole of the first argument (here, Everyone (else) came) is
zeroable, because of the set-vs.-member selection of only. This reduces (3)
to (2). In other cases, e.g. (1) above, no zeroing occurs and only remains
visibly bi-sentential.
Although the meaning of only in (2) seems to differ from that in (1) and in
(3) , that difference is seen on closer inspection to be merely a trace of the
zeroing of the first sentence under only. Indeed, (3), with the meaning of
everyone else not, except John, and (2), with the meaning of John alone,
carry the same information. It is instructive to see here how meaning¬
preserving reductions (here, zeroings) in the first argument of only in (3)
yield such a difference in the meaning of post-reduction only such as we see
in (2).
Next, except. Here too we can find a juxtaposition of structurally arbitrary
component sentences, as in The meeting was fine, except (that) John didn’t
come with optional that, and also of structurally contrasted sentences as in
(4) Everyone came, except that John didn’t come. What is zeroable in the
contrasted sentences around except is just the not (optionally) plus verb plus
that (here, that. . . didn't come) of the second component sentence, more
rarely the verb alone (here, came). Then (4) reduces to (5) Everyone came
except John or (5a) Everyone except John came or the rarer (5b) Everyone
came, except not John. (5b) seems to negate (5), but is actually a paraphrase
of it, differing only in not zeroing the expected not. (The second came must
be zeroed if not is.) Again, the meaning of (4) is preserved in (5).
Then, but. We have any arbitrary component sentences in e.g. The
meeting was fine but John didn’t come, and contrasted ones in e.g. (6)
Everyone (else) came, but John didn't come. Under but, in structurally
contrasted sentences such as (6) one can zero the second verb, and if so also
the high-expectability not. Then (6) reduces to (7a) Everyone came, but not
John or Everyone, but not John, came (separated by commas) and to (7b)
Everyone came but John or Everyone but John came (without commas), all
preserving the meaning of (6). The not and the zeroability of the repeated
verb may be in the first component sentence. Here the optionality of zeroing
the not (compare 7a and 7b) yields the peculiar case of having identical
Overview 11
meanings for a sentence and for its apparent negation: in There weren't hut
two people left and There were but two people left. (The zeroings that
produce the last form are not available under except.)
This analysis shows only, etc., to be a subset of the sentence-introducing
‘particles’ (therefore, however, nevertheless, too, so, even, etc.), which do
not normally appear on the first sentence of a discourse—this last because
they are all derived from conjunctions under which various zeroings can
eliminate parts of the component sentences, including (as under only) the
whole first sentence (G393). In addition to this general derivation, the only,
etc., discussed above have a high likelihood for every matched against a
subset of whatever is referred to in the every, and only, except, but, also have
high likelihood for not on one sentence but not both.
This illustration does more than show that explanatory analyses can be
obtained by the syntactic theory presented here. It shows that regularities
among suitable words can be discovered by comparing their environments.
It also shows that the regularities apply to processes, such as zeroings here,
on those environments. Furthermore, it shows that the analysis of a given
environment has to be made in respect to all the other environments of the
word in question, so that the given one is obtained from the others by the
regular processes. This means taking into account all the environments of a
word rather than saying—unhelpfully—that in different situations it is a
different word or a different use or meaning of that word. It also means
trying to take into account the etymology of a word, i.e. its earlier uses: only
is from one + -ly; except is from Latin ex ‘out of + capere ‘take’; but is from
early forms of by + out + adverbial suffix. The etymology is relevant
because: either the current word in its environment is a descendant (i.e. a
repetition or a continuation) of its earlier source, appearing in the source of
its current environment; or else at some time the old word was extended into
the current environment by some regular or analogical relation to the earlier
environments of the source word. Such etymological comparison is therefore
not a checking of the modern word against its earliest known cognate, but a
tracking of the changing use of the word in various environments, to the
extent possible. And this evidential ‘use’ is a matter of specific environing
words in various sentences, not a matter of meaning, because we have little
knowledge of past meanings except via past word-environments. Further¬
more, appeal to meaning would not suffice to predict the increasing preference
for certain environments, which underlies the zeroabilities within them: why
did the above sources of only, except, but (but not, for example, unless)
develop the likelihoods and zeroings of every and not which they now have
(while unless, not far in meaning, did not)? Why did even not move into a
more contrastive use, with special likelihood for not, as it does slightly in
even so (compare French quand meme)l
12 Introduction
1.1.2. Example: The passive construction
The other illustration is the case of the passive construction in English (G16,
362): The book was taken by John, by the side of the ‘active’ John took the
book. It is first of all clear that the passive is not an independent combination
of words, but rather a transform of the active (i.e. derivable from it), since
for every passive sentence (except for those in subsets explained below)
there corresponds an active sentence, as above (cf. 2.3). It can further be
seen that the passive is not an independent transformation, but rather a
resultant of three elementary transformations. (1) One is the by which arises
in English on the subject (first argument) of a verb when a further operator
acts on that verb, e.g. above in taken by John as argument under was. Note
that this use of by differs syntactically from the ‘prepositional’ operator by
which appears equally on active and passive sentences, as in John read the
book by candlelight and The book was read by John by candlelight. (2) A
second component of the passive which is known elsewhere in English
grammar is the stative -en (-ed) suffix. This suffix appears also in the ‘perfect’
tense (John has taken the book), in the ‘passive adjective’ (a crooked smile, a
learned man), and in a suffix on nouns (a monied family, a two-wheeled
vehicle). (3) The third component is the moving (‘permutation’) of the
object (second argument) of a verb to subject position, more precisely to
being the subject of a further verb (here, is) that is operating on the original
verb. This is seen in The Armenians suffered repeated attacks by (or: from)
Azerbaijanis, where the grammatical (and physical) object of the attack
were Armenians; the form is comparable to the passive The Armenians were
repeatedly attacked by Azerbaijanis, from Azerbaijanis repeatedly attacked
the Armenians.
As to the elementary transformations:
(1) By on the subject of a ‘nominalized’ sentence (i.e. a sentence appear¬

ing as argument of a further operator) is found independently of the other
components of the passive: We opposed the bombing of Dresden by Churchill,
He had a scolding by (or: from) the principal. In contrast, of and's (and also
absence of any such argument-markers) occur on both subject and object,
under various conditions: We opposed Churchill’s bombing of Dresden (’s
on subject, of on object). We opposed Dresden’s bombing by Churchill {'s on
object, by on subject), We heard the singing of birds (of on subject). We
opposed Churchill bombing Dresden (no markers).
(2) The -en suffix occurs independently of the subject-object permutation.

In the perfect tense the -en means ‘state’, and has -en is simply the present
tense of the completed ‘having a state' (and had -en is its past, i.e. attaining
the state at some past time). Thus the present perfect cannot be used in
Goethe has visited Lago di Garda, because the no longer living Goethe
Overview 13
cannot be said to have such a state in the present time. But the present
perfect can be used on the passive of the above, in Lago di Garda has been
visited by Goethe and others of his generation, because the lake still exists
and can be said to have that state. (If the passive were merely a direct
reshuffling of the active, it should have the same range of tenses.) In the
active, the present-tense state (on visit) would have been asserted of
Goethe, but in the passive it is asserted (on be) of the lake. There is also a
limited perfect tense with is, seen for example in The Sun is risen. We are
agreed. We are resolved. She is grown quite tall, I am finished (or: done) with
it. He is possessed of great wealth-, this is a stative or completive of We agree,
etc. Further, there is the passive adjective as in the curved horizon, a learned
scholar, a professed (as well as professing) Christian. Finally, there is the
stative -ed on nouns, as in a monied family (which means not a family which
happens at the moment to have money, but a family in the ongoing state of
having money).
(3) Having the object of a lower verb (e.g. eat, praise below) appear as
subject of a higher verb (e.g. is ready, deserve, is below) occurs independently
of the -en. Among the many forms, consider the following examples: It is
ready to eat. It is ready for eating (where it is the zeroed object of eat); He is to
blame (the missing object of blame is him)-. He deserves our praise (where of
him can be added, as object of praise)-, It improves with (the) telling. It is in
process of dismantling (where it is object of tell, dismantle, and can be said,
e.g. . . . of dismantling it); This is past understanding, Much money is owing
to them, obsolescent The house is building (in all these the subject of is is the
same as the zeroed object of understand, owe, build); This is understandable,
which could be derived from This is capable of one’s understanding it (as
though from This is able for one to understand it). In all these cases the lower
verb can also be said in the passive: It is ready to be eaten. Much money is
owed to them. The house is being built (the last, only in recent English).
To say that the passive includes these three constructions means that it is
the resultant of the application of all three to the source of the passive, i.e. to
the active. Hence the passive (The book was taken by John) results from
taking the simple active (John took the book) and placing it under a higher
operator, is -en, which belongs to the set of (3) above whose subject is
usually the same as the object of the verb under them. We then have is -en
operating on the pair book and John took the book (which under -en is
nominalized to taking of the book by John). Here of the book is zeroed as
being a necessary or usual repetition of the subject of is -en; then the subject
book of is -en looks like a permuting of the object book of take, whereas in
the derivation we see that the former is j ust a repetition of the latter that then
caused the latter (the lower occurrence of the object book) to be zeroed. The
-ing of taking is not an independent accretion (an operator or argument) to
the sentence, but merely a marker, an indicator (G41) that the operator to
14 Introduction
which it is attached has become nominalized as an argument of a further

operator (in this case is -en). Therefore nothing is lost to the sentence when
the addition of -en involves removal of the -ing, yielding taken. These steps,
then, suffice to yield the passive. Using arrows for derivation we have: The
book was in the state of the taking of the book by John —> The book was in the
state of taking by John —> The book was taken by John.
As to the domains: since the passive is the resultant of three components,
the domain of sentences to which it can be applied is the intersection of the
domains of the components. Thus it can be applied only to sentences which
are within the selection (3.2.1) of: (1) the by-nominalization of the lower
sentence, (2) -en as higher operator on it, and (3) the subject of the higher
operator being the same word (with same referent) as the object of the lower
(nominalized) operator.
(1) Hence, there is no passive of The book costs $5 and A cat is a mammal
(no $5 is cost by the book, A mammal is been by a cat, just as there is no I
noted the costing of $5 by the book (but only I noted the book's costing $5)
and no / noted the being of a mammal by the cat (but only I noted the cat's
being a mammal). To say that the passive is available only for transitive
verbs is not precise. There are passives of intransitives, provided these carry
the fiy-nominalization: Everybody laughed at him has both / noted the
laughing at him by everybody and He was laughed at by everybody. But for
He walked near the wall one is far less likely to say I noted the walking near
the wall by him and even less likely to say The wall was walked near by him.
And just as I noted the taking of a ride by the boy is more acceptable than /
noted the taking of fright by the boy, so A ride was taken by the boy is more
acceptable than Fright was taken by the boy.
(2) As is seen in transformation (2) above, -en means roughly ‘state’.
Therefore, when it operates on a pair of arguments consisting of a noun and
a sentence, it selects those pairs in which the noun can be said to be in the
state of that sentence. This is seen in the perfect tense, where the subject of
the lower sentence is the same as the subject of the higher has -en: Goethe
has visited Lago di Garda, equivalent to Goethe has the state of his visiting
Lago di Garda, is no longer sayable because, as noted, the present tense has
cannot now be used of Goethe’s acts; one would have to say Goethe visited
. . . or Goethe had visited . . . ,3 And it is seen in the passive, where it is the
object of the lower sentence that is the same as the subject of is -en (Lago di
Garda is written about by many, equivalent to Lago di Garda is in the state of
the writing about it by many). The fact that the is -en asserts a state of its
subject explains, for example, why one can readily say America was left by a
whole generation of its young writers in the 1920s, but hardly America was left
' But one can say Goethe says this in many places or Goethe is one of the greatest German
writers, because these are continuing properties of Goethe’s work. (He is still saying things in
his books.)
Overview 15
by my neighbor last month; for the expatriation of a generation of writers

could be described as constituting a state of America, but hardly so the
neighbor’s actions.
(3) The condition that the subject of the higher is -en have the same
referent as the object of the lower sentence explains how it is that the main
meaning of Someone was blamed by everybody (in the sense of ‘There was
someone such that everybody blamed him’) is different from that of Every¬
body blamed someone (in the sense of ‘Everybody had one or another
person whom he blamed’). The explanation arises from the following: in the
unreduced language base, as in logic, quantifiers (e.g. every) can be shown
to act not on nouns but on sentences (G252). The referents are first
determined in the source sentence Somebody blamed someone there before
the quantifier is added, and it is on this that the passive Someone was blamed
by somebody is formed; the quantification to everybody is added separately,
on the source and on the passive (G263).
The illustrations above give some indication of how the different construc¬
tions in which a word is found and the different meanings the word has in
them can be derived by a few given reductions from a standard unreduced
(base) set of operator-argument constructions. Thus, a few pervasive
constraints combine to produce the many forms of sentences, including ones
that seem exceptional. These illustrations also show (as follows from 2.3)
that data can be usefully analysed only in respect to a specified whole
structure of a language, and not as selected examples in support of episodic
grammatical rules or processes.4
1.2. Mathematical properties
There are reasons to expect that the structure of language has in some
respects a mathematical character. The grammatical classification of words
is defined in general by relations among the words, rather than by particular
properties (phonetic, morphological, semantic, or logical categories) that
each word has by itself. An indication of this is seen in the simple fact that
4 It is for this reason that the examples in the present hook are almost entirely from English,
although a single language is far from sufficing for a survey of language structure. In the case of
English, there are available above all the multi-volume Oxford English Dictionary, and Otto
Jespersen's Grammar of English on Historical Principles; in addition, the present writer's A
Grammar of English on Mathematical Principles (referred to here as G) presents English
grammar in terms of the syntactic theory of chs. 3-4 below. The latter book not only provides
examples for the discussion in the present volume, but also shows how the methods and theory
stated here suffice to yield an orderly and detailed grammar of a language. The English
grammar of G thus serves for systemic evidence, without which a theory in a field such as
linguistics would not be adequately established. As noted at the end of 12.1, the interrelations
and regularities of the data are essential to any theoretical formulation (on grounds given in
2.2), so that the episodic presentation of examples in support of one analysis or another is of
little relevance.
16 Introduction
every language can accommodate new words. For a language at any one
time, we can think of its word combinations as being determined by specified
lists of words, e.g. the list of words including sheep, child, grass, each of
which can ‘make a sentence’ with a word of a list that includes fall, sleep, eat,
but not with any word of its own list: Sheep eat grass, but not Sheep child
grass. However, when a new word enters the language (i.e. comes to be
used), the question is: on what basis can we say what list it belongs to? If the
new word combined with other new words and not with the previously
existing words, the language would develop separate encapsulated gram¬
matical systems for different lists of words; this is not generally the case,
although something approaching it is seen in the morphological structure of
the English perceive, conceive, persist, consist, etc. If the new word combined
with old words, but differently than the old words did among themselves, we
might in time lose any systematicity over the whole language; this too is not
the case. What we find in each language is that large subsets of words act in
uniform ways in respect to other large subsets.
This means that the late-comers (whatever words they were) combined
with other words in the same ways that existing words did. Such fitting into
existing word sets means that the new words satisfied the condition of being
usable in the existing types of word-set combination. Thus the overall
criterion of classification is the condition that words enter into the same
combinatory relations. Hence the general characterization of word classes is
in terms of their common combining-relations to other classes. It will be seen
that this opens the door to characterizing the entities of language by their
relations rather than by their inherent properties—a necessary condition for
relevant mathematical characterization of language structure.
Beyond the considerations above, language has to have a constructive and
relevantly mathematical structure if it is an unbounded set of different
activities (different utterances) carried out by finite beings (2.5.1), and if it
was not built by anyone on some template or model (the existence of which
prior to language would be inexplicable, 12.4.3-4). Also, if language is
structured as an information carrier (11.5), we can expect it to have
necessary properties related to those known from mathematical Information
Theory (11.7).
This last is of central importance. If we begin analyzing language without
any preconceived guides, we find that the lack of an external metalanguage
in which to conduct the analysis leads us to describe the data in terms of
departures from equiprobability of combination (2.1-2); these are the
constraints of 1.1 and Chapter 3. It is then immediately apparent that these
constraints create the kind of redundancies studied in Information Theory,
so that the structure of syntax is a system of contributions to the total
redundancy—both of language and of its information.
If we now consider the data of language, it indeed seems overtly to have
possibilities of mathematical treatment, as in the transitional probabilities
Overview 17
among the successive letters that make a word in alphabetic writing, or the
successive words that make a sentence. However, the transitional prob¬
abilities do not yield useful characterizations of words and sentences (although
particular stochastic processes can be so formulated as to yield word and
sentence boundaries, 7.3, 7.7). Also, free semigroups are formed by the
sequences of letters and by the sequences of words. But these semigroups do
not represent natural language, and their structure is not relevant to the
structure of language. For example, no simple relation stated on these sets
can characterize which different sequences have which different frequencies
of occurrence. Rather, these sequences constitute sets, of which the words
and the sentences of natural language are undefined subsets. Indeed, to
define the particular subsets which constitute language is the first and
essential task of a theory of language.
The present theory will be seen to have various mathematical properties
(ch. 6). The crucial mathematical property lies in the fact that the essential
determination of operators and arguments is made by a single relation on
their combinations. That is, in each language, the classes of operators and
arguments are not each a collection of words which have been selected on
syntactic or morphological or semantic grounds (such as having a tense, or a
particular suffix, or a particular kind of meaning). Rather, words are
assigned to classes entirely on the grounds of a single co-occurrence relation:
We say that A is operator on ordered classes, say B, C (or: A depends
immediately on B, C), to mean that A is in the set of all words which appear
in an unreduced sentence only if some word of B and some word of C, in that
order, also appear in that sentence; here B or C may be stated to be the same
class as the other, or as A, or different classes, or the null set (3.1. l)r ln turn,
membership in B or C is not given by a list of words or by some property
other than this dependence. Rather, B is itself defined as the set of all words
which depend, in the above sense, on classes D, E\ and so for C; and so on
(3.1.2).5 The word classes are thus defined by their dependence on word
classes which are in turn defined by the same dependence relation. An
infinite regress is avoided only due to the existence of a class of words whose
dependence is on the null set of words. The single relation among words is,
then, dependence on dependence; and a sentence is any sequence of
words—or other objects—in which this relation is satisfied. In accord with
this, there are just a few dependence-on-dependence statuses that a word
can have (i.e. types of argument that a word requires). Hence, languages
have a set of words whose status is that they can be defined (for reasons given
in 3.1) as depending on null (these are ‘primitive arguments’, 3.1.2, e.g.
John, sheep), words depending on (one or more) words that depend on null
(e.g. sleep, eat), and words depending on words that depend on something
5 The reasons for this situation, going back to the lack of an external metalanguage, are
reviewed in ch. 6 and 12.4. Sublanguages, which have an external metalanguage, do not have
the mathematical-object property (10.51).
18 Introduction
not null (e.g. occur, is probable); this last type includes words whose status is
that one of their arguments depends on null while another argument
depends on something not null (e.g. know, whose first argument may be
John while the second may be eat ((a) in 1.1).
It may seem surprising that a practical grammar, able to produce the
actual sentences of a language, can be built out of word classes defined so
generally and indirectly, rather than out of linguistically determined word
lists. This development was perhaps necessary, because if word combinations
were in respect to specified word lists, it would have been difficult to reach
public understanding on how new words are to combine. Greater stability is
attained if once a new word is used in a particular dependence-on-dependence
status, it is publicly clear what types of argument it can occur with (5.4,
12.4). In any case, the fact that we can characterize in this way the word
classes that are sufficient for grammar means that words, whatever phonetic
and semantic properties they have, have also as their defining syntactic
property their position in a network of occurrence-dependences in respect to
each other. Thus the co-occurrences (sequences) of words in unreduced
sentences as created by the first constraint is a mathematical object: a system
of entities defined only by their relations and closed in respect to them, with
operations defined on these entities.
Showing that a given system in the real world is describable as a math¬
ematical object means that any other real system which is also an interpretation
of that mathematical object shares the relevant properties of the first. In
particular it means that any other set of objects or events that satisfied the
first (dependence) constraint would also have sentence-like properties. And
indeed, mathematical notation will be seen below to satisfy the first con¬
straint but not the second (10.6), with the similarities and differences
between language and mathematics being indicated thereby.
In addition, the effect of various constraints has been to produce various
mathematical structures in language—orderings, semigroups, mappings,
and decompositions involving sets of words and of sentences. These are
noted in 6.3, 7.3, 7.6-8, 8.1.5-6, 8.3. What is important is that these
structures are not merely describable, but are relevant to the informational
function of language (11.5). The dependence-on-dependence relation under¬
lies virtually everything in language structure; and beyond that, the math¬
ematical structures noted here suffice to define almost all the semantically
(and grammatically) important classifications and relations; and even the
kinds of information capacity that language has (11.3, 11.6).
1.3. Science sublanguages
Analyzing languages in terms of constraints on word co-occurrence makes it

possible to recognize the existence of discourse-structures and sublanguages
Overview 19
within a language. If the grammar of a language were stated in terms of

objects unique to language, such as a particular list of words constituting the
class 'adjectives’, it would not be possible to apply the entities (classes) of a
whole grammar when investigating any special classes such as might arise in
special conditions of language, unless the special classes were built of the
same entities. In contrast, if the grammar consists in stating a relation among
words, we may obtain, as carriers of the same relation, one set of classes for
the whole language and other, appropriately different, classes for dis¬
tinguished subsets of the language.
One distinguished subset of the utterances of a language is the set of
sentence-sequences, each sequence being a set of two or more sentences
which have been said in succession, as a single discourse. All such sentence-
sequences satisfy, overtly or else in their base reconstruction, a constraint
which calls for one or another kind of repetition or subclassing of words
among the sentences (ch. 9). The repetition constraint expresses—or creates—
a connectedness among the sentences. For many kinds of discourse, this
constraint produces, beyond the linearized partial order of sentence structure,
a double array in which one dimension records each linear sentence while
the other dimension gives the similarities and differences within corresponding
positions among successive sentences (9.3). Even the structure of proof in
mathematics is a special case of this constraint on sentence-sequences
(10.5.4, 10.6).
Another, very important, type of sentence subset is that in which the
connectedness among the sentences lies in their sharing a subject matter.
Here the repetition constraint creates sublanguages, with grammars different
from the grammar of the whole language (10.0).
Not all subsets of utterances lead to different grammars. Quite the
contrary, analyzing various arbitrary collections of utterances in a language
leads to much the same grammar, and the grammars of different collections
increasingly approach each other the bigger the collections are. However, if
the corpus of material which is being analyzed comes from a single, more or
less encapsulated, subject matter, one arrives at a grammar somewhat
different from that of the whole language. The same operator-argument
dependence which creates sentences necessarily appears here too, but with
different resulting word classes and sentence types. In the case of arbitrary
collections of sentences, the word classes that are obtained are constrained
only by the dependence properties of their arguments. But in the case of a
restrictive subject matter, for example a subscience, we may find one
particular small class of nouns (primitive arguments) which occur only under
a particular list of operators, and another noun class which occurs only
under another list of operators; and certain of the operators may occur only
under a particular list of further operators.
For example, in immunology (10.4) we find certain operators including
appear in, are produced by, are secreted by, having antibody, gamma-
20 Introduction
globulin, anti-ferritin, etc., as their first argument (subject) and having

lymphocytes, plasma cells, etc., as their second argument (object): Antibody
appears in lymphocytes. And we find was injected into having antigen, sheep
erythrocytes, virus, etc., as its first argument and having rabbit, left ear, etc.,
as its second argument: Antigen was injected into rabbits. The words within a
class are clearly not synonyms, nor do they necessarily constitute an other-
wise-known semantic class. Combinations cutting across these operator-
argument structures are not just false, nor are they nonsense (as they might
be for the language as a whole—e.g. is produced by as operator on the pair
left ear, rabbit); rather they are simply outside the sublanguage. If someone
inserted The left ear was produced by the rabbit into an immunology article,
it would immediately be recognized by all readers of immunology as not part
of the science, independently of whether it was considered a reasonable
English sentence.
Furthermore, the grammatical analysis here need not go as far as it goes
for the whole language, only far enough to separate out the segments which
fall into operator-argument classifications in respect to each other in this
material—i.e. only enough to obtain a least grammar for the given science
sentences. In some cases grammatical analysis beyond this point would be
detrimental to maximum regularity for sublanguage operator-arguments.
For example, the operators on the pair antibody, lymphocyte include
appears in and is produced by (in Antibody appears in lymphocytes. Anti¬
body is produced by lymphocytes). If we analyzed is produced by into its
grammatical components we would end up with Lymphocytes produce
antibody, which would not readily look (e.g. to a computer recognition-
program) like the same sentence type as in Antibody appears in lymphocytes.
But in 10.3 it will be seen that Antibody is produced by lymphocytes belongs
to the same sentence type as Antibody appears in lymphocytes, in the
immunology sublanguage: both have the same word sets as subject and as
object, though in inverse order.
When we consider the word classes and sentence types obtained from the
operator-argument relations within a subject matter, we find that while they
necessarily satisfy the gross grammar of the whole language to which they
belong they do not do so maximally: compare the kind of classes mentioned
above for immunology with the classes defined by dependence on dependence
in 3.1.2. Rather, what they are found to satisfy in detail is the set of objects
and relations of the science in question. The word classes and sentence types
of a sub-science correspond well to the objects and activities of the laboratory,
the events observed, and the entities and relations of the theoretical
discussion. Thus the grammar of the sub-science is a structural representation
of the knowledge and opinions in the field. In terms of this grammar, the
sentences of articles and reports can be represented in a canonical way as
formulas written in a structured notation, which exhibit the specific inform¬
ation of the sentence and also show its relation to the structure of knowledge
Overview 21
in that science. Utilizing both the discourse and the sublanguage properties
of scientific articles makes possible (in principle) computer processing of
their information.
It will be seen in 10.3 that while the sentences of the science are a subset of
the sentences of the whole language, the grammar of the science is not a
subgrammar of the grammar of the whole language. Rather, it is a kindred
system, different not only in detail but in some basic properties which go
beyond whole languages. In their purest form, the word classes and sentence
formulas of a science (independently of their original language), and in a
different way the double array of a single discourse, constitute not just
sublanguages but new language-like systems of notation. These systems,
intermediate in specifiable respects between natural language and math¬
ematics, show something of the nature of scientific knowledge on the one
hand, and of discussion in general on the other (10.5).
1.4. Information
The relation of language structure to meaning was known to many, especially

to logicians who developed categorial grammar (n.6). When structure is
stated in terms of constraints, this relation can be seen explicitly, in several
respects. First, when one sees what the syntactic relations consist of, one can
see how they come to bear the- meanings they do. In the first constraint,
many operators name properties or relations or acts while their arguments
name objects or events. However, such typical meanings are far from
constituting a semantic criterion for operatorhood or argumenthood, much
less a precise or predictive one: in English, operators include father, while
primitive arguments include time. But when the condition that saying an
operator word depends on saying some argument word is conventionalized
into a requirement (12.3.5), it has the semantic effect that the operator is
being said ‘about’ the argument. The operator is like a predication, on one
or more operands: came is said about John , eat is said about the ordered pair
sheep, grass. The dependence is more than an artefact of speaking: it
presumably arose from the usefulness of interrelating the word meanings in
communicating information. Since major modes of thinking, especially such
as are more reasoned and precise, are mediated by language, there is little
doubt that this relational meaning of predication, which developed in
language, has become a step in the process of thought or cognition. It would
seem that the basic structural features of languages may have implemented
cognitive development, while less universal features, such as tenses or
conjugations, have little or no visible effect on cognitive structure, even in
the languages that have them.
Second, the relation of language to information is more explicit than to
thought. The fact that the constraint which creates sentences creates the
22 Introduction
meaning ‘to say about’ supports other evidence that language developed as a
carrier of information (11.3, 12.3). It will be seen that other structural
properties of language are explicable on this ground, and that, in contrast,
little in grammar is structured to carry subjective expression, emotion and
the like: indeed, such expressive features of speech as the intonations of
anger, hesitancy, etc., are not included in grammar (hence, in effect, in
language), because they do not combine with anything else in a way that
forms further structures. That the content of language is seen here to be the
communicating of information, rather than the expression of meaning in
general, leaves room for the expression of other meanings, such as feelings,
to be carried by other vehicles and behaviors such as music and the other arts
(10.7).
There is another important result concerning information. Those trans¬
formations which do not bring further operators into a sentence ((b) in 1.1)
have been found to be mostly either alternative linearizations of the original
partial ordering of words, or else reductions in the phonemic (phonetic)
composition of words in the sentence, under conditions in which no relevant
information is lost or added in the sentence. The reduced form may lose
some words which seem to have contributed information but did not really,
as in / expect him any minute from I expect him to be here any minute. Or it
may put things in a different grammatical relation than the source did, as in
The conductor, who was in a great rush, did not look back from The
conductor did not look back; the conductor was in a great rush (where only in
the reduced form is the secondary sentence a modifier on conductor). There
is something important which is preserved under transformation: it may be
called the substantive or objective information, which is the same in the
reduced sentence and in its source. The meaning difference between an
unreduced sentence and its reduced form consists largely in the speaker’s
attitude toward the information (e.g. what is the topical part), or in the
hearer’s access to the information, and of course in the compactness of the
information.
When we include among the constraints the choosing of particular words
for a given sentence—in effect this is the second constraint (3.2)—the
methods of the present theory isolate the first two constraints, which are
universal and essential, as bearing all the substantive information and as
creating a base sublanguage from which the whole language is derived.
Within each sentence of the base, each application of these two constraints
contributes accfetionally to the information in the sentence. The other
sentences are derived from the base by constraints which do not change the
substantive information, though they may lose distinctions by degenerately
deriving from two different base sentences the same word-sequence (which
is thereupon ambiguous).
In the theory presented here every grammatical constraint (above all, the
operator dependence) has a stated meaning which is the same in all its
Overview 23
occurrences. And every word has stated meanings; its meaning is the same
whenever it occurs with the same operators and arguments. That is, the
association of words and grammar with phenomena of the world is made in a
regular way. Hence it is possible to work toward orderly processing and
retrieval of the specific information in sentences and discourses, on the basis
of the syntax and words of sentences (cf. ch. 4 n. 1).
A major gain in describing language in terms of constraints—which are a
construct of theory—rather than in terms of word combinations—which are
the observables—is that each constraint can be associated with a meaning,
in a way that each type of combination cannot always be. The relation of
constraints to meaning is not surprising. On grounds of Information Theory
we would expect the total departures from equiprobability, created by all
the constraints together, to determine the total informational capacity of
language. In the present work, it is seen that each constraint, acting on its
predecessors, adds thereto its own departures from equiprobability and
therewith its additional informational capacity (11.3.2). The constraints
determine even what kind of information a system can carry: that math¬
ematics can carry truth but not fact; that natural language cannot distinguish
nonsense, but sublanguages can (10.5-6). A detailed stepwise correlation is
thus created, in ihe base sublanguages, between form and content (11.4).
The constraints—on sounds, on what constitutes words, on syntax—arise
from efficiencies of communicational behavior (12.3.5) and from the speakers’
experience of the world (11.6); but they appear physically as the structure of
language events, as classifications of components of sound-waves. Hence,
the relations within information can be processed on computers by carrying
out the structural operations of language, although otherwise the attempts
to mimic the brain and cognition by computer programs are not at present
based on adequate or relevant methods. The relation of language structure
to information thus makes possible the complex processing of information.
In addition, it provides a stepping-stone toward understanding the structure
of information, by showing that total informational capacity can be reached
as the ordered product of specifiable contributions to information capacity,
and by suggesting that all information involves a departure from equiprob¬
ability acting on departures from equiprobability (11.7).
We see here that there is a particular dependence of co-occurrence which
creates the events which are called language, and that this dependence
necessarily gives these events the power to carry information. More specific¬
ally, the underlying structure of language is the structure of information
(11.5), and in 11.6.1 it is proposed that the co-occurrence constraints which
create information are a reflection of co-occurrence constraints in the
perceived world.
24 Introduction
1.5. The nature of language
The theory of syntax presented here does not constitute a theory of language
in all its manifestations; it is not clear that there can be a single such theory.
However, what is shown here about syntax creates a framework which has to
be taken into account in many other questions concerning language, and
which casts new light upon these: for example, on the relation of language
change to language structure, on the differences among languages, and on
the difference between 'natural' language and language-like systems such as
discourse structures, science sublanguages, mathematics, and computer
languages. More centrally, certain considerations are established here with
respect to the structure of information, and with respect to the nature—even
the early development-of language. (Unavoidably, the sketch below is a
very inadequate summary of the discussion in Chapter 12.)
It must first be noted that the primitive elements are determined by
objective means, not by meaning or even by meaning-differences. The
phonemes, which carry no meaning, are obtained by a behavioral test of
what constitutes repetition in the speech community (7.1), and the words by
a stochastic process on phoneme-sequences in utterances (7.3). Given this,
it has been seen that the syntactic relations, which carry such fundamental
grammatical meanings as being predicate, or being subject, (a) do not have
to be invented or discovered as underivable sui generis properties of
grammar, but (b) can be derived from constraints on word co-occurrence.
In (a) such relations as being subject of a verb are necessarily unique to
language and cannot be compared with anything else. In contrast, for (b),
the dependence constraint is a kind of dependence that we can seek in other
systems and in other types of activity.
The syntactic and informational structure of language contributes to
understanding its development (12.3). The operator constraint gives the
sentential occurrence of each operator in terms of the occurrence of its
arguments, but not vice versa. Thus in the descriptive order of sentence¬
making, the operator classes are 'later' than their argument classes. This
descriptive order has some bearing on the development of language, in
particular on the early development of sentences and syntax. It says nothing
about the existence of words independently of sentences. Many words which
later became arguments, and many which later became operators, must
have been used in pre-sentence times, said separately or in various com¬
binations. But the coming into existence of sentences as explicit kinds of
word combination, with explicit sentential (predicational) meaning for the
combination, arose out of certain words coming to require, for their entry
into a combination, the presence of words from stated sets or else of null.
The order in which the definitions of syntactic classes refer to other syntactic
classes, which is the elementary order of sentence-making, may have a
Overview 25
relation to the historical order of the syntactic classes: words may not have
become established at a particular operator level before words of their
argument levels have become established as being such.
In this developmental picture, the dependence may well have arisen not as
a structure, but as a preference: simply a matter of certain words A being
more likely to be used in association with words of some particular set B than
alone (or than with other words A). However, once this likelihood of
occurrence became conventionalized, with some B present or implicit (via
zeroing) whenever an A appears, the requirement became a (sentential)
structure and gave rise to predication as the meaning of that word relation
(12.3.5). Such conventionalization of a use-likelihood into a requirement,
with an attendant grammatical meaning, is common in grammar, for example
in the distinction between verbs and adjectives (G189).
All other syntactic relations can be obtained from the operator constraint.
And since the meaning of that constraint is ‘saying about’—the operator is
said about its argument (1.4)—one can see that the constraint contributed
a successful form of juxtaposing words so as to give information. This is
of importance for understanding the development of language. For it is
obvious that language has to have developed from a state in which there was
no language. The problem is not resolved by attributing language to a
specifically linguistic mode of neural processing, since it is even harder to
understand how such a mode of processing arose before language existed.
Furthermore there would remain the question of why the neural processing,
and language itself, has the particular grammatical dependence-constraint
that it universally has. In contrast, once that constraint exists, its learning
requires no unique preconditioning: one need only hear a great number of
sentences constructed by this dependence and thereupon construct further
juxtapositions of words that satisfy the same dependence.
The same operation which creates the elementary sentences is then
recursively applied to make extended sentences. Other sentences are
further derived from existing ones by other conventionalizations of language
use.
Sentences are obtained, and defined, as segments of speech within which
the operator-argument relations, and most reductions, hold. Within a
sentence there may be transformations. Phenomenologically, a sentence
transformation appears as a relation (a constraint) among observed sentences.
However, as a process it is merely a change of shape (primarily a reduction)
within a sentence (8.1.7, 9.0), yielding not another sentence but another
shape of the given sentence. Beyond this, there are also constraints among
sentences. These are necessarily quite different from transformations: they
are preferences or requirements for word repetition in certain operator-
argument relations to each other within successive sentences (e.g. discourses)
and within other special aggregates of sentences (sublanguages). Hence
each sentence, discourse, and sublanguage is a structure, the locus of
26 Introduction
specifiable constraints (comparable in this sense to a biological organism).

We may be able to say this also for the successive discourses over time in a
particular science. In contrast, the set of sentences in a language, and the set
of discourses in it, are only aggregates that are not defined by crucial
relations over the members of the set (comparable, in a way, to community
in biology).
Even such recent developments as sublanguages are obtained by existing
linguistic relations operating in new circumstances. Throughout it is seen
that language is not a single object or a one-time creation, but a system
evolving by accretion, one which has grown out of self-organizing factors
and out of the language-survival value of effectiveness in transmitting
information (12.4.2-4).
We see here a few independent kinds of structure common to all languages:
(1) the setting up of particular phoneme-sequences as words or morphemes,
in a way unrelated to syntax; (2) syntax i.e. the partially ordered dependence
in word co-occurrences, which can be broken up into dependence on word-
class presence (which carries the meaning ‘predication’) and the individual
word likelihoods within this dependence relation (‘selection’, which carries
word meaning); (3) having phonemic variants (reductions) within words in
the dependence relation; (4) the metalinguistic capacity created by the
independence of (1) from (2) (cf. 5.1); (5) change; (6) analogic processes;
and (7) room in each individual language (or language family) for special
structures based largely on correlations between (1) and (2), or between (3)
and the domains of words for which these variants hold.
It may be noted that the dependence relation and likelihoods of (2) above
are the crucial steps that create sentences. They are not only universal but
also semantically functional, and are learnable as a generalization or extension
of simple properties which are exhibited in every sentential word combination
(and overtly in unreduced sentences). The rest of syntax is not in detail
universal among languages, not simple, and can only be acquired by learning
from repeated hearing of sentences and sentence fragments. Such are
the specific phonemic sound-sequences which constitute words with their
approximate meanings, and the specific reductions with their specific con¬
ditions in each language. Finally, there are certain properties of language,
generally neither universal nor directly functional or information-bearing,
which may facilitate the speaker’s learning to speak or the hearer’s grasp of a
sentence, but which do not have to be learned as such, as additional
properties, by the users of a language. Such properties are: the distinction
between morphology and syntax (a matter of certain particular kinds of
similarities and of reductions), word-class distinctions other than between
argument and operator (e.g. between verb and adjective, or between words
of the same root), patternings of reduction which create conjugations and
declensions, and the like.
In Chapter 3 it will be seen that given a set of base sentences of a language
Overview 27
produced by the dependence and likelihood constraints (and normal lineariza¬

tion) of (2) above, the reductions and transformations (and alternative
linearizations) of (3) above produce the rest of the sentences in a way that
does not change the dependence and likelihood relations among the words,
and thus does not change the substantive information of the source sentence.
The processes that make sentences are thus separable into the fundamental
informational constraints of dependence and likelihood, and the secondary
paraphrastic constraints of morphophonemics, i.e. of the phonemic shapes
and the relative locations of words without changing their dependence
relations in the sentence. The dependence (and likelihood) relation is close
to the view that language has an operator-argument structure, which is
commonly held because of the pervasive verb-noun relation in the world’s
languages. But from the analysis of the other processes (excluding analogy)
as morphophonemics it transpires that the complexity and patterning of
languages are not due to the fundamental constraint.
First, only a part of the notorious complexity of language arises, as
by-product, from the dependence constraint (i.e. from resultants of succes¬
sive applications of the operator-argument relation), or from likelihood
(i.e. from similarities and contrasts among the selection-domains of each
word). The major complexities in many languages arise from the institu¬
tionalization of differences (as in the verb-adjective distinction in some
languages, 12.3.5), and in morphophonemic similarities as in the different
declensions for different subsets of nouns. Second, the widespread pattern¬
ing that yields much of the special character of different languages turns out
to be not a necessary component of information and communication, nor an
unavoidable by-product of the operator-argument relation, but a systemat¬
ization of morphophonemic similarities and contrasts—something which
has the properties of culturally set behavior and of standardized patterns of
decoration (8.4).
Thus language structure has some universal properties relevant to speaking
(e.g. phonemes, linearity), and some relevant to information (e.g. pre¬
dication), some not-universal properties which may be convenient for
information but are not essential to it (e.g. grammatical structures for time,
quantity, or aspect), and some properties, different for different languages,
which may be irrelevant or even distracting for information (e.g. in some
cases declensions and conjugations). These distinct types of structure arise
in appropriately distinguished ways in the theory presented here.
In addition to all this there are major issues concerning language which
are tangential to language structure as discussed here, but related to it. Such
are: the regularities of language change (comparative, and more generally
historical, linguistics); the relation of time-slice (synchronic) structure to
change (diachronic); the differences between the speaker’s grammar and
the hearer’s; the arising of geographic and social dialectal differences; the
use of language for purposes of art.
28 Introduction
There are also many areas of investigation which are related to language
structure as presented here, but remain to be undertaken. For example:
Ongoing cases of institutionalization of language use;
Specifics on the motive forces and directions for the evolving of language
structures;
The relation of the mechanisms of historical change to evolving structures;
Does the dependence-on-dependence relation indeed apply to all known
natural languages, as it seems to?
How does the operator-argument structure compare with the old deaf
sign-language which is not based on knowledge of spoken language?
Detailed descriptions, in terms of the dependence presented here, of
major language families, and of little-known languages such as those of
Australia and Borneo;
Detailed descriptions of reductions in various language families; many
details remain to be worked out in English too;
Special problems in deriving morphemes from operator-arguments and
reductions: affixes, pronouns, prepositions, particles;
Relatively recent development of such problem morphemes: e.g. articles;
The regularization of the double array in discourses (9.3);
Argumentation and consequence as constraints on sentence-sequences in
science (10.5.4).
In sum, the most general theoretical finding reported here is that the word
combinations which constitute sentences can be characterized not by a
sequential relation but by a partial ordering relation on words, with all
grammatical relations being defined in terms of this partial ordering. The
most general grammatical finding is that if we take, as primitive events on
the sentences of a language, a set of statable reductions in the shapes of
words which have high expectancy in their sentences, then the residual form
of each sentence, after undoing any reductions in it, is an overt fixed
operator-argument relation among its words (the above partial ordering).
The most general informational finding is that every grammatical relation
has a fixed meaning, and that the constraints which create the partial
ordering also create its information.6
f1 At various points, the conclusions which are reached here turn out to be similar to well-
known views of language. For example, the operator-argument relation has similarities to the
predicate structure in Aristotelian logic. The way it functions in the formation of sentences has
similarities to functors (and more generally to the deep understanding of language theory) in
the categorial grammar of S. Lesniewski, followed by K. Ajdukiewicz and others in the Polish
School of logic; cf. particularly W. Buszkowsi, W. Mariszewski, J. van Benthem, eds.,
Categorial Grammar, Benjamins, Amsterdam, 1988; H. Hiz, ‘The Intuitions of Grammatical
Categories’, Methodos (I960) 311-19 and 'Grammar Logicism’, The Monist 51 (1967) 110-27;
J. Lchrberger, Functor Analysis of Natural Language, Mouton, The Hague, 1974. Various
conclusions about structure and meaning here appear similar to, or inverses of (this too being a
similarity), arguments made by L. Wittgenstein. And the analysis of science sublanguages
Overview 29
addresses problems posed by R. Carnap concerning a language of science. Also, various

statements about meaning (in ch. 11) accord with common knowledge or opinion. The intent of
this work, however, was not so much to arrive at such conclusions, as to arrive at them from first
principles, from a single method which is proposed here as being essential because of the lack of
an external metalanguage. The issue was not so much what was so, as whether and how the
essential properties of language made it so.
2
Method
2.0. Introduction
Linguistics has not in general been one of the sciences in which the relevance
and correctness of statements are determined by controlled methods of
observation and argumentation. It is therefore desirable to consider what
methods are relevant in linguistics, in the hope of establishing criteria for
investigation and analysis. Choice of method is not less important than
responsibility in data, and the choice should be determined not by personal
preference or current custom but by the nature and the problems of the data.
Furthermore, it will be seen in 2.2 that it is not enough to go by the separate
observable features of the data, such as their linearity or their carrying
meaning, but rather that one has to consider the regularities of all the data,
and their interdependence.
2.1. The issue to be resolved:

Departures from equiprobability
Chapters 3 and 4 present a theory of language structure, sketched in 1.1.

This theory is a conclusion which has devolved from analyzing language by a
particular method. The choice of method arose out of the choice of problem,
that is, out of choosing what issues were to be resolved. There are various
issues, and types of data, which one might consider in seeking to investigate
language. One is evidence as to the speaker’s relation to what is said—his
intent and how the sentence was put together. Another is the hearer's
(reader's) perception of what is said—how it is decoded and what he
understands in it. A third is the meaning of what is said. A fourth is the
situational and behavioral context—what situations and events correlate
with that which is said, and what are its effects, both immediate and in a
broader cultural background. Lastly, there is the form of language—the flow
of sound-waves which constitute each event of speaking; alternatively, any
adequate representation or equivalent of it, either by instruments such as
the sound spectrograph or by writing. This last type of data, as sequences of
sounds or letters, is observable separately from the intent, perception,
meaning, or context. It is sufficient for carrying the great bulk of information
Method 31
carried by language, as witness the adequacy of writing even though it lacks

the direct support of speech properties or personal contact or extra-linguistic
context. Within present knowledge, the form of utterances, i.e. the sound-
sequences and written marks, is the one aspect of language that is observable
precisely, fully, and independently of other features of speech and language.
There is reason therefore to address first of all the form of utterances, and
to consider what methods are appropriate for studying it. Our objective is a
precise theory which can analyze any sample of utterances (chs. 7-8), and
can predict the structure of utterances (chs. 3-5) in a given language in a
given period—and not predict more than is the case. It will be seen that in
obtaining this we also obtain a record of the information in each analyzed
utterance.
This is not to say that there is not additional knowledge which can be
gleaned from other types of language data. If we knew all there was to know
about the intent, perception, and context of each event of speaking or
writing, we would doubtless learn more about the utterance, over and above
what we can learn from analyzing the form. But we can obtain such
knowledge only episodically and for the most part vaguely. And, when we
analyze form, to use such knowledge only where we happen to have it
interferes with reaching a systematic analysis of the form and the information it
carries (e.g. if we explain the auxiliary verbs by the need to state modalities
of actions, in contrast to combinatorial and likelihood factors as in G298).
It must nevertheless be noted that the data about the form of utterances
are fuzzy in marginal cases. There are utterances for which it is not clear
whether they are in the language’ or not—-speakers may disagree, or not be
sure. And there are utterances which have such different meanings that the
question arises whether they are not different utterances though with the
same form. Here as in general, it will be seen that a theory constructed first
of all to fit all such utterances as present no such problems can be then used
to account for the marginal and problematic utterances.
2.2. The relevant method: Least redundancy in description
To investigate the form of utterances, we consider the entities involved and

their change or combination. We note immediately that no matter what the
entities, some combinations do not occur in utterances: tpiks and musp are
not English words, nor is slept the a the an English sentence. Clearly there
are departures from equiprobability in the combinations.
In language such departures from equiprobability are not only universal
but necessary, because the whole set of human languages has no external
metalanguage. This is a crucial consideration. For in order to describe the
form of a language—or of any system—we have to identify its parts or
properties, their changes, and the relations among them. In the case of logic
32 Introduction
and mathematics, the statements and formulas which are comprised in them
have such explicit forms—-of symbols and their combinations—that it is
clear that statements able to describe logical and mathematical formulas do
not themselves have the structure of those formulas. The statements that
describe a system are in a different system, called the metalanguage—in this
case the metalanguage of logic or mathematics—which is richer in certain
respects than the system it is describing. In the case of natural language, we
have no different system in which the elements and combinations of language
can be identified and described. The identifications for a corpus of language
utterances can indeed be made, but only in the sentences of a natural
language (the same or another)—that is, in sentences which have already
been built out of the same kind of elements and combinations which they are
describing (5.1). We cannot describe the structure of sentences except in a
system which itself has sentences, with predications and word-likelihood
differences. Furthermore, no external metalanguage is available to the child
learning its first language, nor to early man at the time language was coming
to be.
In the absence of an external metalanguage, one could seek to identify the
entities of language by extra-linguistic means, such as the occurrence of
words in life circumstances that exhibit their meaning. Many words may
indeed be identified on such grounds. But many other words, and the
ordering which makes out of them sentences as against ungrammatical word
collections, cannot be thus identified. In contrast, when only a small
percentage of all possible sound-sequences actually occurs in utterances,
one can identify the boundaries of words, and their relative likelihoods,
from their sentential environment; this, even if one was not told (in words)
that there exist such things as words (7.3). And when identified words
combine with each other in relatively few regular ways which are used
throughout the course of an utterance, one can recognize utterances, long
and short, in distinction to the non-occurring sequences of words; and one
can recognize within long utterances segments (sentences) identical to
otherwise-observed short utterances. This holds even if one is not told that
there exist such things as sentences (7.7). Given, then, the absence of an
external metalanguage, it is an essential property of language that the
combinations of words in utterances are not all equiprobable, and in point of
fact that many combinations do not appear at all.
It follows that whatever else there is to be said about the form of language,
a fundamental task is to state the departures from equiprobability in sound-
and word-sequences. Certain problems arise. No one has succeeded in
characterizing the departures from equiprobability by fixed transitional
probabilities between successive words of an utterance; the reason turns out
to be that the overall characterization of word-configurations is a partial
ordering of words (3.1). Furthermore, since the set of all possible word-
sequences in a language is in principle denumerabiy infinite while the
Method 33
grammar of a language must be finite (2.5.1), grammar as a catalog of the

combinations which occur in utterances is not a practical program. Finally,
many combinations (utterances) are marginal: it is not clear whether they
are or are not in the language. All these difficulties can be circumvented by
seeking out not a list of attested combination but a set of constraints on
combination (1.1). A few constraints can together preclude unboundedly
many word-sequences while allowing others; and constraints stated on
different or partially fuzzy word domains can create marginal forms whose
satisfaction of the constraints may then be in question.
If we obtain the occurring combinations, as departures from equiprobability,
by stating the constraints that allow them, we are specifying (in the con¬
straints) the various contributions to non-equiprobability that together
make up the total non-equiprobability of the system. It is therefore clear that
the constraints as formulated should not introduce any additional redundancy,
over and above the least needed to account for what is actually present; for
the non-equiprobability of language is precisely what we are trying to
describe. As a simple example of appealing to fewest constraints, consider
the difference between walk, walked and go, went. In walk, we have one of a
large set of operators which occur with the present, future, and past tenses.
In go, we have a single word having all properties of the operator set except
that it does not in general occur with past-time adverbs such as yesterday,
and in went, we have a word with the operator properties but which occurs
only with past-time adverbs. This case requires us to partition the operator
class into three subsets (go, went, and the others); this partition correlates
with no different properties of that class. In contrast, we can say that all
operators occur with all time-adverbs (except for certain quite other problems
with ‘auxiliary verbs’ such as can, used to), but some words have different
phonemic content in different environments, something which is known for
quite a few words in different cases (5.4). In the latter case, the theory has
to carry an intermediate (‘morphophonemic’) level of elements between
phoneme and word: certain phoneme-sequences (e.g. walk, go, wen-) are
bounded by word or morpheme boundaries, i.e. are word-size segments;
and every word is a set of one such segment (e.g. walk) or of a disjunction of
‘alternant’ such segments (e.g. go/wen-). But this level, of morphophonemes
and word-alternants (7.4), is needed for many words in various situations,
and the constraint involved in specifying it for go affects a far smaller
domain than the constraint involved in partitioning the operator class to
suit go.
If, then, we have for a language different descriptions, adequate to
characterize its utterances, but with different amounts of departures from
equiprobability ascribed at various points in the course of the descriptions,
we opt for the one with least such departures, since that one has clearly
added least to the inherent departures in the language being described. The
consideration of least non-randomness in description is involved in the
34 Introduction
discovery of phonemes, which underlies the whole development of a science^

of language, and it leads to many other methods in the analysis of language.
The effect of the least-redundancy test, when applied at each point in the
language description, is a grammar with fewest possible and most independent
objects (elements and constructions), fewest and least-intersecting classes of
objects, fewest and most independent rules (relations, transformations,
etc.) on the objects, fewest differences in domain for the rules, and finally,
fewest abstract constructs. Such a grammar is least redundant because each
object (other than arbitrary ones defined only by their relations), and each
rule or relation and classification, and each statement of domain for these, is
a constraint on the equiprobability of parts in the utterances which are being
described.
At each analysis of a linguistic form one would therefore seek to add as
little as possible to the system describing the forms thus far analyzed.
Certainly, one would not repeat information already given. For example, if
a given phoneme- or letter-sequence (say, attempt) has been stated to have a
particular meaning and particular other words as high-likelihood (‘selectional’)
arguments, one would avoid giving the information twice: once when the
word appears in verb position (They attempt to stop smoking) and again
when in a noun position (Their attempt to stop smoking failed); the preferred
alternative would be to take the second form as being the same sentence
‘nominalized’ under a further operator.
More generally, occurrences of a word-size segment in two different
environments are not given two different grammatical classifications, if
there is any regularity in these two environment-differences, that is, if
we find other cases of a word-size segment in two comparably different
environments (7.4). For in such cases, one could formulate a derivation that
would carry each of these words from one environment to the other, without
claiming that a given segment constitutes a different word in each different
environment. This avoidance of multiple classification applies to many
situations: one case is the noun-verb difference exemplified by attempt,
above, where one can note that when a sentence (They attempt to escape)
becomes the argument of a further operator (e.g. succeeded) its verb takes
on noun form and the arguments of that verb receive adjectival forms
(Their attempts to escape finally succeeded). That is to say, nominalization
(i.e. occurrence in noun—‘argument’—position) happens not to verbs (or
operators in general) but to sentences: nominalized verbs are not a word
class, but nominalized sentences are the form that sentences take when they
are arguments of a further operator. Thus The attempt succeeded is a
reduction from Someone's attempt succeeded, or the like. In another case,
instead of saying that eat, write, etc., are each both transitive and intransitive
(John writes poetry, (1) John writes), we say that under these transitive verbs
the object is zeroable when it is an indefinite noun (John writes things is
reduced to (1)); this zeroing is widely attested elsewhere.
Method 35
There remain cases where certain occurrences of a word-size segment

have no recognizable similarity to other occurrences in respect to their
selection of high-likelihood neighbors; here we cannot avoid assigning the
different occurrences to different homonymous (i.e. same-phoneme) words.
This can occur when the segments are in the same gross syntactic positions,
as in the two nouns heart and hart, or in different ones, as in the noun sea and
the verb see (or a prune and to prune). Homonymy may be the most efficient
analysis even in some cases where the different occurrences are etymologically
identical but differ grossly in selection (and so in meaning), as in a smart
fellow as against a smart blow and He smarted from the blow, or in a humor
magazine as against the four humors of the body.
The same minimalizing considerations that go into identifying the elements
go also into identifying the relations among them, these identifications being
in any case two sides of the same coin. The analysis of a language form
should make maximal use of analyses used for other forms, avoiding as far as
possible recourse to new (ad hoc) classes, rules, domains. To explain a form,
then, is to find its maximal similarity to the other forms, with the whole
grammar thus becoming a best fit for the data. This means analyzing each
form in respect to the whole relevant grammar—and the grammar of a
recognizable linguistic community, not of an arbitrary speaker (since language
is essentially a public structure).
Making a least grammar for an unbounded set of sentences entails that as
many words and sentences as possible (and their meanings) be derived from
others (by changes and additions), preserving whatever has gone into the
prior making of those others. And it entails that the derivation be in small
steps repeated in various combinations rather than large steps each of which
accounts for a whole difference between starting-point and end-point. A
large step, such as the totality of differences between the active and passive
forms of the verb, can be used only at one point of the grammar—in this
case, to derive the passive. But if we can obtain its effects from a succession
of small steps (where the resultants of one step are in the domain of the
next), we may find that the same small steps occur elsewhere in the
grammar, in different combinations (for the passive: 1.1.2, G362). We thus
seek to have fewest derivational steps, maximally repeated in different
applications. An example is seen in the elementary transformations of 8.1.
Given these considerations, when we find a word—with more or less the
same phonemes and with clearly related meanings—appearing in different
grammatical statuses, we do not simply give it a multiple classification, but
rather, as noted above, use one status as derivational source for all the
others. Examples are the many words which appear in English as noun and
verb (e.g. to house from a house, and the day's take from to take); or as verb
with and without object, or with different object structures (e.g. read from
read things, and expect John from expect John to come); and the many
environments and uses of the prepositions (G80). Also, if words or morphemes
36 Introduction
undergo fixed kinds of meaning changes (lexical shift’) in different gram¬

matical conditions, we would try to attribute the meaning-shift to the
changed grammatical conditions. Examples are abstract nouns as being
nominalized verbs, more precisely as nominalized sentences whose subject
and object have been zeroed (e.g. growth from the growing of things)and
metaphors (4.3) from paired sentences under the connective as (e.g. I’ll buy
your argument, roughly from I'll take your argument as one takes things when
one buys them).
In all this, it should be noted that no assumption is made that there exists a
structure in language, and no appeal is made to any particular principles of
structuralism. Such structure as is found comes out in the process of making
a least description.
2.3. Simplicity
The need to minimize redundancy in grammatical description means that

what may informally be called the overall simplicity of the grammar is not
only a matter of elegance or science policy, but a methodological require¬
ment. Furthermore, since language necessarily has a constructive charac¬
terization (6.0, 2.5), the simpler the grammar, the fewer different processes
are needed in toto to reach all sentences. The total of simplicity is in large
measure the outcome of specific contributions to simplicity, each due to
specific conditions in the data. For example, if a sentence undergoes several
expansions (additions) and reductions (or other transformations), it does
not have to be assumed that they all take place in one order or another.
Some may be defined on the same form and thus take place simultaneously,
independently of each other. Also, since language contains not only long
sentences but also short ones, in which the possibilities of combination
among the entities are fewer, one can seek a simpler system by trying to
analyze the longer sentences in the same way as the short ones, or as
combinations of the short ones.
Yet another example arises from the fact that since languages change
through time, there may always be some forms which do not fit perfectly into
the current grammar. The total system is then simpler to the extent that it
minimizes the disturbance due to such on-going changes. This suggests that
rules be so stated as to be able to house current derivations in repect to
combining-possibilities, phonemic shape, position, etc.—formulations in
which the relevant past or the immediate future of forms would be minor
changes in condition or domain. A major simplification can be achieved in
' In meaning, growth may seem to have also other sources, as when it means ‘amount of
growing’. However, the other meanings can be derived from the same growing with zeroed or
implicit amount (of growing), product (of growing), etc.
Method 37
many cases by considering not just a form but also its ‘behavior’, i.e. the
environment and other factors which characterize its occurrences. It has
been noted here that the general ‘behavior’ of words is such that it is easier to
state the constraints which preclude certain types of word combination than
to state all the types of word combinations that can appear in a language.
However, simplifications in any one part of the grammar have to be
weighed against their effect in other parts of the grammar. To see this in
practice requires a detailed example such as the following: The description
of sounds is simplified when we collect sounds into phonemes, roughly on
the condition that these sounds have complementary environing sounds
(7.2.1). On this basis, English /?, which never occurs at the end of a syllable,
and the ng sound (as in sing), which occurs only at the end of a syllable, could
be considered members of one phoneme, Lng/h\ (Syllables are phonetic,
non-syntactic, restrictions on phoneme-sequences. ) But this would disturb
the phonemic similarity (7.3) of certain words when these occur in different
combinations. For example, long ends in the ng sound, but longer can be
pronounced either with ng alone or with the ng sound followed by g; and the
related word longitude is pronounced only with n followed by j. If we take ng
as a member of an ng/h phoneme, then long and longing and longer contain
the ng/h phoneme, while longer (with pronounced g) contains different
phonemes, n and g, with longitude containing n and j phonemes. If,
however, we take the ng as a sound restricted to syllable end (with no
relation to //), it can then be analyzed as a syllable-end composite of the n
phoneme followed by the g phoneme. Then the two-phoneme sequence ng
at word-end usually has the single sound ng(long, longing, longer), while the
other pronunciation of longer, and longitude, can be related to the presence
of the same n and g (or j replacing g) phonemes. The second analysis fails to
maximize the collecting of complementary sounds into phonemes, but gives
a simpler phonetic characterization of phonemes and a simpler phonemic
composition (‘morphophonemics’) to words such as long in their various
combinations. (Note that the reason against ng/h was not the lack of
phonetic similarity between the sounds but the combinatorial simplicity at
the next level, that of words.)
There are many other situations in which consideration of systemic
simplicity affects the analysis of a given form. In many cases, it is possible to
find an analysis in which both elements and the constraints on their com¬
binations are simplified. Thus the classification of phonemic distinctions into
phonemes, of word-size segments into words, of words into operator classes,
are all done not only for reasons of taxonomy, with the general simplicity
advantages that taxonomy provides, but much more, explicitly, because
these classifications plus the constraints stated on the classes (phonemes,
words, operators) are together far simpler and fewer than the constraints
that would have to be stated on the individual phonemic distinctions, word-
size segments, and individual operator words. The sets that are defined in
38 Introduction
the methods discussed here are thus characterized by relations among their
members (cf. 2.5.5).
Nevertheless, there are also forms for which any simplification in one
respect involves complicating some other part of the description. This means
that the relevant simplicity has to be measured not in respect to the analysis
of each form separately, but in respect to the whole system that describes or
predicts the utterances of the language. It includes the simplicity of the
primitive elements and their properties, and that of their operations or
relations, and of the types of domains on which these latter are defined—
and also the simplicity of the relation between the forms and their meanings,
which is an essential concomitant of them.2
The simplicity of the system is greater if any property or analysis that is
stated at one point in the description of a construction is preserved under
later developments in it. Every preserving of a property saves having to state
it separately for the different occurrences of the words. We try to find
elements and relations such that our analyzing a construction in terms of
them is not affected by further processes acting upon it.
In particular, no analysis of a construction should have to be corrected by
the analysis of a later element operating on that construction. For example,
(1) A whale is not a fish should not have to be derived from an assertion
(under sentence intonation) (2) A whale is a fish to which is added (3) not,
denying that assertion. Also, given the true (4) Lead paint is dangerous, we
would not want this assertion about paint to be derived from the false (5)
Paint is dangerous which is an assertion about paint in general, plus the true
(6) The paint has lead (or the equivalent).2 This does not mean that we
forego relating (1) to (2) or (4) to (5). Rather, we obtain the sentence (1) A
whale is not a fish from (3') One denies (or / say the denial of) operating on
the unasserted intonation-less sentential construction (2') A whale is a fish
(G321). And, for other reasons, we obtain Lead paint is dangerous from
Something, which is paint which has lead, is dangerous from (5') Something
is dangerous to which is adjoined (6') This something is lead paint from This
something is paint; this paint has lead (G124). It is a contribution to
simplicity to have each step in the making and analysis and interpretation of
a language construction be accretional, as in the primed derivations above,
and not require a reconsideration of what has been done up to that point.
A more complicated case of this parsimony of analysis is the avoidance, in
analyzing a form, of going beyond what the language itself has distinguished.
For example, in such pronouns as he, it is not possible, except in special
Since each relation has a specifiable domain (in some cases fuzzy) and each acts only on the
resultants of prior relations (down to the sounds), we can in principle measure the effect of each
relation, or each proposed analysis, in the whole grammar (10.4.5)—and even which proposed
grammar is ‘least’.
1 The truth of a sentence is a special case of its meaning, relating the information of a
sentence (11.6) to axioms or statements believed to be indeed the case (10.5.4). Unlike
meaning in general, it is not a known property for every sentence.
Method 39
contexts, to determine uniquely from what is said who of several preceding

individual human males is being referred to. A simplest grammar would not
try to outdo language by creating a set of determinative readings, one for
each possible antecedent, which may then be narrowed down by later
contextual material. (Each reading would then have a definite antecedent in
the way that the language as spoken does not have.) That such alternatives
would take us beyond the structure of the language becomes evident when it
is found (5.2) that in all cases where grammar has definitely located
antecedents for pronouns, it has done so by special devices (e.g. bringing the
pronoun to immediately after the antecedent). This means that fixed ante¬
cedents are provided for by creating special easily recognized relative
locations; and where this has not been provided, it may be judged that the
uncertainty as to antecedent is not due to there being a disjunction of
determined antecedents, but rather to there not being determined ante¬
cedents at all in the given grammatical case. He then means ‘someone
mentioned’ rather than ‘this or that one or the other’.
By the same token, although language has in principle the addressing
capacity of stating the relative linear location of every phoneme and word in
a sentence, this capacity is not part of a grammar, because universally it is
not used. Indeed, languages use words for referring to easily recognized
locations, e.g. latter, without any system of overall addressing. To reject a
description that overreaches language means that a theory and notation
should at no point produce forms of distinctions beyond what language can
contain, just as it must not produce less than what language contains. This
suggests a theory fitted uniquely to language data (including its marginalities),
neither more nor less, rather than a more powerful system which must then
be whittled down by separately stated conditions.
A final consideration of simplicity: in addition to the internal simplicity of
the grammar, one must require that there be least metalinguistic burden.
This does not refer to the simplicity of the structure of the metalanguage.
The metalanguage will be seen in 10.2 to be a sublanguage of the natural
language, one whose grammar is partly different from that of natural
language. We are here not speaking of its own structure, but of the
discourses in it: how much has to be said in it about the language which it
describes. We can avoid special metalinguistic comments about special
events (forms and their ‘behavior’) in the language if we can obtain these
events from the language grammar (which is in the metalanguage) with no
addition to it, or from an addition which may be useful for other possibilities
in the language. For example, kerb can be written (primarily in British
English) for ‘paving curb’ but not for other occurrences of curb. This
peculiarity can be included in the grammar if we say that by the side of the
environmentally limited morphophonemic (sound) alternants of a word,
such as go and wen- above (2.2), there may also be environmentally limited
spelling variants, such as kerb in particular word-environments. This adds
40 Introduction
nothing to the elementary vocabulary of the metalanguage. For another

example, Edward Sapir had once noted what would seem to be a life-
situation restriction on word use: only a person who had studied mathematics at
school can without dissimulation call it math. Here we can say that math is a
morphophonemic alternant not for mathematics but for mathematics as
known from courses; this provides the same restriction for the word use
without leaving the confines of the ordinary grammar. In both examples the
solution involves specification of the word’s environment, something which
is already known to be relevant.
Considerations of metalinguistic burden suggest cautions against both
reductionist and teleological formulations in a science. In the case of
reductionism, a given science is stated in terms of a prior science of it—one
whose resultants yield the elements of the given science—without indicating
that these elements have wider capacities than the prior science could have
provided for. If indeed such wider capacities exist, that fact cannot be
disregarded, and if not included in the formulations of the given science that
fact would have to be left for informal expression in its metalanguage.
In the case of teleological terminology, the entities of a given science are
named not solely by the processes that create them but rather by their end
product, in particular by what they lead to in a successor science or field, one
to which the given science is prior. While this end product may be of central
import to the successor science, it may not play a role in the processes of the
given prior science. Hence any naming or explanation made in end-product
terms within a given science is a burden either on the grammar of that
science or on its metalanguage.
2.4. Can meaning be utilized?
Language is clearly a carrier of meanings, both in its words and in the

particular ways the words are combined. In considering methods for analyzing
it, one might think that the forms—the words and constructions—of language,
and the meanings which they carry, can be described each independently of
the other. However, there is a feature of language which requires that the
two be interconnected. This is again the absence of an external metalanguage.
It has been seen in 2.2 that this lack makes it impossible to identify all of the
elements in the How of speech without noting the constraints on their
combination. We will see here that this same lack makes it necessary to refer
to constraints on combination also in specifying the meanings of some of the
elements.
The reason is that there is no way of defining the meaning of all of the
words, and of the constructions created by hierarchies of constraints on their
combination, except in language. Other means, such as pointing to the
referent of a word, or understanding the meaning of a word or sentence from
Method 41
the situation in which it is said, are undoubtably adequate, and essential, for
grasping the meanings of many words and short sentences. But they cannot
suffice for every word, nor for its different meanings in different combinations.
The meaning of such words as time, consider, the, of, and the secondary-
sentence meaning of the relative clause structure, cannot be adequately
gathered from the real-world situations in which they are used. Also, any
metalanguage in which their meaning could be given would have to have
already for its own words the kinds of meanings which are being explained.
Hence, these less obvious meanings can be learned—by speakers of the
language and by analysts of language—only by much experience with the
neighboring words and sentences with which they occur. Indeed, when one
learns the meaning of a word from its dictionary definitions, one is learning
from its sentential environment—the words of the definition and the examples
of usage—and not from an extra-linguistic experience with the word.
Furthermore, many words vary the combinations into which they enter, and
with that their meaning. We cannot in general learn the change in co-
occurrents from a prior change in meaning, but we can learn the change in
meaning from seeing the change in co-occurrents: e.g. meat from earlier
‘food’, fond from earlier ‘foolish’, provided and providing in their new
conjunctional meaning ‘if (11.1).
It follows that for an adequate knowledge of meanings in language, we
need to know the constructions, in addition to an unspecified amount of
semantic information garnered from the situations in which earlier-heard
words and sentences had been said. Linguists have discussed when to bring
meanings into the analysis of language—before, after, or simultaneously
with the grammar. However, such a choice does not really arise. This is so
because the meanings that are expressed via language (and even more so
those which are inexpressible in language) are not in general discrete and
definite items which can be mapped onto the words of linguistic constructions.
There are a few sets of meanings which we can know precisely independently
of language: for example, numbers, biological kinship relations, the spatial
coordinates as we experience them. In these cases, we could start with the
meanings and ask how a language expresses them. But this method cannot
be extended to the vast bulk of meanings that correspond to words and
constructions of a language. There, we cannot list a priori a set of meanings
covering the relevant experiences of a given community. We have to see first
what the vocabulary of the language is and then how it is used, in what
combinations and in what constructions. We cannot say a priori what objects
or actions or relations are included in one meaning as against another: we
have to see what words combine with chair as against bench, or with slip as
against slide, or with from as against out of.
In addition to being not sufficiently knowable independently of the word
combinations in a language, the meanings are not precise enough to specify
all combinations of a word, and are not sufficient to predict how words will
42 Introduction
combine. For example, from any purely semantic definition of full, empty,
equal, one would not expect that the language would contain fuller, fullest,
and emptier, emptiest, while more equal does not occur in ordinary English
(though it is used in George Orwell’s Animal Farm, without taking us
outside of the English language). A person, e.g. a foreigner, who selects
words for his sentences purely on the basis of even the fullest dictionary
definitions will produce many combinations that everyone would consider
wrong, and would miss many normal combinations.
Not only is it untenable to use meaning as the general framework for
language analysis, we also cannot use it as an occasional criterion in deciding
how to analyze linguistic sentence structure, for example, at points where
finding a simple combinatorial constraint is difficult. For one thing, the
decision where to use it would be arbitrary and would be different for
various investigators. For another, as Leonard Bloomfield pointed out,
appeal to meaning when formal analysis is difficult cuts off further attempts
to find a formal ‘explanation’ (in the sense of 2.2). When one perseveres in
purely combinatorial investigation, satisfactory explanations are reachable
in various cases where meaning seemed an easier way out—for example, in
the analysis of pronouns (5.3.2, 2.3) and of metaphor (4.2). More importantly,
when the whole language is analyzed consistently in terms of constraints on
combination, with no freedom for the investigator to beg off from that
method, language is found to be built out of a few virtually exceptionless
rules whose force does not appear when we look only at partial analyses of
language, or when constraints on combination are referred to only sporadic¬
ally. Also, the grammatical constructions that are obtained in a purely
formal analysis turn out to have sharper correlation with meanings than do
those recognized in ordinary partially semantic grammars (cf. the last
paragraph of 2.4).
Meanings in language, then, are meanings of words and meanings of
constructions—those words and constructions which are in the language,
and which have been known before we ask their meanings. (And indeed, the
boundaries and identities of words and sentences can be established by
statistical means, with no knowledge of the meanings.) Thereafter, we find
that the meanings are not additional properties unrelated to the syntactic
forms, but are a close concomitant of the constraints on word-choice in the
operator-argument relation and on the participation of words in various
reductions and constructions (11.1, 3).
All this is not to deny the usefulness of considering meaning in formal
investigation, let alone in studies of language use, language arts, and the
like. Meaning, and especially the perception of difference in meaning—
whether between two forms or between two occurrences of a form—may be
used as a preliminary guide where to look for possible constraints. It is clear
from all the above that in no way is meaning being excluded here from the
study of language, but that it is not being used as a criterion in determining
Method 43
the analysis of forms. Indeed, general considerations of meaning will be

obtained from considerations of form (11.4).
A final word about the relation of meaning to the investigation of forms: it
has been seen that a description of the combinations of sounds or of words
may contain two sources of redundancy: redundancy due to the way the
classifications and rules are stated, and residual constraints due to the
language itself. To the extent that meaning is connected with constraints on
combination, it is connected to the residual ones, and not to those imposed
by the form of description. Therefore, in any case, we have to reach the
description with least redundancy (2.2) before we consider meaning and its
interpretive relation to the constraints of form (11.4).
2.5. Properties of the data
2.5.0. Introduction
Language has several universal properties (4.4, 6.1), in addition to the

constraints on word combinations on sentences. These too affect what
methods are relevant for analysis.
2.5.1. Unbounded language, finite speakers
First is the fact that the set of utterances in a language (as describable in a
grammar) is unbounded. There is no reasonable way of specifying what
would be the longest sentence, and almost every sentence can have something
added to it, whereas the speakers and understanders of the language are
finite beings. Many things in language—what sounds are distinguished
phonetically, what phoneme-sequences are pre-set as words or morphemes,
what kinds of word combination are used—depend upon a finite body of
public experiences on the part of speakers and hearers. This means that
every language must be describable by a constructive (recursively enumerable)
system, one in which there is a finite and reasonably learnable core upon
which repetition or recursion without limit—and also change—can be
stated (6.0), with possibly a finite learnable stock of ‘idiomatic’ material
which is not subject to these unbounded operations.
2.5.2 Discrete elements
The success of writing systems, all of which use discrete marks, as carriers of
language suggests that language can be represented with discrete elements.
In addition, it will be seen in 7.1 that in the continuous events and changes in
44 Introduction
speech one can find ways of making cuts and distinctions which create the
discrete elements of spoken language and of structural language change. It
has also been found that the events which remain as continuous elements do
not combine complexly with other elements, and so in effect remain outside
the grammar. Once discrete elements are reached, the choice of relevant
methods for further investigation is greatly narrowed.
2.5.3. Linearity
Speech has extension in time only, and its repesentation in writing is

primarily a sequence of marks, along a line, although concurrent marks are
used in some cases. These facts suggest that language can be described
linearly, or at most multilinearly (with the observed event being a super¬
position of several concurrent linear events). The linear representation,
however, does not mean that the elements of language are linearly ordered
in respect to structure: they may be partially ordered with a linear projection, as
indeed is proposed here. Not surprisingly, those linear analyses of language
which succeed are not directly in respect to word-sequence, but are in terms
of specially defined relations, commonly called grammatical relations, among
the words of the spoken or written utterance. These grammatical relations
represent special cases of an underlying partial ordering of the words.
2.5.4. Contiguity
Talk or writing is not carried out with respect to some measured space. The
only distance between any two words of a sentence is the sequence of other
words between them. There is nothing in language corresponding to rhyme,
meter, or beat, which defines a space for poetry and music, or to the bars in
music notation which make it possible, for example, to distinguish rests of
different time-lengths. Hence, the only directly experienced elementary
relation between two words or entities in a word-sequence is that of beine
next neighbors, or of being simultaneous (especially in the case of intonations).
Any well-formedness for sentence structures must therefore require in the
base a contiguous sequence of (discrete) objects, although later permutations
and operators may intervene. The only property that makes this sequence a
construction of the grammar is that the objects are not arbitrary words but
words of particular classes. But the sequence has to be contiguous, it cannot
be spread out with spaces in between, because there is no way of identifying
or measuring the spaces.
By the same token, the effect of any operation that is defined in language
structure, i.e. the change or addition which it brings to its operand, must be
in its operand or contiguous to that operand. Of course, later operators
acting on the resultant may intervene between the earlier operator and its
Method 45
operand, separating them; but this is a later operation and is in certain

grammar situations avoided. In the description of the final sentence, such a
separation (i.e. interruption due to the embedding of later operators) can be
recognized. But in defining the action of the earlier operator on its operand,
this separation cannot be identified; the separation can only have been due
to a later event.
It follows that if language can have a constructive grammar, then there
must be available some characterization of its sentences which is based on
purely contiguous relations. The contiguity of the successive words is related
to this situation, but does not in itself satisfy this requirement since the
sentence has been found here to be characterized by a partial ordering of its
words of which the linear order is merely a projection (3.1, 4). This
contiguity can be utilized in designing a computer program to analyze the
structure of sentences (ch. 4, n. 1).
2.5.5. Variety within a class
It will be seen in Chapter 7 that the search for a least redundant apparatus of
description leads to many definings of classes, carried out in such a way that a
property can be more simply stated about the class than about the individual
members. This may happen because all the members have the same property,
which we do not want to repeat for each member. Or some members may
have only some of the class properties, so that they are only in some respects
members of the class (e.g. the English auxiliary verbs, G300). And some
members may have some properties while others have others, and it may be
possible to say that two such members between them constitute a whole
(composite, or disjunctional) member of the class (e.g. go and wen- above).
In the last case, we can think of the composite member as being the true
member, with the members that are included in the composite being
regarded just as variant forms that the true (composite) member takes on in
various situations. In all these cases it is enlightening to state a class
property.
The forming of classes is thus not a ‘similarity’-based grouping of the data
but a complex way of finding the most regular way of attributing properties
to entities, that is, of defining properties and their domains (2.3). Therefore,
a consideration in forming classes is the variety, degeneracy, and other
difficulties which are met with. For example, there is the question of how
many and how unusual are the differences among the various sounds
(allophones) grouped into one phoneme: this is equivalent to asking how
similar the sounds of a phoneme are in its various environments. A compar¬
able question is how similar are the phoneme-sequences that constitute a
given word, in the various environments in which that word appears. We can
say that part is a phoneme-sequence constituting a word; and hart is a
46 Introduction
phoneme-sequence constituting in some occurrences a word for deer and in

others the word heart; and the two pronunciations of economics (with e and
with iy) are two phoneme-sequences each constituting the same word; and
is, am, are are three phoneme-sequences each of which constitutes the same
word but over particular complementary arguments. Alternatively, using
classification, we can say that heart and the deer name hart are two words
whose phonemes are the same; in economics, the phoneme-sequence contains
a morphophonemic variation between e and iy in the first position; and be
has different alternant phoneme-sequences in specified complementary
environments.
In classes of words there is the question to what extent they all take the
same set of operations (relations) or environments. From the vantage-point
of the operations or environments, the question is to what extent they have
the same domain of words on which they act. Just as there are many classes
with fuzzy membership, there are many operations and constructions with
fuzzy domains—some being only slightly different from those of other
operations, others being of indeterminate boundaries. In the latter case, one
is not even certain that a given operation acts on a given word that is possibly
in its domain, so that the resultant of the operation is a marginal sentence,
such as The baby took a crawl.
An important question here is how much degeneracy arises when various
operations are carried out on particular members in their domain, i.e. when
one operation on one operand yields the same sequence of words—or of
phonemes—as another operation on another operand. A major example is
the case of sentences which are ambiguous because of different derivations:
Visiting relatives can be difficult from It can be difficult to visit relatives or
from Relatives who are visiting can be difficult; also Frost reads smoothly
from Frost reads in a smooth manner or from Reading Frost goes smoothly.
There are yet other properties which have not so far been investigated in
detail. Such is the distinction between what the speaker is constructing and
what the hearer recognizes, and more generally between the synthesis and
analysis of sentences.
2.5.6. Synchronic single-language coexistence of forms
The fact that there are various languages in the world and dialects within
them, and that languages change through time, means that there are
different possible objects for linguistic description and analysis. One could
try to describe all languages together—either their intersection, namely
what is universal and common to all, or their join or envelope, namely
anything that is found in at least one language. One could also try to describe
a single language or family through time, giving the succesive forms of
particular elements and constructions. However, there is special importance
Method 47
to being limited to a single language community (disregarding mutually

understandable, or hardly noticeable, differences of geographic and social-
class dialects) during one ‘synchronic’ period (during which change is too
slow to be seen as such, hence a time-slice). When language is described
within these boundaries, all its attested forms, meanings, and distinctions
coexist for its users. They can occur together in the same constructions and
sentences and discourses, made and understood by the same community of
users. Not surprisingly, it is under these conditions that the description of
language yields a fully intertwined grammatical system, for it is here that we
survey the forms and the constraints that are used together.
2.5.7. Diachronic change
Natural languages are open: new sentences and discourses can be said in
them. Furthermore, natural languages change in time. As far as we can see,
they do so without at any point in time failing to have a largely coherent
grammatical structure.
At any moment in the history of a language, it is possible to make as
complete a grammar of the language as we wish. No item of the language
need be left out as undescribable; any item which is not a case of existing
rules of the grammar can be listed separately or fit in (as a special case under
special conditions) to some existing rule in respect to which it can be
described. That is, we can describe a language at any time q, giving as
complete a description as we wish and necessarily including a large number
of individual facts (operations whose domains are individual words). At any
sufficiently distant time t2 we can do this again, and in all likelihood there will
be some difference in the two descriptions, in respect to some items X, due
to changes in the language in the intervening period.
Since tx and (2 are arbitrary, and since there are few if any discontinuous
points in a language history (although there may be such in the description of
language history), the description used for X at tx must be valid up to some
period fi, tx<tx<t2, and the description used for X at t2 must be valid from f,
and on, without there being a discontinuity at t,. We conclude that during the
period tn the grammatical items X must have been describable in two
different ways. Before th the item might be described in one way, to fit it into
certain features of the grammar. After th it will have changed sufficiently so
as to require a different description, fitting it into some other features of the
grammar. At th both descriptions must have been possible, i.e. at (, the
amount of change in X must have been sufficient to make X fit the t2
features, but not so great as to make X no longer fit the /, features.4 To this
4 It does not matter whether an individual changes Ins speech during /,, or whether the items
X are used differently hy people who began to speak after /, as against those who began to
speak before
48 Introduction
extent at least, some forms in language may have non-unique analyses. The
situation at f, is indeed often observable in detailed grammars, e.g. in the
case of transformations which are in progress. Thus in He identified it by
the method of paper chromatography, we have two sentences connected by
semicolon (secondary stress): He identified it by a method and The method
was of paper chromatography. In He identified it by the means of paper
chromatography, we can attempt a similar analysis into He identified it by
some means and The means were of paper chromatography. This is, however,
not very acceptable, and a more acceptable analysis would be to take by the
means of as a new preposition similar to by itself. The latter analysis is
already inescapable in He identified it by means of paper chromatography,
since Means were of paper chromatography does not exist.
The possibility of a structural static grammar at time-slices of a changing
system is due to the fact that there are similarities and other relations among
the various domains stated in a grammar and among the various operations
stated in the grammar. In terms of these higher-level systemic relations, the
exceptional domains of an operation can be shown to be extensions or
modifications of one of its regular domains.
This means that a description of language has to provide for the existence
of items which don’t quite fit into the rules for the rest of the language, but
can nevertheless be related to those rules as extensions of their domain or
small modifications of their operation.
2.6. From method to theory
The methods presented above were determined by the issue of constraints

posed in 2.2, and by the nature of language data as seen in respect to that
problem. Much of the nature of the data, such as its linearity and contiguity,
is observable in speech before any analysis is carried out. But the possibility
of defining discrete elements is observable without further ado only in
writing, especially alphabetic; in speech it is not seen until after certain cuts
and partitions are made in the set of speech sounds.
Carrying out these methods does more than provide an analysis of the
data. It narrows the range of possibilities toward forming a theory of the
constraints which were sought in 2.2. Such selective effect of the data
methods upon the ultimate theory happens when the methods are scrutinized
for their relevance (here, to the constraint issue), when they are carried out
over the whole of an adequate sample (here, a sample of the whole
language), and as fully as possible—so that the entities reached by the
analysis are maximally independent of each other (i.e. are least constrained
combinatorially), and when the analyses are made by these methods alone
with no appeal to external considerations (such as phonetic similarity,
meaning, intention, analogy, or history). This is not to say that such
Method 49
considerations may not have great weight, and may not even be the creators
of particular forms and relations; but for purposes of approaching a theory
we must first see to what extent a single relevant method can explain the
forms, and to what (if any) extent forms due to other influences can be
domesticated—reinterpreted—in terms of the central method.
It is especially possible to consider the analysis as laying the groundwork
for a theory, when every set of elements and operations or relations that is
finally reached is structurally related to a single language-wide property,
rather than some of the sets being merely residues created by setting up the
other sets. An example of the latter case would be if the set of inter-sentence
derivations were not characterizable as reductions but were simply the set of
whatever changes one needed to obtain one sentence from another, rather
than a motivated, structured set of reductions and transformations.
The reason that the methods help to suggest a theory is that they are not
simply empirical or descriptive. They necessarily organize the constraints
into a maximally simple system, so that the description of language comes to
consist of maximally unconstrained elements and rules (2.3) T Such a system
may not be identical with the way the speaker produces his sentences or the
way the hearer figures them out. But it has a unique least status in respect to
the data of the language (2.2). And it has a crucial status in respect to the
information carried by language, in that the information is certainly related
to the departures from equiprobability in combination; and a most parsi¬
monious grammar reveals most closely just the individual departures from
equiprobability in a language (11.4).
It should be noted that the method leading to a least system is not
reductive in a simplistic sense. It organizes the constraints, but it does not
lose them. As an example: when constraints on certain entities (say, sounds)
are used to define a ‘higher’ and more powerful class of entities (in this case,
phonemes) which are freed of these constraints, the old constraints still
apply to the members themselves within the new classificatory entities, but
these old constraints do not interfere in the simpler relations that can be
stated among the new entities.
5 This is not to say that we have here a discovery or decision procedure for grammar.
However, given a particular metatheory of language—the need to map the departures lrom
randomness (2.1), which arises from the essential lack of an external metalanguage—we can
propose procedures that lead to a least grammar, which constitutes a best fit of the data in the
niven direction, though not uniquely the best fit. It is nevertheless the case that the departures
from randomness in the various languages are so massive and so similar that all grammars
exhibit a strong family resemblance, no matter in what style they are stated.
II
Theory of Syntax
3
A Theory of Sentences
3.0. Constraints on equiprobability
It has been noted (1.0,2.1-2) that a central problem for a theory of language
is determining the departures from equiprobability in the successive parts of
utterances. It has been found that in natural languages it is not merely that
not all combinations of parts are equiprobable, but that not all combinations
occur. That is, given even very large arbitrary samples of the utterances of a
language, some word combinations have zero probability of being found.
For the given language, then, we make a distinction between the combinations
that have zero probability and those that have any positive probability. It
will be seen below that the zero probability characterizes what is outside of
syntax (3.1), while the differences in positive probability characterize the
differences in word meaning within syntax (3.2).
Hitherto, we have been considering only the directly observable data of
word combinations in a language. However, simply listing the combinations
is not only impossible because of their number, but also fruitless because it
leads to no relevant principled classification. Furthermore, any such listing
would be imprecise since, for a great many word combinations, speakers of
the language are uncertain or do not agree, as to whether they are in the
language: given that take a walk and take a run are in, are take a climb or take
a jog or take a crawl also in? Finally, since the details of word combination
change faster than anything else in language, no exact list would be correct
over any long period.
In contrast with list-making, it is possible to characterize the word
combinations by stating a set of constraints each of which precludes particular
classes of combination from occurring in utterances of a given language
(1.1). These constraints will be found to hold over long time periods. They
can be so formulated as to act each on the product of another constraint
(beginning with a null constraint, i.e. with random word-sequences), so that
a few types of constraint will suffice to preclude very many types of word
combinations and produce the others. It will be seen that the constraints
which produce the combinations found in a language can allow for many
exceptional forms and changes, by minor fuzziness in the domains over
which they are defined: e.g. why America was left by me on July 6 is odd, but
America was left by a whole generation of young writers in the twenties is
54 Theory of Syntax
natural (1.1.2, G365). The regularity of the constraints, and the simple
relations among them, fit in with a mathematical formulation of syntax; and
their direct specification of departures from equiprobability gives them a
direct informational value.
Characterizing sentences in terms of constraints on combinations rather
than of word combinations directly is a matter of method. But when the
sentences are described as resultants of constraints, one can consider the
constraints to be constructs in a theory of sentence construction.
Since sentences have not been defined at this point, these constraints are
not, in the first place, stated in respect to the set of sentences. They can be
said to hold over the set of utterances provided that, first, a small number of
frozen expressions (such as goodbye and ouch), each with unique constraints,
are excluded, and second, the period punctuation between sentences (or, in
speech, between successive sentence intonations) is considered a conjunction
roughly like and. In 7.7 a procedure for finding sentence boundaries will be
presented, permitting us to define sentences. The constraints of Chapter 3
will then suffice for sentences. Further constraints, among sentences of an
utterance, will then be presented in Chapters 9 and 10.
3.1. Zero vs. non-zero likelihood
3.1.0. Introduction
The crucial property of language is that the presence of words in a sentence

depends on how other words in the sentence depend on yet other words in it.
This dependence is the essential constraint on equiprobability of words in
sentences. Though more or less empirically come by, this dependence can be
considered a construct of the syntactic theory. Here, we first define the
dependence (3.1.1), and then state what it means that the dependence is
carried out in respect to further dependence (3.1.2). To a first approximation,
we are speaking here, for example, of how the presence of a verb in a
sentence requires the presence there of some noun as its subject.
3.1.1. Dependence
lo begin, we note that in the sentences of a given language we can find a

distinguished subset—called base—in which this relation of dependence is
overtly exhibited (4.1). We will then see that all other sentences in the
language can be derived in a regular way from those of the base, without
altering the word-dependence, though with some loss of its visibility (4.2).
In the base sentences, most words are morphologically simple (unimorphemic,
i.e. roughly affixless).
A Theory of Sentences 55
We define the dependence as follows: If A is a simple word, and b,..^ 1C,

an ordered set of classes of simple words, then A is said to depend on (or,
require) b, . . e if and only if for every sentence in the base, if A is in the
sentence then there occurs in the sentence a sequence of simple words B. . .
E which are respectively members of b, . . ., e. Within the given sentence,
A may then be said to depend on the word sequence B ... E. If in the given
sentence there is no other word G such that A depends on G and G depends
on the given occurrence of B ... E, then A depends immediately on that
occurrence of B . .. E. A is then called the operator on that B... E, which in
turn is called the argument of A in the sentence; B may be called the first
argument, and so on. When the sentence consists only of A and its B . . . E,
the operator is necessarily contiguous to its arguments (2.5.4)—before,
after, or between them. When another word Fdepends on A as A depends
on B ... E (3.1.2), the constraint would be simplest if the F is similarly
contiguous to (the resultant of) A together with its B ... E; this is indeed
seen to be the case. The contiguity of operator to argument in the base
makes it easy to check if no word G as above intervenes in the operator-
argument relation of A to B ... E. In / know sheep eat grass (1.1), know is
the operator on the pair /, eat as its argument, with eat in turn as the operator
on the pair sheep, grass as argument.
Deriving the remaining sentences from the base sentences is carried out
by means of the reductions and other transformations, which are changes in
the phonemic (alphabetic) shape of words in an utterance. In terms of them
we can make a more general statement about dependence, without referring
to the base set or to sentences at all: A depends on (or requires) b, . . ., e if
and only if for every utterance in the language, if ^4 is in the utterance then
there occurs in the utterance a sequence of simple words B ... E (members
of b, . . ., e), any of which may be in a reduced (including zeroed or
permuted) form as provided by the reductions of 3.3. (The base contiguity
described above is then modified by these reductions and permutations.) As
an example: in John expects Mary to come, which is approximately in base
form, expect depends on a word of the class of John, Mary, coffee, etc., and
on a word of the class of come, go, etc.; come in turn depends on a word of
the class of John, Mary, coffee. In John expects Mary, Mary expects coffee
(which are not base sentences), it can be shown (3.3) that come is still
present in both sentences though in zero shape {John expects Mary to
come, Mary expects coffee to come), so that the dependences remain as
before.
Dependence is a relation between a word and an ordered set of word
classes (or, in a sentence, between a word occurrence and an occurrence of a
sequence of words belonging to those classes). Over a whole language, most
or all dependences hold not for a single word such as A above but for any
word of a set, a; so that dependence can be considered a relation between a
word class (e.g. a) and an ordered set of word classes (b, . . ., e).
56 Theory of Syntax
As examples, we take a few words and their dependences in short

sentences.
First, sleep. We find child (as in The child slept—for the and tense see 4.2),
tree (sleeps all winter), earth (sleeps under the snow), night, universe (will
then sleep until the next Big Bang), rarely algae, vacuum, etc. In contrast, we
do not find go, eat, up, because, etc. (not Go sleeps. Going sleeps. Eat sleeps.
To eat sleeps). At this stage, the distinction between words that may be
found, even very rarely, with a particular word, as withs/eep here, and those
that are never found is imprecise, and any characterization of such two sets
of words would be ad hoc. However, when we compare, in any large corpus
of data, the lists for sleep with the lists for certain other words, such as fall or
sit, we find similarities, though with different frequencies, in what words do
not co-occur. With/a// we find child, tree, book, night, earth (can fall into the
sun), rarely universe, perhaps never vacuum-, and clearly not go, going, eat,
to eat, up, because. With sit we find child, book, night, earth, universe, algae,
almost never vacuum, and not go, going, eat, to eat, up, because (no Because
sits). Even so, at an empirical level some problems remain: we may perhaps
find Going doesn't sleep (though not Go doesn’t sleep) and perhaps even To
eat doesn’t sleep. However, these differ in the circumstances of their use
from This child doesn’t sleep. Algae don't sleep. Vacuum doesn’t sleep. For
example, for the almost-never words in the set that includes child, algae,
vacuum, we find different frequencies among them in respect to sleeps or
falls (where they may be almost never in varying degrees) and in respect to
doesn’t sleep or doesn't fall (where all are appreciably less infrequent). In
contrast, in the set that includes of, eat, because, all have the same frequency
(zero) with respect to sleeps or falls, and all have the approximately same
zero or vanishingly small frequency with respect to doesn't sleep or doesn't
fall.
All this is not quite enough to establish a distinction into a few large
classes: child, vacuum, etc. and sleep, fall, because, up, etc., both with
respect to sleep, fall, etc. But it is enough to suggest a postulate that there is
such a distinction. This postulate, then, holds that there is (a) a class of
words (up, because, sleep, fall, etc., called O below) which have zero
likelihood (or probability, 3.2.3) of occurring with words of a subset of them
(sleep, fall, etc. called On below), and (b) a class of words (child, tree,
universe, vacuum, etc., called N below) which have some positive likelihood
no matter how small of occurring with that subset (O,,). In (b) any member
of N may conceivably be found with any member of On, e.g. The comet slept.
Space slept. Time fell-, and indeed one cannot say which of these sentences, if
any, is absolutely excluded (especially in fairy-tales and the like). It will be
seen that the distinction into these two classes O and N, and such subsets of
O as O,,, will show up throughout the grammar, and will conform to various
informational differences.
Next, wear. We find here for the most part not single words occurring with
wear, but pairs: child, coat {The child wore a coat), and night, hue {The night
wore a dark hue), even tree, bark {This tree wears a heavy bark), or very
rarely child, bark {The child wore bark), but not child, eat {no The child wore
eat, or . . . wore eating), or go, bark (no Go wears bark, or To go wears
bark), or eat, because (no Eat wears because). Much the same exclusion
holds for eat {child, egg or acid, copper, etc. occur, but not the pairs go, child
or go, because: no Go eats because), although more pairs can be readily
found with eat than with wear. One might think that pairs like child, egg or
acid, copper that are not found with wear should be excluded there. But
although we do not find The child wears eggs, we find She wears a set smile.
She wears perfume. She wore roses in her hair and fruit on her hat and we may
find The child wore egg on his face. In any case, the set of words definitely
excluded under sleep, fall (namely because, go, sleep, fall, etc.) is also
excluded under wear, eat. And the words that are not excluded under sleep,
fall, either because they are found there or because they could conceivably
be found there, are similarly not excluded in both positions under wear, eat.
We thus find support for a class of words postulated to have zero
likelihood of occurring with the sleep and wear set in the shortest sentences
of the language—shortest either observably, or after correcting for the
reductions presented in 3.3. This class will henceforth be marked O (operator),
while the residue class, whose members have some positive likelihood no
matter how small under the sleep and wear classes, will be marked N because
most of them are simple (unimorphemic) nouns. The sleep set, a subset of O,
will be marked On, and the wear subset On„, to indicate that they occur with
one or two members of N respectively.
We next note that in any shortest sentence in which is probable or is
unlikely occur, they are found with a member of the O class (including the
On and Onn) and not with a member of the N class alone (except, rarely, due
to reductions). We have wear under is probable in (1) That John will wear a
tie is probable, but we don’t have John is probable. More precisely, the only
members of N that are found in those sentences (e.g. John with probable in
(1)) are the ones already accounted for as occurring with the On, O,,,,
members in question, as in the case of John, tie occurring with wear when
wear occurs with is probable in (1). The case of is unlikely, is probable
appearing only with N (e.g. A boy is probable in guessing about a birth) are
accounted for by the reductions of 3.3 (in this case from e.g. For the baby to
be a boy is probable), as are the cases of an O without its N (e.g. To win is
unlikely reduced from For us to win is unlikely or other such sentences). The
class of is probable is then marked O0, since its co-occurrent (argument) is
an O (which then brings its own co-occurrents into the sentence). Note that
the O argument under 00 may be not only On (like wear above) but also O0,
as in That John's coming is unlikely is quite probable.
Words like assert, declare, deplore are found to have two co-occurrents
with them: first an N and second an O, as in John deplores your arrival,
58 Theory of Syntax
where John, arrive are an ordered pair occurring as arguments of deplore,

while you is accounted for as argument of arrive (an O,,). Hence the class of
deplore is Ono.
Similarly, entail and because are O00: we may find Your arrival entails
John 's departure ox John departed because you arrived. We do not find You
entailed John or Your arrival entailed John, nor John because you or John
because you arrived. As with O0, the Ooa and Ono can occur on Oa or Ono or
O00 as well as on O,,, Onn: in John asserts that Mary asserts that you departed
because l arrived, the first assert (an Ono) has John (N) together with the
second assert as its ordered co-occurrents (arguments); the second assert has
Mary (N) and because (Oao) as its co-occurrents; because has departed (Ofl)
and arrived (O,,); departed has you (N), and arrived has I (N).
The presence of an On thus depends on an N word as argument, since an O
co-occurrent is excluded. And each Ono word depends on the ordered pair
of N, O, since O words and N words respectively are excluded in the two
argument positions. And so for the other words.
3.1.2. Dependence on dependence
We now consider the word classes on which a word depends, i.e. what
classes appear in argument position. The crucial finding here is that each of
these argument classes is characterizable not by a list of members (which
would be difficult, since words can be added to a class or lost from it), but by
what that argument class in turn depends on. For English, each argument
class is characterized by whether it depends on something at all (i.e. on the
presence of yet other words in the sentence) or on nothing: that is, by
whether it in turn has an argument or not. In John asserts Mary came, we say
that asserts (Ono) depends on the pair John (N), came (On); and came
depends on Mary (N). In [urn, John, Mary can be said to depend on nothing;
they can also be said to depend on nothing except the operator that depends
on them (reasons for the first formulation: see below and 3.2.1 end). Thus
the presence of asserts in a sentence depends on the ordered presence of a
word (e.g. John) that depends on nothing further, and of a word (came)
which does have something further on whose presence it depends (in this
case, Mary). No property (e.g. meaning or subclass) of the pair John, came
other than their dependence status is essential for the presence of asserts in
the sentence. The presence of a word, then, depends not on particular other
words but on the dependence properties of the other words.
The status of N is clearly different from that of all the O classes. The
question is what should be the systemic status of this difference. Here it may
be relevant that the ability of an N word to occur alone as an utterance differs
from that of all the O words. When a word occurs alone as an answer, that is
only because its environing words were zeroed as repetition of the question:
Who was there? John from John was there; What did John do all day? Study
from John studied all day. When an O word appears alone, in imperative
intonation, it can be shown to have been reduced from a particular sentence
structure (G351): Stop! from / order you that you stop. But when an A word
appears alone, with or without exclamation intonation, it is far less clear
what are the specific environing words if any that have been zeroed: e.g. the
various uses of John!, or A hawk!, or the numbers (in counting). True, these
grounds do not determine any single way of characterizing N. But they
support the preference for the first formulation above, that an N word be
defined as depending on the null class rather than on the operator over it.
Then by the side of the primitive arguments N, which require nothing, all
O are operators, with non-null requirement. All On „, e.g. On and 0„„,
require only words that themselves require null. All O 0 , e.g. O0, Ono,
Ooa, require in at least one argument position a word that requires some¬
thing, which in turn may require something or null. Such are, in English Oa
(e.g. probable), Ono (assert), Onoo (attribute, in John attributes his success to
my helping him), O00 (entail, because). There is also a class Oon (astonish, in
John's winning astonished me); and within Oon, the prepositions (e.g.
without in John left without me) as operators almost all of whose members
can also appear, due to various reductions, as if they belong to other O
classes.
3.1.3. Pattial order
Here, as at various points in the syntactic theory, we find a partial ordering,

that is a relation (^)ona set, of being ‘higher’ or ‘greater’ (>) or else equal
( = ), in respect to which, for alia, b,c in the set we findaâ (reflexive), and if
both a 2=6 and bâ then a—b (antisymmetric), and if a 2=6 and 62=c then a2=c
(transitive); but there may be members a, b in the set which are not
comparable, where neither a^b nor 62=a.
The dependence relation imposes a partial ordering of word classes in
respect to their definitions. The definition of N (‘zero-level’) includes no
reference to any particular class. The definition of the On „ (‘first-level’)
classes has reference only to N. And the definition of the O 0
(‘second-level’) classes (i.e. operators having an operator as their argument)
includes reference to O (of whatever argument) and in some cases to N. This
yields a partition of the words of the language into the three levels named
above, and within the first and second levels into the argument-requirement
subclasses On, Onn, O0, Ono, etc. (6.3(d)).
When this relation is applied to the words of a sentence we obtain a partial
ordering of them. In each sentence, in the set of word-occurrences A, B,...,
Z we have a transitive relation: A>B, when B is in the dependence chain of
A (that is, B is the immediate argument, or argument of the argument to any
60 Theory of Syntax
depth, of A). When A>B and there is no C such that A>C>B, we say that A
is the operator on B (A covers B), and B is the (immediate) argument of A.
When A covers both B and D, we call B and D co-arguments of A. The co¬
arguments of an operator occurrence are linearly ordered: A on B, D does
not form the same sentence as A on D, B. In John expects Mary to come we
have John, come as co-arguments of expect, and Mary as argument of come.
The ordering is a semilattice; and no word-occurrence can be an immediate
argument of two operator occurrences (6.3).
If one wishes to think of the making of a sentence as a process, one can
view the semilattice, that is the dependence order of the words in a sentence,
as the partial order of entry of the words into the making of the sentence. No
word can enter into a sentence, which is being constructed out of words and
previously made (component) sentences, unless its prerequisite is available—
this means, unless its required argument is present and is (under the lattice
condition) not serving as immediate argument of any other operator in the
sentence. The prerequisite is any one word of each of these classes on which
the given word depends. Then each entry of an operator makes a sentence,
out of a non-sentence, namely N, or out of a sentence (that is, something
that has been made into a sentence by a prior-entering operator). Thus come
made a sentence by operating on Mary, while expect made a sentence by
operating on the co-argument pair John and come. The linear ordering of
the co-arguments means that expect may not be defined (and indeed is not)
on such a pair as come, John (no Mary's coming expects John). And though
both John saw Mary and Mary saw John are sentences, they are not the same
sentence, nor is one derivable from the other or paraphrastic to it. The
similarity of operators to functors in categorial grammar within logic is
evident.
3.1.4. Predication
The major semantic yield of the dependence relation is that it carries

inherently the meaning of‘predication’. If there is a word A which is not said
except in company of one or another word B of a set b, then when A is said
with some B it has perforce the effect of saying something about that B.
Certainly if a word’s only meaning were a property or event relating to
certain things, it would only be used together with the words for those things
(or with pointingor otherwise referring to those things, which can in turn be
covered by metalinguistic predications, 5.2). Sentences are, first of all, word
combinations in which something is said (predicated) about another thing.
Predication inheres not in co-occurrence per se but in the co-occurrences
which are required. The idea of such predicating, or the way of doing it,
cannot have existed in the minds of early speakers without the existence of
sentences. There may well be concepts or mental events that exist in the
mind independently of their expression in language, but this could not be the
case for concepts which are essentially about language. Therefore sentence-
hood can only have arisen not as a vehicle for an otherwise conceptualizable
relation of predicating, but as a concomitant of some other, more behavioral,
relation which then yielded predication as an interpretation. We may
suggest that this underlying relation was the practice of not saying certain
words of a set a except in the company of certain others of a set b, because
the a-words were not specific enough except when used only with, hence
about, a 6-word.
3.2. Inequalities of likelihood
3.2.0. Introduction
Within the dependence relation of operator word to argument class there is

room for a ‘fine-dependence’ relation of operator word to argument word
(or word-sequence), which makes a further contribution to the departures
from equiprobability of word-occurrence. This possibility exists because the
dependence of a word on a class states only that every member of the class
has a positive likelihood, no matter how small, of occurring with that word.
Nothing says that the likelihood has to be a priori determinate, as being
random, or equal, etc., for all members of the class. Hence, for a given On,
different N can have different likelihoods of appearing as its argument;
equivalently, the On word has different likelihoods of occurring on different
N words. (Just what the likelihoods consist of will be considered in 3.2.6.)
For example, sleep has reasonable likelihood of appearing on John, the
child, the dog, but lower likelihood of appearing on the city (as in The city
sleeps), or on trees (in Trees have to sleep each winter, which is not really
metaphoric) or earth (The earth sleeps under a blanket of snow), and lower
yet on window, chaos. If an operator has two or more positions in its
argument (e.g. subject and object), its likelihoods of occurring on the
individual words in one position may be independent of its likelihoods on the
words of the other position; if they are not independent, we have to speak of
the operator’s likelihood of occurring on the word m-tuples formed from the
m ordered classes in the argument. For example, the Onn operator see has
greater likelihood of occurring on John, dogs, fish than on space, water, the
blind (taken as noun, e.g. in French un aveugle), in its first argument
position, and more or less independently it has greater likelihood on John,
dogs, fish, water, the blind than on space, vacuum in the second position of
the NN argument. In contrast, the likelihood of occurrence of the Onn
operator eat can best be described in respect to pairs of individual words in
its NN argument, e.g. with greater likelihoods on children!cereal, goat/
paper, virus! host, acid! copper, revolution/its children, car!gasoline, Venus fly
62 Theory of Syntax
traplinsects, than on children!gasoline, virus/paper, Venus fly trap/cereal.

These likelihoods can change through time, and can differ in different
subsets of speakers, but they are roughly stable over short or long periods,
so that the likelihood inequalities of arguments in respect to particular
operators and vice versa can be used in analyzing language.
3.2.1. Likelihood constraints
If we consider the various words which have the same argument class, e.g.
see, eat, etc. in Onn, we find that they have different inequalities of
likelihood in respect to the individual argument words in the class: see has
greater likelihood, but eat less likelihood, on John/water than on carl
gasoline. In fine, we can state the inequalities of likelihoods of each word in
an argument class in respect to the words in the operator class over that
argument: e.g. the likelihood of water to be in the second N position under
see, under eat, and under other Onn.
The distinction between argument-requirement and likelihood is not
immediately observable in the data. It is a choice of analysis to isolate the
combinations whose frequency not only is zero in the given data, but can be
expected to remain zero in any further data, and whose distinction from the
likelihood-based combinations is expected to be useful for constructing a
theory. The dependence of an operator on an argument is thus a construct. It
is based on a particular criterion, of what arguments they in turn depend on,
which has the merit that the arguments in turn can be characterized by the
same criterion. In contrast to this construct, the likelihoods of an operator in
respect to its argument words are the direct data of word-combination
frequency within the dependence structures, in (a sample of) the sentences
of a language (3.2.3, 3.2.6). These likelihoods form a vast body of data,
which cannot be surveyed except in limited sublanguages (ch. 10).
The inequalities of likelihood of a given operator word in respect to an
argument class amount to a rough grading of likelihood, though not in
general to a specifiable partial ordering thereof. If we check the likelihoods
for sleep, for example, we find many N words which have reasonable
likelihood (much higher than the average of their class) of appearing as its
argument. These have been called, e.g. for sleep, the normal co-occurrents,
or selection, of that operator in respect to its argument. Some of these words
may occur frequently, e.g. you, I, the baby, in How did you sleep?, The baby
slept well, etc. Others may be very infrequent, as in A thrips was sleeping on
the leaf-, but even these are immediately acceptable if they are included
under a classificatory word which in turn is a reasonably frequent argument
of the given operator (see below). For other argument words the operator
sleep may have, in various degrees and respects, less than selectional
likelihood: trees, flowers, bacteria. Yet other N may be even more infrequent as
arguments of sleep, or may appear only when the sleep is under particular
further or neighboring operators. For example. The saucer slept may occur
(not even metaphorically) in a fairy-tale, in which neighboring sentences will
be of appropriate restricted types; The universe slept, The winter slept
involve metaphor, which is analyzed at the end of 4.2 as reduced from a
sentence-sequence under the operator as. And almost any N can be the
argument of sleep when sleep is under not (because it is on say, 5.5.3), as in
Rocks don’t sleep. All the latter N would not be included in the selection of
sleep.
The statuses of the various argument words under an operator do not in
general suffice for setting up any precise subsets within the argument class. It
is even impossible to define a set of ‘human’ or ‘human-like’ N as the first-
position selection under say, know, believe, expect, and other such Ono,
since we find not only I know that he left and The dog knows he left, but also
for example The virus knows when it is in an environment containing material
for replication and The dog said yes with its tail and Here are ten dollars that
say you are wrong (in laying a bet) and These rocks can expect to last another
5 billion years.
Because of its imprecision and instability, as well as its vastness, the data
concerning selection, or likelihood in general, has not been used directly in
defining the structures of language. But there are stable properties and
structures, which can be defined in terms of the likelihood inequalities or
gradings. For one thing, since reductions will be seen not to affect the
likelihood gradings among the words of a sentence, these gradings are
preserved under reductions and transformations (the latter being mostly
products of reductions, 8.1.2). Indeed, transformations can be phenomeno¬
logically defined as subsets of sentences throughout which the likelihood
gradings are preserved (8.1.3). More important, the reductions will be seen
to be based on extremely high likelihood (or, equivalently, low information)
in the operator-argument relation (3.3.2). Similarities and contrasts in
likelihood grading also yield many of the grammatical categories, semantically
important though somewhat fuzzy as to membership, that give texture to
each language grammar: e.g. durative and momentaneous verbs, indefinite
pronouns, etc. The more or less stable likelihood differences—the fine
dependence—thus constitute an additional constraint on word-co-occurrence,
beyond the dependence of 3.1.
While the likelihood constraints differ from the dependence constraint of
3.1, they do not conflict with it, because the likelihood of an operator is
postulated not to reach permanent zero in respect to any word in the
argument class. Also, the likelihood constraints do not repeat any of the
dependence constraints, which are stated on argument classes; the likeli¬
hoods are stated in respect to the argument word, as a further restriction on
what the dependence admits. While virtually everything in langifage structure
will be found to be encased in the relation between an operator and its
64 Theory of Syntax
argument, the likelihood of an operator can be affected not only by its

immediate argument but also by some distinguished word further down in
the dependence chain—e.g. an argument of its argument to some depth. As
an example of one such situation, melt is not a likely operator on house (as in
The house melted) but it has some likelihood on house made of snow. In this
kind of situation, a word with modifier (the latter being reduced from a
conjoined sentence) can carry a different selection, which may approximate
the selection of a different word, in this case igloo, as in The igloo melted,
The house made of snow melted. The consistency of selection breaks down in
negation and question: virus/paper is not in the selection of eat, but one can
readily say Can viruses eat paper? and Viruses do not eat paper. The
explanation is that negation and question are not operators (on eat in this
case), but are part of the metalinguistic performatives in which each
sentence is encased (5.5).
Perhaps every word has a unique selection of normal co-occurrences in
the argument positions under it or in the operator position over it, and
almost every word has unique inequalities or grading, from most likely to
least likely co-occurrents in each of these positions. Nevertheless, words
that are close (or opposite) in meaning have largely overlapping selections,
as to a lesser extent do words that belong to the same universe of discourse
(e.g. walk, run, climb, jump). That is, they have many normal co-occurrents
in common. What is more important, even the divergent co-occurrents are
not unrelated, for their selections, in turn, have large overlaps. Thus the
operator climb but not jump has crocodile, snake in its first-argument N
selection (The crocodile climbed the river bank) and jump but not climb may
have flea, robin in its N selection; all of these nouns N, however, are in the
selection of the operators eat, die, sleep.
Without attempting the impossible task of describing the network of
selections among the words of a language, we can recognize in a rough way
that each word has (or shares in) what we may call a clustered (‘coherent’)
selection. Operators share in a clustered selection if their argument selection
has a large overlap and if many words in the selection of one which are not in
the selections of the others have, in turn, large overlaps in their operator
selections; similarly, many arguments have a coherent selection of their
operators. We see in this that the occurrence likelihoods of arguments under
a particular operator, or of operators over a particular argument, are neither
arbitrary nor unrelated to those of other words. A cladistic similarity
relation here can give coherence to word co-occurrences and to the domains
of reduction; we may also see here some properties of the sociometric
charting of how social contacts cluster.
The importance of coherent selection is that it helps characterize semantic
word classes and also the internal structure of such classes (e.g. of a generic
name and its species names). It also enables us to tell when a given word
consists of two homonyms (namely, when its selection contains two or more
distinguishable coherent selections), or when a particular co-occurrent of

the word is not really in its coherent selection, but is rather a result of
selectional extension or metaphor that happened with the given word and
not with other words.
It is clear that similarity of selection is closely related to similarity of
meaning, and this for two reasons. In part, the meaning of an operator
affects what argument word the speakers will use it on. But in part, the
combination in which a word already occurs affects the meaning it has for the
speakers of a language. These combinations, i.e. the selection of each
operator or argument, arise not only from directly semantic considerations—
what is useful to say about a given argument word—but also from the
complex morphological, analogic, and historical factors that determine
word frequency and word replacement in combinations, as well as meaning-
change.1 Extending the selection of an operator, i.e. increasing its likeli¬
hood on particular argument words, is not based on semantic considerations
alone. For example, when airplanes began to be talked about, why did
people come to say in English The plane flies rather than the plane floats or
the plane glides'? Of balloons one is likely to say both fly and float (but usually
in different circumstances). Consider also that one usually says to boil eggs
rather than to cook eggs, although this changes the meaning of boil from that
in to boil water. Such selectional peculiarities can be found in much of the
vocabulary of a language.
Words often have somewhat different meanings when they are in com¬
bination with different words—a further indication that co-occurrence can
create meaning, not only the converse (11.1). This may be more noticeable
for operators in respect to their arguments than for arguments in respect to
their operators: boil has different meanings in boil eggs and boil water, but
eggs has the same meaning in boil eggs as in buy eggs. The greater variety in
On n meanings supports our formulating the likelihood and dependence
relations as being in the first place a restriction on the operator in respect to
the argument words rather than the converse (3.1); this formulation is
convenient since the restriction is a further constraint within the dependence
relation which had been formulated for the operator in respect to argument
classes.
3.2.2. Likelihood types
In each language there are words whose selection can be summarized in

ways that are of interest for the structure and meaning capabilities of the
language. The major types of selection will be noted here.
One case is that of unrestricted selection.
*
1 Henry M. Hoenigswald, Language Change and Linguistic Reconstruction, Univ. of Chicago

Press, Chicago, 1960.
66 Theory of Syntax
There is a rather small set of words, e.g. say, whose arguments carry no
likelihood restriction. Thus, say can be said of any sentence, hence on every
operator, as in 1 say he’s late or He said she saw me. This holds also for deny
(source of not), ask (source of the interrogative), and request (source of the
imperative), although some stative operators such as exist may be rare under
request (G321,331,351). This set of words will be found to have the status of
metasentential operators (5.5). The further likelihoods, of operators to their
arguments, under deny, ask, request, may differ from those under say itself:
A stone sleeps. Vacuum sleeps may be unusual under say, but less so in A
stone does not sleep (from deny). Does vacuum sleep (under ask). But the
partial-order constraint of 3.1 holds throughout: Going does not sleep and
Does going sleep are about as unacceptable as Going sleeps.
There is also a small closed set of ‘indefinites’ whose union has all
operators in its selection. For example, something is in the argument
selection of very many operators, and any operator under which something
does not normally occur will be found to have somebody (and someone) in
place of something. Since they have to fit under every operator, such words
as the something-somebody set unavoidably carry the meaning of indefinites:
they do not fit one operator rather than another; hence they mean no more
than being the argument of whatever operator is on them. (Although he,
she, it seem also to constitute such a set, this is because they arise as
referential reductions of all arguments (5.3), whose union necessarily has all
operators in its selection.)
Another case is broad selection, when an operator has in its selection not
all the words in its argument but very many of them, far more than most
operators have. This is the case for some Oao (operators on a pair of
sentences) such as because, entail; also operators with mostly space-relation
meanings originally, such as near, to, before. It is also the case for certain
operators of exceptionally broad applicability, such as in a . . . manner (an
earlier English form for which was reduced to -ly as in He walked slowly. He
spoke agitatedly), or such as the condition of (an earlier English form for
which was reduced to -hood as in childhood ‘the condition of being a child’).
Also, operators of time and aspect (i.e. the onset and durational properties
of an event) occur readily on almost all operators, or on all those which are
durative, and so on.
A special kind of broad selection is seen in quantifiers, and somewhat
differently in classifiers. Words for numbers and amounts appear readily as
arguments under almost every operator (Two came late. He ate two. Much
gets forgotten), where they seem first to adjoin and then paraphrastically to
replace the words whose count or measure they are giving (Two men came
late, He ate two apples. Much that is known gets forgotten). But in addition
these words have a selection of their own, such as plus, more than: Two plus
two equals four. Much is more than little. Words for sets, containers, and
fragments are similar: The team played well from The players (in the team)
played well, but in addition The team was formed last year, also The whole
room applauded from The audience (in the room) applauded, but in addition
The room is dark; also This piece is too sweet from This (piece of) cake is too
sweet, but in addition This piece is only one-sixteenth of the cake (G260).
Somewhat differently, classifiers are words each of which has (with certain
modifications) the selection of the union of a unique subset of other words
(its classificands). Thus the selection of operators over snake includes the
selections of operators over viper, copperhead, etc. In addition, however,
there is a small selection of other environments (not applicable to their
classificands), as in Snakes constitute a suborder. When all these classificatory
words are analyzed it is seen that they are co-arguments, under quantifying
and classificatory operators, of the words whose selections they can take
over (e.g. A viper is a snake).
In addition to unrestricted and broad selection, there is ‘strong’ selection,
which is seen when a word occurs exceptionally frequently with a particular
operator over it or argument under it: e.g. verbs under particular prepositions
(as operators on them), in wash up as against the less frequent wash down
and the rare wash under; also the frequent stop in and stop over as against the
rare stop out; and, differently, the frequent occurrence of hermetic with seal
as argument (hermetically sealed from The sealing was hermetic in manner).
An extreme case of strong selection is the narrow or quasi-idiomatic, as
seen for example in the operator loom, where we have by the side of loom
itself (as in It loomed out of the dark) also It loomed large (roughly derivable
from an artificial It loomed into its being large) but not It loomed small nor It
loomed dangerous, etc. We can consider as strong selection also those words
that do not combine grammatically with anything, and so occur only under
metasentential and metalinguistic operators (ch. 5): e.g. ouch, hello in He
said Ouch, Hello is an English word.
The major case of narrow selection is seen in idioms and frozen expressions.
These may have unusual word combinations as in hit upon (He hit upon a
new idea), or unusual reductions as in make do or The more the merrier, or
unusual meaning of the whole expression as in old hat or kick the bucket in
the general sense of ‘die’ (these are often due to metaphor (4.2), which is a
product of reductions).
There is also a strong ‘negative’ selection. Comparable to the demand that
certain words or repetitions appear under a given operator is the demand
that certain words not appear, that is, the rejection of certain argument
choices. Thus while one can readily say He is truly coming. He is certainly
coming. He is undoubtedly coming, one cannot say He is falsely coming, He
is uncertainly coming, and perhaps not He is doubtfully coming. The reason
is that He is truly coming is a reduction of the sentence-pair He is coming;
that he is coming is true; hence (1) He is falsely coming would be a reduction
of (2) He is coming; that he is coming is false, which would generally be
rejected. The difficulty lies not in That he is coming is false, which can readily
68 Theory of Syntax
be said, but in conjoining the two sentences, which is not likely when one
sentence denies what the other sentence asserts. The contradictory com¬
bination can physically be said, as in (2) above; but because of its unlikeli¬
hood (negative selection) it does not participate in the reduction to adverb,
falsely in (1) above (G306).
3.2.3. Reconstructed likelihoods
The types of likelihoods seen above, from unrestricted, broad, and strong to
narrow and negative, are all observable directly in the data. There are
also other situations, in which the regularization and simplification of
the description of the data (2.3, 2.6) make us reconstruct certain word-
occurrences which in turn show that certain word combinations had high
likelihood in the reconstructed source.
For example, as noted above, expect in most cases has the argument pair
N,0, as in / expect Mary to win, where the On word win has Mary as its
argument in turn. When we find a shorter second argument, apparently just
N, as in I expect Mary, the problem is to derive it from the more common O
second argument, since in almost all cases when a word has two forms of
argument one can be derived from the other (2.3). In this case it is found that
any further likely operators on I expect Mary are the same as are likely on /
expect Mary to come (or to be here). Also the likelihoods of the various N to
appear alone as object of expect are graded in about the same way as the
likelihoods of those N when they are subjects of to come (or to be here)
(alone, or when these last are objects of expect): / expect time is less likely
than / expect Mary, as / expect time to be here (or just Time is here) is less
likely than / expect Mary to be here (or Mary is here). We conclude that I
expect Mary is a reduction of / expect Mary to be here, etc. In the source
forms, once every N second argument under expect is reconstructed to N to
come (or N to be here), the locally synonymous set come, be here, etc., now
has greater likelihood as second argument of expect than any other word or
locally synonymous set in that position (G161).
Another example of high likelihood in reconstructed sentences is seen in
the operators that are called reciprocal verbs. We consider John and Mary
met, by the side of (1) John met Mary and Mary met John, whereas there is
no John and Mary saw by the side of (2) John saw Mary and Mary saw John.
We note first that when a sentence like (2) exists in the language then a
sentence like John and .Mary saw each other also exists, with each other
being taken as a reduction of Mary and John respectively. Then from (1) we
have John and Mary met each other. What is special about met is that the
each other is zeroable. This happens under all operators of a subset Orec
(reciprocal verbs) of Otw, which are definable by the condition that given (3)
/V, Orec N2 (where /V, and N2 are in the argument selection of the Orec) in a
text one can always add (4) and N2 Orec A, without affecting the likelihoods
of the further operators, i.e. without affecting what the rest of the text says
about (3). This means that we can reconstruct every occurrence of (3) N\
Orec N2 into an occurrence of (3, 4) A, Orec N2 and N2 Orec A, without
skewing the structural description, or distorting the information, in the
utterances of the language. With Orec operators, the contribution of (4) is
implicit; introducing it does not add information and is indeed not commonly
done (hence the rarity of (1) John met Mary and Mary met John). Every
occurrence of John met Mary (and so for every Orec sentence) can then be
considered as a case of John met Mary and Mary met John, which is
transformable to John and Mary met Mary and John respectively, which is
always reducible to John and Mary met each other. The each other is thus an
alternative which is available on every occurrence of John and Mary met,
and is zeroable (3.3.2). Note that when (4) is not likely, as in John met his
doom, where doom is not in the first-argument selection of met (i.e. not a
normal subject of met), so that His doom met John is most unlikely (though
one cannot claim that it has no meaning), we rarely if ever have the
reduction to John and his doom met each other, and hardly ever the further
reduction to John and his doom met (G314).
A major type of high likelihood in reconstructed sentences is to be found
in the reconstruction of word-repetition. A simple example is seen in
repetition under and, or, but. To see it, we first consider the conditions for
-self. If we compare / introduced myself, You introduced yourself, which
occur in English, with 1 introduced yourself. You introduced myself, I
introduced me. You introduced you, all of which do not, it is clear that -self is
adjoined to a pronoun (which is a reduced repetition of a neighboring
‘antecedent’ argument) as object (second argument) of an operator if and
only if that pronoun repeats the subject (first argument) of that operator:
-self is adjoined to / introduced me and not to / introduced you. If now we
wish to derive (1) / came up and introduced myself, contrasted with the non¬
occurring I came up and introduced yourself, simply by extending the
conditions for -self, we would have to say that the antecedent could also be
the subject of the verb preceding the verb whose pronominal object receives
the -self. This is not a simple extension, because many other words can
intervene between the antecedent subject and the pronoun with -self.
Instead, we can reconstruct a subject for the second-operator (introduced)
by assuming this subject to have been zeroed on grounds of its having
repeated the first argument (I) of the first operator (came up). The recon¬
structed sentence would then be (2) I came up and l introduced myself,
where no extension of conditions is needed to explain the -self.
This reconstruction is justified independently. In all sentences consisting
of a subject followed by two operators connected by and, or, but, both
operators have the class of that subject as their required subject class (e.g.
no NO,, and O0: John phoned and is probable)-, and both operators have the
70 Theory of Syntax
given subject word in their subject selection. If the second operator does not
have the first subject in its own selection the unlikelihood of the sentence is
immediately noticeable, as in John is unmarried but pregnant. Put differently:
the likelihood grading of operators after John is unmarried but ... is
approximately the same as after John is unmarried but John . . . (e.g. is
pregnant would be equally unlikely in both cases). Description of data such
as the occurrence of -self and the likelihood of second operators above
would require additional grammatical statements unless we assume that the
missing subject of the second operator is a zeroed repetition of the subject of
the first operator. By the criterion of least grammar (2.2) we reconstruct this
repetition, obtaining (2). When we do this, (1) / came up and introduced
myself is a case of (2) / came up and l introduced myself. Partly similar
analyses can be made for missing second operators (e.g. plays in John plays
piano and Mary violin), or objects of the second operators (e.g. the new keys
in John misplaced, and Mary found, the new keys), under and, or, but. When
these missing words are reconstructed as repetitions of the correspondingly
placed word before the and, or, but, we obtain filled-out second (component)
sentences after the conjunctions, without any missing words. In the set of
sentences, including component sentences, which have no missing words, it
is then seen that under these conjunctions it is more likely that at least one of
the positions after the conjunction is filled by the same words as in the
corresponding position before the conjunction, than that every position
after the conjunction differs from the corresponding position before. This
property, of expecting some word-repetition in corresponding positions, is
part of the selectional characteristics of and, or, but. As such, it contributes
to the meaning of these conjunctions, which are readily used to collect the
various acts or events done by a given subject or to a given object, or the
various subjects and objects which participate in a given act (G139).
Another type of repetition preference, not restricted to corresponding
positions, holds under many other conjunctions. When we consider bi-
sentential operators (0„„), we find what seem to be semantic differences,
between arbitrary-sounding or nonsensical sentences such as I sharpened my
pencils because the snow was soft on Tuesday and ones that sound ordinary,
like People are corrupted by (their) having access to power. It is notable that
in many cases the latter, more ‘acceptable’ kind of sentence has a word
repeated in its two component sentences (e.g. people and they) while the
former kind does not. The relevance of this observation is supported when it
turns out that the former lose their eyebrow-raising effect when a third
component sentence is adjoined which creates a word-repetition bridge
between the original component sentences: e.g. / sharpened my pencils,
having promised him a pencil sketch of a snowy tree, because the snow was
soft on Tuesday. True, there are many bi-sentential sentences which lack
repetition but are nevertheless immediately acceptable, such as 7 will stay
here when you go. But in such cases we can always adjoin a third component
sentence which creates a repetitional bridge, and which is zeroable by 3.3.2

as involving merely dictionary definitions or other knowledge common to
speaker and hearer. For example,/ will stay here and not go, when you go, or
/ will stay here, which means that l won't go, when you go. We see here that
the selectional characteristic of these bi-sentential operators is that their
component sentences be bridged by repetition of some word, with common-
knowledge bridges being then zeroable. Indeed, it is this selection that gives
them their connective meaning (9.1).
Yet a different repetition-preference is seen in many Ono and Onno
operators, whose first one or two arguments are N and last is O (this lower O
carrying its own arguments in turn). When one of the arguments of the lower
O is the same as one of the earlier argument words under the ‘higher’
operator (the Ono or O„„0), the argument of the lower operator is zeroable.
Which lower argument is zeroed in respect to which higher argument
depends on the particular higher operator. For certain higher operators
there is apparently a greater likelihood that a particular lower argument
should be the same as a particular position of the higher argument, and it is
this lower argument that is zeroed. When one of the arguments (the subject
or the object) of the O argument (the ‘lower’ O) is missing, the question is,
then, whether or not the missing lower argument is a preferred repetition of
an N argument of the higher Ono or Onno. Under some of these higher
operators, e.g. observe, hear, there is no omission for repetition: we have I
observed John shaving himself and In the mirror I observed myself shaving
myself but no / observed shaving myself. Similarly, when one says I heard
shouting and I heard the shouting at me the missing subject of shout is not a
repetition, but a zeroing of indefinite N (3.3.1) from I heard someone
shouting at me. Note that if we start with / heard myself shouting at myself,
there is no repetition-omission to yield (1) / heard shouting at myself-, the
latter is non-extant presumably because of the non-existence of I heard
someone shouting at myself, from which alone (1) could have been derived.
As to the cases where lower arguments are zeroed:
First, omissions of the subject of the lower operators: under prefer, want,
report, admit, etc., we have (2) I prefer for John to shave himself and (3) I
prefer for me to shave myself and also I prefer to shave myself (with omitted
repetition of the subject of prefer in (3)), but no I prefer to shave himself
(there being no grounds for omitting John in (2)). Similarly, there is I
reported having introduced myself falsely, which can be reconstructed only to
I reported my having introduced myself falsely, also I admit to having
introduced myself falsely. Under the Onno offer, request, ask, we have such
sentences as / offered (to) John to introduce myself with the subject of
introduce omitted as repetition of the subject / of offer, but also in some
cases / offered John to introduce himself with the subject of introduce
omitted as repetition of the second argument of offer. However, under
order, tell, advise, prohibit, prevent, we have only I prevented John from
72 Theory of Syntax
introducing himself, I ordered John to introduce himself (from I ordered John

for him to introduce himself), and no I ordered John to introduce myself; and
if we can say 1 ordered John for me to be included in the list we apparently
cannot omit the for me as a repetition of the subject I (to obtain I ordered
John to be included in the list as meaning ‘for me to be included’).
Second, omission of the object of the lower operator: under the Ono
deserve, merit, accept, avoid, escape, we have You deserve (for you) to see
yourself in this costume, and John merits (his) receiving the prize, where the
omittable (parenthesized) subjects of see, receive repeat the subjects of the
higher operators deserve, merit, etc. But we also have John deserves punish¬
ment and John escaped sentencing for his crime and John merits our help (to
him) or . . . our helping him, where it is the omitted objects of punish, etc.,
that repeat the subjects of the higher operators. In contrast to deserve, under
the Ono suffer, undergo, and is -en,( 1.1.2), we have only Napoleon suffered
defeat (or. . . Wellington's defeat of his forces), John underwent an operation
(on him) by a team of three surgeons, I withstood their attack (on me), where
it is only the object of the lower operator (defeat, operate, attack) that is
omittable as being a repetition of the subject of the higher operator. In
contrast, under the Onno defend, the object of the lower operator is omittable as
repetition of the object of the higher operator: / defended John from attack
(on him). Note that the apparent ambiguity in French defendre, in defense de
la patrie compared to defense de fumer, lies in defendre admitting both the
omissions of the English defend and the omission of the English prohibit,
prevent, order. Should the question be raised as to the evidence that the
verbs under suffer, undergo, and defend indeed have objects (which are then
zeroed), note that only transitive verbs (which have objects) appear as lower
verbs under these: there is no / defended John from sleep. In contrast, under
prefer, etc., the lower verb can be transitive or intransitive: 1 prefer for John
to wear it, / prefer for John to sleep.
In the Ooa, Ono, and Onno cases here we have seen that various omissions
in the lower sentence have to be reconstructed (by such evidence as -self) as
being repetitions of an earlier argument of the higher operator. When we
add these frequent omissions to the occurrence of observable repetitions, it
is seen that the (original) repeated material has a high likelihood, higher in
the positions noted than any one other word in those positions. The
positions of this high-likelihood repetition differ under different higher
operators, in a way that conforms to (or affects) the meaning of the higher
operators (and that does not in general vary with the meanings of the
lower operators). Thus under the higher operator withstand, the object of
attack (and of every other lower operator under withstand) has high likeli¬
hood of being a repetition of the subject of the higher operator, but under
defend (in English, but not in French) it often repeats the object of the
higher operator; under prefer the subject of the lower operator often repeats
the subject of the lower operator, while under prevent it often repeats the
object of the higher operator. One could say that this is just a reflection of
the meanings of these higher operators. But almost all selection can be
considered a reflection of meaning. Nevertheless, if we start with the
meanings—especially given our lack of an effective universal framework by
which to formulate such subtle meanings—we could not reach the particular
classification of words which we have found here, in respect to particular
grammatical processes such as the position of omission: how zeroings under
offer differ from under order, and how French defendre has the repetition
likelihoods and zeroings (and meanings) of both defend and prohibit in
English.
3.2.4. Likelihood restricted to zero
One type of likelihood remains to be discussed. All the likelihood types

noted up to now were assumed to describe combination probabilities larger
than zero. This enables us to say that the likelihood differences act on the
resultant of the operator-argument relation, since the latter excluded all
combination probabilities equal to zero. Flence no matter how strong the
selection of particular arguments under a given operator, the other words in
the argument class of that operator were assumed to have more than zero
probability there; and indeed the occurrence of very low-likelihood argu¬
ment words does not make an impossible sentence, only a nonsensical one
(e.g. The vacuum conversed with the ether), or a'non-reducible one (e.g. He
is coming; that he is coming is false not reducible to He is falsely coming,
3.2.2 end).
Nevertheless, borderline cases arise. For example, under order the omitted
subject of the lower operator can be reconstructed as a repetition of the
second argument of order, as in I ordered John to shave himself reconstructed to
/ ordered John for him to shave himself. But it is questionable whether any
other subject can appear here for the lower operator: can anyone say /
ordered J ohn for you to shave yourself ? True, one can perhaps say I ordered
John for you to be admitted, but this appears limited to those lower sentences
which may be transforms of a source in which the subject is the same as the
object of order, e.g. I ordered John (for him) to admit you or . . . (for him) to
allow your being admitted. It may therefore be that all other subjects of the
lower operator have zero probability under order. Similarly, it is difficult to
find deserve and especially merit (above) without an argument of the lower
operator repeating the subject of the deserve or merit.
This situation exists explicitly in English, in the auxiliary verbs can, may,
shall, etc.: John can play violin. We could analyze this as John plays violin
with an auxiliary on the verb, or else as a two-verb expression; with John as
subject. These analyses would involve the serious cost of introducing new
74 Theory of Syntax
and ad hoc entities into English grammar, and would violate the fundamental
partial ordering of 3.1. Alternatively, we may say that can and the others are
Ono operators of the prefer, want, type above, in which the subject of the
lower operator is omitted when it repeats the subject of the higher operator.
Then the only serious difference is that under can, etc., for the subject of the
lower operator to be different from the subject of the higher operator has
not just a lower probability but zero probability: one does not say I can for
John to play violin as one can say / prefer for John to play violin and even /
want for John to play violin. In earlier periods of English, some of the
auxiliaries indeed could have different subjects for their lower operator, so
that we have here not a principled restriction on these higher operators but a
lowering of probability to zero (G298). The result in modern English is the
same whatever the history, but the history provides support for saying that in
addition to the massive zero-probability constraint that creates the operator-
argument classes (3.1), a few words within an established operator class may
lower to zero the probability for part of their argument class (this being then
viewed as a limiting case of the likelihood constraint).
The fluidity and secondariness of this zero grade in the likelihood con¬
straint, in contrast to the large-domain zero-probability constraint that
underlies the argument relation, is seen in a few words such as order, merit,
which approach the zero grade seen in can, but fall short of the definite
grammatical exclusion that we find under can. The fluidity of the can
situation is also seen in the attempts to circumvent the exclusion, as in the
case of try. Many speakers of English would consider subject-repetition to
be required under try: / will try to shave myself, but only questionably (1) /
will try for him to shave himself; however, the questionable form is circum¬
vented by a virtually meaningless use of have: I will try to have him shave
himself can mean the same as (1) in the sense of l will try that he should shave
himself, aside from its other meaning of T will try to arrange for him . . (the
latter from I will try on l will have him shave himself).
3.2.5. Word subsets
The various special selections noted above yield certain subsets of words,
mostly such that the words in a subset have selections that are similar,
complementary, etc., but also of those words which together comprise the
selection of particular interesting operators (3.2.2). For example, there are
operators whose argument selections are virtually identical. Such are the
various operators expressing duration, onset, and the like, such as continue,
last, throughout, which appear on approximately the same selection of
arguments (which are themselves operators, such as sleep, exist, grow): He
slept throughout the night but not He arrived throughout the night (G77).
This situation has the effect of creating a set, very fuzzy in the case of
English, of durative verbs.
There are other words whose selection is similar to (mostly, has a large
intersection with) the selection of some other word, or includes the other
selection, e.g. classifier words as compared with their classificands. Also,
there are words whose selection contrasts with, or is virtually complementary
to, the selection of certain other words: sleep, grow, are said normally under
was throughout, lasted, etc., and not so readily under was sudden, occurred
at, etc.; but arrive, depart, say etc., are readily arguments of was sudden,
occurred at, and rarely of was throughout, lasted. Such contrastive selections
may partition the words of a class into two (or more) selectional subsets
(e.g. in some languages imperfective or durative as against perfective or
momentaneous verbs, G275). There are also words such as speak, think,
which occur as normally with lasted as with was sudden. Hence ‘aspect’ in
English is a graded property, which does not easily create subsets of
operators, and does not partition them as having sharply either durative or
else momentaneous selection.
In almost all cases the selectional subsets are fuzzy—imprecise or in
various degrees unstable as to membership; some words have some but not
all the properties of the subset, or new words enter the subset, and so on.
Thus the pair something, somebody (or someone) can be considered as
variants of a single indefinite noun; the word people has some of the
properties of somebody but not all. The auxiliaries can, may, etc., seem at
first like a well-defined English subset; but, for example, need, ought have
only some of the properties common to the set (G300). The property of
falsely above (3.2.2 end) is shared by uncertainly, improbably, doubtfully,
and a few others, but not with equal definiteness. Setting up such subsets is
not necessarily precluded by such complexities of conditions. For example,
an operator which is not in the selection of throughout may enter that
selection if it has a plural subject (first argument): there is no The guest
arrived throughout the afternoon, but there is The guests arrived throughout
the afternoon. This does not mean that the ‘momentaneous’ operator arrived
has become ‘durative’, but that a momentaneous operator with a plural
subject {The guests) can constitute a durative sentence. Durativity is selec¬
tional; it is thus a property of a word together with its arguments; the
definition of selection allows it to be affected by lower parts of the argument
chain. But the collecting of operator words into (fuzzy) subsets is still
possible if we choose controlled conditions with maximal distinction: English
durativity would be defined on operators with subjects in the singular (with
certain durativity restrictions on their objects too, such as between the
continuous or ‘imperfective’ line and the completive or ‘perfective’ circle as
objects of draw).
All the word-class names used above (other than the N, On, Oa, etc.
defined in 3.1) are definable in respect to selectional (3.2) or morphological
76 Theory of Syntax
(3.3) properties on the N, <9„, etc. This holds, for example, for metasentential
operators, indefinites, pronouns, adjectives, prepositions, auxiliaries, re¬
ciprocals, and durative-momentaneous verbs.
The difficulty in establishing word subsets shows that in general languages
have few if any definite subsets in respect to likelihoods (3.2). In science
sublanguages (10.3-5) it will be seen that selectional subsets are possible
and of great importance, when subject matter is restricted.
3.2.6. Frequency; low information
We have seen (3.2.2) several types of special likelihoods in operator-

argument occurrence: unrestricted and exceptionally broad selection, and
preference and rejection for certain combinations, down to the zero grade of
likelihoods (3.2.4). It remains to see just what features of likelihood are
needed in the present theory. The different likelihoods within the operator-
argument relation, which are roughly stable over short periods, would be
needed only for describing the data of actually occurring sentences within
the set created by the partial order of 3.1, and for indicating the (mostly
fuzzy) word subclasses formed by similarities of selection (3.2.5). However,
the upper and lower extremes of likelihood determine the domains of
reduction (3.3), which is a universal process in language. In order to
characterize—one might almost say motivate—reduction it is necessary to
specify likelihood difference in a way that is measurable at least in principle.
If there are n words in an argument class, say N, then for a given operator
on it, say sleep, some members of N will occur, in many non-specialized
collections of sentences, in much more than Vn of the occurrences of sleep,
while other members of N occur in much less than Vn. The occurrences for
each N will not tend toward a common Vn ‘average’ as the size of the
collection increases. We thus have not a random distribution of N as
argument words of sleep, but inequalities of likelihood, and to some extent a
grading, of N under sleep. These inequalities of likelihood are much the
same for given N in any very large corpus of data. The set of higher-than-
average-frequency members of N constitutes what was called (3.2.1) the
selection (in N) under, for example, sleep. A comparable formulation can be
made for the likelihood, and the selection, in the various O,, (includingsleep
and all other members of On) as operators on a particular member of N. In
some cases the likelihood of each member (or of certain members) of an
operator class to occur on a particular argument word depends on the
identity of the latter’s arguments words in turn.
Another way of characterizing the selection under a given operator is by
seeing the variety of further contexts (i.e. further operators and co-arguments)
when the argument is on a particular argument word. The occurrences of,
say, sleep on a particular argument word may have many different and
unrelated further operators on sleep, or different co-arguments of sleep

under these further operators, up to considerable distances in the text from
the position of sleep. Thus child as argument of sleep may occur in a personal
letter, in a hospital report, in a newspaper account of daily life, in a novel,
etc.—all with different further contexts in the same sentence and with
different neighboring sentences. In contrast, saucer as argument of sleep
may occur in fairy-tales, stories told to young children, possibly some
science fiction (many of these sources having similar kinds of further
operators and co-arguments on sleep). Also, this and many other combina¬
tions occur primarily under not, and question (/ ask whether-, I wonder if
-), and similar further operators (e.g. Saucers don't sleep). We can say
that argument words which occur under a given operator in a wide variety of
further contexts are in the selection (or normal likelihood) for that operator,
but ones such as saucer under sleep, which occur under that operator only in
a narrow range of further contexts, are not.
To a first approximation, likelihood is relative frequency. It is not
absolute frequency in utterances of the language, since the total rarity of a
word (e.g. thrips, an insect) has no general relation to its meaning—or to its
reducibility (as in The thrips crawled out and flew away from The thrips
crawled out and the thrips flew away). Nor is the likelihood measured relative
to the whole sentence; if it were, we would have the impossible task of
considering each sentence as a different environment in which to count the
word likelihoods at each position. Rather, likelihood as used here is the
frequency of each word in a given argument or operator class (N, On, etc.)
relative to a given operator word over it or argument word under it (or co¬
argument under be, as in the classifier is an insect, below). In the great bulk
of cases, it is relative to the immediate operator over it or argument under it,
but in some cases an operator or argument further up or down the chain may
be relevant to a word’s likelihood. One example was noted above in the
further operators which characterize children’s stories. Another is the
expectable frequency of melt, above, as operator on house, which, as noted,
is very low (e.g. in A house can melt), but higher if is of snow operates
secondarily on house (in A house of snow can melt). (In the source, melt does
not operate on house alone but on house plus modifier: the source here is A
thing which is a house which is of snow can melt, from A thing can melt; the
thing is a house; the house is of snow, G124). The appeal to a secondary
operator (i.e. a modifier, e.g. of snow) is not needed if we have a single word
whose selection is that of house of snow, e.g. igloo.
The considerations listed above are corrections on our first approximation
to what constitutes likelihood. In addition, the frequency has to be counted
in respect to the reconstructed sentences (3.2.3): when the -self in (1) I prefer
to shave myself shows that (1) has to be reconstructed to (2) / prefer for me to
shave myself, we count I as occurring twice in (2) but also twice implicitly in
(!)•
78 Theory of Syntax
Finally, certain words of low frequency relative to a given operator or

argument may have a classificand or synonymy or definitional relation to
words that are more frequent relative to that operator or argument. For
example, we may have a word which is so infrequent that no grading of its
observed argument or co-argument words is of interest: e.g. the number in
The rock weighed 3.073 kilograms. But while rock—and everything else—
may be very rare with this particular number, it has selectional likelihood
under normally frequent words of the same subclass, numbers, as in The
rock weighed 2 kilograms. There are many N which we may suspect were
never said before with a particular operator or co-argument, but once said
are recognizable as comparable to normal selection: e.g. (1) The very young
thrips maneuvered itself into a narrow crack in the gray bark. The immediate
acceptance of such infrequent material as a normal sentence is due to its
component sentences having special relations of classification, synonymy,
and the like to sentences of ordinary frequency. Thus, is an insect is in the
selection of the infrequent word thrips, as in the classifying sentence A thrips
is an insect; and the common move is in the classificatory selection of the
somewhat less frequent maneuver oneself. These selectional relations classify
the word combinations of (1) under the word combinations of a sentence
which is far less infrequent or unusual than (1), namely The insect moved into
a crack.
The term likelihood is used here, then, for the frequency of occurrence of
a word as specified above: relative to a particular operator or argument word
of it, in respect to the reconstructed base form of any sentence which is in
reduced form, and with allowance for classifier relations and the like. This
frequency is not in principle unmeasurable, since it is primarily a word-pair
relation, between argument and immediate operator, rather than a relation
between a word and its whole sentential environment. In addition, what
makes this relative frequency manageable is that its use in defining word
subclasses and reduction domains requires only that a few grades be
distinguished—roughly, frequency estimated to be higher than for any other
word in the given position (i.e. highest likelihood); selectional (higher than
average); selection that is exceptionally broad or narrow (in various ways);
and rejection (‘negative’).
These various likelihood grades carry different amounts of information,
in the sense that is known from the statistical structure discussed in Infor¬
mation Theory. In this sense, a word that has broad selection and high total
likelihood carries less information than other words, and a specified word
(or locally synonymous set words) that has highest likelihood in a given
environment can be considered to contribute no information in that environ¬
ment. Since various kinds of likelihood are characteristic of different subsets
of words, and since these have informational properties, one might wonder
if some other factors, specifically semantic ones, might not underlie the
various frequencies observed. This is a reasonable supposition, but the
semantic factors that we can distinguish are more varied yet, and are not
organizable in any objective way, and not measurable. Among such factors
in determining frequency are: the informational usefulness of particular
word combinations and the non-language occasions for them to be said, the
particular aspects of experience that it is useful or customary to note (time,
place, evidentiality, etc.), euphemisms and avoidances of particular expres¬
sions, expressive and communicational exaggerations (e.g. a preference for
stronger terms for time, such as may have led to development of the
periphrastic tenses, e.g. from took to has taken), preference for argot and in¬
group or colorful speech, and so on. Aside from these more or less semantic
factors determining likelihood and the change and extension of likelihood,
there is the massive effect of analogic extension and change in selection
which is determined by complex grammatical factors (of similarity, etc.),
only part of them semantic.
3.3. Reduction
3.3.0. Introduction
In reaching the syntactic theory presented here, the core of the work was to
find a small set of relations between sentences in terms of which we can say
that one observed sentence has been derived in some explicit and non ad
hoc way from another. Here, to ‘derive’ is to find a constant difference
between two formally characterizable sets of sentences. The intent in
establishing such derivations is that the meaning be preserved in corresponding
sentences (or sentential segments) of the two sets: the meaning of John
phoned and wrote is the same as in John phoned and John wrote, and the
meaning of John phoned is the same in the sentence John phoned as in I deny
that John phoned. It was found that such derivation can be effected by
reductions and transformations which have an important property: if we
undo these, the residue, consisting of the pre-reduction (‘base’) sentences
(4.1), has the transparent structure of dependence and likelihood (3.1-2).
Furthermore, it was found that the reductions which suffice to produce this
residue are not simply an arbitrary conglomeration of inter-sentence relations
which had been so chosen as to leave this residue of transparently constructed
sentences. Rather, most of the reductions fall into a coherent and not very
large set of reductions, merely in the phonemic composition of words, which
take place when those words contribute little or no information to their
sentence. The fact that language is the resultant of two coherent systems,
each with its own structural definition (and not one being j ust whatever is left
as the residue of the other) gives credibility to each of the systems. It is
therefore important not merely to note a few examples of reduction, but to
80 Theory of Syntax
present a survey of the major reductions of a language (English), sufficient

to show the character of the reduction system. -
The constraints of 3.1-2 create both the form and also the information¬
bearing grammatical relations of the base sentences. The reductions neither
add nor remove information, and are thus information-preserving. In
addition, while reductions create the vast bulk of sentences outside the base,
a language may have non-reductional transformations of the base sentences
(8.1); and other non-base sentences may be formed by analogic processes
and by borrowing.
3.3.1. The reduced shapes
The most widespread reduction is to zero. In John plays violin and Mary
piano or in John left, and Tom too it is clear that plays has been zeroed after
Mary, and left after Tom; similarly I will go if you will is produced by zeroing
go after you will. If we compare I plan to see for myself with / plan for Tom to
see for himself it is clear that for me had occurred after the first I plan and had
been zeroed there (3.2.3). It can be shown that Never send to ask for whom
the bell tolls is in principle from Never send someone to ask . . . ; and I don’t
have time to read from I don 7 have time for me to read things (since elsewhere
read has a second argument, its object, as in He reads comics). The zeroing
of to come, to be here under expect, as in I expect John from / expect John to
be here, has been noted in 3.2.3. Complicated forms, such as metaphor, the
passive (1.1.2), and the zero causative (3.3.2, as in He sat us down from He
made us sit down), can be derived from attested kinds of zeroing. That these
are indeed zeroings, not merely 'implicit meanings’, can be seen in languages
which have gender agreement. Thus in a cafe one asks for un blanc, from un
verre de vin blanc, but in a wine shop for une blanc, from line bouteille de vin
blanc.2
Another widespread reduction is to affixes. These are short morphemes
attached to words. In English and in many languages most prefixes have the
meaning of an operator whose second argument is the word to which they
are attached, their 'host’: He is anti-war comparable to He opposes war. In
some cases the prefix is visibly a reduced form, as in mat- (in maltreat) from a
word meaning ‘badly’, or be- (in because, beside) from by. As to the suffixes
of English, almost all have the status of operators, or rarely of co-arguments
on their host, as in earthly from earth-like, from like earth (leaving aside here
the specialization of meaning in earthly). For most English suffixes, there is
no evidence of their having been historically reduced from free-word
operators; but some suffixes are visibly such reduced forms, as in sinful, or in
kingdom from Old English dom 'law, jurisdiction’, or in childhood (older
2 Noted by Maurice Gross.

child-had) from Old English had ‘state, condition’. Note also the adverbial
-ly, as in slowly, from the early lice (the like above) in dative case, meaning
originally ‘with form, with body’. Reduction to affix involves reduction in
the stress relation to the host and in phonemic distance from it (called
grammatical juncture), as in -dom, -hood, or in childlike from like a child.
(For the position of affixes see below.)
Less common in English than the above are shortenings of exceptionally
widespread words. This is seen in the prepositions, e.g. of from off, and
beside from by (the) side of, about from earlier on plus by plus out. Most
English prepositions have been exceptionally short words throughout their
known history, but some have been reduced, even in recent times. Other
recent colloquial shortenings are, for example, will in He'll go soon, is in he's
here, has before verbs in He’s taken it (and before nouns, only marginally in
America, in He’s two jobs), going to before verbs in They’re gonna make it
this time.
Another kind of reduction is seen in the pronouns. These have been
considered to be words whose meaning refers the hearer to the meaning of
some other word in the discourse. But such a characterization would
constitute an addition to the grammatical apparatus, these meanings being
different in kind from those of non-pronominal words. For a simpler system,
we seek to derive pronominal meaning from the kinds of meaning found in
other words. To this end, we note that in any case we have to recognize the
existence of ‘metasentential’ words (ch. 5), and among these such as cite
something in the same discourse in which the metasentential words occur;
this holds both for words like latter which refer to a location in the discourse,
and for words like say followed by quotes, which refer to the material in the
quotes. The pronouns can be derived from words of this kind. As noted in
5.3, we would consider, say, l met John, who sends regards to be a reduction
from, roughly, (1) I met John; John—the preceding word has the same
referent as the word before—sends regards. Here John—the preceding word
has the same referent as the word before— is reduced to who (or rather to -o,
the wh- being a connective, 3.44). The word referent here simply carries its
dictionary meaning, and is not itself a referential. Of course, no one ever
said (1). This complex sentence is not the historical source of the pronoun.
However, (1) shows that the referential relation of who to the preceding
word (the first John) can be represented by a repetition of that preceding
word, plus a secondary metasentential ‘sameness statement’ (between the
dashes above). All the apparatus needed for (1), namely for the secondary
sentences, and the metasentential words and sentences, is attested else¬
where in the language (5.2-3).
All the above are reductions in situ: prefixes are in many cases reductions
of operators which had preceded the host word (the host being their second
argument). Most suffixes can be obtained as reductions or equivalents of the
last word in a compound word (G233); in many cases the suffix is originally
82 Theory of Syntax
the operator or predicate classifier on the word to which it is attached (now

its host), as in waterproof from water-proof from proof against water,
childhood from the state (had) of a child. Many of the changes of word-order
in sentences are not due to any permutation process on the words within a
sentence. Rather, some are alternative linearizations of the partial ordering
which created the sentence (3.4). Others are complex additions and reductions
which produce the appearance of a permutation process: The letter was
written by him from He wrote the letter (1.1.2).
Nevertheless, English has a few constructions which seem to be due to
transposing a segment of a sentence from one position to another (3.4.3).
An example of this is seen in the modifiers which appear (in English) before
their modificand. A modifier is a relative clause (G120), i.e. a sentence-like
construction having at or near its beginning a relative pronoun (which, who,
etc.) which refers to the word or sentential construction that preceded the
relative clause, and is the modificand of the given modifier: e.g. I met John,
who sends regards (above), or The money which is needed is unavailable, or I
phoned Mary, which was after I met you (here the modificand is the sentence
I phoned Mary). Here the which is, who is, can be zeroed, whereupon in
English certain types of modifiers can or must be moved to before the
modificand (G115, 229): The money needed is unavailable. The needed
money is unavailable-, I phoned Mary after I met you. After I met you l
phoned Mary, but from The money which is new is not counted only The new
money is not counted and not The money new is not counted. The full
relative-clause modifier with its which is never appears before the modificand.
Presumably it was not formed there; and the shortened after I met you,
needed, new, etc. before the modificand were presumably transposed to that
position after the which is was zeroed.
The same transposition is found in certain word-sequences which are
constructed like modifiers. If we consider He sells books which are (specifically)
for school, we have a zeroing to He sells books specifically for school (in
which both books and school have full stress). Here a transposition can
replace the preposition (specifically) for, yielding He sells school-books.
This is the important compound-word structure, in which the first word
carries full stress and the second has secondary stress. The last word in the
compound is the modificand (host): school-books are books, not schools. A
similar processs appears in his writing of letters (from He writes letters), the
fall(ing) of rain (from Rain is falling). This lock is proof against burglars, and
his state (earlier: had) of (being) a child, yielding respectively his letter¬
writing, the rainfall. This lock is burglar-proof, his childhood. In certain
cases the last word in the compound is reduced in situ from having secondary
stress to zero stress, thereby becoming an affix (e.g. childhood, kingdom,
above). In rare cases the last word is reduced to zero, as in red-cap, which
has to be derived from something like a red-cap-person since a red-cap is not
a cap but a person.
Another transposition is seen in residues of zeroing under and: John and

Mary (too) left early is from John left early and Mary (too) from John left
early and Mary left early (too).
We have seen that certain word-occurrences are reducible to zero, others
to short words, yet others to affixes (when their juncture, i.e. their distance
in time or space, or their stress status, in respect to their argument or co¬
argument—their host—is reduced). Also, that certain modifier-like occur¬
rences of words are transposed (‘permuted’) from being after their arguments
or co-arguments to being before them. These few changes create fixed
differences in shape, accompanied by no difference in substantive meaning,
between certain sets of sentences: so in the examples above, or as between
I'm expecting John, I’m expecting dinner, and I’m expecting John to come,
/ m expecting dinner to arrive. Many languages also have complex trans¬
formations between certain sets of sentences. These are larger fixed changes
which are accompanied by no difference or by fixed differences in meaning
(8.1): e.g. the passive, between John will send the letter, Mary will find a
stamp and The letter will be sent by John, A stamp will be found by Mary
(1.1.2); the relation (‘zero causative’, 3.3.2) between He caused them all to
sit down and He sat them all down, or between He treated the salad with
pepper and He peppered the salad (G317); and metaphor, between He
treated his speech with jokes as one peppers things and He peppered his speech
with jokes (G405). In many cases, these are composed of successive reduc¬
tions and transpositions, in some cases with further operators acting on the
original sentence. Thus the large transformations do not in general constitute a
new process in sentence formation, over and above the reductions surveyed
here.
The net effect of all the reductional and transformational possibilities is
that in many languages the grammatical relations among words in a given
sentence can be recast into many other grammatical relations among them,
preserving only the fact that those words are, at a given depth, grammatically
related, i.e. related in the partial order. This, without change in substantive
information (or in the word-semilattice of the sentence): compare He
interdicted their transmitting the story with The interdiction was his in the
transmission of the story by them.
3.3.2. Likelihood conditions
3.3.2.0. Introduction
Reductions are made in particular circumstances. The necessary but not

sufficient condition is that the word or sequence being reduced have
exceptionally high frequency of occurrence, or contribute little or no
information at the point it occupies in the making of the sentence. Only
84 Theory of Syntax
rarely is reduction occasioned by absolute high frequency independently of
the sentential environment: e.g. in goodbye. More common is the reduction
of words whose high total frequency is due to having an exceptionally broad
selection, i.e. having normal (selectional) frequency to an exceptionally
large number of other words. Lastly, there is the reduction of words which
are the most frequent in respect to the environment (usually the specific
operator or operand) with which they are appearing.
In these terms, we can understand that in the word-sequences of sentences, a
word-choice which has very high likelihood (in the environment in which it
occurs) needs less phonemic information to make it identifiable by the
hearer, and so may be shortened (including reduced in stress, or reduced to
affix, both of which mean reduced in phonemic distance from its environment).
A word-choice which is determinable from its environment can even be
reduced phonemically to zero. This last applies not only to words that have
recognizably the highest likelihood in a given environment, but also to a
word (or to any of a set of words whose difference is irrelevant in the given
environment) which is conventionally accepted as the zeroable word there:
e.g. come (or also arrive, be here, etc.) under expect. The issue here is not
even whether the come, etc., can be shown to be the most frequent there but
merely whether any hearer will supply one of these words as what was
zeroed there.
These various high-frequency conditions for reduction are related to low
information in terms of Information Theory. In the sequence of symbols that
constitutes any language-like message, each sequential choice of a symbol
(out of the total set of symbols) is a contribution to the information in the
message. In a given environment of symbols, the more expected (i.e. likely)
a particular choice is, the less that choice contributes to the information in
the message. In the extreme case, a choice that is determinate from the
environment contributes no information in that environment. In such a case,
if such a symbol is omitted (with evidence that it has been omitted), it can be
reconstructed from the environment: we can say that the choice had been
made (by virtue of the environing symbols) but zeroed—i.e. its physical
shape is zero, so that it cannot be seen except by inference from the
environment.
The low-information status of these reducible words is seen in their
meanings. In the case of words whose selection is broad, the increase in their
selection is accompanied by a decrease in the specificity of their meaning.
Compare the meaning of about, close to the meaning and selection of its
etymological source, in He walked about the room, with its meaning and its
far-extended selection, as in He spoke about this problem (G80); and the
meanings of give, take in He took a book, He gave me a book, with those in
their extended selection, as in He took a good look at it, He gave it a good
look (G302). When the extension of selection, whether by metaphor or
analogy or otherwise, is recent, the decreased specificity of meaning in the
extended selection is obvious, as in the conjunctional use of provided,

providing (as in I will go, provided you agree, G195).
In the case of words which are zeroed for having high frequency in a given
environment (which in most cases is the immediate operator or argument
under it), the fact that the word can be reconstructed by the hearer, on the
basis of the environment, shows that the word contributed no information at
that point.
3.3.2.1. Broad selection
A major example of broad selection is the set of prepositional operators

(G80), extended from their original selection of time and place words
(indicated by their etymology) to general relations, as in about above. Many
of these words have been shortened, as in beside, off (3.3.1). A further
reduction in English and generally in Germanic languages is replacing
the preposition by the ‘compound word’ construction (G233). In this
construction we begin with a preposition, which is an operator, between two
words, its arguments: e.g. books for school, from books which are for
school. The operator is zeroable if its second argument (school) has special
likelihood in respect to the first (books), primarily as stating a characteristic
property of the first argument. When the preposition is zeroed the second
argument has the status of a shortened modifier of the first argument, and is
transposed to before the modificand, as adjectives are in those languages
(3.4.3). The trace of the zeroed preposition is having the main stress on the
first word of the resultant: school-books. The process can be repeated,
yielding public-school-books from books (specifically) for public schools
(where public schools is not itself a compound), and repeated as in earth-
moving-machinery from machinery for earth-moving from machinery for the
moving of earth. Another example is stone-grey from grey like stone.
Other words having very broad selection get forms that relate them more
closely to their arguments. This is seen in the interrogative and imperative
intonations over a given sentence. It can be shown that these are derived
from saying that one is asking or demanding the given matter, i.e. from I ask
and I request over the given sentence: the / ask or / request, which can occur
on almost any sentence, can be replaced by the respective intonation (G331,
351). A different situation is seen in the tense suffixes (specifically, the past
-ed) which can appear on every verb (disregarding its morphophonemic
variants such as in stood, sang), and the plural suffix, which can appear on
most nouns. These can be analyzed as though derived (not historically, but
in the grammatical system) from operators on the whole sentence: roughly, I
state that his going is before my stating it to I state that his going is before to I
state that he went, to He went (G265); and I bought a book and l bought a
book, up to five to I bought five books (G252). Such derivations are offered
here only to show that the operator-argument structure of 3.1-2 suffices to
86 Theory of Syntax
state all the information, within its grammatical universe, in the sentences of
the language, including such categories as time and number which are not
expressed overtly in operator form. In any case, these categories, which
have very broad selection, are expressed by short morphemes attached as
affixes to the words on which they act.
As to the other affixes, it is relevant that almost all have one or another of
just a few very general meanings, such as the verbalizing set including -ize,
-ify, -ate, and en- in enrich (all of which mean ‘cause to be’) or the nominalizing
set including -ness, -ment, -ion, -ure, -hood, -ship (all of which mean
‘condition’ or the like). In the case of the past and plural, it is generally
considered now that the various forms of each are morphophonemic variants of
a single morpheme: a single past in leaned, went, stood, sang, cut (here,
zero) and one plural in books, children, sheep (here, zero).
In -ness, -ment, -ion, etc., it is commonly thought that there are various
nominalizing morphemes. However, it is very hard to state a difference in
meaning, or in host domain (which words get which affix), or in sentential
environment (i.e. further operators) that would distinguish all occurrences
of one of these from all of another. Even in the relatively few cases of a word
taking two different affixes of a set, yielding two different meanings, as in
admission and admittance, the general meanings of the two affixes (in all
their occurrences on admit) are close. The main difference is the gram¬
matically uncharacterized choice of word per affix: advancement, improve¬
ment, but progression, admission-, or purify, solidify but realize, materialize.
The selection in respect to higher operators is roughly the same for all affixes
within a set, except for special problems as in -ion, - ance on admit. Hence we
can consider -ify and -ize, for example, to be complementary variants of a
single variant-rich morpheme A whose combined selection (i.e. all the stems
on which all variants of A occur) is much broader than the selection of any
one verbalizing affix. Seen in this way, virtually all English affixes either
themselves have exceptionally broad selection of host (their argument), or
else are variants of a suffix-alternation (e.g. the plural or the verbalizing
above) which in total has such selection. It is this broad selectional property
that gives the variant-rich morpheme a low information status in the sense of
3.3.2.0, which justifies its reduction to affix (in varying forms). Many problems
nevertheless remain in analyzing the selection and source of the individual
affixes.
The broad-selection word set and its synonyms can also be zeroed in
certain situations, when they occur with words for their members. For
example, (1) Gilbert and Sullivan wrote operettas, cannot be formed from (2)
Gilbert wrote operettas and Sullivan wrote operettas, since anyone acquainted
with the matter would not say (2); this, differently from Mozart and
Beethoven wrote operas. However, from (1) we can reconstruct (3) The team
(or set) containing Gilbert and Sullivan wrote operettas, from A team wrote
operettas; the team contained Gilbert and the team contained Sullivan. In (3),
the team containing is zeroable, as being something that could be said in

almost every sentence in which Gilbert and Sullivan is said (especially in that
order, with the stress pattern of a single name). Then, in the reconstructed
sentences, the team of (or the team containing) is very frequent before
occurrences of Gilbert and Sullivan; it also adds no information, being
common knowledge to people who speak about Gilbert and Sullivan,
especially in the stress pattern stated.
Zeroing of broad-selection words in fixed positions is seen in the indefinite
nouns. Something and someone are, between them, in the selection of every
operator: I ate something; I read something; Someone spoke'. Someone, or
something, fell. In the second-argument position, the indefinite is zeroable
under many operators (/ ate from I ate something', I read all night from I read
things all night). But it is not zeroed under certain operators, such as wear,
where perhaps the indefinite is less frequent as second argument (no / wear,
which would be from a perhaps less common I wear things).
Further, these indefinites, and classifiers (i.e. words naming a class) with
more limited but still very broad selection (and hence broad meaning), can
be zeroed when they are the last word in a compound. This case was seen in
such compounds as red-cap, which differs from other compound nouns in
the status of its last word (the modificand): as noted, a red-cap is not a cap.
But if we assume the source to be one (or: person) who specifically (by
definition) is with a red cap reduced to red-cap-one (or red-cap-person), then
the properties of compound nouns would be satisfied, and a zeroing of the
low-information, broad-selection indefinite or classifier host (last word of
the compound) would then produce red-cap.
A similar zeroing would explain the ‘zero causative’ seen for example in
The nurse walked the patient. He leaned the broom against the wall, which
must be from, roughly, The nurse enabled (or: helped) the patient to walk. He
caused the broom to lean against the wall. The transformation would have to
be via length permutation (3.4.3), yielding e.g. He caused-to-lean the broom,
and then compounding (3.3.1), yielding a non-extant lean-cause, and then
zeroing the broad-selection cause (3.3.2.1), leaving lean and -ed. The trace of
this transformation is seeing an intransitive verb (O,,, walk) appear as
transitive, i.e. with a second argument, and one which could be the subject
of that verb.
3.3.2.2. High frequency
Reduction also occurs in some high-frequency set phrases, such as Goodbye

from earlier God be with you, and ’s, ’m, ’ll, ’d from is, am, will, would',
compare also the Spanish usted.
The most varied and complicated reduction is to zero, always in
word-occurrences that are reconstructible from the environing material
88 Theory of Syntax
(occasionally with some ambiguity due to degeneracy, i.e. to different

possible sources). In almost all cases the high-likelihood words that can be
zeroed, and their positions, are fixed. The difference between reduction to
zero and reduction short of zeroing is seen in the treatment of expected
word-repetition. In situations where the occurrence of repetition is frequent,
and expected, but not specifically in any one position, a repeated word is
reduced to a fixed form, called pronoun. Within a connected discourse there
is a considerable likelihood that some of the referents will be repeated (9.1),
though mostly in unpredictable positions: if some person or things are
referred to, it is highly likely that one or another of them will be referred to
again later on. When such repetition occurs, i.e. when a word is repeated
with a statement that it has the same referent as another occurrence of that
word (cf. who in 3.3.1), then the whole sequence—the repeated word
together with the sameness statement—is reduced to a pronoun which states
only the fact of repetition. Thus John came in, but I was busy. So he left
would be considered a reduction of (roughly) John came in, but I was busy.
So John (which has the same referent as a preceding occurrence of John) left.
It is this kind of derivability that shows he to be a reduction and referential,
not a new word.
In contrast, when a repeated word is in a fixed position where repetition
has high frequency and where the word is reconstructible in situ by the
hearer, then it is reducible to zero. This was seen in the Ono, Onno operators
where the zeroing depends on the higher operator—specifically on whether
the first or second arguments of the lower operator is likely to be the same as
one of the noun arguments of the higher operator. As was seen in 3.2.3:
under hear, observe, where there is no special likelihood of repetition, there
is no zeroing of an argument if it is repeated (e.g. no You heard shouting at
yourself as a reduction of You heard yourself shouting at yourself). Under
prefer, the subject (first argument) of the lower operator (which may be
intransitive or transitive, i.e. with either one or two arguments) is zeroable if
it is the same as the subject of the higher operator (/ prefer to stay from I
prefer for me to stay). Under order, the lower subject is zeroable if it is the
same as the higher object (second argument) (I ordered you to shred it from I
ordered you for you to shred it). Under suffer, the object of the lower verb
(which is here necessarily transitive, i.e. has an object) is zeroable as
repetition of the higher subject (/ suffered defeat from I suffered someone's
defeat of me). Under defend, the lower object is zeroable as repetition of the
higher object (/ will defend you from attack from I will defend you from
anyone's attack on you). The high frequency of fixed-positioned repetition
permits these repeated words to be reduced to pronouns in most cases, but
they can also be reduced to zero, which was not possible when the positions
were not fixed, as they are here.
Another zeroing of word repetition in fixed position was seen in 3.2 when
a word in the second sentence under and, or, but, more than is zeroable when
it is the same as the word in the corresponding position of the first sentence:
John plays piano and Mary violin.
A less obvious case of zeroing the repetitions in fixed position is seen in the
bi-sentential operators (O00) whose selectional preference is that in their
arguments there be some word in the first sentence occurring also in the
second. It is seen in 3.2.3 and 9.1 that such conjunctional sentences which
seemed nonsensical became acceptable in an ordinary way if we interposed a
bridging sentence that related a word from the first component sentence to
one from the second: if in I sharpened my pencil because the snow was soft we
interpose, say, since / had promised a pencil sketch of a snowy tree. These
observations suggest that when we see conjunctional sentences that are
acceptable despite their lacking word-repetition we should postulate an
interposed bridging sentence that was common knowledge to speaker and
hearer and hence reconstructible and zeroable: e.g. for I'll stay if you go we
reconstruct an intervening component sentence like which means, I won’t
go-
There are various fixed positions in English in which we find zeroing of a
particular word which may be the most frequent and is certainly the most
expected, in that it, or in certain cases also synonyms to it in the given
position, can be readily reconstructed by hearers of the sentence in agree¬
ment with the speaker. Among the examples, presented below, are: which
is; also for and its synonyms before time and distance words; come and its
synonyms under expect; each other under ‘reciprocal’ verbs.
When a sentence under a secondary intonation (which can be marked by
semicolon) follows a preceding sentence, or enters as interruption after a
word in the preceding sentence (3.4.4), the first word of the interruption is
frequently which, who(m), etc., as below. This wh- word is a pronoun (i.e. a
repetition) of the preceding noun. The great frequency of the wh- pronouns
means that more interruptions had begun (before reduction) with a repetition
of the word or sentence before them than with any other word. This high
likelihood explains why the repetition is reduced to a pronoun, and also why
the which, whom, is zeroable when not separated from its antecedent by a
comma: A boy whom I knew ran up to A boy I knew ran up. One verb, is,
occurs much more commonly in interruptions than does any other verb, and
the sequence which or who plus is (even if introduced by comma) is
zeroable: The list which is here is incomplete reduced to The list here is
incomplete.
Given that the zeroability of which is arises as noted above, the decision as
to whether a particular occurrence of which is is zeroed depends on an
interesting further consideration of likelihood, which creates the ordering of
adjectives before a noun. The consideration here is that given A„ . . . A{N,
i.e. a noun with ordered preceding modifiers (mostly adjectives or adjectival
nouns) that have accumulated before the noun, a following which is is
zeroed only if the adjective following which is has good likelihood of
90 Theory of Syntax
occurring as an additional modifier, i.e. as interruption after the whole

A„ . . . A,N. This is an understandable condition, since upon zeroing this
which is the n+ Yh adjective which it carries moves automatically (in
English) to before the A„. .. A ,/V and thus enters a closer modifier relation
to the whole sentence. A new red car is said of a red car that is new (a new one
of the red cars), while a red new car is a reduced form of a new car which is
red (a red one of the new cars). This is clear semantically, and also from
checking the relative permanence as against accidentality of modifiers which
occur closer or farther before the noun (the generally most accidental or
changeable are the farthest). 1 f A„N, is Ah is more likely than AhN,is At, then
Aa is more permanent than Ah, at least on Nr Since no property or modifier
is more permanent than its modificand, the fact that each adjective A „+1 is a
modifier of the whole following sequence An . . . AXN means that the
adjectives before a noun recede from the noun in monotonically descending
order of permanence and of likelihood. This is the adjective order wherein
the adjectives closer to the noun are more durative or relevant classifications
at least in the given occurrence, as we see for example in a dented cheap old
red car. This permanence order decreasing with distance results simply from
restricting the zeroing of which is to the case when the adjective following it
has high likelihood to be an interrupting predicate, i.e. a further modifier
(3.4.4), on the whole A„ . . . A{N before it.
On this analysis, the pre-noun (attributive) adjective order gives indirect
support for the attributive adjectives being reductions of relative clauses (by
zeroing the which is), and for the informational grounds of reduction (of the
which is). For here the unusual ordering seen in the pre-noun adjectives is
obtained simply by repeated application of the which is zeroing on each
n + llh relative clause if it has high likelihood in respect to the host noun
together with its previously reduced relative clauses. If a person wants to
refer to the red one among the new cars, and has already said red car, he will
not reduce a following which is new (from ; the car is new) to obtain new red
car (except perhaps by using contrastive stress on red).
The analysis given here builds up the adjective order by reducing each
successive relative clause to an adjective as long as it is a likely further
modifier on the noun as modified so far. This does not mean that all
sequences of secondary sentences must be reduced to relative clauses and
further, in this order. Quite the contrary, it is possible to build up two
sentences, each with one or more relative clauses (from secondary sentences
on it), and make the whole second sentence (with its relative clauses
attached) into a large relative clause on the first sentence. An example was
seen in the derivation of the 'restrictive' relative clause in 3.2.6, where
A house of snow can melt was derived from A thing can melt carrying as a
secondary sentence The thing is a snow house from The thing is a house which
is of snow (where the is a reduction of the sameness statement on the second
thing).
There are various other cases in which words with highest expectability
in respect to their argument or operator are zeroed. For example, operators
expressing duration have exceptionally high likelihood on arguments which
refer to periods of time and are then zeroable in appropriate situations:
He talked during (or for) a full hour is reducible to He talked a full hour,
and I saw him during one morning (or this morning or mornings) to / saw
him one morning (or this morning or mornings); but I saw him during the
(or a) morning is not reduced in English. Similarly for space-measuring
operators on measured-space arguments, which is a high-likelihood or
preferential combination: He ran for a mile reducible to He ran a mile, and
He paced through the halls to He paced the halls', but He circulated through
the halls is not reduced, since circulate is less likely to be here a measured
activity.
More specifically, certain words are conventionally the favored or ‘ap¬
propriate' arguments or operators of certain others, and are zeroable when
they occur with those others. An example is come, be here, or the like, under
expect discussed previously. Another example given in 3.2.3 is the zeroing of
each other under ‘reciprocal' verbs, as in John and Mary met each other
(where if John met Mary it is almost certain that Mary also met John), but
not elsewhere, e.g. not in John and Mary saw each other (which is not
reduced to John and Mary saw). Also scale-operators (G245) are zeroable in
the presence of scale-measurements: e.g. These chairs begin at $50 from The
costs (or prices) of these chairs begin at $50, or This chair is $50 from This
chair costs $50 (where the operator cost is zeroed, leaving the tense -s which
is then carried by be). Such favored-word zeroing can also arise marginally in
nonce-forms, colloquialisms, or special subject-matters: e.g. at the chemist's
from at the chemist's place or the like, She’s expecting from She's expecting a
baby. Coming? from Are you coming?, and Coming\ from I'm coming (the
latter two despite the fact that the subject of a sentence is not otherwise
zeroed in English). These words are reconstructible by the hearer from the
environment, whether or not they are the most frequent there. Their
meaning is assumed even in their absence, hence their presence contributes
no information.
A different kind of low-information zeroing is seen in sentence-segments
which engender in the sentence an event that expresses their information,
making them superfluous in the sentence. This zeroing is seen in the
performative segments Isay (but not / said ox He says) on a sentence, where
the saying of the sentence carries the same information as / say on it (5.5).
There are many kinds of evidence that a sentence as said is derived from its
form plus a preceding / say or the like. In the case of other performatives,
such as I ask, where again saying I ask on a sentence constitutes asking it,
one can consider ask to be derived from say plus as a question or the like,
with I say zeroable and as a question imposing the interrogative intonation.
Note that the performative is zeroable only in its performative occurrences:
92 Theory of Syntax
He will be angry if l ask: Are we losing? has no associated He will be angry if

are we losing?
In contrast to all the high-frequency reductions above, exceptional low
frequency can block a reduction that would otherwise take place. Thus it was
noted at the end of 3.2.2 that the secondary-sentence intonation (the
semicolon conjunction) has negative selection toward having its second
argument directly deny its first argument (e.g. He is coming; that he is
coming is false), with the result that the usual reduction of such secondaries
to adverbs does not take place here (hence no He is falsely coming), even
though the unreduced form is sayable within English.
The survey above of the main types of reductions shows them to be
based, roughly, on exceptionally high frequency or expectancy, or excep¬
tionally broad ranges of higher-than-average frequency (selection), within
the operator-argument relation. Irregularities and mistakes in the application
of this criterion can arise. For example, in He's away, because his car is not in the
garage there is indeed a zeroing of I say, but that / say was merely a part of
the first argument under because, and it should not have been zeroed; the
source is Isay he's away, because his car is not in the garage, where I say is not
zeroable. (His being away was not caused by the absence of the car, unless
he went to report it stolen.)
Aside from these likelihood conditions there may be, in one language or
another, positional conditions for reduction, especially for zeroing. Thus in
English the first argument of the highest operator is not normally zeroed,
although this is done in the colloquial Coming! above. As to repetitions,
what is reduced is usually the occurrence that is second (linearly), or lower in
the partial ordering of words in the sentence; but reductions in the first
occurrence can be found, as in He had liked it from boyhood, so John
majored in geology. However, by the side of John was emboldened by his
winning the argument we do not find (for the same meaning) He was
emboldened by John's winning the argument. The positions in which pro-
nouning and repetitional zeroing take place differ considerably, as also the
positions of repetitional zeroing in contrast to other high-expectancy zeroing.
For example, the final (host) noun of a compound can be zeroed, though
only rarely, if it is an indefinite, as in red-cap above. But it is not zeroable on
grounds of being a repetition, as can be seen if we compare He is manager of
the office and she is manager of sales (where the second is manager is
zeroable) with He is office-manager and she is sales-manager (where neither
manager nor the,second is .. . manager is zeroable).
The likelihood condition for reduction is practicable because it involves
not the absolute likelihood of occurrence, nor even the general inequalities
of likelihood, but a few grades: extreme expectability, extreme unexpected¬
ness (rejection), and higher-than-average likelihood.
3.3.3. Required reductions

To speak of required reductions might seem to contradict the demand for a
least grammar since it means that in addition to positing a particular form we
have to say that the form does not exist as is but must be reduced. And
indeed the great bulk of reductions are optional. That means that the
unreduced forms exist, and can be constructed in the grammar by 3.1-2, and
that the reduced forms also exist and can be constructed by making changes
in the unreduced sources (3.3). Obtaining the reduced form out of the
unreduced, by a reduction satisfying specified conditions, makes for a far
less redundant grammar than constructing the reduced sentences de novo,
since the reduced and unreduced forms are otherwise similar.
In various languages, however, there exist forms for which the least-
grammar demand makes us first construct a non-extant form (which is
similar to all other forms of its type) and on it impose a required reduction
(of some known type) which yields the form that is extant. The most obvious
case in English is that of the auxiliary verbs can, may, etc., where we first
form the non-extant I can for me to shave myself with required repeated
subject (3.2.4)and required zeroing ofthe repeated subject (with/or. . . to),
comparable to the preferred but not required zeroing of for me in I want for
me to shave myself and the optional one in I prefer for me to shave myself.
Had we constructed I can go directly (as verb acting on verb), not by
reduction (from verb acting on sentence), we would have added to the
grammar a set—fuzzy, at that—of operators whose second argument (e.g.
shave) does not keep its own arguments (here, I) with it; it would then be
impossible to define a sentence as a sequence all of whose words have their
argument dependence satisfied.
Another required reduction in English is zeroing a repetition of the word
that carries the comparative: John is older than Mary is (from . . . than Mary
is old), John bought more books than Mary bought (from . . . than Mary
bought books)-, but John likes books more than Mary does (from . . . than
Mary likes books) where the more is on the whole component sentence John
likes books, hence on its operator like, so that the second like is zeroed
leaving its tense -5 to be carried by do. Yet another required reduction in
English is moving a post-which is adjective to before the preceding noun
(more exactly, the preceding A .. . AN) when the which is is zeroed: a new
red car from a red car which is new (we do not find a red car new). Also,
under and, or, the moving of a verb whose object has been repetitionally
zeroed, to before the antecedent object: John composes pop-music and John
plays pop-music reduced to John composes and plays pop-music, without
the intermediate John composes pop-music and plays-, also John composes
light classical music and Mary plays light classical music to John composes,
and Mary plays, light classical music.
Other than this, almost all reductions in English are optional.
94 Theory of Syntax
3.3.4. How reduction takes place
The fact that reduction is based on operator-argument likelihood rather

than on features of the whole sentence means that its applicability depends
on a subset of word-pairs rather than on the unmanageable set of all
sentences.
A reduction or other transformation cannot take place before the con¬
ditions specified for it are satisfied, i.e. are in the sentence. For example,
plays in Mary plays piano cannot be zeroed before that sentence is con¬
nected by and to John plays violin. Furthermore, the zeroing of a cor¬
respondingly positioned repetition under and (3.3.2) does not take place
when the and interrupts its first component sentence too early. Since this
zeroing compares the partial-order status of a word in the second sentence
with that in the first, it can only be carried out after the partial order of the
first sentence is available for comparison. In John—and John is very
insistent—thinks we should not go the second John is not zeroable (even
though the first has already occurred), because the partial-order (i.e.
grammatical) status of John in the first sentence is not yet known at the point
when the second John is said. Note that in John thinks—and is very
insistent—that we should not go the second John is zeroable (from and John
is very insistent) because the preceding John thinks suffices for the status of
the first John. ITad the first sentence been John I don't like, the interruption
and John is very insistent would not admit zeroing (no matter where it
entered) because the two occurrences of John do not have the same status:
from John—and John is very insistent—I don’t like we cannot obtain John—
and is very insistent—I don’t like.
It follows from this that words are reduced only in their reducible, i.e.
high-expectability, occurrences. Consider the colloquial reduction of going
to to gonna. The going to has selectional (normal) likelihood on virtually
every verb, and hence very high total likelihood on verbs: They are going to
sleep, They are going to reject it, or to go, or to roam. But it does not have
selectional likelihood on every noun: They are going to is normally likely on
Rome, the river, the attic, a bit less on the chair, the fox, still less on the
teaspoon, vanishingly little on the electron, time. Hence the reduced form
gonna occurs before verbs (They're gonna come soon. They're gonna
stampede), but not before nouns (no They’re gonna the attic. They're gonna
Stampede, Washington). More generally, this shows that it is not a particular
word but a particular word-in-combination that gets reduced.
Given these considerations as to when a reduction cannot take place,
there is also evidence that if it does take place, it does so before any further
operators act. For example, given John knew that Mary came quickly, we
also have paraphrastically John knew that Mary quickly came (or. . . quickly
Mary came), but not John knew quickly that Mary came or John quickly
knew that Mary came: in the latter cases, quickly cannot refer to came. The
only simple and non ad hoc explanation is that the permuting of quickly
(from Mary came quickly to Mary quickly came and Quickly Mary came)
took place before knew operated on the pair John, came; hence this quickly
could not move into the then non-present John knew segment.
Furthermore, John slowly knew that Mary would recover is rare, and if it is
said its slowly refers to John's knowing, and not to Mary’s recovery; it is not
a permutation of John knew that Mary would slowly recover or John knew
that slowly Mary would recover. That is, if slowly is said on (1) John knew
that S, (where S, is Mary would recover), it can be permuted among the
arguments and operator in (1) but cannot enter inside an argument of (1).
Reductions can thus be located between fixed points in the semilattice of
word-entry—i.e. of operator-argument relations—in the sentence (6.3,
8.3).
Since reductions are carried out before any further operator acts on the
sentence, any further operator acts on the reduced sentence. This does not
affect the content on which the further operator is acting, since the reduc¬
tions are substantively paraphrastic, and indeed any zeroed material is
reconstructible and can be considered to be still present though in zero form.
But it may alter the formal conditions for the further operator, and indeed
some operators cannot act on particular reduced forms. If there are cases in
some language of an operator not being able to act unless its operand
sentence has been reduced, that would compromise the ability of the base to
carry all the information in the language (4.1, 11.3).
As to how reductions are ordered with respect to each other:
Two or more reductions may be applicable as a result of a single operator
entry. For example, when and operates on the pair John phoned Mary, John
wrote Mary (with sameness statements, as in 3.3.1, on both/0/2/7 and Mary),
it simultaneously makes both John and Mary zeroable in the second sentence,
yielding John phoned and wrote Mary. These two zeroings are independent
of each other; neither, either, or both may take place, with no indication of
their being ordered in respect to each other. Somewhat similarly, in (1) John
came home, which was because John was tired (with sameness statement on
John) we have available a pronouning of the second John (to John came
home, which was because he was tired), and independently of this a zeroing
of the which is (to John came home because John was tired or . . . because he
was tired), followed by a possible permutation to Because John was tired
John came home or Because he was tired John came home. In the former
case, pronouning of the second John is still available, now yielding he in a
different position: (2) Because John was tired he came home. (Pronouning
the first occurrence, as in They were as yet unsure of themselves, when the
doctors first suspected a viral cause, is so limited as to require more stringent
conditions than the second-occurrence pronouning discussed here.) It is
not necessary to state a time order for these events, which would require that
96 Theory of Syntax
pronouning be able to take place both before and after the permutation. A
least-restriction grammar can obtain the existing forms by saying merely that
both the pronouning and the which w-zeroing plus permutation are avail¬
able in (1) above until some further operator acts on the whole of (1). Such a
further operator is seen, for example, in Johns coming home, which was
because he was tired, interrupted our conversation. In (1), the permutation, if
it takes place, creates a new pronouning-possibility in John comes home, if
pronouning had not already taken place in because John was tired, yielding
(2). All we need say, then, is that the conjunction which brought together
John came home and John was tired, expressed in (1) by commas, admits two
independent reductions which are available, unordered among themselves,
until a further operator enters the sentence.
In contrast with these, there are cases of reductions which are ordered,
simply because the conditions for one of them to take place require the result
of the other. An example is moving an adjective to before the host, after its
which is has been zeroed, as in costs which are allowable to costs allowable to
allowable costs; the permutation requires zeroing of which is (there is no
which-are-allowable costs). Another example is zeroing a high-expectability
operator, or a classifier noun, after it has become the last (weak-stressed)
member of a compound: a quick drinking of coffee to a quick coffee-drinking
to a quick coffee (in Let’s stop for a quick coffee)-, a fly like a blue bottle to a
blue-bottle-fly to a blue-bottle.
A reduction is described as altering the shape of a word, even if to zero,
but nothing is said about its altering the presence of the word in the given
occurrence, and indeed the word continues to be present in the sentence.
This is evident in the meaning, where a blue-bottle is a fly. It is also evident in
various grammatical features, such as the presence of modifiers of zeroed
words, as in the momentarily of I expect John momentarily from / expect
John to come momentarily, momentarily is clearly on come.
In view of all this, it is not merely a figure of speech to say that reduction
does not affect the basic syntactic status of the words which have entered
into the partial order that creates a sentence. When we consider reductions
as variant forms of a sentence, then, the reductions do not add constraints on
equiprobability except, slightly, in the case of permutations. But if we want
to account directly for the observable combinations of words that constitute
sentences, then pronouning (which replaces many words by the same he, or
it, etc.), and especially zeroing (no matter what traces are left), have
complex effects on the departures from word-equiprobability. In any case, a
transformed sentence has to be described by describing its source sentence,
plus the change; if we try to describe the transformed sentence de novo, we
have to repeat for it almost everything that was said for the source sentence.
3.4. Linearizations
3.4.0. Introduction
In all languages, speaking is an event in time only, and thus overtly linear,
even though the linear form can contain simultaneous components, and can
even contain long components which are simultaneous with a sequence of
short ones/ The partial order established in 3.1 is descriptively prior to the
linear order; that is, it is independent of the linear order in a description of
language structure, whereas the linear order is not useful toward a structural
description without reference to how it embodies the partial order (or the
grammatical relations, which can be defined in terms of that partial order).
This is not to say that the partial ordering exists in time or in human activity
prior to the linear. There is no evidence against saying that sentences are
constructed as word-sequences, with the partial ordering bringing in an
additional degree of freedom (and so an additional dimension) in the choice
of words, recognized now by educated speakers as ‘grammatical relations’
among the sequenced words. Within the sequence, the partial order is
realized by two linear orderings: that of an operator to each part of its
argument, and that between co-arguments {John ate a fish being not the
same sentence as A fish ate John).
The relation of linear to partial order differs for speaker and hearer.
The speaker chooses certain words in a partial order which satisfies their
dependences, and says them linearly in time. Since much of the partially
ordered set, particularly the arguments of operators said so far, is available
to him when he is saying the linearly earlier words, there is room for his
making the kind of slip called anticipatory contamination, such as Spoonerisms
(as notoriously in our queer old dean for our dear old queen). The partial
order, which underlies the linear, suggests that slips which are unintentional
should appear chiefly in operator words drawing upon their linearly later
arguments, or in argument words drawing upon their linearly later co¬
arguments; this is indeed found to be the case. We would not expect slips in a
word to draw as frequently upon linearly later words that are also higher in
the partial order, for these higher words are not necessarily in the speaker’s
mind when he says the earlier word. Intentional (jocular) contaminations do
not show this distinction, since the speaker is then drawing upon a complete
sentence or segment.
The hearer receives the words in linear order of time or writing and has to
reconstruct their partial order as grammatical relations. In this he meets the
well-known problems of degeneracies, especially those due to homonymy
and zeroing. Upon occasion he also has to reinterpret an earlier word on the
3 Z. Harris, Methods in Structural Linguistics (pb. title: Structural Linguistics) AJnw. of Chicago
Press, Chicago, 1951, pp. 45, 50, 92, 117, 125, 299.
98 Theory of Syntax
basis of its linearly later operators or arguments, as in She made him a good
husband because she made him a good wife. Here the expectation that the
second made has the same partial-order relation to its following him as does
the first one is counteracted by the gender of wife: the first made him is from
made him into, the second is from made for him. Somewhat similar is
K. S. Lashley’s example, 'Rapid righting with his uninjured hand saved from
loss the contents of the capsized canoe’, where the hearer normally assumes
rapid writing until he hears the last words.4 In reconstructing the partial
order, as grammatical relations, out of the linearly ordered words, the
hearer uses the residual traces of each reduction which has taken place in the
partially ordered material. In addition, certain changes are made in respect
to the linear order: phonetic assimilation and dissimilation in neighboring
words.5
3.4.1. Normal linearization
Thus the major departures from word-equiprobability in the making of

sentences are made in a partial-order relation. This means that speech
linearization has yet to be stated, over and above the partial order. One
might think that the linearization could constitute a complex change in the
constraints on equiprobability, possibly with different linearizations for
different parts or kinds of the partial order. This is not the case. It seems that
each language has simple normal linear projections of the partial order. In
English, the operator is said normally after its first argument: the On sleep
appears in John sleeps, the On„ eat appears after John in John eats fish. In
short sentences, which contain only one operator, the arguments have to be
next to their operator, and this is the order in which they appear. In longer
sentences, if the same projection is to be retained with least modification,
the further operators Ono, O00, etc., acting on O arguments, would have
the same linear order in respect to their arguments; and this is indeed the
case. The Ooa cause acting on two O, e.g. phone and leave (each necessarily
with its arguments next to it), produces Mary phoning John caused his
leaving; and the Onoo attribute ... to on the triple /, leave, phone would yield I
attribute John's leaving to Mary phoning him. (If we expect the partial order
in long sentences to be the same as in short sentences, it follows that if there
are several N in a sentence under Onoo, the N which is its first argument is
likely to be the one before it, just as is the first-argument N under Onn in the
short sentence John eats fish.)
4 K. S. Lashley, ‘The Problem of Serial Order in Behavior’, in L. A. Jeffress, ed.. Cerebral

Mechanisms in Behavior, The Hixon Symposium, Wiley, 1951.
The relation of speaker’s grammar to hearer’s will be touched upon, though very slightly in
4.3 and 12.1.
In English and in many languages an operator (with its arguments) which

is appearing as an argument of a further operator receives a marker (affix or
word, e.g. -ing, s -ing of, that, for . . . to) to indicate its new status as
argument: That John arrived surprised me, John's buying of the book
surprised me, / know that John arrived (where that is omittable), / prefer for
John to stay, I prefer John’s staying. This argument indicator obviates certain
ambiguities. Consider three sentences connected by two conjunctions:
SCSCS;. In He came over because she was leaving because they had quarrelled,
there are two ways of grouping and understanding the sentence: SC(SCS) is
not the same as (SCS)CS. But in His coming over was because of her leaving
being because of their having quarrelled, which is SC(SCS), as against His
coming over being because of her leaving was because of their having
quarrelled, the two groupings above are differentiated. Spoken language has
no parentheses, but the argument indicator does their work. If the linearization
places the operator before or after its full argument, instead of between the
parts of the argument, this ambiguity does not arise, and the argument
indicators are then not needed (though they would still be a convenience), as
parentheses are not needed in Polish notation of logical formulas.
3.4.2. Alternative linearizations
The fact that there is a normal linearization of the partial order does not
preclude there being alternative linearizations of it at different points or in
certain situations. Indeed, one of the advantages in describing sentences in
terms not of a direct linear ordering, but of a partial order which is then
represented linearly, is that a limited freedom in the linear representation
can account for important differences in sentence structure without con¬
stituting a difference in the grammatical relations which characterize the
sentence: In many languages certain parts of a sentence which are normally
not at its beginning have an alternative linearization at the front of the
sentence, without changing the grammatical relations: This I will say (or I
this will say) instead of / will say this. There is no evidence that this
alternative linearization (‘fronting’) is a permutation taking place after the
normal one has been formed; rather, it may be simply a different linear
representation of the partial order (here say as operator on I, this), used for
stylistic reasons or to make the front word stand out as the ‘topic’.
That which is subject to alternative linearization (i.e. that which is
linearized) is not simply particular words of the sentence, but rather
particular items of the partial order, as will be seen. Furthermore, certain of
these items have their adjuncts (modifiers, 3.4.4) and even their second
arguments with them in any alternative position. In English, a noun always
has its adjuncts (Long books I cannot read-, never Books I cannot read long);
an operator optionally (Read I will immediately, Read immediately l will).
100 Theory of Syntax
Prepositional operators always have their second arguments (Near the lab he
has a house, never Near he has a house the lab)-, verbs optionally (Sign I will
any and all petitions, Sign any and all petitions I will). The distant items do
not conflict with contiguity: arguments and their operators and their co¬
arguments all remain contiguous (except for later inserts). Since adverbs are
secondary-sentence operators on their host verbs, they remain next to either
the verb or the arguments of the verb (which were present with the verb
when the adverb acted on it).
The main alternative linearization in English is the front position of non-
first arguments and of tenseless operators: specifically, the second argument
(This I say), the third argument (as in On the table I put the book), or the
operator with its non-first arguments (Explain this l will not. Tired he was).
There is a rarer placing of the second argument before the operator (/ this
doubt), and a limited interchange of the two arguments without losing their
identity in the partial order (i.e. in the grammatical relations, as in This say
we all). In longer sentences there are certain partial-order positions which
do not have the alternative available to them: they are too far grammatically
(too complexly down in the partial order) to appear in the front. This applies
to certain positions in sentences which are second arguments of certain
conjunctions: for example, the last argument (Hiroshima) does not have the
choice of occurring up front in I didn't mention Nagasaki because of your
detailed description of Hiroshima-, but it has that choice when it is deep in a
chain of arguments, as in I remember his describing how they made the
decision to bomb Hiroshima, where one can say, if awkwardly, Hiroshima
I r^nember his describing how they made the decision to bomb.
Another set of alternative linearization has certain adverbs, especially
those (such as scarcely) which operate on the metasentential I say (5.5.3),
appear at the front of the sentence, followed by the tense: Scarcely had I
taken it. Slightly different is the order in Tired though he was, and in Here
arises an important question, and also in the moving of tense when certain
front words are dropped: Were he there l would go as compared with If he
were there l would go, and I ask: Will he return? from I ask whether he will
return f
3.4.3. Permutation
There is in English one other important type of alternative linearization.

This is the permuting of neighboring long and short constructions at the end
of a sentence. For the major case, in English: an adjoined (secondary)
6 Both fronting and interruption (3.4.4), together with their semantic effects, can perhaps be
obtained without assuming a special added freedom by deriving them from adverbs on say
(5.5.3): e.g. This I will do from / say of this that / will do this, with zeroing of the second this
(3.3.2) and of Isay (5.5.1).
sentence which has been shortened can move to before a linearly preceding
item from the partial ordering (with any adjuncts, and possibly with its
arguments) usually on condition that the preceding item is longer, and above
all is not shortened to a pronoun. Thus John clamped the lid down, where
down is reduced from a secondary sentence John's clamping the lid was
down(ward) (secondary to John clamped the lid), has an alternative order
John clamped down the lid, provided the lid is not pronouned to it (no John
clamped down it). In contrast John slid down the pole, where down is simply
an Oon operator on the pair slid, pole, has no permutation for down, and
permits pronouning of pole (to John slid down it).
Somewhat differently, when the car which is new (where which is a
pronoun for the preceding car) has its which is zeroed, the residual new
moves necessarily to before car: the new car. Permuting the residue after
zeroing which is holds for all uni-morphemic adjectives and for stated other
words, including the compound-word structure (after zeroing the charac¬
teristic/or or the like, as in books for school to school-books, and grey like
stone to stone-grey). When which is is zeroed in / have decided to leave now,
which is because my time is up, we obtain I have decided to leave now,
because my time is up; and I have decided, because my time is up, to leave
now; also Because my time is up I have decided to leave now. Another
moving of a shortened conjoined sentence can follow zeroing under and: in
John phoned and Mary phoned, zeroing the second phoned (yielding John
phoned, and Mary) permits (and virtually requires) and Mary (the residue of
the second sentence) to move to a point where all that follows it is the
antecedent for the zeroing (namely, phoned): hence, John and Mary phoned.
This condition accounts for the other a/id-residues, e.g. John composes, and
Mary plays, light piano music; and it explains why there is no permutation in
John plays piano and Mary violin (because the above stated point, which has
to be before the antecedent plays but not before the non-antecedent piano,
does not exist).
A slightly different length-permutation permits a final verb to move to
before the interruption (3.4.4) on its subject: My friend whom I told you
about is coming to My friend is coming, whom l told you about. And under
some higher verbs, a verb can move to before its subject: He let the book fall
to He let fall the book. Calling these permutations is a descriptive convenience;
their histories can be different. And some apparent permutations are not
that at all (e.g. the passive, 1.1.2).
3.4.4. Interruption: Modifiers
There is one last freedom of linearization which is of great importance

because it creates the semantic relation of being a modifier (adjunct). This is
the possibility of interrupting one sentence by a secondary sentence bearing
the ‘aside’ intonation, usually one referring to the first. It is directly

observable in, for example. Science fiction—science fiction I won’t touch—is
his favorite, or in You’re in danger—and I’m speaking seriously—of failing
math (or You—and I'm speaking seriously—are in danger offailing math). In
terms of the partial order, the aside-intonation (which can be written as
semicolon) is the highest operator, an Oao whose first argument is the
preceding primary sentence and whose second argument is the following
secondary sentence: Science fiction is his favorite; science fiction I won’t
touch. In terms of linearization, the semicolon conjunction plus its secondary
sentence can interrupt the primary, instead of coming after it.
In the case that the first word, or word-sequence, of the interruption is the
same as the primary-sentence word or sequence (with minor adjustments)
that immediately precedes the interruption, a new situation is created, for
the interruption is then placing a secondary sentence after specially relevant
words in the primary. The words in the primary (here science fiction) become
the antecedent; and the repetition of those words at the beginning of the
secondary is reduced to a ‘relative pronbun’—who, which, where, etc. The
whole interruption is then called a relative clause, and its semicolon aside-
intonation is generally reduced to the weaker intonation between commas:
Science fiction, which I won’t touch, is his favorite. In this situation it is
semantically felt to be a secondary sentence ‘about’ the particular preceding
word or words which are the antecedent of the relative pronoun. The
relative clause is then a modifier, called adjunct, on the antecedent, which is
its host. This is the source of the adjunct-host relation, the major syntactic
relation other than operator-argument.
There are several considerations which favor the above explanation of the
development and nature of the relative clause. First, if one tries to state the
structure and actual word-content of all relative clauses, one finds that they
have no unique composition, but simply are identical with the set of all
sentences, each lacking one word or word-sequence which is such as could
be referred to by the wh-word.7 It follows that the relative clause must have
been a sentence that repeated the antecedent words of the primary sentence.
Second, if one asks what words in the secondary sentences can be replaced
by initial wh-words, one finds, surprisingly, that they are precisely those
words in the secondary sentence which can appear at its front either
normally or by the fronting linearization (3.4.2, G109, 121). It must be,
then, that the secondary sentences in which certain words were reduced to
w/z-words (relative pronouns) had those words already in the front position
(rather than that the tv/z-word was formed wherever the words would
normally have been and then brought to the front). This is not hard to
understand: the front position of a word gives it the status of being the topic
of the sentence, which as noted above is indeed apt for a secondary sentence
7 Z. Harris, ‘Co-occurrence and Transformation in Linguistic Structure', 3.6.2, Language 33
about the antecedent: Science fiction—science fiction I won't touch—is his

favorite exhibits the relevance of the interruption more directly than Science
fiction—1 won't touch science fiction—is his favorite.
Finally, the interruptions: one might think that the secondary sentence
which is the source of the relative clause came after the primary, in ordinary
sentential succession, and was thereafter moved to the antecedent. Flowever,
the fact that the relative-pronouned material was precisely that which could
be at the front of its sentence before it was pronouned means that the fronted
secondary sentence already had special relevance to particular words in the
primary—those which its front words repeat, and after which it indeed
appears in the primary sentence. This makes it desirable to assume that
the secondary sentence was not successive but rather an interruption in
the primary, coming immediately after the antecedent host words. More
importantly, taking it as an interruption removes the need for an addressing
apparatus—which is lacking in language—in terms of which we could locate
the host. We would need this addressing apparatus in order to know to which
primary words the secondary words were referring, e.g. if The big boy hit the
little boy were to be derived from successive sentences A boy hit a boy;
the boy was big; the boy was little. And we would need it in order to move the
secondary sentence up to the antecedent in the primary. Languages do not
have an addressing apparatus which can be attached to a secondary sentence
to move it precisely to those words in the primary to which its front words
refer. However, if the secondary sentence interrupts the primary we do not
need to locate (or address) the antecedent host. We merely say that when a
secondary beginning with certain words interrupts a primary right after the
same words, relative pronouning can result. In other situations, the secondary
sentence simply does not develop into a relative clause.
Indeed, in English grammar, interruption by a front-repetition secondary
sentence removes the need for a full addressing apparatus for all cross-
reference. The reason is that it brings two referentially related items (having
various partial-order statuses) into a fixed distance in the linearization.
Specifically, first, having the repeated words come at the front of the
secondary sentence and, second, having the whole secondary sentence come
right after the antecedent means that the two occurrences of the given words
are next to each other in the sentence. To locate in the linearized form the
antecedent of the words to be pronouned requires no more than the word
next or preceding (in respect to the linearization), and the reduction to
pronoun requires a metasentential statement that the second occurrence of
these neighboring (next) words refers to the same objects (or in certain
situations to members of the same class of objects or events) as the
antecedent occurrence (5.3.3). With no more than this apparatus, it is
possible to derive the relative pronouns and relative clauses. In English, it is
then found that with the aid of the relative clauses all other referential
situations can be derived (5.3). There are also other avoidances of the need
for a full addressing system to locate antecedents: most repetitional zeroings

depend on definable relative positions in the partial order, which can be
stated a priori without locating the antecedent in the individual sentence:
e.g. zeroing of words in a second sentence under and if the same partial-
order position in the first sentence is filled by the same words. And many
other pronouns do not depend on any particular relative position in the
partial or linear order (e.g. he, they). The addressing requirement is thus
circumvented.
In summary, we find that the partial ordering of words which creates
sentences has a normal linear representation in the word-sequence of a
sentence, but that certain items in the partial order have optional alternative
positions, of particular types, in the linear representation. One of these
types, the front position, makes a particular word or sequence the topic of
the sentence. Another, the interruption by a secondary sentence, makes the
secondary sentence a modifier of the primary sentence, or of a particular
word in the primary. The alternative linearizations are almost entirely
optional, and do not otherwise change the meaning of the sentence or the
grammatical relations (i.e. the partial order) within them. They do, however,
have an effect on the constraints on equiprobability of the words making
sentences; but this effect is of a second order, consisting of a few kinds of
alteration in the relative position.
4
Language Structure:
The System Created by the Constraints
4.0. Introduction
The methods of 3.1-2,4 yield a base set of sentences from which the other
sentences of the language are derived. The dependence of 3.1, together with
the linearization of 3.4.1-2,4, characterizes the set of possible sentences; it
eliminates what cannot occur in the language. The likelihood gradation of
3.2 creates not-well-defined sets of ‘normal’ and marginal sentences, on
which the reductions of 3.3 operate. Separating out the different constraints
that contribute to the departures from equiprobability in the word-sequences of
all sentences distinguishes the base subset of sentences, in which reductions
have not taken place. This subset is structurally recognizable, because the
operator-argument relation in it has not been obscured by zeroing, reduc¬
tion to affix, length permutation, and other transformations. It is also purely
syntactic, without having morphology carry part of the syntactic functions,
since all affixes that can be derived from free words have been separated off,
leaving the vocabulary to consist of syntactically indivisible words. That is
not to say that one may not see internal construction, as in the con-, per-,
-ceive, -sist of conceive, perceive, consist, persist or even the -le of dazzle,
settle, etc., but that such constructions cannot be usefully derived from
separately functioning morphemes or words, joining in the operator-argument
relation.
Since reductions are optional, except as in 3.3.3 (where the missing
sources in the base can be created), the unreduced sentences remain in the
language by the side of their reduced forms: / prefer for me to go first and
I prefer to go first. And since reduction is paraphrastic, changing only the
shape of a word but not its existence or its operator-argument status in the
sentence, the sentence that results from it does not have different substantive
information than the sentence on which it acted. Hence the information of
reduced sentences is present in the unreduced ones from which they have
been formed. It follows that the base set of syntactically transparent
unreduced sentences carries all the information that is carried by the whole
set of sentences of the language.
4.1. The base set of sentences
In the base, the syntactic classification of vocabulary is into the N, On, Ono,
etc. categories of 3.1. The syntactic operations in it are the operator-
argument partial ordering, with optional interruption by secondary sentences
and optional front-positioning of words. In English, operators carry operator-
indicators (the ‘present tense’ -s), which are replaced or overruled by
argument-indicators (-mg, that, for. . . to) when they become the argument
of a further operator. This holds for all sentences. Similarly, all sentences
can carry the metasentential I say with certain appended material on it such
as not, and, or, in query, in command, and also the ‘aside’ intonation which
attaches a sentence as secondary to its primary one. When an operator is
defined as acting on an operator (as its argument), it acts on the whole
sentence that the lower operator had created, since all operators in a
sentence carry their arguments with them (by 3.1), so that each creates
a sentence of its own. Given destroy on the pair bomb, atoll (yielding (1)
Bombs destroyed the atoll), then think on the pair we, destroy, yields We
think bombs destroyed the atoll (for plural, tense, the, cf. 8.2). Further
operators and transformations on the higher operator (here, think) do not
enter into the lower sentence, which remains as an unchangeable component
sentence in the larger one. Thus in Secretly we think the bombs destroyed the
small atoll completely, the completely, operating on destroy, can move only
in the contained sentence (1), and secretly, on think, can move only outside
the contained sentence. The completely cannot move into we think (no We
completely think the bombs . . . ), for any moving had to be done right after
completely entered its sentence (the component one), and we think had not
yet acted upon that sentence then.
In the base, few words are multiply classified, i.e. belong to more than one
of the N, On, etc. classes; most of the apparent multiple classification arises
in the derived sentences, and is due to zeroing. And few base sentences are
ambiguous, other than for homonymy and for unclarities in the meaning-
ranges of words; the great bulk of ambiguous sentences are due to degener¬
acies in reduction—to two different sentences being reduced to the same
form.
It follows from the above that this relatively simple grammatical apparatus is
adequate for all the substantive information which is carried in language.
This would not be so if the meaning of further operations depended on
whether the operand had been reduced, for then the meaning that required a
reduced form would not be available in the (unreduced) base. But this
apparently does not happen.
Certain situations present difficulties for the base, but can be resolved.
For example: as has been seen, the auxiliaries in English differ from all other
operators in that they act only on a subject-less verb: John can go first, but no
Language Structure 107
John can for him to go first, and no John can for her to go first, in contrast to
John prefers to go first by the side of John prefers for her to go first. The
description of can in the base was regularized in 3.2.4 by saying that like
other Ono operators it operates on a whole sentence, but one in which the
subject must be the same as that of can, and must then be zeroed: a non-
extant but well-formed John can for him to go first is then necessarily
reduced to John can go first. The cost of this derivation is that the source
sentence does not remain in the language, hence is not in the base, since its
reduction is required; the information of the reduced sentence cannot then
be expressed in the base. A resolution of this difficulty can be reached by
finding some word or sequence which is approximately synonymous to can,
and selectionally close to it, which we can consider a variant of can in this
form, and under which the above zeroing is not required. We can consider
that some such sentence as (1) John is able for him to go first, which is in the
base, occupies there the place of the non-extant (or no longer extant) (2)
John can for him to go first, while the reduced forms of both, John is able to
go first and (3) John can go first both remained in use. Such a resolution is
supported by the fact that historically some auxiliaries indeed were able to
have a repeated subject under them. A more general support is that change
in word use consists in principle not in a word being dropped or replaced
within a sentence, but in sentences which contain that word competing with
approximately paraphrastic sentences containing another word, and losing
out to them. Given (3) we have to posit a source (2), and if (2) does not exist
it must have been replaced by an informationally roughly equivalent sentence
such as (1).
This situation is particularly common in the case of affixes, which are
analyzed, to the extent possible, as being reductions from free words, mostly
operators, in the base. Even when such a reduction is historically evidenced,
the free word may have meanwhile been lost; affixes are less amenable to
historical replacement than free words. Thus, as noted, the -hood of
childhood was once a compound form of a free word had ‘state, condition’
which was lost from the vocabulary, whereas the affix occurrences of it were
not lost. But had was not simply lost. Sentences containing some other word,
such as state, came to be used more than corresponding sentences containing
had. When free had was no longer said, some word such as state remained in
its place, like a suppletive variant of it, so that the unreplaced -hood is now
the affixal reduction of state rather than of the old had. Such established
historical cases are rare. For most cases in English, saying here that a given
affix is a reduction of a particular free word in a particular position is not a
claim of history but a statement that the operator-argument relation of that
affix to its host is the same as that which the given free word would have in
the stated position, and that the selection (relative likelihoods) of the affix,
as to host and to further operators on it (and so its meaning), is approx¬
imately the same as that of the given free word in the stated position.
With the above statement as criterion, we can reconstruct a less restricted

variant for every word or sequence whose occurrence is restricted. This does
not place in the base something that is appreciably different in information
from that which is outside the base. Of course, such methods are acceptable
only if they apply to a few special situations (even if many words are involved
in one or another of these situations), so that the methods are used only to
regularize the aberrant situations to make them conform to the main
character of the language.
Such reconstructions could also be used in any rare situation in which a
reduction in a given word admits a class of further operators that would not
otherwise come into the sentence, or admits a selection of further operators
different from the selection that the given word has when not reduced. In
each case we would say—artificially—that the further operator indeed
requires the reduction, but that we can assign to it in the base some
synonymous variant that does not require the reduction.
The three difficulties above, which detract from the informational adequacy
of the base (and potentially detract from the applicability of the theory),
arise only in encapsulated situations in a language. One can more readily
find cases of reductions and transformations which block a class of further
operators, but this does not interfere with the base carrying all the informa¬
tion which is carried in the language. One example is that the fronting
linearization of a sentence (3.4.2) blocks the operator-indicator -ing (but not
that) which appears when the given sentence comes under a further operator:
by the side of 1 like this we have This I like; but there is no I reported this my
liking by the side of I reported my liking this, while there is / reported that this
I like as well as I reported that I like this.
The base words are generally uni-morphemic, i.e. affixless, since affixes
are analyzed, as far as is possible, as reductions of free words in the given
position. The only morphemes in English which are not operators or
arguments (or reduced forms of such) are: indicators of argumenthood (e.g.
-ing, that, G40) or of operatorhood (‘present tense’ -5, G37), carriers for
these when these are suffixes (be, 8.2, G54), and words which can be
analyzed as constructed from regular word sequences (e.g. the, G237).
Within N and the various O classes, the similarities and gross levels of
likelihood grading create a few fuzzy subsets (3.2.5). Such are for example
the argument-repeating Ono operators which have a high likelihood that the
subject of the O under them be the same as their own subject (John tried to
phone, rarely tried for Mary to phone), or that the object be the same as their
subject (John suffered defeat), and so on (3.2.3). Some second-level (i.e.
O . . . o . .) operators have the operator under them almost always future
in respect to themselves; under some of these higher operators the future
tense on the lower operator is omitted (e.g. the ‘subjunctives’: e.g. order in
The commander ordered that they be shot, G283). Some arguments (the
indefinites) occur freely under virtually all operators; others (the classifiers)
under the same operators as occur on any of particular other arguments
(their classificands). Such likelihoods common to sets of words yield in the
base many subsets, such as indefinites, quantifiers, prepositions, reciprocal
verbs; these are of major or minor importance in the grammar of a language
(3.2.2). They may have few or many members, and one or several charac¬
terizing properties. In almost every such subset there is a common element
of meaning, related to its special syntactic constraint (or lack of constraint),
though it is doubtful that one could guess a priori which shared elements of
meaning will be represented in a likelihood-similarity subset (cf. 3.2.3, end).
Some words may occur with the subset property only irregularly or with
marginal acceptance; how acceptably can one say John tried for Mary to
phone? (The equivalent in French is quite grammatical; in English one
would have to say John tried to have Mary phone.) And in cases where a
subset is characterized by more than one property, one can usually find
words which have some but not others of the properties (3.2.5).
In addition to the subsets in respect to likelihood, there are some
individual words which have peculiarities of likelihood of such importance
as to have a syntactic character worthy of remark. For example, one does not
ordinarily find a sentence such as Mary is here and Mary is not here except as
part of a joke or a metalinguistic discourse; but one can say Well, he's a
politician and he isn't. One also does not find a sentence such as Mary is here
and John knows that she is not here or John knows that she is here and I know
that she is not (but the sentence occurs with just, in John just knows that she is
here and I know that she is not here). If the vocabulary is factored on co¬
occurrence grounds into maximally independent vocabulary-elements, know
may be replaced by a combination of words that would reduce the syntactic
(combinatory) property above to the property of not above; but such
factoring is beyond the present confines of syntax.
4.2. Sentences derived from the base
From the base sentences, the reductions (3.3) create the other sentences of
the language, which can thus be considered to be derived from the base.
These reductions are changes in particular word-occurrences, not recastings
of the whole sentence. As noted, they are made on the basis of the likelihood
of the word in respect to its argument or operator in the sentence, and they
are made as soon (in constructing the sentence) as these conditions are
satisfied. Reductions are thus located at specified points inserted in the
semilattice of the source sentence which they contain (8.3). Since reductions
are only changes in shape, the unreduced sentence is not merely a source,
but is actually contained in the reduced sentence. Although the types of
reduction are similar in many languages, the specific reductions differ; and it
is here that we find the main difference among individual language structures.
When a reduction is optional, each sentence it creates is an image of a
source in the base. When a reduction is required, each sentence claimed to
result from it (e.g. one containing can) has a source in the base only if we can
assign to the reduced word (e.g. can) a suppletive word in the base (e.g. is
able to). There are also cases where the reduced form and its source both
exist, but intermediate reduced forms do not. Thus He rotated the globe on
its axis (by the side of The globe rotated on its axis) can be derived regularly
from a source He caused the globe to rotate on its axis, going through length-
permutation to He caused-to-rotate the globe on its axis and compounding-
permutation (3.4.3) to the non-extant He rotate-caused the globe on its axis,
where zeroing of the broad-selection last element of the compound (cause)
yields He rotated the globe on its axis. In such cases we need only say that
intermediate steps in the succession of transformations are required, in this
case the zeroing after the compounding; the compounding itself was not
required. That is, what is optional here is the product of steps, not each step
alone.
The reductions leave room for degeneracy, producing ambiguities: two
different base sentences, each undergoing different reductions, may end up
with identical visible word-sequences. For example, the ambiguous John has
shined shoes is obtained from two sources: (1) John has the state of John's
shining shoes, via repetitional zeroing of John's and reduction of the state. . .
ing to -ed, this being roughly the historical source of the ‘perfect’ tense
(G290); (2) John has shoes; these shoes are in the state of one’s shining them,
via zeroing of the indefinite one's and of the repetitional them, and reduction
of the state of.. . ing to -ed (this is offered as the source of the passive, 1.1.2),
and reduction of semicolon plus these shoes to which (yielding John has
shoes which are shined) followed by zeroing of which are and permutation of
the residual shined to before the host. All these are widespread and well
attested English reductions. The two derivations admit different reductions
of further operators: for example, (1) appears in John has shined his shoes;
(2) in John has well-shined shoes.
If a given sentential word-sequence has more than one source from which
it can be derived by known reductions, then for each derivation the given
sentence has the meaning of the corresponding source. The ambiguity of
such a sentence is not a matter of vagueness, nor of a broad range of
meaning, but of a choice between two or more specific meanings, those
of the specific source sentences.
There may be other disturbing situations in deriving sentences from the
base. One is that the reduced sentences may lose their semantic relation to
their source. This can happen if a reduced word, for example a word with an
affix, has come to be used with different likelihoods for the operators on it,
no longer the likelihoods (selection) of its component words: e.g. romance
‘story’ from ‘in Roman (language)'. To the extent that this happens, the
reduced sentences can carry information that is unavailable in the base
sentences, although one can usually replace the semantically changed

derived word by a definition given in base vocabulary (largely uni-morphemic).
In any case, for the great bulk of sentences their derivation from base
sentences preserves meaning: this is because the reductions do not change
the partial ordering, and the likelihoods of further operators are usually not
materially changed.
How to analyze each sentence follows from the above. Given the con¬
straints which characterize sentences, we know that base sentences can only
contain words in operator-argument relations, specifically N, Oin etc. In
each sentence, we know that there can only be as many N and O as are
required by the various operator classes in the sentence, and that for each O,,
there must be one N (possibly reduced) that is not being used as the
argument of any other operator, and for each Otm there must be two N, for
each Oa there must be an O, of any kind, not otherwise used as argument,
and so on. In each base sentence the words are all visible, and their few
possible positions relative to their operator are given by the linearizations,
including the possibility of interrupting sentences. In a derived sentence, on
the other hand, one has to look for the traces of reductions—evidence of
zeroing, pronouns, affixes, and coalesced words like the (G237). Given the
domains and conditions for the reductions, it is thus possible to construct an
effective procedure, complex though it be, for computing the derivation and
base composition of a sentence. The procedure would of course have to
allow for degeneracies as mentioned above.1 An effective procedure might
not be assured if sentences were characterized as being transforms of whole
other sentences rather than as being due primarily to reductions on words as
the words enter the partial order (as they are here). Nor would it be assured
if for some sentences the only characterization was that they were due to
analogical processes on some other sentence—which they may certainly
be—rather than that they also, even if analogical, must have the form of
operator-argument and reduction relations among their words.
The reductions in a language might be such that in many cases a derived
sentence would be reachable from the same base sentence by more than one
path of different successions of different reductions. In such a situation we
may have a grid of reductions; to be a sentence a word-sequence might then
need only be reachable along one or more paths of the grid from some other
sentence. This would be a grammar of elementary differences rather than of
1 The sentence-generating system ofch. 3 does not suffice directly for an analysis (‘recognition’)
procedure, which requires an orderly accounting of all similarities among the generated forms.
A computer program using a detailed grammar of English, based on the contiguity of
grammatical constructions, is in operation, analyzing scientific texts in English and some other
European languages, and transforming them into informational data bases: N. Sager, Natural
Language Information Processing, Addison-Wesley, Reading, 1981 An algorithm, with the
dictionary written in the form of the algorithm, has been produced by D. Hiz. An algorithm for
the operator-argument recognition of sentence structure has been developed and implemented
by S. B. Johnson.
elementary reductions. However, no case of such a situation has been

established.
In addition, obtaining sentences by the constraints of Chapter 3 makes it
possible to characterize each sentence not only by its operator-argument
and reduction orderings but also by the sentence-subsets and the sub¬
language to which it belongs. This will be discussed in 8.1 and in 10.1.
The reductions are not always the grammatical events that historically
produced the sentence, for, as noted above, some sentences are formed on
the analogy of others, or borrowed, and the like. And even when the
reductions are historical, they are not necessarily recognized as such by the
speakers of the language, who may have constructed their own perceptions
of how certain sentences relate to each other. But in any case, establishing
these derivations for every sentence structure, and so presumably for every
sentence, shows that a single kind of relation, namely reduction based on
high frequency or low information as between operators and arguments,
suffices to derive the forms of all sentences, and in general to preserve the
information, from the base sentence.
The consistency in the whole set of derivations provides a single measure
of complexity in terms of which we can find the ‘cost’ of each derivation. We
may judge that some sentences can be derived directly by the stated
reductions even if in historical fact, or in popular perception, they have a
different origin. A case in point is the Modern English status of -ing as an
argument indicator on virtually any sentence, and the status of the ‘progress¬
ive tense’ is . . . -ing as derived from the above -ing under on or the like; the
historical sources of these two were phonemically different, and not etymol¬
ogically related (G41,295, 373). We may also judge that some forms require
so complex a derivation, passing through such strange intermediate sentences,
that we may prefer to sacrifice the reductional derivation—especially in
computer processing—and say that at the given point the language is using a
form or process outside the operator-argument universe (e.g. for tense,
cf. G265); this, even if in fact the sentence went through this derivation
historically, but with some of the intermediate sentences having gone out of
use. But if we want an apparatus for organizing language-borne information,
then the operator-argument and modifier relations which are reconstructed,
based on this one method of derivation, from the sentences of a language
will suffice in all cases; the reconstruction may also reveal subtleties of
meaning and use.
Several of the reductions create new subsets of words: those which accept
a particular reduction. Some of these are large and important, and may even
be sharply bounded: e.g. the distinction between verb and other operators
(adjectives, prepositions, and even predicate nouns). This distinction arises
because those operators which are less stable in respect to tense and time-
adverbs (e.g. have high likelihood of changing their tense within a discourse)
receive the operator indicator -s (which serves as the present tense, and can
be replaced by the past tense -ed) directly on them: walks, sleeps. In

contrast, the other operators receive it as a separate word, the indicator -s
being then carried by the word be: is ill, is here, is an uncle. The difference in
durativity which is involved here is graded over all operators; but the
structural difference of carrying the -5 directly or, alternatively, on be
presents only two choices. Thus certain operators are marked off from the
others as being more actional and receiving the -5 directly: the verbs. This
division in English is strongly conventionalized: few words have both forms
except by one being derived from the other.
Another example of a reductional subset is the prepositions. These Oon
operators have already been recognized as a subset on likelihood grounds,
because they occur between very many O and N: John fell from the stage,
John sat near the tree, etc. In addition, their high frequency and the broad
selection of their widely applicable originally spatial and temporal meanings
enabled them to spread by zeroing or metaphor into having many kinds of
argument (G80): between two N as in Mary is in the house from Mary is
present in the house; after one N as in Mary is in from Mary is in the house;
after one O as in Mary came in from Mary came in the house, The boss is in
from The boss is in the office. As in many subsets, membership is not sharply
defined: some prepositions have developed from a preposition plus a noun
(before); others have developed from other classes (during from a verb).
And not all members have all the subset properties: before, during do not
occur with one N (as subject: no John is before, normally), and during also
not with one O (as subject: no He wrote this during, but one can say He wrote
this before).
Similarly, consider the auxiliary Ono verbs can, may, will, ought, etc.,
under which the subject of the lower operator, and the to on it, must be
zeroed, and which do not carry the present-tense -s (3.2.4). These are found
to have neither a closed and definite membership nor fully common properties
for all members. Typically, we have John can go by requiring zeroing from a
non-extant John can for him to go, and the set of verbs that have this
required zeroing is commonly thought of as being well defined. But will
occurs also as an ordinary operator, with a related meaning ‘to desire’.
Ought is erratic about zeroing to: John ought to go but Ought John go? as
well as Ought John to go?, and John ought not go, John ought not to go. Need
has John need not go, and in special environments John need go only
occasionally; but need is also an ordinary verb in John needs to go, John
doesn’t need to go, John needs Mary to go, etc. And let, make zero to while
admitting different lower subjects: John let it go.
One important successive application of reductions can create not a subset
but an apparent escape route for words from the N class to the O class and
vice versa, and less dramatically can create non-coherent jumps in the
selection and meaning of words—all these being changes that would not
result from ordinary extension of selection or from analogical processes.
This is metaphor, which can be derived from two sentences under the
conjunction as, provided everything beyond what is needed for one sentence
is zeroable (indefinites or classifiers). It is seen for example in / can see his
wanting to go (where see appears as Ono instead of its base status as Onn)
from (roughly) I can treat his wanting to go as one’s seeing things (via the
artificial / can treat-as-seeing his wanting to go, G405); in the source, see is
correctly Otm. Such a derivation explains for example how one can say The
car smashed into the wall without getting smashed (or: without smashing): we
obtain The car smashed into the wall from The car moved into the wall as a
thing's smashing (via the non-extant The car moved-as-smashing into the
wall), to which without (the car) smashing can be adjoined. It also explains
how one may say He peppered his lecture with asides but not He peppered his
lecture, though one can say He peppered his soup (from He put pepper into
his soup similarly to the zero causative, 3.3.2). The source is He put asides
into his lecture as one's putting pepper into things, which is reducible to the
non-extant intermediate He put-as-putting-pepper asides into his lecture,
thence into He peppered asides into his lecture, whence He peppered his
lecture with asides (G369). (All these are elsewhere attested zeroings in
English.) The fact that as is involved in the derivation can be seen in such
forms as the French Le juge a ravacholise X, where one would not say
(except perhaps as a joke) (1) Le juge a ravacholise Ravachol: we see that
ravacholiser is ‘to treat as one treated (thfe executed anarchist) Ravachol’
(which cannot be said of Ravachol himself as in (1)); it is not ‘to do what was
done to Ravachol’ (which could be said to Ravachol, and which therefore
would permit (1)). The source here would be in English The judge treated X
as one's treating (i.e. as in the treatment of) Ravachol (G406).
4.3. Systemic properties following from the constraints
From the operator-argument dependence of 3.1 above, it follows that there

must exist a set of words (zero-level arguments, N) whose argument is null
(or which are defined in respect to their own operators), for otherwise there
would be an infinite regress of word classes defined by their argument
classes. And there must exist operators (first-level, On „) all of whose
arguments are zero-level, since no other arguments have been defined at this
point. The first-level operators create elementary sentences. If thereafter we
do not define second-level operators (O0 . . . 0 . . ), at least one of whose
arguments is itself an operator, the system would not provide for any
sentences beyond the elementary. It further follows that every sentence
must have at least one zero-level word, in order to start forming, and no
more first-level operators than it has zero-level words to satisfy their
argument requirements; if it is more than an elementary sentence, it
contains second- level operators in addition. The base words of the language
(excluding whole-sentence words such as Hello) fall into a three-level
partial order in respect to what they require in a sentence. And the words of
base sentences constitute a semilattice in respect to their arguments in their
sentence (6.3, 8.3). The reductions in a sentence are locatable as additional
points in that semilattice.
Since operators appear in sentences only given the presence of their
arguments (before reduction), it follows that when a second-level operator
appears with its argument, which is itself a first- or second-level operator,
the arguments of that first- or second-level operator are also present. Hence
the sentence which is made by the first- or second-level operator is imbedded
in the sentence made by the higher second-level operator. Thus every
sentence consists ultimately of one or more elementary sentences, possibly
with second-level operators on them. And since reduction and transformation
only alter the phonemic shape or the position of words, a reduced sentence
contains its (usually longer) unreduced source. The contained sentence is
the same as it is in its independent occurrences, since the further operators
on the words and thus their meaning is largely unchanged—aside from the
fact that the meaning of certain words, and of certain reduced word-forms,
differs under different operators (11.1). Various word-sequences may remain
as idioms: irregular reduction (the more the merrier), specialization of
meaning (kick the bucket), etc.
It should be understood that saying that every non-base sentence contains, or
is derived from, a base sentence does not mean that such a base sentence was
in use first and then reduced. / like tennis is derivable from I like for me to
play tennis or from I like people’s playing tennis (by zeroing play as high-
frequency operator on tennis, and people as indefinite argument, or me as
repeated argument). This is so not in the sense that the base forms were said
first, or that the speaker thought first of the base form, but in the sense that /
like tennis says only what the two base forms say more regularly (unless the
environment supports other zeroings). And to say the four-color problem
does not imply that anyone first said the problem of four colors, only that the
latter is sayable in the language with the same meaning, and that if the
former was not indeed reduced from the latter then it was built on the model
of compound words which are reduced in this manner (a four-door sedan
from a sedan with four doors).
All sentences created by the partial ordering and reductions can in
principle be said in the language. Some are extremely rare, either because of
very low likelihood of a word combination, or because a reduction is heavily
favored as against its source (e.g. / want to win as against / want for me to
win). Since the constraint of likelihood (3.2) is defined on the resultant of
3.1, it follows that no matter how low is the likelihood of a particular word in
respect to a given operator on it, the likelihood is presumed never to reach
zero;2 for if it did, we could not say that the occurrence of words in sentences
depends only on the dependence properties of other word-occurrences
there. As was argued in 3.2.1, this accords with the data about language,
since even those members of an argument class which are least likely to
occur under a particular argument (say, vacuum under cough) cannot be
entirely excluded from what is possible. (But outside the base, required
reductions may preclude the visibility of words, as under the auxiliaries,
3.3.3.)
Certain other properties of language are due to the reductions. First, the
base, i.e. unreduced, sentences have the least constraints on word com¬
binations. This arises from the fact that each reduction, defined as operating
on some operator-argument or sentential condition, has as domain either
all sentences containing that condition or else a proper subset of these,
where stating a proper subset constitutes a restriction on word-occurrences.
Hence the set of sentences resulting from a reduction cannot be less
restricted than the set existing before the reduction is defined.
Second, the phonemic possibilities in reduction leave room for the
creation of morphology, as a structure (and hence a restriction) beyond what
is found in the base. In many but not all languages, certain reduced words
are attached as affixes (mostly operators attached to their arguments), thus
creating pluri-morphemic ‘derived’ words, which do not exist in the base.
Third, reduction of certain operators to zero makes their arguments seem
to take over the operator’s status (and vice versa): e.g. the apparent noun
take in today's take, derived from (roughly) the amount of today’s taking
things, or the apparent verb piece in to piece together, reduced from
(roughly) to put the pieces together. Such reductions also cause an apparent
‘lexical’ shift in the meaning of the words (as between the take and to take, or
between to piece and a piece), although the word can be considered to have
the same range of meanings in both environments, with the difference being
merely the meaning of the zeroed word (amount, put)—zeroed in its
phonemics but still present in grammar and meaning.
Lastly, many reductions create degeneracies (ambiguity) in sentences,
when two reductions, acting on two sentences which differ in form and
meaning, produce the same sequence of words.
There are also certain properties which follow from linearization. The
alternative linearizations create the syntactic relation of modifier. As noted
in 3.4.4, if we interrupt a sentence after a particular (‘host’) word, and if the
interrupting sentence begins with the same word, the two occurrences of the
word are successive (neighbors) so that the second can be pronouned
As noted in 3.2.4 and 4.1, there may be cases of an argument which is excluded under a
given operator (e.g. can). The generality of 3.1 may then be salvaged by finding synonymous
words in the base which are free of this exclusion and which serve as variants of the restricted
word.
without a general addressing system, leaving the rest of the interrupting

sentence as a modifier on the host to which it is adjoined.
From all the constraints together there follows the major property of
language: that not all combinations of words occur as sentences of a
language (2.1). This is a universal and necessary property. In every language
not all sequences of the phonemes, or the morphemes (words), in the
language occur as sentences, or as discourses.
Various features of the constraints, and especially the fact that each is
defined on the resultants of others, have the effect of creating a system
(12.4.2). The sentences all have a family resemblance, not only because all
elementary sentences are made in one way, but also because the second-
level operators act on their arguments (ultimately, elementary sentences) in
the same way that the first-level ones act on the zero-level words to make the
elementary sentences. Both the deformation of sentences (by argument
indicators) to make them into arguments, and also the reductions, create the
situation of sentences having somewhat varied paraphrastic forms. The
existence of a system into which sentences fit has made it possible to recast
sentences paraphrastically in ways that enable almost every part of a base
sentence to become an argument or operator in a paraphrastic transformation
of it: e.g. John varnished the violin; The varnishing of the violin was by John',
This violin is John’s varnishing (doubtful, but compare This dish is my
cooking)'. Varnishing was what John did to the violin, The varnishing was
John's (doing).
It should also be noted that the system provides for certain kinds of
classification. There are classes of sentences in respect to their creation by
operators (elementary and not), and in respect to reductions and other
transformations (8.1). There are many classes of words and sentence-
segments established as domains of various reductions. And there are the
never fully statable and slowly changing similarities and gradings of words in
respect to their selection of individual operators over them and arguments
under them: examples of such classes are the indefinite pronouns, quantifiers,
classifiers (3.2.2).
In all, it is seen that the few constraints of chapter 3 determine a
considerable amount of structure.
4.4. Properties following from additional conditions
In addition to the properties which follow directly from constraints on

equiprobability, there are various ones which are due to other conditions of
language, some of these essential and universal, but others non-essential.
One condition that is universal is that no space is definable for speech
other than its sequence of phonemes and words. The various junctures
that may be found between phonemes or between morphemes are sound
differences or pauses which establish boundaries and reduce distances

phonetically, but do not create different base distances between neighboring
phonemes or words (such as are available for example in the beat of music,
or in the meter and rhyme of poetry). Hence, in linearization, an operator
has to be contiguous to its operand: it can be before, between, or after its
arguments, but no pre-existing distance is definable such that it could be at a
distance from them (2.5.4) when it acted.
Another universal condition is the finiteness of the users of language and
of their public experience with language. Because of it, everything that they
have to learn or to agree on in order to be able to speak to each other has to
be finite, even if recursively applied, and the stock of entities and processes
in language has to be finite or at most recursive. Hence, the set of independent
items of vocabulary has to be finite, though changeable. As to the word
classes: if they are defined by their relation—e.g. zero-level argument (N),
first-level operators (e.g. O,,),second-level operators (e.g. Oa)—then we
have a recursive construction for word classes. This also means that the
grammar is finite, and that for every language it is possible to formulate a
constructive and not merely descriptive grammar—that is, a grammar which
lists entities and processes sufficient to form and decode precisely the word-
sequences which are sentences of the language (2.5.1).
A third condition met in language is that in many cases the grammatical
constructions are the least that are necessary within the constraints (2.2-3).
One instance is that many, perhaps all, languages do not have clear cases of
third-level operators, i.e. operators which can be used only given the
presence of second-level operators. A second-level operator, such as cause,
carries information that cannot be carried by first-level operators such as
angry, heckle: e.g. Their anger caused them to heckle, but not Their anger
heckled them or Their anger heckled them to cause. But it appears that any
information that could be carried by a word operating on a second-level
operator could be carried (though less explicitly) by a word which is itself of
second level, namely a word that can also operate on first-level operators.
For example, heighten or increase can operate on correlate (The emergency
increased the refugee flow’s correlation with local repression), but these can
also operate on flow or repress directly {The refugee flow increased).
A similar case is that many languages have few operators with three-part
arguments. The information carried by such an operator can in many cases
be carried by an operator on two operators each with two-part arguments:
e.g. John put the hook on the table (where put has a three-part argument,
there being no John put the hook) can be roughly paraphrased by John
placed the book so that the book was on the table).
Yet another instance of least structure is the fact that to the extent possible
second-level operators have the same morphology (due especially to tense
and person reductions), and the same linearization relative to their operands,
as do first-level operators in respect to their own operands. For example.
first-level operators appear immediately before, between, or after their

arguments because no other (distant) position is definable for them. In
contrast, second-level operators have one other possible position: they
could appear between the arguments of their own operand (e.g. sufficed on
He wrote a letter could have produced His writing sufficed a letter instead of
His writing a letter sufficed). They do not do so; rather they in general appear
in the same positions relative to their operand (with its attached arguments)
that this operand has to its operands (arguments) in turn. This creates the
well-known imbedding of whole sentences inside larger sentences, as against
inserting higher operators into the lower sentence (except as interruptions,
3.4.4). Similarly, just as a secondary operator (adverb, preposition, or
conjunction) on an imbedded operator cannot appear out of contact with the
imbedded sentence, so an adverb on the higher operator does not in general
appear inside the imbedded sentence. As noted in 4.1 for English, if later
operates on the higher explain, while immediately operates on the imbedded
come, we can find later only in positions 2, and immediately only in positions
1 of the sentence:
2 He 2 explained 2 that 1 he 1 had 1 come 1 2
(i.e. Later he explained that he immediately had come, or He explained later

that he had come immediately, etc., but not He explained immediately that he
later had come, or He explained that he had come later immediately, etc.).
Thus the adverb on each sentence is always contiguous to that sentence, and
the adverb on the higher operator is always outside the lower sentence. This
creates the well-known nesting of adverbs, and so for other nestings in
language. Together with the basic contiguity properties noted above, this
creates the possibility of analyzing sentences in a constructive and computable
manner into contiguous strings of contiguous strings of words, with each
string referring informationally to its imbedded or imbedding string (ch. 4
n. 1, and 7.6).
Another instance of least structure is the limitation of each word, by and
large, to a single kind of argument. If a word had one kind of argument in
some cases (say, cause requiring a zero-level first argument in John caused
our car to swerve) and another kind of argument in other cases (say, cause
requiring a first-level first argument in The truck's skidding caused our car to
swerve), the resulting multiple classification of words, by their various
argument-requirements, would be very complicated. It turns out that in such
cases we can generally obtain one occurrence as a reduction from the other
kind of occurrence (e.g. John caused . . . from John's action caused ... or
John’s presence caused . . . ).
A result related to least structure is the fact that each language has only
one or a few sentence structures. In English the sentence structure of
elementary sentences (operator after its first argument) is preserved in more
complicated sentences: (1) the morphology (e.g. tense, plural) and relative
position of higher operators are generally the same as for lower operators, as
noted above (so that, for example, the subject precedes the verb both in The
refugees crossed the border and The refugees' crossing of the border angered
the government); and (2) a secondary sentence, given that it contains an
argument that has already appeared in the primary, can interrupt the
primary right after the repeated word, so that it creates a modifier (relative
clause) that is contiguous to the antecedent (host) word (so that, for
example, the phrase refugees from El Salvador occupies the same position as
the single word refugees in The refugees crossed the border, The refugees
from El Salvador crossed the border). In contrast, the interruptions that do
not form relative clauses do not merge into the previously existing sentence
structure, as in Tomorrow—so I hear—will be cold, Tomorrow will—so
I hear—be cold.3
A fourth condition met in languages, which is related to the least-structure
condition, is that in various respects the structural possibilities of a language
are not fully utilized. For example, both the partial ordering and the
linearization make it possible to state an address for each word in a sentence.
However, languages do not use complete addressing systems, and the only
cases in which the location of a word is cited are those where the location can
be identified with the aid of a few simple addressing words such as ‘prior’ (for
relative pronouns), ‘co-argument’ (for -self, as in John washed himself),
‘same partial-order status’ (for corresponding-position zeroing, as in John
plays piano and Mary violin). Another example is the fact that every
language admits only a small choice of reductions, leaving unreduced many
a word that could have been reduced, as having little information in respect
to a given operator: thus to be here or the like is zeroable under expect (/
expect John to be here), and books or the like is zeroable under read (John
reads a lot), but clothes is not zeroable under wear (no The natives wear).
A fifth condition, which may be due to the independence of selection from
phonemic form (5.4), is the possibility of various similarities among words,
in their meaning-range, phonemic form, or reduced form, sufficient to
create structural details in various languages, beyond the major structure
created by the fundamental constraints. There are frequently used reductions
(or combinations of reductions) in particular kinds of word-sequences. Such
situations create specialized sentence types: e.g. those that result in the
passive form, or in metaphoric uses of words (4.2, end), or in new conjunc¬
tions such as because,provided,providing (below). Similarities in treatment
of I, you, he, etc., or of secondary sentences that indicate time relations, can
create in many languages ‘person’ (subject, object) and ‘tense’ affixes which
get attached to operators. Such constructions as tenses and subject-object
affixes in verb conjugations (or possessive affixes on nouns) can become
' A widespread condition is the saying of incomplete sentences. These fragments generally
turn out to be not merely arbitrary combinings of words, but specifiable segments of sentence-
construction.
required combinations (on ail verbs, or on all nouns) and thereby take on
great importance in the grammar of a language.
The high likelihood of particular operators to occur with particular time
modifiers can lead to subclassification of operators into verb, adjective,
preposition (4.2), and within verbs into perfective (durative) and imperfective
(momentaneous) verbs (3.2.5)—morphologically distinguished in some
languages but not in others. Similarly, prepositions, quantity words, and
other frequent additions can be reduced to such common affixes as cases and
plural on elementary arguments, and even on higher-level arguments,
creating nouns and ‘nominalized sentences’ (sentences appearing as argu¬
ments). Here too differences of likelihood in respect to plural, etc., can
make a distinction in some languages between ‘mass’ and ‘count’ nouns.
There are also similarities and differences in the (morphophonemic) form
of these important affixes on various verbs, or nouns, yielding different
conjugational and declensional subclasses of verbs and nouns; some may be
associated with meaning-differences (e.g. declensions for gender), others
may have no semantic association (e.g. conjugations). In many languages,
these morphophonemic similarities do not cover all verbs and nouns, leaving
various exceptions (‘strong’ verbs, etc.). Such syntactic constructions can
have great weight in the structure of an individual language or language
family.
Finally, two processes are found in language which are quite different
from those involved in chapter 3. One is analogy: the saying of a word in an
operator-argument combination on the basis of another word, which has
some correspondences to the first, appearing in such a combination. This is a
widespread process, restrained by its own conditions, and almost always
dependent on the prior existence of a grammatical construction on the
model of which the new form is made. Because of this dependence, many
forms which historically have resulted from analogy can be analyzed as
though derived by regular reductions from existing sentences: the only cost
may be extending the selection of a word, or extending the domain of a
reduction.
The other process is change through time. All languages change. These
changes are primarily of certain types. One type is a slow shift in the basis of
articulation, which changes certain sounds, and in particular situations
changes their phonemic relations. Another type is of a word or affix
extending its selection, sometimes into that of another, so that it competes
with the other for the same selectional and semantic niche, and may replace
it. Yet another type is extension of the domain of a reduction (i.e. which
words are subject to it), or of its conditions or frequency of application. And
there are changes due to the borrowing or invention of new words. These
changes are largely related to the structure of the existing language, so that
major alterations in the structure of language are rare indeed.
The change in selection, i.e. in which words have higher than average
frequency as operators or arguments of a given word, is widespread, and

important both for meaning and grammar. Many words change their meaning
in this way, especially when with particular other words. When a word
combination, including word plus affix, changes its selection considerably it
may have to be considered a new word, no longer a combination of its
component words (11.1). This is not necessary so long as the meaning
difference is regular. For example, a great many verbs have both intransitive
and transitive uses and meanings (The disc rotated, He rotated the disc), but
one form can be derived from the other: e.g. by zeroing in the causative (He
caused the disc to rotate), or by the ‘middle’ (Theglass broke, The book sold
well, G368). More widespread are the many metaphoric uses of words, as in
I see what you say, but these can all be derived from the conjunction (Oao) as
plus indefinites and classifiers (roughly, l react to what you say as one's seeing
things, 4.2, end). However, selection changes in individual words or in small
sets of them can change the grammatical status of the words or introduce the
possibility of new classes into the grammar. We see the first case in provided,
providing, where a great extension in their arguments took them from a
secondary-sentence verb, as in You can make the trip, provided with enough
money to a conjunction as in You can make the trip provided (or providing)
there is no objection. Possibility of a new grammatical construction is seen in
the extension of take, give, have, etc. from their original use in take an apple,
give an apple, etc., to take a look, give a look, have a look, etc., where most
of their original meaning is lost and where they could be analyzed as a new
class of ‘preverbs’ with a modifier-like relation to the verb that follows them.
(Within the uniform theory presented here this analysis is avoided, 8.2.)
The net effect of all these conditions is that languages have various
detailed structures over and above what the constraints of Chapter 3 create,
going beyond the corollaries of those constraints as surveyed in 4.3. This
further structure includes subsets of words within the basic argument and
operator classes, and various grammatical paradigms and standardized
sentence types such as the passive, the interrogative, the imperative. These
forms can, however, always be derived from the fundamental constraints
plus one or another of the conditions of 4.4. Anckhe information can always
be found in the base structure, in the base sentences corresponding to the
derivations.
5
Metalinguistic Apparatus
within Language
5.0. Introduction
Certain features of language structure create a family of metalinguistic

devices within language, i.e. of sentences talking about sentences and their
parts. These metalinguistic devices include: metalinguistic sentences (5.2,
9.2), and a sublanguage that serves as metalanguage of the whole language
(10.2); cross-reference (pronouns, 5.3); the use-mention distinction (5.4);
the self-actualizing (performative) properties of Isay (5.5); and metascience
operators in science languages (10.4).
5.1. The metalinguistic capacity
Two features underlie this capability of language. One is the fact that
phonemes are types and not merely tokens; that is, the set of word-sounds is
effectively partitioned (by distinctions, which create types) into phoneme-
sequences (7.1-2). The other is the lack of any relation, in respect to
meaning, between phoneme and words; that is, the meaning of words is not
referable to any meaning of their phonemes.
As to the first: the status of phonemes as types means that every occurrence
of a word-sound can be named, as being a case of one known phoneme or
another. We should not take this property for granted, as something which
holds for every system that could seek to represent (as language does) the
information of the world. The objects and events of the world could be
referred to not only by words but also by, say, pointing at them, or drawing a
picture of them. However, the latter kinds of representation cannot be
reused, to represent the representations themselves. True, we can point at
(someone) pointing, or draw (someone) drawing; but we do not have a
general way of uniquely pointing at all pointings which are specifically at a
given set of things, or drawing all drawings of a specific set of things. In
contrast, we can take all word-occurrences that name a given set of things
(say, the word chair for certain seats or the word bench for others, or walk
for certain perambulations or stroll for others) and we can uniquely identify
these occurrences by the name of their phoneme-sequence, e.g. the names

of the successive letters c, h, a, i, r. Since phonemes (and phonemic letters)
are characterized by their distinction from other phonemes, every occurrence
of a sound satisfying the properties of a particular phoneme distinction can,
as a token, be called by the name of its phoneme, as a type. The name of a
phoneme—say, vee or doubleyou—is thus itself a word, with its own
phoneme or letter composition (e.g. v, e, e), and with a set of objects (word-
sounds) to which it refers; and that name can be used in sentences, e.g. in
sentences that give the composition of each word. This makes it possible to
use words—that is, to talk—about phonemes, word-occurrences, and about
words as sets of word-occurrences. Such talking is called metalinguistic.
Once phoneme-sequences are associated with experiences of the world (and
so can be used to talk about them), nothing prevents their being associated
with a particular phoneme-sequence (and so being used to talk about it).
As to the second feature above, the lack of semantic relation between
phonemes and words: that the phonemic components of a word are not its
semantic components is a necessary property of language (5.4). Because of
this property, the operators which are selected by a word when that word is
said as a phoneme-sequence differ in general from the operators that are
selected when the same word is said as a carrier of meaning, of a dictionary
definition. It will be seen that this property underlies the metalinguistic
constructions. More explicitly, it gives rise to the distinction between use
and mention, and also to the distinction between direct and indirect discourse,
and in particular the effect of / say operating on a sentence.
The fact that the metalanguage sentences are sentences of the language
makes it possible to talk in the language about words and occurrences and
structures of language (speech), and also about the meaning of words and
their relation to the world observed by the speakers.
5.2. Metalinguistic sentences
The basic metalinguistic sentences, from which all others can be constructed,
are made of the names of phonemes with operators on them. The operators
specify the sound-differences among the phonemes, their location relative
to each other in utterances of the language, and which phoneme-sequences
constitute words in the language. These metalinguistic sentences are not
about sounds as such, for individual sounds (tokens) are in general evanescent
events which cannot be named—that is, cannot appear as words—unless
they have been classified into fixed types. But the names of the phonemes
and their classes (e.g. the word phoneme itself), and the names of those
phoneme-sequences which satisfy the conditions of being a word, are all
words or word-sequences, of zero level (i.e. primitive arguments), and they
occur under a particular selection of operators. We then have sentences such
Metalinguistic Apparatus within Language 125
as: Speech sounds with property A (or differing by A from other speech
sounds) are called occurrences of em; Em is a phoneme, written m; Some
phoneme-sequences constitute words', Mary is a word'. The letter-sequence m,
a, r, y constitutes the word Mary. The metalinguistic sentences are thus
sentences of the language itself: C precedes a in the word cat (or: Cee
precedes ay in the word composed of the letters named cee, ay, tee) is not
structurally different from The band precedes the marchers in the parade.
The metalinguistic sentences are a separate subset, and a part of a meta¬
language, only in that their elementary arguments refer to segments of
utterances, ultimately to phoneme-sequences. Thus the arguments used in
metalinguistic sentences are mentions (i.e. names) of the words (and
phonemes) used in the sentences of the language. Other than this the
metalinguistic sentences are themselves sentences of the language.
Given sentences which identify particular phoneme-sequences as con¬
stituting words, we can have sentences whose arguments are the words thus
constituted. Such sentences can state (morphophonemically) that different
phoneme-sequences constitute the same word, as alternative forms of it
(e.g. is and am, or initial e and initial iy in economics), or that a single
phoneme-sequence can, in different occurrences, constitute different words
(homonyms). Such sentences can also state the location of particular words
in respect to each other in utterances of the language, and also their
meaning, in terms of other words, either in isolation or in particular
combinations.
A phoneme is defined for the utterances of a given language (or possibly a
specified set of languages), but—crucially—not for sounds outside language
or in other languages. This restriction of domain holds necessarily for
metalinguistic sentences, which have been described above as sentences
whose arguments ultimately are phonemes, or phonemic distinctions. When
the domain of utterances is unspecified, the metalinguistic statements can
only assert the existence of phonemes, words, and their combinations.
When the domain is all utterances of a language—all that are observed or all
that are predicted—or a specified subset of them, we can state regularities of
form and of occurrences within the domain. These regularities may be
of phonemic shape, as in similarities or transformations among word-
occurrences; or they may be of word combination, as in defining operators
and arguments. We thus reach statements about phoneme classes, word
classes, grammatical structures, and types of meaning for words and sentences.
A new situation arises when the metalinguistic statement is made about a
particular utterance. Such statements can be used freely, as in John is
brash—and that last word is an understatement. The utterance in question
has to be identified—that is, cited—within the metalinguistic statement; the
utterance cannot be referred to unless it (or a pronoun for it) is grammatically
included in the metalinguistic statement itself. This means that the utterance
in question and the metalinguistic statement about it form together a larger
operator-argument sentence: e.g. The utterance which consists of John

bought a book has the word John as first argument of the word bought, or The
word John is the first argument of the word bought in the utterance which
consists of John bought a book. In particular, the instruction to carry out a
reduction (3.3) can be a metalinguistic statement attached to the sentence in
which the reduction is to be made: (1) Sheep are voracious and sheep eat all
day, in which sentence the second occurrence of the word sheep is zeroable.
The same holds for the sameness statements of 5.3.3, which are sentences
conjoined to the sentence of which they speak, whether or not they are
followed by an instruction to reduce the repeated word: as in John lost some
time and Mary lost some time, to which one can conjoin (roughly) where the
second argument of the first operator consists of the same words as the second
argument of the second operator', also in John deserves John s being appointed
as chairman, to which one can conjoin (roughly) where the first argument of
the first (higher) operator has the same referent as the first argument of the
second (lower) operator.
This situation makes it possible to consider reductions in the form of a
word, and other transformational variants in the form of a sentence, to be
simply morphophonemic variants of the attached instruction to carry out
that reduction. In an exaggerated way, one can consider that in (2) Sheep are
voracious and eat all day, the lack of the required first argument for eat all
day (i.e. the absence of the second sheep) is simply an alternative form for
(3) in which sentence the second occurrence of the word sheep is zeroable. If
we said that (2) was a variant form of Sheep are voracious and sheep eat all
day, we would still need a grammatical statement (in the metalanguage)
saying that which is said in (3); and so for all other reductions and trans¬
formations. But if (2) is taken as a variant form of (1), then all we need for
the language is a single grammatical statement, namely that the phonemic
addition which comprises a transformational instruction adjoined to a given
primary sentence has as phonemic variant the particular phonemic change in
the primary sentence which is stated in the adjoined instruction. To regard
the absence of the second sheep in (2) as a phonemic variant of (3) in (1) is
not different in principle from regarding the initial sound e of economics as a
variant of the morphophonemic instruction in In the word economics (with
initial sound iy), the initial iy is replaceable by e.
The effect of sentences carrying metalinguistic information about them¬
selves is even more far-reaching than the case above. In 9.1 it will be seen
that conjoined sentences have the structural property of being typically
connected by word-repetition; that is, there is likely to be some word that
occurs in both sentences. (1) Mary went home because Mary (or: she) was
tired is more likely to occur without question than (2) Mary went home
because Jane was tired. And (2) itself raises no question if we add (3) Jane is
Mary's daughter, which supplies a repeated word for each of the component
sentences in (2). And if (2) without (3) raises no question, it is because the
hearer knows (3) or the like, so that for him (3) can be zeroed. Indeed, any
two sentences can be acceptably conjoined it we add a repetition-supplying
sentence. We can now consider a similar treatment for the structural
properties of single sentences. Clearly, all the properties of a given sentence,
including definitions of the words, can be stated in metalinguistic statements
about that sentence which are adjoined to it: e.g. (1) John bought a book,
where the word John is first argument of the Om, word bought and the word
book is second, with John being the name of a man, bought meaning
purchased, and book meaning a printed volume. (An n-fold ambiguous
sentence would have n alternative sets of metalinguistic statements adjoin-
able to it.) When the grammatical status and the definition of each word is
understood by the hearer, the conjoined metalinguistic sentences which
state them are zeroable. Evidence for their presence and their zeroing may
be seen in the following observation: virtually any arbitrary phonemic
sequence satisfying English syllabic constraints becomes an English sentence,
acceptable and understandable to the hearer, if it carries adjoined meta¬
linguistic statements identifying its parts as words satisfying their argument-
requirement: e.g. stating that its initial sub-sequence is the name of some
foreign scientists, its final sub-sequence is the name of some new gas, and the
sequence between these (which would have to end in 5, or ed, or t) is a
scientific term for some laboratory operation perhaps derived from the
name of the scientist who developed the operation. In this last case,
differently from (1), the grammatical and definitional statements are pre¬
sumed not known to the hearer and hence not zeroed, as they can be in (1).
It follows that all sentences can be thought of as originally carrying
metalinguistic adjunctions which state all the structural relations and word
meanings necessary for understanding the sentence, these being zeroed if
presumed known to the hearer. Every sentence, such as in (1) and (2) above,
is thus a self-sufficient instance of what is permitted by the metalanguage
(the grammar and dictionary definitions of the language). We can thus
append to a sentence in a language all the metalinguistic statements necessary
for accepting and understanding it, with the whole still being a sentence of
that language. This is possible because the grammatical and dictionary
descriptions of a language are stated in its metalinguistic sublanguage (10.2),
somewhat as in computer programs from von Neumann’s original plan and
on, the instructions have the same computer structure as the data on which
they operate.
One might ask where these metalinguistic sentences about a sentence are
adjoined, since many of them disappear in the course of constructing the
sentence. The observations of 3.3.4 and the considerations of 8.3 suggest
that they enter into the making of a sentence as soon as the words they are
speaking about have entered the sentence (since the reductions which some
of these engender are carried out then). This clearly holds for those
metalinguistic adjunctions which are made known to the hearer only by their
being carried out, namely permutations and reductions, including those that
are based on a sameness statement (e.g. pronouning): these adjunctions are
replaced morphophonemically by the act of carrying them out. In the case of
the relevant grammar and dictionary statements, which are in most cases
known to both speaker and hearer, the issue of their location is moot since
these are zeroed on grounds of being known.
5.3. Reference
5.3.0. Introduction
All referentials (e.g. pronouns) can be derived from cross-reference, i.e.

from having one word in a discourse refer to another occurrence of a word in
the same discourse. The capacity for cross-reference arises from three
properties: having repetition, having words for stating the relative location
of words in the discourse, and having words for stating sameness of the
repeated words.
We have seen that metalinguistic statements can refer to the relative
location of elements in utterances of a language. This applies to the easily
perceived linear order of phonemes or of words (especially to being before,
after, or next), and to the partially ordered ‘grammatical relations’ of words
(e.g. subject of, i.e. first argument of). When an utterance is cited in a
metalinguistic sentence, the utterance segments can be located linearly or
grammatically in respect to the metalinguistic sentence which contains it and
refers to it.
5.3.1. Repetition
Under all two-argument operators, there is a possibility of word-repetition

in the arguments, i.e. of a word-occurrence in the second argument being a
case of the same word (and possibly referring to the same referent) as an
occurrence in the first argument. The grammatical importance of such
repetition (in high-likelihood situations) is seen in its reduction to pronouns
or zero. Under 0„„, we see this in The boy washed himself, which is
derivable (G134) from a reconstructed The boy washed the same boy (where
we have the same referent). Repetition is frequent under certain Orm, as in I
prefer to go now, derivable from I prefer for me to go now (me having the
same referent as /); and under Oao, as in John plays violin and Mary piano
from . . . and Mary plays piano (where plays is the same word with the same
meaning). Repetition has high likelihood also in successive sentences of a
discourse (9.3).
Repetition may occur in a position where it has high likelihood but where
the first occurrence (the antecedent) is not in a fixed favored position in
respect to the repetition. In such cases, the repetition can be reduced to a
pronoun rather than to zero. For example, given Oao under which repetition
is expected, but without relative positions being specified (9.1), we find e.g.
The mail did not come though everyone was awaiting the mail reducible to. ..
though everyone was awaiting it (but not to everyone was awaiting). And
under and, where repetition has highest expectancy in the corresponding
grammatical position, any repetition in other positions is reducible only to
pronoun: The play was good and the audience liked it. But when repetition
occurs in the fixed high-expectancy position in respect to the antecedent, it
can (and under some operators must) be reduced to zero. Under and this
happens when the grammatical position of the repetition is the same in the
second sentence as that of the antecedent in the first: The play was good and
the play was applauded is reducible to The play was good and it was
applauded and also to The play was good and was applauded. Under O00,
everything after the tense is zeroed in the second sentence if it is the same as
in the first sentence: / will support John for the job if you will support John for
the job is reducible to I will support John for the job if you will. Under O,,,, (as
above), if the second argument has the same referent as the first, it is
reducible to pronoun plus -self: The man heard himself on the tape (from
heard the same man). But under particular Onn where sameness of the two
arguments may be especially frequent, it is also reducible to zero: The man
washed himself and The man washed. In the case of the Ono, each operator
has a particular relative position for zeroing repetition (3.2.3). For example,
/ promised him to go is from for me to go, but I ordered him to go is from him
to go, while 7 offered him to go can be from either. But under Oon, nothing is
reduced to zero: the second argument of surprise is not zeroed either in His
meeting her surprised him or in His meeting her surprised her.1
5.3.2. Locating the antecedent
When a pronoun such as he or it appears in a discourse, or in most positions

under Oao, the hearer does not know what preceding word it refers to
(unless there happens to be only one appropriate word, e.g. masculine
1 The analysis of pronouns as repetition (mostly within a sentence) may explain why
referential pro-verbs are hard to come by. By the semilattice property, only one word-
occurrence in a sentence can be a free operator (the main verb , the highest operator in the
semilattice). Hence if an operator word occurs twice (as repetition) within a sentence, at least
one occurrence, and in many cases both, would be as argument of a higher operator, there
would never be two occurrences as free operators, one pronouned as repetition of the other. In
contrast, the pronounings that are met with are between two argument occurrences, either of
N, e.g. in John washed himself, or of nominalized operators, as in Mary's happiness is more
important than John’s, or in I won, which he had hoped to do (where which . . . do is as though a
pronouning of do winning, from He had hoped to do the winning).
singular for he). However, in particular positions under certain operators, a

pronoun or zero is recognized by the hearer as referring to a word in a
particular antecedent position in the sentence: in John phoned Mary and
wrote me, the zero before wrote refers only to John. To state the information
required for this recognition, it is necessary to have an addressing apparatus
which is communicated to the hearer within the sentence itself. In the
hearer’s learning of his language, this recognition arises without addresses
being said: the hearer, in hearing whole sentences and gathering their
meaning, recognizes that in particular positions the pronouns or the absences of
an expected word (e.g. of the subject of a following verb) always refer to an
antecedent in a particular other position. Nevertheless, it is important for us
to know that the information needed by the hearer in order to recognize
the sentence and to locate the antecedent can be stated in metalinguistic
sentences (of ordinary grammatical construction) adjoined to the given
sentence. To state the relative location of the antecedent, these sentences
need just the existing vocabulary of the language.
Languages have such words as latter, former, earlier, before, after, next,
following, inverse, first, second, etc. In the metalinguistic sentences which
talk about utterances, these predicates can be used about word- and
phoneme-occurrences: e.g. He was influenced by Stravinsky and Prokofiev,
more by the former than by the latter (meaning ‘former said’ and ‘later said’),
The first three across the finish line were Smith, Jones, and Bannister, in that
order (i.e. the runners were in the order of saying their names here; also . . .
in reverse order). Other locational words have arisen which refer explicitly to
word-occurrences, e.g. in English aforementioned, respectively, vice versa,
and the common use of latter.
The utterances or word-occurrences that are being talked about must be
in the same sentence as the metalinguistic statement about them, or at a
stated distance from it in the same discourse, and this for two reasons. One
reason is that, as in the examples above, the material cannot be talked about
unless it is presented to the hearer: we would not say c precedes a unless we
add in the word cat, etc. The other reason is that it cannot be identified unless
it is at a stated grammatical distance from the metalinguistic statement or at
a stated linear distance from the beginning of that statement. If we want to
express within the metalinguistic statement the hearer’s recognition of the
position of the antecedent, we can do it by having the metalinguistic
statement interrupt right after the word-occurrence that it is talking about.
Thus (1) A small boy disappeared, which is equivalent to (2) A boy who is
small disappeared, is composed of (3) A boy disappeared interrupted by (4)
A boy is small plus (5) the information that the two occurrences of boy refer
to the same entity. For the hearer, this information is carried here by the
who. To express this information by existing language material in an
adjoined metalinguistic sentence, we need only to consider wh- as an O00
(conjunction) introducing a secondary sentence (i.e. a sentence under the
secondary intonation marked by dash or semicolon, as in A boy disappeared;

the boy was small) and to consider -o, -ich, etc. as reductions of a repetition
of the preceding argument. I hen in A boy—a boy is small—disappeared we
need a further interruption after the second boy stating (5) that the two
successive occurrences of boy have the same referent. Thereupon -o replaces
the second a boy plus the sameness statement, and wh- replaces the dash, all
of which yields (2) which is reducible to (1).
Of course, neither the history of language nor the recognition process of
the hearer goes through this reconstruction. But the reconstruction shows
what apparatus is needed for cross-reference. And it shows that language
developed in a way that always permitted an as-though reductional derivation
from a regular predicational source, and that all the information carried by
language—in this case, that contained in pronouns—can be expressed in a
purely predicational—in this case metalinguistic—source.
Aside from the relative pronouns in wh-, there are certain other cases of
reduced repetition with fixed antecedent. Here too, given the reduced item,
we have to say that it is the same as a word at a fixed position relative to it. In
English, the main casesof this are zeroings under O00, Ono, and Onno. Under
Ono and Onno, it is when the first or second argument of the second (lower)
operator is the same as the first or second noun argument of the first (higher)
operator, depending on what is the highest-likelihood position for repetition
under the particular higher operator. As has been seen, under promise we
have I promised him to go from / promised him for me to go; under leave we
have I left him feeling sad from either/ left him, I feeling sad, or / left him, he
feeling sad; under order we have 1 ordered him to go from l ordered him for
him to go; and so on (3.2.3, G148).
We see that in all cases where the antecedent of pronouning or zeroing is
in a fixed relative position, that position is easy to state, either linearly or in
the partial order. In the case of the relative clause, if the secondary sentence
interrupts the first immediately after the antecedent, and the repeated word
is at the beginning of the secondary, then the two same words are next to
each other so that the antecedent can be identified simply as predecessor. To
this end, it is relevant that in English the words that can be reduced to a
relative pronoun are precisely those words in the secondary sentence that
can be said at its beginning even when that sentence is said by itself: words
that are never said (even for stress) at the beginning of a sentence are never
replaced by a wh- pronoun (when that sentence appears as a secondary one).
This means that the presence of the relative pronouns (who, etc.) at the
beginning of the relative clause is due not to permuting a pronoun but to the
repeated word (before being pronouned) being already at the beginning of
the secondary sentence before the secondary sentence was brought in as an
interruption of the primary: e.g .John, whom I asked, refused is from John—
John I asked—refused (originally from John—I asked John—refused).
Since, then, the relative pronouning does not take place except when the
antecedent and its repetition are already next to each other, we must judge
that what makes it possible to reduce the repeated word to a wh- pronoun is
its nextness to its antecedent, where the fact of repetition is most obvious
and the naming of the antecedent location is easiest: the preceding word.
Somewhat similarly, in the zeroings, under Oao, Ono, and Onno (above),
which have a fixed position for their antecedent, that position is easy to
recognize and state in terms of the basic grammatical relations (the partial
order of dependence): the corresponding position under O00, and a fixed
semilattice-relation of lower to higher argument under O,lo and Onno.
Thus the only referentials whose antecedent is always at a fixed relative
position (-self, relative pronouns, and zeroed repetitions) are ones for which
the antecedent position is easily indicated, linearly or in the partial order.
This suggests that the ability to reduce the repetition in these cases was based
on this ease of locating the antecedent. That means that in the formation of
referentials, i.e. in the reducing of repeated material, the development of
the language and also the recognition by the hearer do not use the general
partial-order and linear addressing systems that are inherently possible in
language. If either or both of these addressings were usable over their full
range by the speaker and the hearer, every phoneme- and word-occurrence
could be identified and referred to, including the position of any antecedent
relative to the referential. There would then be no reason for referentials
with fixed antecedent to occur only in situations where the location of the
antecedent relative to the referential is especially easy to state and to
recognize.
Conversely, we must judge that the other referentials, whose antecedent
is not limited to a single position, indeed do not refer to any particular
position (which, as seen, could not be addressed or named within the use of
language as we see it). For example, in John prefers that he be appointed
now, the he can be a repetition of or refer to the preceding John or any other
human masculine singular noun in a nearby preceding sentence. One might
think that these free pronouns refer to a disjunction of all preceding
appropriate nouns: either to John, or else to the masculine singular noun
before it, or to the one before it, and so on. But each one of these
alternatives (except the first John) would have to be identified in a complex
address utilizing the apparatus of a general addressing system. Since no fixed
antecedent uses the (in any case, non-existent) large addressing apparatus,
we have to judge that the non-fixed antecedents are not simply a disjunction
of fixed antecedents. All this means that we have to understand such free
pronouns as he and it to be not a reduced repetition of one or another
preceding noun but a word meaning 'aforementioned (human) male' (or:
‘male mentioned nearby’), 'nearby-mentioned thing’, etc. This, even though
the speaker knows which word is the antecedent, and the hearer may guess it
on some grounds.
5.3.3. Sameness
We have seen here that the fixed-antecedent referentials are reduced

repetitions (and thereby cross-references), whereas the other pronouns are
words with their own cross-referential meanings. The information which
triggers the reduction of repeated words is that a given word is the same as
the word at the fixed antecedent position, and when the hearer comes upon
the referential this is the information it gives him (5.2). It remains to be seen
what ‘same’ means in this case. When the antecedent carries a ‘determiner’
(e.g. demonstratives, article, quantifiers) the same usually means ‘has the
same individual referent’, as in I saw a bird, which flew off immediately or
Someone, who shall remain unnamed, goofed. Note that A man came and a
man left may refer to a different man, but A man came and left refers to the
same man, and hence would have to be derived from A man came and the
same man left. But compare I brought him a cup of tea, which he needed
badly, where -ich indicates the words cup of tea but not its individual
referent. Otherwise, same means ‘is a case of the same word’ (which includes
having the same selection, hence meaning) as in: All day, men came and left
(not necessarily all of them the same men); He is educated, which they are
not; John plays violin and Mary piano (but hardly John plays chess and Mary
piano, where different selections of play are involved, except as a joke); /
will go today because he did yesterday (not the same act of going); This one
grew more than the others (with each doing its own, possibly different,
growing).
Also in he, she, it, they, there is a distinction between ‘same word’ and
‘same individual referent’, except that here these sameness properties are
part of the meaning of the pronominal words, rather than being part of the
metalinguistic instructions about the antecedent. He and she almost always
refer to the same individual referent as in some appropriate nearby word.
But it, that, and they may have merely the same meaning as some nearby
word: Eat less bread! Do it now! (Wartime governmental billboards in
England); He left early, and I did that too; People may succeed, or they may
fail.
5.3.4. Demonstratives: Reference to the external situation
Up to this point, 5.3 has analyzed the cross-reference of words and zeros
(word-absences) referring to word-occurrences in the same discourse. There
remain a few pronouns which are demonstrative or deictic, i.e. refer overtly
to an object or situation not mentioned in the text but present to the speaker
or hearer: e.g. That is a beautiful tree, This tree is beautiful. These can be
derived from the already-established intra-text cross-reference pronouns by
assuming a zeroable indefinite noun attached to the zeroable metalinguistic /
say (5.5.1): e.g. I say of something here that this is a beautiful tree (where this
is a cross-reference to something here), or / say of something at which we are
looking (or of which we are speaking) that that is a beautiful tree.
Such words as here and now can be derived from at the place in which I am
speaking and at the time at which l am speaking. And /, you can be
considered reductions of speaker, hearer when these words occur in a
discourse as repetitions from a zeroable The speaker is saying to the hearer,
which one might presume to introduce every discourse (5.5.1): / will phone
you as though derived from a reconstructed The speaker is (hereby) saying to
the hearer that the speaker will phone the hearer.
Before analysis, the kind of meanings of all these words differ from the
kinds of meanings of the other words of the language. In the other words, the
meanings or referents may vary with the environing words (the meaning of
child in a child of his time as against a child prodigy), but they do not vary
with the time or place in which the sentence is said. But in the words
discussed in 5.3.4, the referents vary with the particular bit of speaking:
which tree is this and who is meant by I depends on the speaking situation.
The information of demonstratives, including the tenses (below), is for this
reason unique also in not being transmissible beyond face-to-face com¬
munication. The derivation from an Isay component underscores, and does
not eliminate, this property. Nevertheless, it is a gain for the compactness
and simplicity of grammatical description if we can eliminate the new kind of
meaning required for 5.3.4; and this can be done by assuming a zeroable I
say of some thing that that... (as source of demonstrative that, this), or The
speaker is saying to the hearer at the place and time of speaking that.. . —» here
and now. It will be seen in 5.5.1 that there are many grammatical reasons for
assuming such an introduction to be implicit in each discourse or sentence
that is said (not merely composed). Note that in sentences that are composed
but not really said (e.g. in grammar exercises), I, now, and that tree, and the
like do not have specific referents, as they do in each said sentence.
The derivability of tenses from before, after connecting a said sentence to
the I say operator on it (G265) means that the tenses are equivalent to
demonstrative time-adverbs.
5.3.5. Summary: Reference from cross-reference
In 5.3.4, it was seen that words which are deictically referential in an

utterance can be derived from cross-reference by reconstructing a source
that contains a zeroable The speaker says, i.e. that refers to the speaking (or
writing) of the utterance. Once this is done, we do not need to know
any extra-linguistic context, i.e. anything outside the sentence itself (with its
adjoined zeroed metalinguistic statements, 5.2), in order to understand it as
a sentence.
As to cross-reference, this was analyzed as a reduction of a second
occurrence of a word stated to be the same word (or to have the same
referent) as another occurrence (the antecedent) within the utterance. Here
a distinction was found between two types: first, there are pronouns whose
antecedent was not at a fixed relative position (e.g. he)', here the antecedent
is characterized merely as having occurred nearby in the discourse. Second,
there are pronouns and zeroings whose antecedent is at a fixed (and simple)
position relative to the pronoun or zero, in the linear order of words or in the
partial order that creates a sentence (e.g. who)', here the position of the
antecedent and its sameness is given in a metalinguistic adjoined statement.
5.4. Use/mention
We now consider the second feature that underlies the internal meta¬
linguistic capacity of language (5.1), namely, the fact that the meaning of a
word is not referable to any meaning of its phonemes. The independence of
word meaning from its phonemes is necessary, for language, for if the
meanings of words were composed regularly out of the meanings (or the
effects upon meaning) of their phonemes, the number of distinguishable
word meanings would be limited by the combinations and permutations that
one could construct, in reasonably short sayable sequences, out of the
meaning contributions of the twenty-odd phonemes that the language has.
Given this necessary independence, what we-say about a word (i.e. its
selection) when it is taken as indicating its referent differs from the selection
of operators which appears on the phoneme-sequence of that word: e.g.
Mary phoned me, but Mary consists of four phonemes. Note that one cannot
pronoun a word’s occurrence under the one selection from its occurrence
under the other (5.3.3): we do not pronoun John and Tom to make / saw
John and Tom whose second phonemes are identical, but we do in l saw John
and Tom, in whose names the second phonemes are identical.
The sameness statements do not support pronouning when the selections
and meanings of a repeated word are sufficiently different, as in play of 5.3.3
or heart below. Going beyond that, when certain occurrences of a word have
a coherent selection totally different from that of other occurrences of
that word, we say that those two sets of occurrences belong to different
homonymous words: e.g. the occurrence of sound in the senses of ‘noise’ as
against ‘strait’ and ‘healthy’; but no homonym for the occurrences of heart in
the senses of ‘organ’ and of ‘center (as of lettuce)’ (which we consider to be
different meanings of a single word). In the case of the two sets of operators
on Mary, we hesitate to say that, for example, (1) Mary in Mary has (or:
consists of) four letters, which is not pronounable from (2) Mary in Mary
phoned me, is simply a homonym of (2) in the way that applies to the various
occurrences of sound, for then every word would be a homonymous pair.
That is, we do not want to say that Mary is written with a capital m and Mary
phoned have two homonymous words Mary. A more special treatment is
clearly required here.
Aside from the difference in selection between the phoneme-sequence of
a word, and a word as bearer of a referent, there are occasional other
differences. On the one hand, there are cases of different occurrences of the
same phoneme-sequence constituting different words as referent-bearers:
the homonyms, above. On the other hand, there are cases of different
phoneme-sequences constituting the same word. This happens not only in
what are recognized as suppletive variants of a word (e.g. is, am as variant
forms of be) but also in what are considered alternative pronunciations of a
word. For example, the semantically single word written spigot consists of
two alternative phoneme-sequences, one with medial phoneme g and one
with medial phoneme k (and with older spelling spicket). Both this problem
and that of the homonymous pairs can be avoided if the underlying sentences
are taken as, for example. The phoneme-sequence that constitutes the word
Mary consists of (or has) four letters, reducible to Mary has four letters, as
against The object referred to by the word Mary phoned me (or, for that
matter, has four letters to mail), reducible to Mary phoned me. (The
evidence that what was reduced was indeed the phoneme-sequence is that the
predicate on it, consists of four letters, is in the selection of phoneme-
sequence.) In the reduced forms, the distinction is expressed by saying, in
Quine’s terminology, that in the latter case the word Mary is ‘used’ in the
sentence, while in the former case it is only ‘mentioned’: ‘mentioning’ a
word is ‘using’ its phonemes. These useful terms isolate and name the
distinction; to derive or explain the distinction grammatically, we need
underlying sources as above, while noting that the zeroing of the phoneme-
sequence that constitutes leaves quotation marks as its trace (in writing:
‘Mary' has four letters). This derivation explains why a mentioned plural
noun is singular: Bookworms is on p. 137 in this dictionary (from The
phoneme-sequence that constitutes the word bookworms is on ... ); contrast
Bookworms are all over in this dictionary (the latter from The objects
referred to by the word bookworms are . . . ). The derivation above also
explains why one does not say Mary has a little lamb and four letters, which
would be a possible zeroing from Mary has a little lamb and Mary has four
letters, but not from The object referred to by the word Mary has a little lamb
and the phoneme-sequence constituting the word Mary has four letters. The
importance of this explanation is that it remains within the universe of word-
occurrences and their reductions, without appealing to such considerations
as whether Mary is ‘taken’ as a meaning or as a spelling. When the source
sentences are reconstructed, the use/mention distinction, and the Bourbaki
‘abus de langage’, becomes redundant.
As to the failures in one-to-one correspondence between phoneme-
sequence and word-referent, they can be covered by appropriate underlying
statements. Thus if a word has alternative phonemic forms, the sentence

underlying its mention would be something like One of the (alternative)
phoneme-sequences that constitute the word economics begins with the phoneme
e.
In the metalinguistic sentences of 5.2, the operators speak about the
occurrences of a word, hence about the occurrences of its phonemes. This
applies to such a sentence as The phoneme-sequence t, e, n constitutes a
word, or Ten means the successor of nine. Thus the metalinguistic sentences
speak about a word as ‘mentioned’ even when they give a definition of the
word or indicate its referents.
5.5. Perform a ti ves
5.5.0. Introduction
A distinction related to use/mention appears under the word say, which has
a unique status in language structure. In various languages, say appears with
two forms of second argument (object): He said (or: Isay) ‘Mary is here’-, He
said (or: I say) that Mary is here. There are syntactic differences between the
two forms. Under the bi-sentential operator but, which requires in principle
two contrasts between its arguments, we can find I said 'Franklin Roosevelt
died in 1945' but he said ‘President Roosevelt died in 1945’; but we would not
normally find (even with contrastive stress on President) I said that Franklin
Roosevelt died in 1945 but he said that President Roosevelt died in 1945. We
ask if these two forms of the argument of say can be derived in a way that
would explain why the two forms differ in respect to but. With an eye to the
use/mention sources of 5.4, the obvious sources to propose are: for the first,
1 said the words (or word-sequence) Franklin Roosevelt died in 1945 but he
said the words President Roosevelt died in 1945 (with the words replaced by
intonation of quotation marks); and for the unlikely second, 1 said words
which mean that Franklin Roosevelt died in 1945 but he said words which
mean that President Roosevelt died in 1945 (with zeroings of words which
mean), where any form of that name and indeed any synonym of it or of died
could have been what was actually said. In the second source, but is clearly
less likely than in the first. In both of the proposed sources, the second
argument of say is word, thus eliminating the problem of having two forms of
the argument (above).
The relation between a quoted sentence and that plus the sentence is also
seen in the difference between Yesterday in Boston the candidate said to the
strikers 7 have come to support you here and now' and ‘ Yesterday in Boston
the candidate said to the strikers that he had come to support them there and
then. Such differences as here/there hold not only underlay, but also under
tell, state, repeat, assert, claim, insist,promise, think, agree, and other words.
Among all these there is one, Isay or One says, in the present tense, that has
a peculiar property which serves to explain the relations noted above, and
also several other exceptional forms in language.
This property is a metalinguistic analog to the performatives discussed in
logic, as when a judge says / (hereby) sentence you to . . ., in which his saying
that he is sentencing constitutes the act of sentencing. For a speaker to
compose the sentence I say ‘John left' or I say that John left (or possibly One
hereby says that John left) is the same as for him to actually say John left.
Hence, the I say or One hereby says contributes no information, and is
zeroable. Indeed, the sentence John left, and any sentence that is said (aloud
or to oneself) or written (except e.g. as an example of a sentence) can be
considered a reduction from: I say, plus that or quotes (or equivalent marks
such as a colon), plus the sentence in question. The Isay is not zeroed when
it is not used performatively, as in I say it’s spinach and Isay to hell with if, or
when it is used contrastively, as in I say John left but Mary denies it, or I say
that he left, but I don 7 mean it. In such cases, it may even be repeated, as in I
say that I say that he left, in order to make it clear. But a sentence cannot be
suspected of carrying an infinite regress of zeroings of / say, from an
infinite repeating of the performative I say, because that zeroing takes place
upon the occasion of the saying, and once the sentence has been said, it
cannot repeat the performative I say and its zeroing.
5.5.1. I say
There is varied evidence of the original presence of I say or One hereby says
and their subsequent zeroing. To any sentence one can add such implicitly
metalinguistic phrases as to tell the truth, or to be specific, or not to make too
fine a point of it. The zeroing of the subject in these phrases means that there
must have been an earlier occurrence of that missing subject (of tell, etc.)
which had served as repetitional antecedent for the zeroing of that subject.
This other occurrence of the subject could not have been present except as
an argument of some operator; and the combination of that earlier subject
with its operator had to have been zeroed on some grounds, since they do
not appear. The only candidate for such zeroing that would have been
possible in every sentence and that would have left no trace (except for the
very saying of the sentence) is the performative I say or One hereby says; no
other subject and tense of say would make it performative, nor would any
other verb with different meaning.
In many cases there is direct evidence of a metalinguistic I, as when the
insertable metalinguistic phrase is not to repeat myself (where myself could
arise only from for me not to repeat me), or to make myself clear-, also in I
hasten to add or at least I think so or and I say again. Here, the repeat, say
again, add, and at least each implies in different ways an I say or One
(hereby) says on the primary sentence: John left and, 7 hasten to add, not
secretly from / say John left and I hasten to add. . .; John left, at least l think so
from 7 say John left, at least I think so; He's wrong. I say again: He’s
absolutely wrong from 7 say he's wrong, I say again ....
The widespread zeroing of / say explains certain forms where the analysis
shows that I say has been zeroed by mistake. For example, sentences such as
He's back, because his car is in the garage, which do not make sense as said,
clearly have been reduced from / say he’s back, because his car is in the
garage, where the 7 say should not have been zeroed, because it operated
only on the first sentence and is in turn operated on (as first argument) by the
because. Another example is in sentences such as The Times says that our
stupid mayor will be speaking tonight; since the newspaper said neither
stupid nor our, the source here has to be 7 say that the Times says that the
mayor who I say is our stupid one . . . Here, zeroing the second 7 say left the
who is stupid to appear as modifier on mayor, which is incorrect since mayor
is under the Times says. Consider also the sentence in which a reporter
writes, about a woman speaking to her husband. She said Mary, their
daughter, would go to the police station. Had the woman said to her husband
Mary, our daughter, . . ., this would have been reported as She said Mary,
their daughter,. . .; but of course she did not say our daughter. She must have
said Mary will go in which the reporter inserted something like who I say is
their daughter, where 7 say should not have been zeroed because (as with
stupid above) it is one speaker’s insert in reporting another person’s remark.
5.5.2. Imperative and interrogative
There is grammatical evidence that the imperative and interrogative are

constructed like the performative say. In the first place, even languages such
as English, in which the imperative has no subject, carry evidence of a
zeroed subject. In Look! we find no direct evidence, but in Wash yourself!
the -self indicates that the second argument you is a repetition of a first
argument you under wash. This second-person first argument is visible in
such forms as You stop that!, (which may well be different from You (there)!
Stop that!) and in the older Go thou! As to third-person subject, there is no
Wash himself!-, but Let him wash himself! is a way of getting a pseudo third-
person imperative on wash from what is grammatically a second-person
imperative on let-, note also Let's go as a second-person imperative form for a
first-person imperative .meaning. A rare third-person imperative may be
claimed in Somebody call the police!
The invisible existence of a higher metalinguistic operator is indicated by
the rare English That he go! (common in French: Qu'ilparte!). Here we can
reconstruct 7 ask (or request) that he go, with the subjunctive of go under ask-,
in effect, for modern English, this consists in zeroing the will which has high
likelihood in the last argument of such operators as request, beg, demand,

command, desire. Given the high likelihood of repeated yon here, as in / ask
that you go, I beg of you that you go, the you of you go is zeroed in the direct-
quotes form / beg of you: ‘Go!’, where the imperative intonation (the
exclamation mark) is a phonemic extension of the beg, request, operators.
Given the imperative intonation, the I ask you or / beg of you, etc., can
now be zeroed, as having probability of 1, and hence contributing no
information, in the presence of that intonation: this yields Go! (In Would
that he went, usually without an imperative-like intonation, only the /
subject of would has been zeroed.) Furthermore, given the zeroability of /
say above, it is possible to think of / request as composed of / say plus an
adverbial imperatively, as request (as in 5.5.3), where the adverbial portion
imposes the imperative intonation and is then zeroable, while the I say
portion is zeroable performatively: this admits of the existing intermediate
form I say (to you): Go!2 (G351).
The interrogative is far more complex, and its metalinguistic structure is in
English more evident. It is generally considered that grammatically there
are two kinds of questions: (a) the yes-no questions (Will John leave? Yes.)
and (b) the wh- questions (Who will leave? John.) However, the yes-no
questions can be readily derived from (c) disjunctions of questions—or, for
that matter, a querying of a disjunction of possible answers: Will John leave?
from Will John leave or will John not leave? It is derived by optionally
zeroing any material that is repeated under not (if one of the disjuncts
contained repeated material under not), and then optionally zeroing the or
not. Hence Will John leave or not leave?, Will John leave or not?. Will John
leave?
It is for this reason that Will you?, which is reduced from Will you or won't
you?, means (aside from nuance) the same as Won't you?, which is reduced
from the equivalent Won't you or will you? (or and and being semantically
commutative). The extended form (c) is a source, and an always-possible
constituent, of the short (reduced) one, and is a special case of the fact that
every question offers a disjunction, and any disjunction of questions con¬
stitutes a question: Will John leave or will Mary leave? Will John leave or will
Mary return or will everybody stay put?
When we now consider (b) the w/j-question, we find that, e.g.. Which one
will leave? is derivable from a distributive For each one, will this one leave or
not? This last form, in turn, is derivable from I ask, for each one, whether this
one will leave or this one will not leave. Note that the source disjunction is not
of questions but of possible answers. Even in such a question as What
~ The sentence on which / request operates can be a disjunction (or conjunction) of

sentences, as in Take it or leave it! from / request that either you take it or vou leave it (and not
from or operating on two occurrences of / request). However, this differs from the essential
occurrence of a disjunction of sentences as the only and complete operand of I ask, I query, in
the interrogative form below.
numbers are prime?, where no finite disjunction of answers seems possible,

this analysis offers a finite source: / ask, for each number, whether that
number is prime or that number is not prime. The same source is then found
grammatically adequate for (a) the yes-no question too: in the example
above, I ask whether John will leave or John will not leave. We thus arrive at a
single source from which all interrogatives can be derived by established
reductions: namely, I ask (or: enquire, wonder, query each answer, etc.) plus
whether (in English wh- plus either), all operating on two or more non-
interrogative sentences. In this structure, the disjunctive either imposes or
between the operand sentences (like what and, or do, 5.5.3). Like the
imperative, the ask can be considered as composed of say plus an adverbial
querying or in request to choose either, where this adverbial generates a
question intonation on the operand sentence-disjunction and is thereafter
zeroable, and the residual / say is thereafter zeroable.
This analysis is not merely a matter of finding a source that has operator-
argument form. It produces and explains many peculiarities of the inter¬
rogative. In English, it yields the question form out of the ordinary form of
the operand sentences (This one will leave under (b), and John will leave
under (a), above). For English, it takes the wh- of whether to be the same
morpheme as the wh- of the relative clause. Also, conclusively, it is able to
indicate which words in an operand sentence can be asked about—namely,
those that can be pronouned with wh- in the relative clause (in effect, those
words that can be brought to the head of their sentence). It provides for
multiple questions (Who saw whom?) and various special forms such as
You'll go, won't you? (G331).
5.5.3. Adverbs on say: And, or, not, any, tense
There is a final utilization of performative say, which provides explanatory

derivations for important but peculiar grammatical constructions, but presents
problems—or as yet presents problems—for an orderly analysis of complex
sentences.
There is grammatical evidence that the performative I say carries certain
adverbial modifiers (secondary operators on it) which produce various
special words whose grammatical properties are peculiar and not otherwise
explicable. Such are the logical not, and, or, and certain exceptional English
forms such as any; cf. also the source of the demonstratives (5.3.4). In many
languages, the words for not, and, or have peculiar grammatical properties
(G387). For example, in English they, alone of all operators, cannot take the
form of verbs, and cannot carry tense either directly or on an associated be:
for He phoned and she wrote, there is no His phoning was and she wrote
comparable to He phoned because she wrote, His phoning was because she
wrote. Hence they are not ordinary operators, but neither do they fit any
other word category. They are also unique in language as being basic
operators of set theory and logic. In addition, not and the interrogative do
not preserve the selection of the words under them, whereas operators in
general preserve such selection: thus the stone spoke and I want the stone to
speak are rare and peculiar combinations, but The stone didn't speak and
Did the stone speak? are much less so. Instead of disregarding these
peculiarities and also the logic status of these words, we can to some extent
explain them by deriving the words from modifiers on the I say (e.g. for each
one in the source of the interrogative, 5.5.2, and in the source of any,
below). This can be expressed in artificial restructured sources, such as I say
negatively (or I deny, gainsay, contradict, disclaim, claim as false) that John is
here, reduced upon zeroing of I say (or -diet or claim) to John is not here-, I
say both (or jointly, or co-state) Mary is here, John is here, where / say both
operates on a pair of sentences and is reduced upon zeroing of I say or / state
to Both Mary is here and John is here with both then zeroable; also I say
either (or alternatively) Mary is here, John is here, reduced to Either Mary is
here or John is here with either then zeroable. (Possibly, also, the intonation
that marks a secondary sentence, written as semicolon or dash, could be
obtained from /say with an aside operating on a sentence-pair: e.g. Mary is
coming, She phoned, yielding Mary is coming; she phoned or Mary—she
phoned—is coming, ultimately Mary, who phoned, is coming.)
As to any: in Anyone can go the word means roughly ‘every’, whereas in 1
didn’t see anyone go or in Will anyone go? it means ‘even one’. Also, there is
the peculiarity that any rarely occurs with only a single operator. In spite of
the popularity of phrases like Anything goes, the more customary form has
two operators, as in Anything small will do, or Anyone who can goes there.
All these forms and distinctions can be obtained within the existing grammar
if we take any to be an adverbial form of one, in particular (one) by one
(which is roughly its etymological source) operating specifically on the say.
Then Anyone can do it is from, roughly, / say for each one that this one can
do it (hence ‘everyone can’); and / didn't see anyone do it is from I deny for
each one that / saw this one do it (hence ‘not even one did'); and Will anyone
do it? is from / ask for each one whether this one will do it (hence ‘will even
one do it?’). The selectional preference of any for two operators is thus in
order to relate the two predicates in respect to the one-by-one matching:
For each one if he can then he goes there, above (G327).3
Finally, one of the most important grammatical advantages of recognizing
the metalinguistic say operator is the explanation which it offers for the
tenses. If we consider a sentence such as A person who registered as late as
' There are also less elear cases of words whose meanings or positions can he explained by
saying that they operate on the performative say. One such is seen in hardly, scarcely, whose
occurrence in many sentences such as John hardly deserves it, John scarcely deserves it would be
more understandable if the hardly, scarcely referred to the saying of the sentence: that is, it
would be hard for the speaker to say that John deserves it.
tomorrow will still be able to vote on Tuesday, we see a known phenomenon,

in which the secondary operator is tensed to express its relation to the
primary operator as taking place before or after it; the source is A person
who registers as late as tomorrow before that (i.e. before voting) will still be
able to vote on Tuesday, with before that reduced to -ed on register (G265).
Similarly, if a verb operates on another, the lower is tensed as being before
or after the higher: Yesterday I heard that Mary will (or: would) come later
that day, from Yesterday I heard that Mary comes later that day, after my
hearing it, with my hearing it zeroed as repetition, and after reduced to will
(G269); in the now less usual form would, the lower operator adds to its
tense the tense of the higher operator. (If, under a particular higher verb,
the lower verb is usually future in respect to it, the future tense on that lower
verb can be zeroed: / request that he go from l request that he will go.)
We can now say that in the same way, the highest operator of a sentence,
i.e. the operator of a primary sentence, can be seen as tensed in respect to
the / say that operates on it (G267): I say: 7 will go' ox I say that l will go from
I say (that) I go; my going is after my saying it, reduced to I say (that) I go
which is after my saying it, then to I say (that) I go after, to I say (that) / will
go, to I will go (G268). In English, in many situations, when a lower verb is
tensed in respeci to a higher one, then any tense that the higher verb gets, in
its turn, from say is added to the lower tense: e.g. in the would above. The
source was / state my hearing yesterday, which is before my saying it—> l state
my hearing yesterday, before —> I state I heard yesterday —> I heard yesterday.
(Note that one may prefer to have before and after as adverbs on the higher
verb (including say) rather than after and before, respectively, as modifiers
on the lower verb: I say beforehand that I go —» / say that I will go.)
In this way, the ‘absolute’ tense of the highest operator (/ will go, I heard,
above) and the ‘relative’ tense of those below it (e.g. come)4 are seen to be
the same: the highest operator’s tense is relative to the I say above it. Thus
tense in general can be derived, paraphrastically and by attested reductions,
from sentences with time-order conjunctions but no tense. Specifically, it
can be obtained as a reduction of an adjoined statement giving the time
relation of two operators in the sentence, and finally the relation between
the time of speaking about an event and the time of the event spoken of.
This, rather than an external semantic instruction about tense in sentences
(G265).5
4 Note that since a relative clause is a secondary sentence, the lower operator registered
above receives -ed from being ‘before’ in respect to the primary operator able to vote. The
higher able to vote receives will from being ‘after’ in respect to the time of saying the sentence,
i.e. in respect to the say on the sentence.
5 There are many subtle grammatical considerations in the syntax of tense and of before and
after which support the convoluted derivation offered here. In addition, taking tense not as an
ad hoc primitive construction (outside the operator-argument framework) but as a derivation
from time or order conjunctions between a sentence and the statement that it is being said
leaves room for various non-temporal uses of tense, deriving from non-temporal uses of those
In sum: say and the like can carry certain adverbial modifiers (as false,
both, either, demand, querying, 0/something (5-3.4), o/ze by one, before,
a/ter, etc.) which can impose respectively not, and, or, imperative and
interrogative forms, deictic pronouns, a/zy, tense, etc., on the operand
sentence. Of these, both and either (the latter includes the interrogative
whether) require the operand to be a sequence of sentences rather than a
single sentence. Uniquely, the and and or differ from the two-sentence O00
operators of grammar in that they are not restricted to being binary, and in
that they are semantically associative and commutative: He left because she
phoned him because it was late means different things in different binary
groupings, but He went and she phoned and I stayed does not; also She stayed
because he left and He left because she stayed mean different things, but She
will stay or he will leave and He will leave or she will stay mean substantially
the same, though not in nuance. Also the disjuncts of the interrogative are
commutative. All of these adverbial modifiers relate the operand sentence
to the performative operators on them. In particular, not, both, either are
relevant to logic and set-theory because they deal with the saying of
sentences.
Most generally, some of the operators on sentences are metalinguistic in
that they can operate on the words of the sentence as phoneme-sequences
and not only on those words as indicators of particular referents. Of these,
one operator, say, or a small set {say, ask, etc.), with I or one as subject, has
performative effect and is therefore zeroable upon the saying of the operand
sentence.
As indicated at the beginning of 5.5.3, the adverbs on say require more
investigation before they can be adequately formulated in a theory of
language structure. The least of the problems is that the reconstructed
sources are clearly non-historical. Whereas the imperative and interrogative
can be seen to be related to their proposed sources, the sources for tenses
and and/or/not are convoluted beyond what would be said in natural
language. What remains relevant, however, is, first, that all of these (and
such words as any in English) have combinatorial properties and levels of
meaning that differ from the rest of language, and, second, that these
grammatical and semantic properties can be obtained, as suggested above,
from entities and relations of a syntactic theory that suffices for the rest of
the language. The serious problems arise in accounting for all the environ¬
ments in which these proposed adverbs on say occur and do not occur, and
above all in tracing how these derivations function together in the kinds of
complex sentences that many languages have.
ordering-conjunctions: e.g. using the past for ‘contrary to fact’, as in If he were here l would
know it; or using the future for reducing the speaker's responsibility for what he is saying, as in I
think what you want will he in the fourth aisle. These considerations are surveyed in Z. Harris,
Notes du cours de syntaxe, ed M. Gross, Editions du Seuil, Paris, 1976, p. 158.
6
On the Mathematics of Language
6.0. Why a mathematical approach?
As was noted in 1.2, there are reasons to expect language to have a

mathematical structure. One is that if new words entering language are to ht
into the existing system of word combinations, the existing words have to be
characterized by their combining-relation to each other rather than by any
individual properties: the new word may have a kind of meaning or of
phonemic composition different from the existing words, but it will have
to fit into one or another of the kinds of word combination which make
sentences. Another reason is that in the absence of an external meta¬
language (2.1), the elements of language can only be identified by their
departures from equiprobability, which is the kind of property that can
be studied mathematically (differently from meaning or phonemic com¬
position). A third reason is that the speakers of a language, who have limited
(though generalizable) public experiences of recognizing speech entities and
their combination, can understand an unbounded number of sentences
(more than are ever actually said), since there is no structurally relevant
specifiable upper bound to the length of a sentence: hence there must be
some constructive or recursive characterization of sentences.1
Ultimately, the question is whether language can be defined as one or
another set of entities (elements, or particular relations among elements)
closed under stated relations or operations. As noted in 1.2, this does not
1 The expectation of useful mathematical description of the data of language stems from
developments in logic and the foundations of mathematics during the first half of the twentieth
century. One main source was the growth of syntactic methods to analyze the structure of
formulas, as in Skolem normal form and Lowenheim’s theorem, and in the Polish School
of logic (as in the treatment of sentential calculus in J. Lukasiewicz, and the categorial grammar
of S. Lesniewski and later K. Ajdukiewicz, cf. ch. 1 n. 6), and in W. V. O. Quine's Mathematical
Logic (Norton, New York) of 1940. Another source is in the post-Cantor-paradoxes con¬
structivist views of L. E. J. Brouwer and the Intuitionist mathematicians, and in the specific
constructivist techniques of Emil Post and Kurt Godel, in recursive function theory, and from a
somewhat different direction in the Turing machine and automata theory. Cf. S. C. Kleene,
Introduction to Metamathematics, van Nostrand, New York, 1952, and Alonzo Church,
Introduction to Mathematical Logic, Princeton Univ. Press, Princeton, 1956. In linguistics, the
‘distributional’ (combinatorial) methods of Edward Sapir and Leonard Bloomfield were
hospitable to this approach. Cf. also Nelson Goodman, The Structure of Appearance, Bobbs
Merrill, New York, 1966.
mean the set of phoneme-sequences or word-sequences under concatena¬

tion, for language is only an unstated subset of the set of all the phoneme-
sequences or of all word-sequences. Rather, the search has to be conducted
on the basis of whatever characterizes the particular combinations that occur
in language. It has been found here that this can best be done not by listing
the combinations themselves (1.0) but by stating constraints on combination,
and that one can define new less-constrained higher-order elements either as
classes of more-constrained alternant elements or else as sequences of more-
constrained elements. As one does this, the listing of elements is increas¬
ingly replaced by stating relations among (the newer) elements. All this
brings the description to where mathematical formulation is more readily
usable: a hierarchy of (newer) elements with simpler and more regular
relations both among the higher classificatory elements and among the lower
elements which make up each class. It was found that almost everything in
grammar could be characterized in terms of three such relations. Two of
these are complex: the different phoneme-sequences which constitute the
same words (by morphophonemic variants or by syntactic reduction); and
the particular probabilities for those word co-occurrences whose probability
is higher than zero. The third is simpler: the partial ordering of words whose
co-occurrence probability in sentences is higher than zero, i.e. whose
combinations produce sentences. It is here that a mathematical formulation
was most immediately indicated.
Closer scrutiny of this partial ordering of words in a sentence showed that
it did not start with some particular list of words, but w as based entirely upon
a single relation among words in respect to their co-occurrence.2 The ability
of words to enter into all sentence-making combinations depends on nothing
except this relation among them, so that the set of combinations (sentences)
constitutes a mathematical object. The importance of this relation is that not
only does it create the set of possible sentences, but also that all further
grammatical events and mathematical structures in language (6.3) are
defined on resultants of this relation. When this linguistic structure is studied
as a case of applied mathematics, we have here not an example of applying
known formulas to calculations relevant to a given science, but an example
of finding particular mathematical structures holding for particular aspects
of the real world—again, holding precisely and not by over-generating and
then whittling down.
More generally, Chapters 3 and 4 show that the whole structure of syntax
is determined by a system of constraints each of which states a departure
from equiprobability in word co-occurrence. Syntax and the information it
expresses (or distinguishes) and engenders is thus characterized by a partially
2 This is called the autonomy of syntax, i.e. that no phonetic or semantic property of words is
considered in determining what constraints on word combination create sentences. The syntactic
structure nevertheless conforms to meaning-differences, but the phonemic structure of words
does not in general.
On the Mathematics of Language 147
ordered set of departures from equiprobability, something which is usefully

subject to mathematical investigation.
6.1 Language properties having mathematical relevance
There are various more or less directly observable universal properties of

language which make the data amenable to particular mathematical treat¬
ment.
(a) Discreteness. In the analysis of language, it turns out that the only
entities which have usable grammatical regularities are discrete. In particular,
phonemes are discrete elements, and words and sentences are specifiable
sequences of them. Although phonetically there may be a continuum of
difference among sounds, what assigns particular ranges of sound to particu¬
lar phonemes is a system of relative distinctions, which can be represented
by discrete marks or names (7.1). And although the successive sounds of an
utterance are in general a continuously changing sound, the sequence of
phonemes which represent them is identified by successive cuts in the
continuous utterance sound.3
Measurement of length does not have to be considered a property of
phonemes. The sound between successive cuts may be of different lengths,
and different phonemes may differ in length, but these lengths do not
distinguish or characterize the phonemes. In some languages some phonemes
have distinct long and short sound segments: the long ones can be con¬
sidered double-length occurrences of the phoneme, or simply two occur¬
rences of the phoneme. There are also certain long phonemic intonations
which extend over a sequence of words, e.g. in the interrogative type of
intonation and the imperative type; but each of these can be represented by
a single discrete entity, since the grammar does not distinguish continuously
differing grades of each of these intonations. Those linguistic phenomena
which are not representable by discrete elements may indeed carry informa¬
tion (e.g. intonations of hesitation, or of exaggerated matter-of-factness),
but they do not enter in any regular way into grammatical relations to other
elements (and are generally restricted to face-to-face situations, being not
precisely repeatable), and are thus not included in grammar.
(b) Pre-setting. The phonemic distinctions of a language, and the sequences
of phonemes that constitute words, have to be learned by users of the
language, and are hence pre-set in advance of use. When a person hears a
phoneme or a word, he need only hear enough to distinguish it from all
others that could occur in the given environment, and to know that it is the
‘same’ as a repetition of it.
3 Z. Harris, [Methods in] Structural Linguistics, Univ. of Chicago Press, Chicago, 1951,
pp. 26-33, 60-4, 70-1.
(c) Repetition, [f the hearer thereafter transmits what he has heard to

someone else, he pronounces his own rendition of the phonemes. It is this
that makes the transmission of utterances a repetition, somewhat as in the
reforming of pulses at gates in a computer. In contrast, if the hearer tries to
transmit to a third person a sound which has not been pre-set to him, he can
only approximate it in an imitation. 4
(d) Error compounding. The discreteness and pre-setting of the elements,
with the result that transmission is by repetition rather than imitation,
reduces by far the danger of error compounding in transmission. If hearers
transmit continuous data which they have received (e.g. intonations of
anger, or gestures), especially if those are not categorized in respect to pre¬
set elements, there is a considerable likelihood of errors being compounded
by succesive imprecisions in one direction, in imitating the signals at
successive transmissions. In contrast, when the signal is characterized by
discrete and pre-set elements, it can tolerate a considerable amount of noise
in transmission from one hearer to the next without producing error.
(e) Arbitrary signals. The fact that the grammatical elements must be
pre-set makes possible what has been called the arbitrary character of
linguistic signals, i.e. the fact that the sounds out of which words are
composed do not in general indicate the meanings of the words, and differ
arbitrarily as among languages. A relation between word sounds and word
meanings would be useful primarily if it eliminated the need for language
users to learn a fixed and jointly accepted (pre-set) stock of elements. But
the fixed stock, with phonemic distinctions, is precisely what is required for
transmissibility without error compounding.
Also, a regular relation between sounds and meanings would limit the
possible word meanings to what could be obtained out of short sequences of
20 or so phoneme meanings (if such meanings existed), something which
would be grievously inadequate for language. Furthermore, if the linguistic
signals were not arbitrary, but inherently related to their sound composition
or their meanings, they could not in many respects be identically pre-set for
all participants. For if the meanings of sound-sequences were not learned,
but left for each speaker and hearer to judge on the basis of his individual
experience, the speaker and hearer would not necessarily be in agreement,
since experience varies in detail for different individuals.
Instead of going by the phonemes individually, the hearer hears certain
phoneme-sequences (or broken sequences, or replacements) as words, and
then judges from non-language context, and also from the range of combina¬
tions in which each word or morpheme appears, what can be the meaning-
range associated with each word or morpheme. The discrete and pre-set
character of linguistic signals therefore leaves them arbitrary, and indeed
4 For the distinction between such repetition and imitation, cf. Kurt Goldstein, Language
and Language Disturbances, Grune and Stratton, New York, 1948, pp. 71, 103.
requires that there be no general correlation between sound and meaning.

This does not preclude limited cases of onomatopoeia, i.e. of meanings in
particular words being attached to particular sounds in those words.
The fact that the grammatical elements (characteristic sounds, and the
collecting of these to form morphemes or words) are arbitrary symbols, which
are related to meanings only by convention, makes the word-sequences
available for study as mathematical sets; for anything which we establish
about language as composed of these sequences of elements will hold for any
physically different set of elements (e.g. letters) onto which the original set
can be mapped.
(f) Linearity. Speech, and following it writing, is in a rough way linearly
ordered, although in some languages certain phonemes may be pronounced
simultaneously (e.g. vowels and nasalizations); and intonations with phon¬
emic or morphemic status are in general simultaneous with words (e.g.
contrastive stress) or with sentences (e.g. the interrogative intonation).
Certain elements or sequences are nested within others, but even then we
can consider the initial segments of these to be linearly ordered (e.g. the
component sentences So he claimed and It was not loaded in respect to He
believed that.... in He believed—so he claimed—that it was not loaded).
The linearity of physical elements is not to be confused with linear
orderings in mathematically defined sets for the description of language. In
the latter case, we are describing language in terms of a set which is closed
under some operation; and we ask whether the set is linearly ordered. This
turns out to be not quite the case, although we can come close to this result.
In the former case, of the physical elements, we can only say that language
events are a proper part of the set A of linear orderings of words (or
morphemes, or phonemes). Going beyond the phonemic composition of
words: as has been noted, the set A that is closed under the word-concatenation
operation is not natural language, and we need another relation between
words to characterize what subset of A a natural language is. But it remains
true that in each language event the phonemes, words, and sentences are
linearly ordered (at least in respect to their initial segments), or can be
readily mapped homomorphically onto a linear ordering. In addition, it is
seen in 9.3 that repetitive structures within the linear ordering can create a
double array.
(g) Sub-sequences. The sequences of phonemes and of words that make
utterances are segmentable into sub-sequences: the phoneme-sequence of a
word into syllables, the word-sequence in a sentence into (transformed)
component sentences (e.g. clauses). Given a large set of sequences, it is
found that each element—phoneme or word—has a particular frequency of
occurring with other elements, not in terms of fixed transitional probabilities
but in more complicated dependences within a sub-sequence which can be
captured in stochastic processes and in local and global partial orderings on
the sequence.
(h) Contiguity. It was seen in 2.5.4 that all relations and operations in
language are on contiguous entities. This is unavoidable, since language has
no prior defined space in which its phoneme- or word-occurrences could be
located. This condition of contiguity does not preclude successive segments
of an operator from being placed next to different (but already present)
parts of its operand (as in the case of discontinuous morphemes, 2.5.4), or
between parts of its operand as in the nested sentence example of (f) above.
(i) Finite length. There is nothing to indicate that utterances, i.e. occur¬
rences of speaking or writing, must in principle be finite in length. However,
the sentences, whether these are defined structurally by some relation
among segments of the utterance, or stochastically by sequential regularities
of the segments (7.7), are necessarily finite in length; for until a segment of
an utterance is finished we cannot be sure that it has satisfied throughout its
length the conditions for being a sentence. In actuality, sentences are
reasonably short except for tours de force of various kinds.
(j) Number of sentences. For a given number of phonemes, and also for a
given number of words, the set of sequences is recursively denumerable. But
for the subset of these sequences that constitutes a language, the question as
to their number is more complex, and depends on how we characterize the
criteria for their being in the subset/ In any case, the set of possible
sentences as fixed in the grammar is not known to be finite, since to every
sentence one can in principle add without limit some further clause.
(k) Finitary grammar. Aside from the non-finiteness and recursivity of
language as a defined object, one should also consider the finiteness of the
speakers and hearers. The phonemes and vocabulary, and the ways of
combining words into sentences, all have to be used by the speakers of a
language in a way that enables them to understand each other. The users,
and their experiences with the language, are all finite. Hence it must be
possible to characterize a language by a finitary grammar: a finite stock of
elements and also of rules (constraints) of combination, with at least some of
the rules being recursively applicable in making a sentence.
We now consider the number of event types (as against tokens) in a
language, i.e. of occurrences which are linguistically distinguishable from
each other. The number of items that have to be described is affected by the
fact that languages change. In particular, new words and new combinations
enter the language, and some go out of use. Over a very long time, with
successive generations of speakers, the vocabulary of the changing language
may be large to an extent that we cannot judge. However, there are no doubt
limits—barely investigated so far—on the growth of the vocabulary, and on
the rate of change in the structure of language. The number of elements and
Whether a subset of a denumerably infinite set is itself known to be denumerable depends

on the availability of methods of enumerating the members of the subset. This is a crucial
difficulty in characterizing language.
of rules of combination available at any one time is not very large. It will be
seen in 12.4.5 that in respect to an appropriate theory of language one can in
principle count everything that a user has to know in using a language.
(1) Computability. It follows, from the discreteness, linearity, contiguity,
and finitary grammar, that the structure of sentences is computable. That
is, finite algorithms can be created that can synthesize, and much more
importantly analyze, any sentence of the language. Although constructing
such a language, especially for analysis (recognition), is a daunting task, it
has not proved impracticable (ch. 4 n. 1).
6.2. Increasing the amenability to

mathematical description
It was seen in 6.1 that many properties of language data are amenable, with
various exceptions, to description in mathematical terms. This makes it
desirable to regularize the properties not only for efficiency and elegance of
theory, but also in order to reach a description of language in terms of such
elements and relations as permit interesting decompositions, mappings, and
constructions that are closed over relevant domains. The criterion is that any
mathematical structure which is posited should apply in a specified way to
observables in the data and to their stated relations, and not to abstract
concepts whose precise relation to the regularities of the data has yet to be
established.
Such regularization means finding effective ways to treat each ‘exception’
as a non ad hoc special case of the regular relation. This may require a
modified formulation of the relation. For example, not all words need to be
pre-set, i.e. known to the hearer, in order to be understood: a sentence can
contain an unknown word whose boundaries are given by its neighbors and
whose meaning can be guessed from the rest of the sentence. Or the
regularization may be accomplished by finding special conditions that
together create the apparent exception. Thus in linearization of phonemes,
the nasalized vowel in the French don (‘gift’)—overtly a single sound—is
linearized as vowel plus n (in the absence of following vowel), in order to
agree with the vowel plus n sequence when the word appears with following
vowel, as in donner (‘to give’). Also, in defining words (or morphemes) as
being each a sequence of contiguous phonemes, allowance has to be made
for cases where the phonemes are not contiguous. An extreme case is in
what is called ‘agreement’ (‘government’, ‘concord’) in grammar: e.g.
plural-agreement in The stars rise as against The star rises. Instead of
defining here an ad hoc grammatical relation between the verb and its
subject, additional to the partial-ordering relation which creates sentences,
we can say that in English the plural morpheme (and also and) on the subject
of a sentence has a discontinuous segment located on the verb (consisting of

dropping the -s on it): The stars rise, The sun and moon rise (G37, 169,263).
Occurrences of the plural morpheme when not in subject position, or with
other tenses, do not have this detached segment (The star emits photons, The
star rose, The stars rose). This then becomes a correction on the contiguity of
phonemes in a morpheme (the plural morpheme), but avoids any special
subject-verb relation.
In addition to cases in which exceptions can be treated as special or
marginal cases, there are some properties of language data which present
essential difficulties for mathematical treatment. Even after the welter of
sentences is made more manageable by recognizing the different likelihoods
of word-choices in operator-argument relations and the existence of reduc¬
tion in sentences, we find that the co-occurrence likelihoods differ for each
word in a way that limits classifiability, and we find that the criteria for
reducibility are in many cases so complex as to preclude predictability of
reduction. As a result, the set of sentences is not well defined: there are a
great many sentences which are marginal either in that speakers disagree as
to whether they are sentences (e.g. A man who I met called me) or in that
speakers agree that they are borderline (e.g. The baby took a crawl to the
chair).
All of these difficulties involve encapsulated, although in some cases
major, aspects of language (e.g. likelihoods and reductions), or special cases
of general properties. It is therefore possible to deal with them within a
precise and regular structure, and this not by ad hoc deviations but by
arguably justifiable additions to the structural relations or to their domains.
A major step is to try to reach independence of the elements. This includes
separating out, in a given occurrence, features that appear elsewhere in
different combinations. For example, consider pairs of words which have
the same selection except in respect to time words. The fact that the
difference in some neighboring time-adverbs between is and was is the same
as between walks and walked (and so in many other verb-pairs) shows that is
and was contain one element in common (as most of the other verb-pairs
do), and that was contains a variant of the -ed (past-tense) morpheme seen in
the verb-pairs. And comparison of He is, I am with He walks, l walk, etc.,
shows that am is a post-/ obligatory variant of is. Such analyses lead to
systemic regularization only if they are based on full consideration of the
‘behavior’—i.e. the relative occurrence—of the entities. Thus given the co¬
occurrence of plural markers on both subject and verb (above), we say that
the morpheme occurs on the first argument (the subject) and has a detached
segment on the verb, rather than the other way around, because its occurrence
on arguments can be independent of its appearance on the verb (e.g. when
the argument is second rather than first, as for photon in The star emits
photons), whereas its appearance on the verb is dependent on its occurrence
on the subject.
In general, then, expressing the properties of entities by relations among

them is a precondition tor mathematical description, e.g. in seeking a least
grammar with most regular domains for the operations of the system. The
regularity of domains is a matter not only of avoiding small ad hoc domains
as in the examples of 1.1.1—2 but also of large-scale regularizations. One
such is the avoidance of any distinction between operators on a sentence and
operators on a verb, which requires evidence that what seemed to be
operators on a verb are derivable from operators on a sentence (with zeroed
arguments in the lower sentence). Another large regularization of domains
is achieved by separating (a) the permanent partition of words into argument-
requirement classes (A, On, Ono, O0, etc.) from (b) the changeable graded
likelihoods of operator-argument selection within (a). Isolating (a), which
constitutes the mathematical object underlying language, from (b) leads to
many advantages in structure-description and in meaning-correlation. Of
course, all the regularizations, great and small, have to be explicitly justified
in terms of the linguistic data.
A more fundamental step toward mathematical treatment of language is
replacing the observable combinations of entities by the constraints that
limit the combinations to some generalization of those that are observed,
and by the transformations (including reductions) that replace certain
combinations by others. These replacements are important because, as
noted, mathematical combinatorics on the letters and words as directly
observed is inapplicable to language, since the latter essentially falls short of
all combinations.
These preparings of language for mathematical treatment have the addi¬
tional effect of making the structure of language more similar to that of
mathematics and logic. Thus, the not, and, or, and in a different way the
quantifiers, which are special here in that they operate both in language and
in mathematics-logic, are seen here to operate on sentences (rather than on
words) in a manner closer to mathematics and logic than appears on the
surface (G244, 321,387). And the imperative and the interrogative, which
are absent from the syntax of mathematics and logic, are found to be
derivable in language from assertions (G331, 351). Furthermore, the structure
of the metalanguage (as a science sublanguage) is closer to the syntax of
mathematics than is the structure of natural language as a whole (10.2).
6.3. Mathematical structures in language
The properties of language surveyed in 6.1, and the processings mentioned

in 6.2 and in Chapter 2, make it possible to state various mathematic¬
ally definable operations and various sets closed under mathematically
definable relations, which together characterize almost all the grammatic¬
ally and semantically important universal structures of language. Beyond
this, individual languages and language families have special structures

which are less amenable to mathematical characterization: e.g. syllable
types, morphological systems, morphophonemic systems such as conjuga¬
tions and declensions, genders, verb ‘aspect’ (of duration), etc. The major
operations and sets will be noted here; brief descriptions of the most
important structures are given in 7.3, 7.6-8, 8.1.5-6, 8.3, 9.3.
(a) First, we consider those operations which do not involve the syntactic
analysis presented in Chapter 3. Phonemes are a partition of sounds (after
minor adjustments for independence of sound-occurrences) based on a set
of distinctions which is constant over the population even if individual
pronunciations differ (7.1-2). There are then stochastic procedures on
utterances for finding word (and morpheme) boundaries as sub-sequences
of phonemes (7.3), and sentence boundaries as sub-sequences of words
(7.7) . These arise because there are far fewer distinct words than possible
short sequences of phonemes, and far fewer n-word sentences than possible
sequences of n words. It is possible to define a hierarchy of rewrite rules in
such a way that an elementary sentence ((g) below) is found inside each
given sentence by a system of recursive equations that expand the parts of
the elementary sentence into corresponding parts of the given sentence
(7.8) . And it is possible to define a recursive expansion of the elementary
sentence (as a single entity) into the given sentence, by defining the structure
of a type of ‘center’ word-sequence and of a number of types of ‘adjunct’
word-sequences, such that the adjoining of any sequence, as indicated by its
definition, to any other sequence A creates a new sequence of the same type
as A (7.6). To the extent that each word can be marked by what are its
possible neighbors in the segments of a sentence, it is possible to use a
cycling free group cancellation to test whether a word-sequence is a sentence
of the language (7.8).
(b) Second, the set of word-sequences under dependence on dependence
constitutes a mathematical object. We start with the fact that not all
combinations of phonemes make words or sentences, as follows from the
effectiveness of the stochastic procedures above. We now consider for each
given word the set of words which have frequency zero of occurring with it,
in a regular way, in any sentence (or utterance) of the language, including
the shortest, in any corpus of utterances no matter how large.6 The relation
of the given word to every word in the complement to the above set, i.e. of
every word that has positive frequency no matter how small of occurring
with it in a regular way, is called dependence (of operator word on argument
words). For mathematical interest, the crucial step is that the dependence is
not on a fixed list of neighboring words but on the dependence properties of
the neighboring words in turn (3.1): eat can be the (‘first-level’) operator on
In a regular way here means in the same relative position in all sentences, including the
shortest, after correcting for any reductions (3.3) and linearization differences (3.4.2-4).
(and only on) any pair of words which in themselves do not depend on any
words (equivalently, that depend circularly only on the first-level operators),
namely two ‘zero-level arguments’ N\ ‘second-level’ operators (e.g. because)
act on (and only on) any pair of words that in turn depend on other words.
Every word which depends on any stated dependence properties can form a
sentence, by combining with words whose dependence properties satisfy it.
Every sentence is a sequence which satisfies the dependence-on-dependence
of every word in it; and every such sequence is grammatically a possible
sentence, even if nonsensical (e.g. Vacuum chides infants) or artificially
complex (e.g. the proposed restructured source for tense, as in / state his
leaving; his leaving is before my stating it as source for He left).
As noted previously, the dependence relation creates the meaning ‘pre¬
dication’ (11.2), or carries that interpretation, which holds for every system
having this dependence, such as mathematics, even if the system lacks the
other features of language (e.g. likelihood differences). The predicational
meaning does not require any of the other features of language, nor does it
require the particular physical content (sound, ink marks) or meanings of
the words involved. The dependence on dependence is defined on arbitrary
objects, not necessarily words at all; and the abstract system it creates is thus
a system of categories.
(c) This mathematical object is not the whole of language, but it underlies
language in that the other sets and relations which are essential to language
are defined on its products. To obtain the structure of natural language we
need two additional relations: the differences in frequencies of co-occurrence,
and reductions in phonemic content.
The distinction between zero and non-zero frequency of combination, in
terms of which the dependence was defined, leaves open the question of
randomness, or stable or variable differences, in frequencies of each word
relative to the various words on whose set it is dependent. What gives
language its unique informational capacity is that there are indeed such
differences and that they are roughly stable, although marginally change¬
able over time; and they correlate with—and to some extent create—the
individual meanings of each word involved (3.2). These stable differences
can be expressed as inequalities, or even a grading of likelihood (or
probability), of each operator in respect to its possible arguments (and vice
versa).
The reductions are constraints whereby word-occurrences in a given
sentence which have highest likelihood (or otherwise no informational
contribution) in respect to their argument or operator there can be reduced
in phonemic content (3.3). The reductions are relations between sentences
because every occurrence of an operator on an argument constitutes a
sentence, and the reduction changes that into a physically (phonemically)
different sentence. There may also be some stated other grammatical
transformations of sentences, not based on these considerations. And there

may be morphophonemic variants of particular words, by which the word
has different phonemic content in different operator-argument or linear
environments (7.3). In meaning, all of these changes are paraphrastic; they
leave the substantive information of the sentence unaltered.
(d) The dependence-on-dependence relation imposes a partial ordering
(of just three levels, in most cases) on the words of a language in respect to
their co-occurrences in the set of all sentences; some (second-level) words
depend on at least one (first-level or second-level) word that depends on
something not null, with other (first-level) words depending on one or more
(zero-level) words that depend only on null (equivalently, only on the first-
level words). This ordering induces an equivalence relation on the words of
the language, which partitions them into second-level, first-level, and zero-
level. The dependence also imposes on the word-occurrences in each
sentence a tree structure, specifically a join (union or cup) semilattice:7 an
ordering in which for any two word-occurrences there is a word-occurrence
in the sentence which is their least upper bound (l.u.b., union). For all a, f>,
the greatest lower bound of a, b. if defined, is either a or b. This means that
no one word-occurrence can be an argument of two different operator-
occurrences.8 Thus John phoned and came is derived from John phoned and
John came {John in the first example is not simultaneously the argument of
both phoned and came); l know John came is not from / know John and John
came but from I know on John came; I asked John to come is from I asked
John for John to come (with zeroing of for John); The man I wrote answered
is from The man whom I wrote answered from The man answered; I wrote the
man.
In the semilattice, for word-occurrences a. b, c, with b>a, in the sentence,
b covers a (i.e. is the immediate operator on a) if there is no c (cJ=a, c±b)
such that b>c>a; we call a the argument of b. Two or more word-
occurrences which are covered by the same word-occurrence as operator are
co-arguments (of each other) under that operator. Each set of co-arguments
is linearly ordered; and this ordering is semantically relevant, distinguishing
the subject and object of an operator. The word semilattices are the
semilattices of the base sentences.
Equivalently to the semilattice, the word-occurrences of a sentence
constitute a set with one operation: to each «, b in the set there is associated
an element a\!b (corresponding to the l.u.b. of a, b). The operation is
7 The common uses of the term ‘tree’ include cases that are not applicable to language. For
example, in genealogical trees, the relation of two cousins to their six instead of eight
grandparents does not occur in the dependence relations among words of a sentence: as noted
below, a word-occurrence cannot be the argument of two different other word-occurrences.
The sameness statement (5.3.3) presents a problem here, because it refers to two words
which are already the argument of other operators. The problem is avoided by a solution
proposed by Danuta Hiz, which excepts the metalinguistic statements from the lattice of the
sentence (8.3).
idempotent: a\!a=a\ associative: for all c, («V/>)Vc = r/V(M/c); and com¬

mutative: a\/b=b\/a. This last means that the semilattice is unchanged by
different linearizations of the base partial order, in particular by permutation
(3.4.3).
We now consider the reduced sentences which are derived from the base
by reducing particular words in the operator-argument structure that makes
those sentences. Here a related, ‘word-and-reduction’, semilattice of a
sentence is formed by ordering the reduction instructions as points in respect
to the word semilattice.This ordering is available because the description
of reductions holds that they are made as soon as the conditions for them are
satisfied in the constructing of the sentence. For example, in I like teaching,
from I like my teaching something, the zeroing of my is carried out right after
like operated on the pair /, teach: this, since teach brought with it the pair I
(my), something (the latter having been zeroed as being an indefinite), and
like permits the zeroing of the subject of the operator under it if it is the same
as its own subject. This semilattice of a sentence (e.g. in 8.2.2, 8.3) is a
precise description of it as an ordering of the entry of words and reductions
into the making of the sentence—and of each component sentence of it (as
sub-semilattices).
(e) The denumerability of the sequences of phonemes of a language, and
of the words in a language, cannot be applied to the particular subset of them
that constitutes the utterances of that language. However, given the partial
ordering in sentences, one can say which parts are iterable without specifiable
limit, thus creating for the theoretical structure the seemingly unbounded
set of possible sentences out of a finite vocabulary.
(f) In the word-sequence of a sentence, concatenation is semantically
non-associative and ambiguous: The yellow and green cards can be derived
both from The cards which are yellow and green and The cards which are
yellow and the cards which are green. In contrast, the successive entries of
operators, which (together with reductions) describe the same sentences as
concatenation would, are associative and non-ambiguous. Mappings and
operations on sets of sentences can therefore be more conveniently carried
out on the entries, and in particular on the operators, in these sentences,
than on the overt word-sequence of the sentences.
We consider first the unary second-level operators, i.e. those whose
argument-requirement dependence includes precisely one operator (with
possible zero-level arguments too). The set of these is a semigroup with
successive application (i.e. next later entry in the making of the sentence) as
y There may be one or more simultaneous (independent) reductions located at a particular

point, i.e. immediately above a particular point in the word semilattice. In case more than one
reduction is located above the highest operator (the maximum, or universal, point) of the word
semilattice, we obtain correspondingly more than one maximal point. It is therefore necessary
to define an artificial new universal point, the sentence intonation of that sentence, which
provides the l.u.b. for the more than one maximal reduction points if such exist.
operation; one can consider the null operator as identity. The object here is
not the operator itself, but the application of it to its argument. For present
purposes, given A applied to (operating on) an argument X, and then B
applied to A (with its argument), we consider B applied to A without
argument as a complex operator BA applied to the argument X. A product
of two operator-successions in the semilattice is itself an operator-succession;
the multiplication is associative.
The binary second-level operators, i.e. those whose arguments include
precisely two operators, form a set of binary compositions on the set of
sentences. In the base set, each binary second-level operator can act on
every pair of sentences, although its likelihood of occurrence is higher on
sentence-pairs which contain in their base form a word in common. In the
whole set of sentences, certain reduced forms (e.g. question) do not appear
under certain binary operators: not in I am late because will you go?', this
may be resolved by defining here some required transformations which
would replace these sentences by their non-prohibited source, in this case I
am late because l am asking you if you will go. Products of these binaries are
in general not associative in their semantic effect, and correspondingly in the
history of making the resultant sentence. Here the structure of the base
semilattice, which is contained in the word-and-reduction semilattice of the
resultant sentence, is crucial.
(g) The most important algebraic structures in the set S of sentences are
those which arise from equivalence relations in S in respect to the particular
base (operator-argument) semilattice in each sentence, and in respect to the
highest operator (the upper bound of all words in the semilattice), or to
the word-and-reduction semilattice. These equivalence relations identify
the informational sublanguage (the base) and the grammatical transforma¬
tions, as will be seen below.
We note first that the resultant of every operator is a sentence. Every
unary second-level operator acts on a sentence (and possibly also an N) to
make a further sentence; and every binary second-level operator acts on two
sentences (and possibly an N) to make a sentence. Every reduction acts on a
sentence to make a (changed) sentence. All of these, in acting on a sentence,
preserve the inequalities of operator-argument likelihoods in the operand
sentences (except in statable circumstances). The unary second-level operators
are a set of transformations on the set of sentences: each maps the whole set
S of sentences into itself (specifically, onto a subset of sentences which have
that second-level operator as their latest entry, i.e. their universal point);
and the binaries map SxS into S. The reductions and other transformations
are a set of partial transformations on S, each mapping a subset of S
(sentences whose last entries are a particular low-information operator-
argument pair) onto another subset of S (sentences containing the reduction
on a member of that pair).
The preservation of inequalities of likelihood under transformations of

the set of sentences, i.e. under the second-level operators and under the
reductions, is of great importance. Without it, there would be no semantic
connection between a sentence and its occurrences under further operators
or reductions. The second-level operators and transformations preserve the
likelihood inequalities and the meanings in their operand sentences, although
with a reasonable number of specified exceptions. The second-level operators
also add their own meanings and likelihoods in respect to their argument.
However, some metasentential operators such as deny (the source of not)
and ask can change the likelihoods of operator-argument pairs under them;
hence the inequalities in the resultant sentences (with their new higher
operator) need not be the same as in the corresponding operand sentences.
As to the reductions, these preserve with few exceptions the inequalities of
likelihood and the meaning of their operand sentences. In some cases, a
reduction may raise (or lower) likelihoods of occurrence in the resultant
sentences, but for the most part equally on all its operand sentences, or only
stylistically and idiomatically.
We consider the set S', where each word-sequence which is grammatically
ambiguous in n different ways is considered to be a case of n different
sentences, each with its own base semilattice. S' is a semigroup under the
binary operator and: for any two sentences A, B we have A and B as a new
sentence C. (This, after adjustments are made for A, B pairs which do not
take and.)
We present now a structure which isolates the minimal subset of the set of
sentences as a residue of the second-level operators and reductional trans¬
formations. It has little importance when the great bulk of transformations
are products of such simple reductions as have been established for English,
and when the minimal (elementary) sentences can be characterized, as in
Chapter 3, as the resultants of first-level operators on zero-level arguments.
However, in a language in which we do not have so clear a picture of the
structure of the set of transformations or of the set of base sentences, a way
of identifying the elementary sentences is useful; and it indeed was useful in
the early stages of the development of the transformational theory (8.1). To
obtain this structure, we form the set S" of sentences, which consists of all the
sentences of S' that contain no 00(, (including and, or). We then take an
equivalence relation in S", whereby two sentences are in the same equivalence
class if they exhibit the presence of the same succession of unary operators
and the same orderings of particular reductions. In the natural mapping of S'
onto its quotient set E of these equivalence classes, the kernel of the
mapping, i.e. all the sentences which are mapped onto the identity of £,
contains only the elementary sentences.
A different and more important structure is obtained if in the set S' we
take an equivalence relation by which two sentences are in the same
equivalence class if they have the same ordered word-entries (i.e. the same
operator-argument semilattice). Since almost all reductions are optional,
each equivalence class contains (with possibly certain adjustments) precisely
one reduction-less sentence. The set of these is the base set (which includes
the elementary set), from which the other sentences are derived by reductions.
The base set is closed under the word-entry operation: any word-sequence
satisfying this form is such a sentence, and the composition of two sentences
of the base under and is again a sentence of the base. Hence the base is a
sublanguage (10.1).
(h) The sentences in the language or in the base cannot be listed: they are
constructed as sequences of words in Oin On, etc., or of proper subsets of
these, the subset being due to extreme likelihood properties or to reductions.
Each reduction is defined on a particular domain of words—all the words of
On, or of O0, etc., or else a proper part of one of these classes; it cannot be
defined on more than a whole argument-requirement class. Hence if there
are successive reductions on some word class in a given base structure, the
domains of the successive reductions are monotonically decreasing. Hence
the base set of unreduced sentences is the most unrestricted in S, i.e. is least
limited in respect to particular subsets of words.
The distinction of dependence, likelihood, and reduction in the construc¬
tion of sentences makes possible precise characterization of several other
grammatical entities or relations. For example, one can obtain in an orderly
way all the component sentences of a given sentence. One can present the
unambiguous forms of each ambiguous sentence: either the base form, or
the given form associated with distinguished reductions or component
sentences, sufficient to separate one meaning of the ambiguous sentence
from another. One can design a procedure for translating between two
languages by decomposing a sentence (via its reductions) into operators and
their arguments, thereafter translating these (a far simpler task than trans¬
lating the original sentence), and then carrying out equivalent reductions, if
desired, in the new language. Alternatively, one can create a common
grammar of all that is identical or similar in the two languages, and use it as
an intermediate translating or grammar-describing station.10
111 Such constructions are presented in Z. Harris, Mathematical Structures of Language,

Interscience Tracts in Pure and Applied Mathematics 21, Wiley-Interscience, New York, 1968.
Ill
Grammatical Analysis
*.
7
Analysis of Sentences
7.0. Classes and sequences
The theory of syntax presented in the preceding chapters has words as its
primitive elements. Since one might hesitate at a theory whose primitives
are undefined or established only circularly, it is necessary to consider
whether these words are determinable by objective and effective methods,
independently of the syntactic theory. To do so we have to see how linguistic
methods have segmented speech to obtain phonemes as classes of sound
distinctions, words as sequences of phonemes, and sentences as sequences
of words.
Traditional grammar was an informal description of the shape and meaning
of words and sentences. In contrast, descriptive and structural linguistics
recorded the shape of successively larger segments of sentences. The search
for coherent and regular analyses of sentences in terms of their segments was
satisfied by successive steps of classification and sequencing. At various
stages, the sets of entities obtained were partitioned into equivalence
classes. The set of equivalence classes then constituted a new set of 'higher'
entities whose behavior in respect to the sentence-segments was far more
regular than that of the earlier entities. Thus: the phonemically distinct
sounds are collected (as complementary variants, or free variants) into
phonemes (7.2); the word-size segments are collected as complementary or
free variants into words (‘free morphemes’) or bound morphemes (7.4); and
words are collected, by similarity of environments, into word classes (7.5).
Particularly constrained sequences of the entities that have been obtained
are defined as constituting new ‘higher’ entities, which have new regularities
in respect to each other. These are phoneme-sequences constituting word-
size segments (which have the status of words or bound morphemes in their
sentences, 7.3), word-class sequences constituting either constituents or else
strings (by two different styles of grammar, 7.6), and string sequences (or
constituent sequences) constituting a sentence as a regular segment of
utterances (7.8). In certain important cases the sequences can be characterized
by stochastic (hereditary) processes that proceed through them, stating for
each first n members of a sequence what are the possibilities for the n+ 1th
position in it. Such methods can define certain kinds of sequences as
constituting next-higher entities (words, sentences, 7.3, 7).
164 Grammatical Analysis
It will be seen that all the entities can be identified by objective repeatable
procedures, and defined on the basis of the relations in which they participate,
without appeal to other properties. It is of interest that all the major
segmentations are recognizable in writing, in particular in alphabetic writing.
Phonemes correspond roughly to letters of the alphabet. Words are marked
by space, and bound morphemes can be considered word forms in which the
bounding-space has been zeroed. Virtually all punctuation will be seen to
mark the boundaries of sentences; the non-sentence segments marked by
comma and dash will be seen in 8.1 to be reduced forms of sentences. But
many sentences, reduced or not, which are contained in larger sentences are
not bounded by punctuation: e.g. small in A small stone fell. We sleep in He
works while we sleep.
For the methodological considerations of 2.3, it may be noted that at each
stage the classifications allow for simpler statements of the sequencing
constraints (on the domain of the equivalence classes) than would have
applied to the domain of the entities themselves, before the partitions. And
in the case of special problems, such as the exceptions and special forms so
common in language, the peculiarities are moved down from the domain of
the sequencing relations to the domain of the partitions where they can be
treated much more easily.
7.1. Phonemic distinctions
The use of language takes place not only in speech but also in writing and
thinking; but these are in different ways secondary to speech, and in
particular writing is relatively recent. A search for the ultimate minimal
elements of language, therefore, necessarily turns to the sounds and sound-
features that play a role in speech. Phonemic analysis holds that in every
language certain sets of sounds, called phonemes, are distinguished from
others by users of the language, with each language having a fixed set of
phonemes.1 The existence of hearers' linguistic distinction of sets of sounds,
and the actual list of sounds thus distinguished in a language, can be seen in
the following simple test with users of the language.
We begin with two speakers of the language, and two utterances, prefer¬
ably short and preferably similar to each other, e.g. heart (an organ) and hart
(a deer), or heart and hearth (a fireplace), or economics with initial e and
with initial iy. We ask one speaker to pronounce randomly intermixed
repetitions of the two words—the organ and the deer, or the organ and the
fireplace—and ask the other, as hearer, to guess each time which word (or
1 Sec chiefly Ferdinand dc Saussure, Cours de linguistique generate. Paris, 1910; Edward
Sapir, ‘Sound Patterns in Language', Language 1 (1925) 37-51, reprinted in David G.
Mandelbaum, Selected Writings of Edward Sapir, Univ. of California Press, 1958, pp. 33-45;
Leonard Bloomfield, Language, Holt, New York, 1933.
Analysis of Sentences 165
pronunciation) was being said. It is found that the hearer’s guesses are
correct about 50 per cent of the time for some pairs of words, such as heart-
hart, and about 100 per cent for other pairs such as heart—hearth, or for the
two pronunciations of economics-, and the results are much the same for any
users of the language. In the cases where the hearer’s guesses are correct
close to 100 per cent of the time we say there is a phonemic distinction
between the two words (heart, hearth, ox the two pronunciations of economics);
otherwise not. In some cases the results of this pair test are problematic, and
the decision as to whether a phonemic distinction holds is adjusted on the
grounds of later grammatical considerations. But the direct results of this
pair test provide a starting-point, a first approximation to the set of ultimate
elements adequate for characterization of language.
The ability to distinguish utterances in the pair test does not depend on
any recognition or agreement by speaker and hearer in respect to the
meaning of the utterances, or even just whether the meanings are the same
or not-—the more so as a single utterance may have more than one
meaning, and there are cases of homonyms (e.g. heart and hart), which are
considered different words with different meanings, but are phonemically
identical; also, the two pronunciations of a single word, economics (with
initial e or else iy sounds), are phonemically different and are distinguished
in the pair test. Rather, then, the distinguishability depends upon, and
illustrates, the behavioral fact that when a person pronounces a word he is
not imitating its sound but repeating it (ch. 3 n. 4), that is, saying his own
version of the phonemic distinctions which he has heard. The role of
repetition as against imitation will be discussed in Chapter 12, but the
clustering of pair-test responses near 50 per cent and 100 per cent shows that
utterance-recognition is based on repetition: an utterance either is or is not a
repetition of another. The ability to recognize what is a repetition means
that the phonemic distinctions of a language are pre-set by convention. They
differ from language to language but can be acquired—in many cases
perfectly—when a person learns a new language. Even a predisposition to
recognize phonemic distinction in the abstract, or repetition as a relation
between behaviors, need not be assumed as an inborn or a priori preparation
for a person’s being able to recognize phonemic distinctions when he meets
them. It may be simply that various pronouncings are maintained at such
distances or contrasts from other pronouncings that a person learns to
maintain the distance or contrast even if his own pronunciation (or that of a
new person he hears) is somewhat different.2 3 This is an unusual relation in
human behavior, so that the search for analogs to phonemic distinction in
other human behavior has not been fruitful.
2 The pair test will equally well distinguish tez from taz, where no meaning or meaning-
difference is present.
3 Cf. E. Sapir in n. 1 above. Phonemic distinctions are distinctions among timbres, i.e. as to
which overtones have particularly high amplitudes.
We have yet to connect the phonetic distinctions between words with a

phonetic segmentation of words. Given two words or very short utterances
which have been shown to be phonemically different, we can locate the
points at which they seem to differ; the difference can be measured instru-
mentally, although present knowledge does not yield an adequate phonetic
characterization of phonemic differences. Thus between write and ride one
may notice a difference (of length) in the vowel and a difference (of voicing)
in the final consonant; in this case compactness of description favors taking
the vowel difference to be dependent on the consonant difference, rather
than the converse. It does not matter if the observations mislocate the
relevant differences, or note too many or too few, because comparison with
differences in other word-pairs will correct such errors. The phonemic
difference can then be stated between two phonetic segments rather than
between two utterances. If one of the paired words is phonemically different
from yet other words at the same point, we can begin forming a distinction
between its segment and all other segments that differ from it in the various
distinct words, e.g. between the last segment of heart and those of hard,
harp, etc.
If we compare the results from various pairings of a few short utterances,
we can describe some utterances as being distinct from others at more than
one point. Thus we can make a segmentation such that hart differs from
hearth in the last segment, hart differs from dart in the first segment, heart
differs from hurt and dart from dirt in the second segment, and heart differs
from dirt in the first two segments, and heart differs from dearth in the first,
second, and last segments (despite the spelling). We can then try to cover the
length of each short utterance in some adequate sample by segments that
distinguish it from other short utterances. For each utterance, the minimal
number of distinctions necessary to distinguish it from every other utterance, in
terms of pair tests, determines the number (and location) of segments which
express these distinctions. These are the phonemic segments of the utterance.4
A phonemically distinct segment (or segment-type) is, then, a set of
segment-occurrences (in one position of utterances) which are not dis¬
tinguishable one from the other (by the pair test), but each of which is
distinguishable from all segment-occurrences not in the set.
7.2. Phonemes and letters
7.2.0. Phonemes
Many phonemically distinct segments of utterances have complicated restric¬

tions'on how they combine with others in the utterance. Some of these
4
Cf. Z. Harris, /Methods in] Structural Linguistics, Univ. of Chicago Press, Chicago, 1951.
constraints apply to several phonetically similar segments (e.g. to all voice¬

less stops such as p, t, k)\ others are unique to one segment type. The use of
segmentation in identifying words—which is its major use in grammar—is
made far more regular, and simpler to state, if we impose a partition on the
set of segments whereby no two segments in one equivalence class (one part
in the partition) are phonemically distinct from each other. Each equivalence
class is called a phoneme. The main advantage in this step is to collect
complementary phonemic segments, i.e. those which occur in different
positions of utterances and with different neighboring segments, into one
entity: for example, in English the aspirated stops whose phonetic value is
ph, th, kh which occur in initial position of words (pin, tin, kin) and the
unaspirated stops p, t, k, which occur only after s (spin, stint, skin) can be
considered to be cases of the same morphemes respectively, p, t, k. Also, the
two vowel segments in write and ride are complementary variants of the
same phoneme (before voiceless and voiced consonants). Some segments
have more than one segment complementary to them (but distinct from each
other) in a particular position: for example, not only ph but also th and kh in
initial position is complementary top afters. In such cases the assignment as
to which complementary segments are in the same phoneme is made on the
basis of various considerations of simplicity, including having least phonemic
change in the different occurrences of the same word (‘morphophonemics’).
For example, if we want least change in the phonemes of connect or of
position when they appear after dis (in disconnect, disposition), we would
take initial kh as ‘complementary variant’ of post-s occurrence of k, and
similarly for ph in respect to p. For each occurrence of a phoneme, say p, we
can always tell from its phonemic neighborhood what phonemic segment it
represents, e.g. whether ph or unaspirated p.
The sounds in the segments are continuous, and vary during their course,
and may dissolve in a continuous way into the next segment, and may have
various lengths; but the phonemes have none of these properties. Since
they are simply the minimal segment-distinctions necessary for utterance-
distinctions they are discrete and have no length. For example, it is meaning¬
less to look between any arbitrary two phonemes for some other phoneme
half-way between them (either in the difference scale, or in the sound-
sequence of the word).
7.2.1. The phonemic system
The phonemes of a language fall into patterns of similarities and relations.

There are phonetic similarities, such as among/?, t, k and b, d, g, degrees of
phonetic closeness to other phonemes, and degrees of audibility (e.g. the
difference between unreleased p, t, k may not be very audible), particular
positions in words and similarities in these respects (e.g. English h and the ng
phoneme occur in limited—and complementary—positions). These inter¬

relations of phonemes help to decide to what phonemes to assign particular
phonemically distinct phonetic segments when the complementary environ¬
ments do not suffice for unique assignments. (Note that phonemically
distinct tones, stresses, and intonations are found by further analysis of
distinctions among segments.) The final assignments are made in such a
way as to obtain least constraint on combination for each phoneme, and
maximum regularities and for the phonemic composition of words (7.3).
Some phenomena remain marginal: e.g. the phonemic status of grammatical
intonations such as those of a sentence (period, question mark, exclamation
point), which extend over a phoneme-sequence but can be linearly located
at the end of that sequence. In some cases it may be possible to replace the
phonemes by elements which are combinatorially less restricted, at the cost
of these latter elements having more degrees of freedom, such as being able
to occur simultaneously or having multiple lengths. Aside from such solutions,
which may not be worth the cost of a new dimension over a whole grammar,
phonemes can be considered as having unit length (aside from the intonations)
and as being linearly ordered in the composition of an utterance. In those
cases where a single sound is assignable to the combination of two phonemes
(e.g. assigning the nasal flap in Midwest American winter to the successive
phonemes nt in that position) the phonemes remain linearly ordered.
We thus arrive for each language at a set of phonemes, as the most freely
combining set of phonetic-segment distinctions sufficient to distinguish
every utterance of the language from all utterances which are not recognized
by speakers as repetitions of it. (The speakers, however, need not be aware
of phonemic distinctions and phonemes.) The phonemes give a unique
spelling for each such utterance (but cannot distinguish homonyms); and
they do this in a precise and objective way that is virtually free of substantive
differences of judgement. On this depends the discreteness and precision
that can be obtained in the analysis of everything else in language, which is
ultimately defined in terms of phonemes.
7.2.2. The phonemic alphabet
The basic status of phonemics in representing language may be seen when

we consider its contribution to the origin of the alphabet, and to the success
of the alphabet as a system of writing. The creation of the alphabet was
somewhat of an accident in culture history. As far as is known, the various
true writing systems which were developed in early civilizations consisted of
marks—generally pictographs—for words (not directly for objects of the
world). In some cases, a mark came to be used as a phonogram for any
similar-sounding word, in addition to the one pictured by it ('rebus' marks).
Some systems developed further,- with marks that were originally picto-
graphic coming to represent particular syllables (of the originally represented

word); these marks were then used for segments of words, in whatever word
those syllables occurred. The alphabet, however, was not a further stage,
beyond syllable marks, toward which various writing systems developed. Its
formation was a single historical set of events—a fact which needs explaining—
and all alphabets are either later stages of the original one (in many cases
preserving the order of the letters), or else are modelled upon an already
existing alphabet.
The single creation of alphabetic marks apparently developed out of
Semitic (Canaanite) speakers' contact with Egyptian hieroglyphic writing
around the middle of the second millenium bcT In Egyptian as in Semitic,
each word began with a consonant; and in the various words which a single
Egyptian hieroglyph pictured (all of which, in general, shared the same
grammatical root) the consonants were generally constant while the vowels
varied with the grammatical status of the word in its sentence. Furthermore,
hieroglyphs could be used for other words or segments of words which
sounded like those pictured. The pictographs for very short words (having
one-consonant roots) came to be used, rebus-style, for the successive
syllables of names, especially foreign names, which naturally had no picto¬
graphs of their own. As a result, the names, or rather their successive
syllables, were written with a sequence of these one-syllable pictographs,
each of which represented the sound (initial consonant plus vowel) of
the pictured word. The root consonant was in general the same for all
occurrences of a given hieroglyph (and for the.given syllable of the name
being written), while the vowel (belonging to different Egyptian case-
endings of the pictured word) could be taken as whatever was the vowel at
that point of the name which was being written. Early Semitic alphabetic
inscriptions in Sinai and nearby suggest that Semitic speakers acquainted
with this Egyptian rebus use of one-consonant hieroglyphs saw in it an
acrophonic relation. That is, they apparently saw it as pictures being used to
represent the first sound of the pictured word—when actually the pictures
were representing (non-acrophonically) the full syllable sound of their
word, of which the constant part was the first sound alone (the consonant)
only in the case of uniconsonantal words. The earliest alphabetic inscriptions
show crude acrophonic pictures—new marks used for the initial phonemes
of the pictured (not necessarily uniconsonantal) Semitic (not Egyptian)
5 Kurt Sethe, Vom Bilde zum Buchstaben, Untersuchungcn zur Geschichte und Altertum-
skunde Agyptens 12, Leipzig, 1939; Johannes Friedrich, Geschichte der Schrift, Heidelberg,
1966; W. F. Albright, The Proto-Sinaitic Inscriptions and their Decipherment, Harvard
Theological Studies 22, Cambridge, Mass., 1966. Acrophony, as a new relation between the
picture or name of a symbol and the sound it represents (its phonetic value), was reached by a
misunderstanding; the people who first used the method for making a set of writing marks
apparently thought, erroneously, that that was the basis for the Egyptian hieroglyphs used for
names. Acrophony can naturally be alphabetic and phonemic, because it recognizes the initial
distinguishing (hence phonemic) sounds of the words of the language, whereas writing systems
do not otherwise naturally reach single sound segments.
word—such as a bull-head (’alpu) used for the glottal stop a house (betu)
for b, a camel (gamlu) for g (the greek gamma), etc.
That the original alphabet was only the application of a technique, rather
than a full understanding of sound-segments, is seen in the fact that it did not
develop marks for the vowels. The acrophonic reading of the pictographs
yielded only the sounds that were word-initial, which were always consonants
(in Semitic as in Egyptian). The marks were not thereupon extended to
vowels. This was not because of the usual explanation, namely that ‘vowels
are unimportant in Semitic’—they are essential for distinguishing gram¬
matical forms—but because the acrophonic method which was used to
obtain the alphabet could not yield the vowels, since no Semitic or Egyptian
word began with a vowel. Thus the acrophonic basis of the original alphabet
explains not only the introduction of alphabetic marks but also their original
vowellessness. The Greek borrowing of the alphabet, which retained its
acrophonic property by keeping the Semitic letter-names (alpha, beta, etc.),
yielded the vowels indirectly because Greek disregarded the Semitic laryngeal
consonants, leaving their marks for the vowels which followed the unheard
laryngeals. (Intentional additions of vowels—and additional consonants
and even consonant clusters—took place later in certain extensions of the
original alphabet, and at the end of the Greek alphabet.)
To summarize: the alphabet was not made by any phoneme-like de¬
composition of words into their successive sounds: such decomposition does
not appear as a normal final stage of writing development. It was made by
utilization of phonemic distinctions in word-initial position. The successive
sounds of each word—to the extent that pictographs were available for
them—were written with the pictograph for a word whose beginning was
that sound; in the original alphabet, this supplied only the consonants of the
word.6 Pictographs became established to the extent that they had distinct
beginnings under repetition; and the retention of their names (alpha, beta,
etc.) was essential to their acrophonic use, which is why they were retained
in Greek.7 The vowels, in the words being written, were not written,
because there was no picture whose word began with those sounds. Com¬
plementary variants of phonemes were not come upon, because they did not
occur in the initial position. But as many marks were established as there
were phonemically distinct word-initials in the language. The fact that
successive segments of a word were marked by the beginnings of various
other words was thus not a new recognition of phonetic segmentation: the
Sinai writers were doubtless doing only what they thought the Egyptians
6 We cannot tell whether the early users of the alphabet assumed that the required vowels for
a given word could be drawn from the various vowel values of the Egyptian signs (which did not
hold for the specific Semitic letter-names 'alpu, betu, gamlu, etc.), or whether they simply
wrote for a word as many relevant marks as were available to them, which were only those that
represented word-initial sounds (consonants).
7 Later borrowings of the alphabet lost the acrophonic names while keeping much of the
original order: e.g. the Latin alphabet and the Arabic.
were doing. And the Egyptians continued with their non-alphabetic system,
just writing some long words with the successive marks for short words
whose sounds were similar to successive segments of the long words.
The alphabet was thus neither an invention nor a discovery, but a
development whose product was novel because existing materials had
unused potentialities: the apparent acrophony of marks for one-syllable
words. This development has a certain distant similarity to aspects of the
evolution of language (12.4.4).
7.3. Morpheme and word boundaries
Each utterance is a sequence of phonemes (7.2). However, not every

sequence of phonemes of a language constitutes an utterance of the language,
and it has been seen that a central problem of language analysis is to find a
way of characterizing the sequences which are utterances as against those
which are not (2.1). We cannot expect to find an effective procedure that will
do this directly by stated constraints on phoneme combination alone. First,
it is impossible to describe (or generate) all the phoneme-sequences of the
language and only these in a non-hereditary way, for example as a Markov
process. For instance, the entry of a new word into the language by
borrowing, or by coining technical vocabulary, can create in its intra¬
sentence environments many new phoneme-sequences. And finite-state
transitional probabilities can be disturbed arbitrarily when, for example, a
word of limited use, with a rare initial phoneme, is used in a new environment:
the phoneme th (of this) followed by zzw may be virtually absent from
English, but it is immediately recognized as English if one chooses to say He
bathes Zouaves there.
Secondly, a single hereditary process cannot be expected to predict
directly all the phoneme-sequences in the language and no others. Evidence
for this is, for example, that into almost any sufficiently long utterance it is
possible to insert, at various points, arbitrarily many segments of utterances,
repeated arbitrarily many times, and still have an utterance of the language
as a result: e.g. insertion of very very inadequate and in my opinion into
A performance was given last night, to yield A very very inadequate per¬
formance—in my opinion—was given last night. In contrast, the charac¬
terization of the phoneme-sequences constituting utterances can be done in
two passes, if we first establish words and bound morphemes as ‘word-size’
sequences of phonemes, and then utterances as—in a complex way-
sequences of words. The word-size segments are those phoneme sequences
(or sets of alternant phoneme-sequences) whose combinations in utterances
are regular.
The boundaries of morphemes within a word, and more sharply of words
within a sentence, can be found by a particular stochastic process on
phoneme-sequences in utterances. In a given sample of utterances, we

consider all those whose first n phonemes are a particular phoneme-
sequence q(n). In the «+llh position of these utterances, after q(n), the
number of different phonemes which are found will be called v(n). The v(n)
thus measures the variety of followers after all occurrences of a given n-
phoneme sequence in the sample. It is found that in each w-phoneme
utterance, if we record v(n) successively for itsinitia! n phonemes,
the value of v(n) has repeatedly a gradual fall and sharp rise, as n goes from 1
to m. This sawtooth regularity is not found for arbitrary phoneme-sequences
which do not constitute utterances of the language. The sharp rises will be
seen to mark what we would want to consider (if we know the language) as
the ends of morphemes and words. As an example, in Figure 7. Is we take the
results for a short English sentence, written phonemically.
FIG. 7.1. Word boundaries in a sentence. Dogs were indisputably quicker. Dots
have been inserted between the phonemes, to show where a morphemic segmentation
would have been made on syntactic grounds. These dots were, of course, not
involved in the test.
Such results have been obtained by this procedure, going both forward
and backward, in various disparate languages. Similar results were also
obtained for the more difficult case of morpheme boundaries within one
word at a time, each initial sequence being matched for its next-neighbor
variety, both forward and backward, in a sampling from Webster’s unabridged
English dictionary, together with various science dictionaries, alphabetized
both forward and backwards. In this case, spelling was used unavoidably
instead of phonemic sequence, but the results remained in good accord with
grammatical segmentation, as can be seen from Figure 7.2.
s The figures, and various sections of the text, in 7.3-8 and 8.1 have been taken from
Z. Harris, Mathematical Structures of Language, Interscience Tracts in Pure and Applied
Mathematics 21, Wiley-Interscience, New York, 1968, where more details are given.
1 2 1 1 2 20 5 13 25
1 s * t u r b • a n c e
15 24 24 8
2 2 4 2 1 1
1 1 4 3 18 15 11 25
i s • e m • b o d y
15 24 24 11 5 6 5 2
4 7 1 1 5 24 12 26
i • s u 1 f • i d e
15 24 24 5 2 4 2 1
1 4 2 2 9 19 17 25
d e • f o r m • i t y
15 26 9 5 4 4 3 1
5 9 24 26
apple
26 14 7 5
FIG. 7.2. Morpheme boundaries in words. An example of the dependencies in this

recurrent stochastic process is given in the list above of individual words, each tested
against all the words in English. The probabilities for each outcome, i.e. each
possible sucessor, are of no interest here, and are not given. The process states only
the number of different outcomes, at each state, which have any positive probability.
As in the preceding figure, the dots indicate syntactic boundaries of morphemes, and
were not used in the test.
This procedure reveals a difference between the phoneme-sequences that

constitute utterances of a language and those that do not. Specifically, all
utterances in a language have a constraint that increasingly restricts the
possible phonemic successors (hence the gradual decrease of v) up to a point
at which the restriction abruptly ceases (the v peaks). This peaking can be
interpreted as a segmentation of the utterance into phoneme sub-sequences
within each of which the increasing successor-restriction holds. It will be
seen that these ‘word-size’ segments, which constitute words or morphemes,
are associated with meanings in a way that phonemes or sequences of
phonemes other than word-size are not, and that certain regularities of
combination of these word-size segments characterize utterances. The
word-size segments between peaks are the ones which have stable selections
or other grammatical relations to following or preceding segments. (Note
that the initials/-,#/-, segments of English do not have such relations, and do
not have a peak after their /.)
The effect of having this intermediate word-size segmentation between
phonemes and utterances (or, rather, as will be seen, sentences) is that the
constraints on freedom of combination which we seek as characterizers of
utterances, and the meanings which these constraints carry, can be stated on
a far larger body of elements—the morphemic segments (including words)
—than just the twenty to thirty phonemes that each language has. This
makes it possible to expect a much simpler system of constraints than if
everything language does had to be stated in terms of the few phonemes it
has. By the same token, the segmentation means that phoneme combination
short of the peaking constraint (e.g. in syllables), is not directly relatable to
the main, and information-bearing, constraints of utterances—those of
Chapter 3. That is to say, the phonemic composition of morphemes is
arbitrary, as are phonemes themselves, in respect to most of the rest of
grammar; and phonemes do not carry meaning. The peaks of v which
determine the segmentation result from two facts: that not all phoneme
combinations are utilized in making morphemes, and that the number of
morphemes in a language is some three orders of magnitude larger than the
number of phonemes. The boundaries of a morpheme are therefore points
of greater freedom for phoneme combination.
We thus have not only a distinguishing property of those phoneme-
sequences which constitute initial segments of sentences, but also evidence
that sentences (differently from non-sentences) contain a certain segmentation
into ‘morphemic’ sub-sequences of phonemes; and we have the boundaries
of these morphemic segments.
For the user of a language, the morphemes are known not, of course, from
this comparison of phoneme-sequence successors, but from being learned,
some in isolation and some by their regularity of combination in utterances.
They are thus pre-set as the user becomes able to speak or understand the
language, except insofar as the meaning of an unknown morphemic segment
in an utterance can be guessed from knowledge of the environing ones.
Various adjustments can be made in this procedure, particularly in order
to get a morphemic segmentation that has simpler syntactic regularities.
One adjustment eliminates the occasional case in which the rise, which
marks the end of the morpheme, is gradual rather than abrupt as it should
be. For this, we take v not as the number of phonemic followers to each
initial sequence ., p,„ (where pm is some phoneme a and pm_x is a
phoneme b) but as the ratio of that number to the number of different
phonemes that follow the phoneme a (or the sequence of phonemes ba) in all
neighborhoods in which a (or ba) occurs.9 In this ratio of hereditary to
Markov chain dependence, we obtain the hereditary effect more sharply;
i.e. the restriction on the followers of pm=a due to that particular pm being
'' We may consider the pair ba, rather than just the last phoneme, because many syllable
restrictions depend on pairs of neighboring phonemes. The way in which the phonetic property
of phonemes restricts their successor does not depend on the predecessors back to the
beginning of the utterance.
located in a complete initial sequence/;,, . . „pm_x,pm, rather than simply

in the phonetic neighborhood b=pm_l.w
In general, the successor count can show syllable boundaries, because
of the greater number of consonants than vowels and because various
consonant clusters can be followed only by a vowel. However, these syllabic
differences in successor count become far less pronounced as we proceed
through the word (e.g. in body in Fig. 7.2), because the hereditary effect
within the word becomes more prominent.
Certain peaks in some sentences do not correlate with a morphemic
segmentation of that sentence. This happens when the first part of a
morpheme in one sentence is homonymous with a whole morpheme which
can occur in the same initial neighborhoods. Thus in He was under. .., there
would unavoidably be a peak after He was un because of the prefix un- which
occurs here with many followers (He was unreserved, He was uninformed,
etc.). Such wrong segmentations can almost always be corrected by carrying
out the corresponding count (of predecessors) backward on the same
sentence.
7.4. Words and morphemes; complementary alternants
Once we have reached morphemic (i.e. word-size) segments, we can proceed

to regularities of their combination in utterances. Before attempting this, it
is useful to seek out any special constraints which hold for just one or a few
morphemic segments. In particular, there are many cases in which two
or more segments are complementary in respect to some feature in the
environment, so that the sum of their environments roughly equals the set
of environments of many single morphemic segments. Each set of one
morphemic segment or of more than one complementary segment is called a
morpheme, bound if it occurs only attached to one or a small set of other
morphemes, free (or: a word) otherwise.
One kind of complementary variant is seen in a bet, two bets (with added s
phoneme) as against a bed, two beds (with added z phoneme) and a child,
two children (with added -ren), where after two the -ren appears only after
child, s only after voiceless phonemes such as t, and z only after voiced
phonemes such as d or vowels. Clearly, it is preferable here to define not
three plural morphemes (not different in respect to further operators on
them), each restricted to different phonetic environments, but one un¬
restricted plural morpheme with three environmentally complementary
phonemic shapes. The spelling of this morpheme would then not be the
10 In the examples above, the two cases in which v rises gradually to the peak, 9-14-20 for dis
and 11-6-9-28 for kwik, would thus be replaced by decreasing ratios. Fewer phonemes can
follow a consonant, especially a pair of two consonants, than can follow a vowel. Thus 9 is a
smaller percentage of the total followers of i than 6 is of the total followers of w.
phoneme s, etc., but a new entity, the morphophoneme S defined as a set of

several alternants (including a few additional one's, as in foot, feet below).
Similarly, we find a knife, two knives, the knife’s edge, where instead of
having two restricted morphemes, knive before the plural morpheme S and
knife before all other morphemes (including the possessiveT), we can define
a single morpheme nayF where the last phoneme has been replaced by a
morphophoneme F defined as a set of two phonemes/, v with morphemically
complementary environments. (In contrast, we see from a fife, two fifes that
the morpheme/z/e is phonemically and morphophonemically fciyf, and from
a five, two fives that the morpheme five is fayv.)
We see here that cases in which two environmentally complementary
morphemic segments have different phonemes at a particular point can be
treated by giving them a single morphophoneme, which is a set (a disjunction)
of those phonemes: this leaves us with but one, more regular, morpheme,
with the complementary alternation being now stated on the phonemic
members of the morphophoneme. Morphemes can then be defined as
sequences of phonemes and morphophonemes. They could be considered to
be sequences of just morphophonemes if the unchanged phonemes in the
sequence are taken as morphophonemes with only one, non-alternating,
phonemic merfiber. One might then think that we can define morphemes
directly as sequences of morphophonemes, each of these defined directly in
terms of sounds. But this would be circular, because if we did not have
phonemes as utterance-distinctions independently of morphemes, we would
not be able to state precisely what are the alternant forms of each morpheme
and morphophoneme. More generally, we would miss the importance of
phonemes as a rare case of fixed distinctions among behaviors, creating
repetition in contrast to imitation, and yielding a system of discrete elements
which assures the discreteness of everything further in language.
There are some complementary morphemic segments for which morpho¬
phonemes cannot readily be used. In l am, you are, he is, to be we have four
morphemic segments restricted as no other English segments are. We can
form a single morpheme BE as a set of complementary suppletive phoneme-
sequences, without being able to define successive morphophonemes as
its components. (This, too, argues against morphophonemes replacing
phonemes as components of morphemes.) Such a solution would be valu¬
able if, as in this case, the set of all environments in which the comple¬
mentary segments occurred was roughly the same as that in which many
single morphemic segments do (go, come, walk, talk, etc.), so that BE
becomes a regular member of a known set of morphemes (the word-class
‘verbs’, 7.5), whereas each of am, are, etc. (as morpheme-alternants)
have unique environmental restrictions and lack the regularity of whole
morphemes.
Another situation is seen in They break records. They broke records
before too as against They brake the car too fast. They braked the car too fast
before too, where we find one morphemic segment brake (phonemically

breyk), having the added segment -ed before before, while the homonymous
other one, break (phonemically breyk, for the homonymity see below), has a
change of phonemes in that position. The change of phonemes from ey to ow
is not a variant of break, since it occurs with the same likelihoods (‘selection’)
of further operators as does -ed on brake; hence it is a post-break alternant of
-ed. Even though we do not obtain it as a morphemic segment in the
phoneme-successor procedure of 7.3, we have to accept ey -> ow as an
alternant member of the -ed morpheme, occurring with the morpheme
written break but not with brake (while wake has both waked and woke).
Similarly, in a foot, two feet, we have an additional a —»iy alternant of the
plural S morpheme, occurring with foot but not with boot.
Once we see that morphemes can have alternant members or forms
consisting of phoneme replacements, we can consider cases in which phonemic
replacement is the sole form of a morpheme, not just a variant form. A need
to recognize such morphemes arises in Semitic word-roots, where words
having a common meaning share the same consonants while such grammatical
meanings as past, imperative, place of action, etc. are indicated by different
intercalated vowels. These non-contiguous phonemic sequences can be
fitted into the system of internally connected morphemic segments which are
more common in the world’s languages, by accepting some basic connected
segment, e.g. the root with past tense vowels, as the lexical morpheme, and
considering all the other vowel-arrangements—future, nominalization, etc.
—as morphemes whose phonemic composition consists mostly in replacing
the vowels of the lexical morpheme. In other languages, the meanings and
grammatical functions carried out by the Semitic intercalated phoneme-
sequences are carried out by ordinary connected phoneme-sequences such
as words and affixes.
There are also cases where one has to say that zero is an alternant of a
morphemic segment. This is seen in a sheep, two sheep, where we either have
to say that sheep is virtually unique among morphemes in not accepting a
plural affix (while accepting a plural verb, as in The sheep are here), or else
that S has zero as its alternant after sheep. A larger issue of this kind arises
when we compare The trees fall. They fell the trees, and The children sit. They
seat the children, and We wake at 8, They waken (or wake) us up at 8, with
The umbrellas stand against the wall. They stand the umbrellas against the
wall. In fell, seat, waken we have a morphemic segment or a phoneme
replacement, with a causative meaning, which accompanies the moving
of the subject (e.g. trees) to object position. Given that the unchanged
morphemes (stand, wake, also walk, run, rest, etc.) can by themselves carry
this change in subject plus causative meaning, the simplest description is that
these words are then carrying a zero alternant of that causative morpheme,
an alternant whose phonemic composition is zero.
A somewhat more complicated case is that of He presents the thesis
tomorrow compared with his presentation of the thesis tomorrow, and He

defends the thesis tomorrow compared with his defense of the thesis tomorrow,
and They will write up the thesis compared with their write-up of the thesis, all
in contrast to He will delay in completing the job compared with his delay in
completing the job. When a sentence is ‘nominalized’, i.e. becomes the
subject or object of a further verb (8.2), the verbs in the above examples
have, respectively, an added morphemic segment, a phoneme replacement,
a stress change, or zero, all with a following of or the like. These have all to
be taken as alternant members of a nominalizing morpheme (an ‘argument
indicator’).
Occasionally one finds two or more morphemic segments which occur
together, and which can be taken as parts of a single morpheme even though
not contiguous. An important case is what is called agreement in grammar
(6.2). This exists, for example, in languages in which the adjective matches
its noun as to number and gender: hortus amplus ‘large garden’, puella
pulchra ‘beautiful girl’. Instead of adding to the apparatus of the grammar a
special relation of ‘agreement’ between a noun and its adjective, we need
merely say that the -us, -a, -um, and their respective alternants which come
at the end of each Latin noun are repeated on each adjective in predicate
(including modifier) relation to it. That is, the phonemic shape of -us is really
-us, -us,. . ., -us, the number of portions in each occurrence being one more
than the number of associated adjectives and predicates. This is taken as
holding even when no gander suffix appears on the noun, as in lux aeterna
where lux, ‘light’, ‘has feminine gender’, i.e. ends in a ... a ... a . . . (even
though with zero alternant for the initial a), because its adjectives receive
the repeated portions of the feminine suffix -a.
In this survey of major forms of morphemes, primarily such as are found
in English, we see phoneme-sequences, with sets of complementary alternants,
and also phoneme replacement and zero. We also see that many morphemes
are sayable independently, and in many languages carry a separate main
stress, and enter into a complex spectrum of combinations: these are words
(free morphemes). Other, bound, morphemes lack these properties and are
always pronounced, under a main stress, with one or another morpheme of a
particular set: these are affixes or phoneme-replacements or zeros, and they
form words only together with the morphemes of the set to which they are
restricted. In most cases they are attached to words, as in -ation on present;
but there are cases of particular bound morphemes combining with each
other to make a word, which then takes on its own stress: per-, con- and
-ceive, -mit forming perceive, permit, conceive, commit.
In summary: we started with morphemic segments, many of which
function as independent morphemes. Some morphemes, however, were
defined as disjunctions of segments. In other cases, a segment was broken
into two morphemes, as in the case of phoneme replacements. These and all
the other modifications which we make in the results of the phoneme-
sequence successor procedure are always made on the basis of comparing

the further constraints on a given morphemic segment with those on other
segments having comparable environments, in order to arrive at a simplest
set of entities and constraints for all the data. For example, in the case of
complementary members of a morpheme, there are three criteria, all
simplifying the description: that the members be complementary in respect
to some stated near neighbor; that the members be very similar in respect to
their farther neighbors (operators) and thus have the same meaning (e.g. I
am, you are, he is occur in much the same farther environments); and that
the set of environments of all the members together be roughly the same as
that of some—preferably many—single morphemic segments (e.g. be in
respect to verbs).
When the criterion of maximum regularity is applied to the selection of a
word, that is, to the set of words with which it occurs with greater than
average likelihood (3.2), we can distinguish homonyms, i.e. cases where
a single phoneme-sequence occurs in the kind of separate grammatical
positions and separate coherent selections that are generally covered by two
or more different words. On such grounds, occurrences of heart in a heart
attack and in the heart of the matter and even a hearty meal would all be
considered the same word (the latter uses having the properties of meta¬
phoric and other known extensions of a word’s selection); but the same
phoneme-sequence in a hart and a dove would be a homonym rather than the
same word. Conversely, slightly different phoneme-sequences which have
the same (synonymous) selection would be considered free alternants of the
same word, as in the two pronunciations of economics. This assignment
might even be made ‘suppletively’ when there is no phonemic similarity, but
there are other special grounds, as between e.g. and for example. The whole
question of marginal cases of being the ‘same’ word is not crucial and may
ordinarily be ignored in many languages: is case in briefcase, a small case (of
wine) the same word as case in a small case (taken by a legal-aid lawyer)! The
issue affects zeroing of word-repetition: one cannot say He got two small
cases today, one of wine and one in law. One can say He plays violin and she
piano, but can one say except jokingly He plays violin and she chess?
The criterion of maximum regularity does not suffice to decide whether
certain uniquely restricted segments are morphemic (word-size) at all. One
problem is seen in small sets of phoneme-sequences which combine only
among themselves. A situation roughly of this type exists in per-, con-,
-ceive, -mit, etc., in English perceive, permit, conceive, etc. Another situation
is that of phoneme-sequences which do not combine freely enough to be
considered morphemes, but which have a bit of the meaning-properties and
combinatorial properties of morphemes: e.g. the si-, gl-, of slide, glide, slow,
glow, and slither, gleam, glimmer, etc. (11.1). Somewhat more morphemic is
the -le of dazzle, drizzle, jiggle, joggle, etc., which historically was a
morpheme, but most of whose attached segments are not recognizable as
morphemes now: if the -le is to be a morpheme, that to which it is attached

must also be one. Another type of complexity is seen in curb, meaning
‘restrain; a bit; a street border’, where British English has the spelling kerb,
but only for the third meaning. The question whether kerb is a different
word does not arise in speech; but it arises in writing. Somewhat similarly,
stanch has two meanings, one of which is the same as one of the meanings of
staunch, which in American English has two pronunciations, one the same
as stanch.
The phoneme-sequences identified by the morpheme-boundary procedure
have regularities of combination with each other in utterances of the
language, but we have seen here that various ones of them have to be
adjusted to fit the further regularities. The specific phonemic content of
morphemes is not related in any essential way to their further grammatical
behavior, and indeed phonemes are a requirement of speech rather than of
syntax: the same syntax as in a spoken language, with the same informational
interpretation, can be created in a one-one correspondence between mor¬
phemes and indecomposable marks such as pictographs. The phonemic
composition of morphemes is needed primarily because speech cannot make
as many single distinct sounds as the distinct morphemes of a language. It
incidentally follows from this that any classification or redefinition of a
syntactic problem that reduces the syntactic irregularity to a phonemic-
composition irregularity (morphophonemics in a broad sense) is a real
descriptive gain, since the syntax is seen to be a system independent of the
phonemic composition. An example is moving the dependence of am, are, is
on I, you, he from a restriction on verbs in respect to their subjects, to a
restriction on the complementary phonemic compositions of the verb be.
7.5. Word classes
To go from words and bound morphemes of 7.4 to the set of all possible word
combinations in utterances is a large step. This is especially so in view of the
great number of words, with almost every one being unique in some of its
combinations and transformations, in keeping with its unique meaning and
its stylistic character. In 3.1-2, it was found that this large step could be
decomposed into two crucial successive stages for each word. The first stage
determined what word-occurrences the given word never combines with in
all base utterances including the shortest (where observation is easier);
this creates the argument-requirement word classes. The second stage
characterizes the graded likelihood of each word in respect to the various
other words in whose environment it has probability >0 in the shortest
utterances.
Meanwhile, in traditional grammar, an approximation to these word
classes is obtained by related if less global considerations. The traditional
classification is based on similarities in combination. It is well known that for

almost every word there are many other words whose environments are
largely similar to those of the given word. For example, we can collect many
words into a class of nouns (N), not only because they can all occur in some
particular and more or less immediate morphemic environment such as after
the (or, for most of them, before plural S), but even more because they can
all occur with one or another of some other collection of words: e.g. every
member of N can be found before some verb, though not before every verb
(F). It is then possible to say that the sequence NV is a sentence structure,
where NV indicates some member of N followed by some member of F, but
not all NV combinations. The other classes can be characterized comparably
(sometimes a bit roughly): each verb occurs with a tense affix and can have
one or another preceding N; adjectives (A) can occur between the and N;
prepositions (P) occur before N\ co-ordinate conjunctions (C) can occur
between any two like-classified words, and so on. Most of these classes are
found in many languages, and serve for the combinations found in con¬
stituents or in strings (7.6), on the way to stating the structure of sentences.
In English, each of the large classes N, V,A, contains many single morphemes
as words (e.g. man, think, large); but each class also contains words which
belong to some other class but which carry affixes that make them members
of the class in question: manhood, consideration, largeness in N; to man, to
reconsider, to enlarge, to up in F; man-like, considerable, largeish, uppity, in
A. These are thus classes of words rather than of morphemes; they are such
because the criteria for each class are satisfied not only by its special
morphemes but also by the morphemes of other classes when carrying the
appropriate affixes.
These classes do not constitute partitions of the set of words of a language.
For one thing, some words have the properties of one class with an
occasional property of another: e.g. rich, in A, can occur by itself after the,
like N; and the color and number words in several languages have both A
and N properties. Indeed, some languages have a great quantity of regular
or irregular multiple classifications, whereby some occurrences of a word
have the properties of one class, say A, while other occurrences of it have the
properties of another class, say N. Furthermore, there are some single words
or small sets of words which satisfy none of these criteria: e.g. yes, or in
English a set containing only, hence, etc. which have a peculiar bi-sentential
environment (ELI, G395). In any case, these classes say nothing about the
relative likelihood of finding a particular member of one class with a
particular member of another. Thus, tensed F are stated to have a preceding
N; but whereas we can find, as a case of NV, The earth slept, we will be
hard put to find The vacuum slept, and will not at all find His departure
slept.
In the final analysis, it will be seen that these word classes, based on
observable combinations in the word-sequence, can be replaced by a
sharper classification of morphemes based on a more fundamental relation

of argument-requirement (3.1, 8.2).
7.6. Word-class sequences:

Continuous and contiguous strings
Given the word classes, it is possible to isolate a few types of word-class

sequence which suffice as the structural components of utterances. In
traditional grammar, what came to be formulated was primarily the near
neighborhood of a word, which was the most sharply fixed: above all its
affixes, its morphology; and in addition, its immediately neighboring modi¬
fiers, constituting such local constructions as noun phrases and verb phrases
which were codified by Leonard Bloomfield as immediate constituents.
The structure of the sentence was then described as a combination of
immediate constituents, primarily noun phrase (as subject) followed by verb
phrase. However, not all modifiers are in the immediate neighborhood of
their host. One example of this is Finally the man arrived whom they had all
come to meet (where the whom phrase modifies man, and would normally be
next to it as in Finally the man whom they had all come to meet arrived).
Another example, in Latin, is the ability of adjectives to be at various (but
not unlimited) distances from the noun they modify.
One can still describe the constituent,.even with non-contiguous parts: the
subject above is man with adjoined the on the left and whom . . . after the
verb on the right. But the difficulty lies in formulating a constructive
definition of the sentence. For if we wish to construct the sentence by
defining a subject constitutent and then next to it a verb (or predicate)
constituent, we are unable to specify the subject if it has a non-contiguous
part because we cannot as yet specify the location of the second part (the
modifier at a distance). At least, we have no general and complete way of
specifying the location of the distant modifier until we have placed the verb
constituent in respect to the subject; but we have no general way of placing
the verb in respect to the subject as a single entity unless the subject has been
fully specified.
A similar constructive difficulty arises with respect to all other non¬
contiguous phenomena such as discontinuous morphemes, and the special
case of them known as grammatical agreement (above, as between subject
and verb in respect to the plural: The name appears below, The names appear
below). Arbitrarily many different word-sequences can appear in utterances
between the subject and its predicate, and the specification of plural
agreement between them is hard to formulate, especially when the predicate
precedes the subject as in Below appear, as John requests, the names of all the
participants. More generally, not every sentence has a consistent analysis
into the kind of constituents described above (e.g. what is the status of that
he go in It is important that he go?).
A more adequate and more regular decomposition of sentences into sub¬
sequences is to find in each sentence a minimal component sentence (the
'center string’: generally the tensed verb with its subject and object) and to
show that everything else in the sentence is adjoined as ‘adjunct string’ to a
‘host string’ (word or sentence) which is in the sentence (i.e. to the center
string or to a previously entering adjunct). In each sentence, the strings
consist of sequences of one or more words in the sentence. In a theory of
sentence structure, the strings are defined as particular sequences of word
classes, each originally continuous, which can adjoin particular other types
of strings at stated points—before them, after, or next to particular host
words in the string, i.e. in all cases contiguously at the time it adjoins. Each
adjunct string is defined as capable of adjoining any host string of a given
type, the resultant being again a string of the class of that host string. Thus A
is (in English) a left adjunct string of a host N (within a string containing that
N). We consider the A as a string adjoined to the string containing the N,
rather than as an adjunct to the N alone. D (adverb) is a left adjunct string of
A or D (or an adjunct of V in some string) producing DA as an A string
(entirely empty), DD as a D string (almost entirely), as well as DV (almost
fell). The DD on A is an A string, as is D on DA (both as in almost entirely
empty)-, AN and DV are not in themselves strings, since the N and V are
merely words within a string.
That the adjunctions are to a string and not to a word in it can be seen in
the cases when an adjunct referring to a word is at a distance from the word,
but not at a distance from the string of which the word is part: e.g. in Finally
the man arrived whom they had all come to meet, where the adjunct whom . . .
depends on the occurrence of man (i.e. occurs only in the presence of certain
N including man)-, or in Latin, where the adjective (adjunct) that agrees with
and refers to a noun can be at a distance from the noun. In both cases, the
adjunct is always contiguous to the string of which that noun is a part. Thus
adjunction is a relation between strings, even though the grammatical
dependence and semantic connection can be between one string and a host
word in the other string.
The main string types sufficient for English utterances are found to be:
Center strings (S): chiefly noun plus verb plus tense plus the object
required by that verb; also the interrogative and imperative forms of
this;
Adjuncts on center strings, entering at right or left or at interior points of
the host string and often separated by comma: e.g. D (perhaps,
certainly), preposition-plus-noun PN (in truth), certain idiomatic strings
(no doubt), and C conj unction followed by a center string (because they
rioted) or P followed by a 'nominalized’ center string (upon the army's
arrival, with defeat certain);
Right adjuncts on virtually any string or string-segment, consisting of and,
or, but, or comma, followed by a string or string-segment of the same
type: e.g. N or N (pen or pencil), S but S (He went but l didn't).
Left adjunct on host nouns: e.g. A (wooden expression), N (woodstove)-,
or on host adjectives, prepositions, adverbs: e.g. Don A (very wooden), D
on P (completely under), D on D (almost entirely)'.
Right adjunct on nouns: e.g. PN (of matches after a box), relative clauses
(whom we met), APN (yellow in color after a box)', on adjectives: e.g.
PN (in color after yellow)-, on verbs: e.g. PN (with vigor after oppose)'.
Left or right adjuncts on verbs: e.g. D (vigorously before or after oppose).
(By right or left adjunct on a noun or a verb is meant an adjunct to the
string that contains that noun or verb, the adjunct being placed to the
right or left respectively of the host noun or verb in the string.)
These strings are not merely constructional artefacts, but an aspect of the
total structure of utterances. This is seen in the fact that many grammatical
phenomena are describable in terms of the internal structure of strings or in
terms of a host and its adjunct string. This holds for the restrictions as to
which word subclass (within one class) occurs with which other word
subclass, or which individual word of one class occurs with which individual
word of another, or the locating of a grammatical agreement on a pair of
classes. Examples are the restrictions between the members of a string (e.g.
plural agreement between its N and its V) or beween part of a string and an
adjunct inserted into that string (the restriction of a pronoun to its antecedent,
or of an adjunct to the particular word in the host string to which the adjunct
is directed), or between two related adjuncts of a single string (as in People
who smoke distrust people who don't where the zeroing of smoke in the
second adjunct is based on its parallel relation to the first adjunct).
The string property of language arises from the condition that language
has no prior-defined space in which its sentences are constructed (2.5.4).
Within this condition, there is room for a certain variety. For example, a
string b\b2 could be inserted at two points of its host a{a2 instead of at one,
yielding e.g. axbxa2b2 instead of the usual bxb2axa2 or axbxb2a2 or axa2b\b2,
this happens in the rare intercalation He and she play violin and piano
respectively. But one adjunct c of a host axa2 could not be inserted into
another adjunct bxb2 of that host: we would not find axbxcb2a2 or the like
(unless c was an adjunct of bxb2, or an interruption between dashes). And a
part of a string cannot be dependent on an adjunct to that string, or on
anything else which is outside that string itself. Any sequence of classes
which satisfies the definitional conditions above is called a string; and what is
empirical is that it is possible in language to find class-sequences (of rather
similar types in the various languages) which indeed satisfy these conditions.
A good approximation" to the strings of a language can be obtained by

successively excising, from each set of what are tentatively judged to be
similarly constructed sentences, the smallest parts that can be excised
preserving independent utterancehood (i.e. where the residual utterance
does not depend on the excised portion, and retains certain stated formal
and semantic similarities to the original utterance), until no more can be
excised: each excised segment is an elementary adjunct string, and the
residue is a minimal center string. It follows that the successive word classes
within a string are required by each other, while each excisable string is an
adjunct, permitted but not required by the residue to which it had been
adjoined. Thus in Crowds from the whole countryside demanding their rights
ringed the palace, we can first excise whole and their and then from the
countryside and demanding rights. In constructing the sentence from ele¬
mentary strings, we first adjoin the two left adjuncts of N, whole and their, to
the two elementary right adjuncts of N,from the countryside and demanding
rights; and then the two resultant non-elementary right adjuncts of N.from
the whole countryside and demanding their rights, are adjoined to the
elementary center string Crowds ringed the palace (to the right of a stated N
in it), yielding the desired sentence.
The result of applying the criteria mentioned here is that we are able to set
up a small number of types of strings, with not many specific word-class
strings in each type, which suffice for the construction of all sentences,
and which satisfy various additional properties. For example, no further
operation on the strings can introduce intercalation of them (beyond what
may be given in the original insertion conditions for particular strings, as
with respectively), because any permutation can only be on the word classes
within a string, or in moving a whole string from one point to another in the
host string (as in permuting adverbs to new positions). No non-contiguity
arises, because whereas an adjunct, e.g. of a noun, may be at a distance from
that noun, it can never be at a distance from the string which contains that
noun. For such reasons, string analysis is convenient for computation.
7.7. Sentence boundaries
What string analysis characterizes directly is not the set of all utterances of a
language, but the set of all sentences, these being a segmentation of
utterances. The reason string analysis produces only sentences is not that we
started intentionally with sentences—we started with whatever was a short
utterance—but that we did not include sequences beginning with a period
among the adjunct strings: such a sequence as and perhaps she left is an
adjunct string (as in He left and perhaps she left) but period + Perhaps she left
I I
Proposed by Henry Hiz.
(as in He left. Perhaps she left) is not. If we look only at the succession of
these sequences, we may see little difference, e.g. between He left and
perhaps she left and He left. Perhaps she left. But if we check for restrictions,
we find important differences. For example, under and there is a rare
zeroing to He left and perhaps she, but hardly under a period (He left.
Perhaps she. is dubiously accepted). More importantly, sequences under
and can be the subject or object of further words, but not sequences under a
period: His leaving and her leaving are clearly related, but no His leaving.
Her leaving are clearly related; similarly That he left and she left surprised us,
but no That he left. She left surprised us. For such reasons, the period is not
included in the set C of conjunctional strings. Hence, although the minimal
‘center’ strings include any minimal utterance, the result of string adjunction
is not the set of all utterances but the set of certain segments of utterances
called ‘sentences’.
The existence of sentences as a unique grammatical construction is shown,
differently for different languages, by various considerations such as above.
In a more general way, string analysis can be used for a recurrent stochastic
process which locates the sentence boundaries within a discourse. Just
as the process proposed in 7.3 shows that a segmentation into morphemes
exists (and locates the specific boundaries), so the process proposed below
shows that a segmentation into sentences exists (and locates the sentence
boundaries).
We consider the word-class sequences which constitute sentences, and we
think in terms of distinguishing in them all positive transitional probabilities;
from those which are zero. We can now state a recurrent stochastic depend¬
ence between successive word classes of each sentence form in respect to the
string status of each of those words.12 In this hereditary dependence of the
nth word class, w„, of a discourse on the initial sequence wq, . . ., w„_,,
sentence boundaries appear as recurrent events in the sequence of word
classes. This is seen as follows: if a discourse D is a sequence of word (or
morpheme) classes and jc, y, are strings (as in 7.6) included in D, then:
a. the first word class of D is: (la) the first word class of a center
string,
or (2a) the first word class of a left
adjunct which is defined as
able to enter before (la), or
before (2a);
12 Cf. N. Sager, ‘Procedure for Left-to-Right Recognition of Sentence Structure’, Trans¬

formations and Discourse Analysis Tapers 27, Univ. of Pennsylvania, I960. We take the choice
of a particular word class w„ in the n"' word position of the sentence as the outcome which
depends on the selections w, . . ., w„ , in the successive preceding word positions. As before,
we disregard the probability weightings of each outcome, and note only which outcomes have a
non-negligible probability.
b. if the nth word class of D is: the m,h word class of a string x con¬
taining p word classes, p>m
then the n+lth word class
of D is: (3) the m-f- 1th word class of x
or (4) the first word class of a right
able to enter after the mth
word class of x,
or (2b) the first word class of a left
able to enter before (3), or
before (4), or before (2b);
c. if the nth word class of D is: the last word class of a left-adjunct
string x, where x is defined as enter¬
ing before the m,h word class of a
string y
of D is: (5) the m,h word class of y,
or (6) the first word class of a left
adjunct defined as able to enter
before the mtb word class of
y, and such as is permitted to
occur after x,
or (4c) the first word class of a right
after x,
or (2c) the first word class of a left
before (6), or before (4c), or
before (2c);
d. if the /zth word class of D is: (7) the last word class of a right-
adjunct string x which had
entered after the mlh word
class of a string y which con¬
tains p word classes, p>m,
or (7') the last word class of a right-
adjunct string x which had
entered after (7)
of D is: (8) the ra+lth word class of y,
or (6d) the first word class of a right
after the mth word class of y.
and such as is permitted to

occur after x,
or (4d) the first word class of a right
after x,
or (2d) the first word class of a left
before (8), or before (6d), or
before (4d), or before (2d);
e. if the /z,h word class of D is: the last word class of a right-adjunct
string x which had entered after a
string y
then the n + lth word class of
D is: (le) the first word class of a center
string,
or (6e) the first word class of a right
after y, and such as is per¬
mitted to occur after jc,
or (4e) the first word class of a right
after x,
or (2e) the first word class of a left
before (6e), or before (4e),
or before (2e),
or (2e') the first word class of a left
before (le), or before (2e'),
or (9e) null;
f. if the ntb word class of D is: the last word class of a center string
then the n+ 1th word class
of D is: (lf) the first word class of a center
string,
or (4f) the first word class of a right
adjunct defined as able to occur
after a center string,
or (2f) the first word class of a left
before (4f) or before (2f),
or (2f) the first word class of a left
before (If) or before (2f')
or (9f) null.
Thus we are given two possibilities for the string standing of the first word
class of a discourse; and given the string standing of the nth word class of a
discourse, there are a few stated kinds of possibilities for the string standing
of the n+lth word class. The possibilities numbered (2) are recursively
defined. This necessary relation between the nth and n+lth word classes of a
sequence holds for all word-sequences that are in sentences, in contra¬
distinction to all other word-sequences.
We have here an infinite process of a restricted kind. In cases le and If, 2e'
and 2f', 9e and 9f, the nth word class of D is the end of a sentence form,13 and
the // + 1th word class of Z), if it exists, begins a next putative sentence form.
The transitions among successive word classes of D carry hereditary depend¬
encies of string relations. But le, f and 2e', f' are identical with la and 2a,
respectively. The dependency process is therefore begun afresh at all points
in D which satisfy le, f or 2e\ f'. These points therefore have the status of a
non-periodic recurrent event for the process.
In this way, sentence boundaries are definable as recurrent events in a
dependency process going through the word classes of a discourse; and the
existence of recurrent events in this process shows that a sentential segmenta¬
tion of discourse exists. Given a sequence of phonemes, we can then tell
whether the first word is a member of a first word class of some sentence
form, and whether each transition to a successor word is a possible string
transition. That is, we can tell whether the phoneme-sequence is a sentence
—is well formed in respect to sentencehood. In so doing we are also stating
the string relation between every two neighboring words, hence the gram¬
matical analysis of the sentence in terms of string theory.
7.8. Sentence structure
Traditional grammar referred to sentences, but did not provide a general

and precise structure for sentences, because this could not quite be done
with the ‘immediate constituent’ entities of language that were recognized in
grammar (e.g. verb phrase). However, one can construct various devices for
recognizing the structure of a sentence on the basis of grammatical relations
among its successive words, and also for deriving other sentences from
elementary ones. For example, it is possible to make an orderly constituent¬
like expansion of word classes into word-class sequences by a type-theory¬
like hierarchy of rewrite rules such as (using L here for plural): AN' = N' if
one can substitute AN' for Nl (e.g. dark sky or ominous sky, for sky)
13 This holds also if the n+ 1th word class of D does not satisfy any of the possibilities for cases
e, f, since only 6e, 4e, 2e, and 4f, 2f can continue a sentence form beyond a point where it could
have ended. In this case the n+ 1th word class of D is part of no sentence form. Aside from this, if
the mth word class does not satisfy the table above, then D is covered by sentence-form sections
only up to the sentence-end preceding m.
(obtaining AAN1 (e.g. dark ominous sky) in the class TV1); and NlL=N2 if
one cannot substitute N'L for Nl throughout (we cannot put a plural noun
N'L for Nl before L to obtain NLL, i.e. a noun with double plural). At each
point where recursion does not hold, we raise the power of the host. The
hierarchical system of equations presents several stages of substitutability:
the combining of morphemes to take the place of single words (mostly level
1), the repeatable adjoining of modifiers (remaining at the same level), the
non-repeatable expansions (e.g. plural, the, and the compound tenses)
forming levels 2 and 3; and it keeps track of how a whole sentence is pushed
down into being a modifier (e.g. relative clause). For a given sentence, the
formula which is obtained, e.g. the sequence N4V4 (in English), points to a
particular ‘central' elementary sentence within it (the level 1 material which
is contained in that N4, V4). And the successive equations which reach this
formula for the sentence point to the way N1 and V1 of that elementary
sentence were expanded to make the given sentence. In this manner, it is
possible to house all structural sub-sequences of a sentence in one relation,
and also to show that each sentence contains a central elementary sentence
underlying in an orderly way the hierarchies of expansion and replacement.14
In string analysis, these expansions are then found to constitute a regular
system of contiguous adjunction of originally continuous strings. Each
sentence is obtained from its sentential string plus various adjunct strings on
it or on previous strings.
It is also possible to construct a cycling automaton with free-group
cancellation to determine sentence well- formedness. For this purpose, each
word is represented by the string relations into which it enters. We then have
an alphabet of symbols indicating string relations, and each word of the
language is represented by a sequence of these new symbols. On any
sequence of these new symbols we can, by means of a simple cycling
automaton, erase each contiguous pair consisting of a symbol and its
appropriate inverse, in the manner of free-group cancellation. Then each
sequence which can be completely erased represents a sentence.
There are three string relations (i.e. conditions of well-formedness) that
hold between words of a sentence: they can be members of the same string;
or one word is the head of an adjunct and the other the host word to which
the adjunct is adjoined; or they can be the heads (or other corresponding
words) of two related adjuncts adjoined to the same string. Furthermore,
since sentences are constructed simply by the insertion of whole strings into
interior or boundary points of other strings, it follows that the above
14 Z. Harris, ‘From Morpheme to Utterance', Language 22 (1946) 161-83; reprinted in the

writer’s Papers in Structural and Transformational Linguistics, Reidel, Dordrecht, 1970, where
there are fuller treatments of the material presented here throughout Chapters 7 and 8. The
hierarchy of rewrite rules is sketched for various languages in the paper cited here, and in the
writer's ‘Emeneau’s Kota Texts’, Language 21 (1945) 283-9, and ‘Structural Restatements, I’,
International Journal of American Linguistics 13 (1947) 47-58; also in pp. 285 ff. of Z. Harris'
[Methods in] Structural Linguistics, Univ. of Chicago Press, Chicago, 1951.
relations are contiguous or can be reduced to such. For each string consists
of contiguous word classes except in so far as it is interrupted by another
string; and each adjunct string is contiguous in its first or last word to a
distinguished word of its host string (or to the first or last word of another
adjunct at the same point of the host).
This makes it possible to devise the cycling automaton. In effect, we check
each most-nested string (which contains no other string within it), i.e. each
unbroken elementary string in the sentence. If we find it well formed (as to
composition and location in host), we cancel it, leaving the string which had
contained it as a most-nested, i.e. elementary, string. We repeat, until we
check the center string.
In view of the contiguity, the presence of a well-formed (i.e. string-
related) word class B on the right (or: left) of a class A can be sensed by
adding to the symbol A a left-inverse b' (or: a right-inverse 'b) of B. If the
class A is represented by ab\ the sequence AB would be represented by
ab'b; and the b'b would cancel, indicating that the presence of B was well
formed. For example, consider the verb leave in the word class Vna (i.e. a
verb having as object either NA, as in leave him happy, or NE, as in leave
him here). It would have two representations: va'n' and ve'n'. Then leave
him happy could be represented by va'n'.n.a,15 which would cancel to v,
indicating that the object was well formed.
Specifically: for any two word classes (or subsequences of word classes) X,
Y, if (a) the sequence XY occurs as part of a string (i.e. Y is the next string
member to the right of A'), or (b) Y is the head of a string which can be
inserted to the right of X, or (c) Y is the head of a string inserted to the right
of the string headed by X—then, for this occurrence of X, Y in a sentence
form, we set X—>y' (read: Xis represented by y', or: the representation of X
includes y') and Y—>y, or alternatively A"—>x, and T—>'x.16 Herey' is the left
inverse of y, ’x is the right inverse of x, and the sequences y'y, x'x (but not,
for example, xx') will be cancelled by the device here proposed.17
It should be stressed that what is essential here is the fact that the inverse
symbols represent the string relations of the word in whose representation
they are included. When a most-nested symbol pair is cancelled, we are
eliminating not a segment of the sentence but a string requirement and its
satisfaction.
It follows that if a word class Z occurs as a freely occurring string, i.e.,
without any further members of its own string and without restriction to
particular points of insertion in other strings, then its contribution to the
well-formedness of the sentence form is that of an identity (i.e. a cancellable
15 The dots separate the representations of successive words, for the reader’s convenience,
and play no role in the cancellation procedure.
16 Correspondingly, if Y is the end of a string inserted to the left of X, etc., then Y—>x\ X^>x;
or Y—>y, X—>'y.
17 In this notation, it will be understood that (xy)'=yV, '(xy) = 'y'x and (x')'=x='('x).
sequence). The representation would be Z—>z'z. (While such classes are

rare in languages, an approximation to this in English is the class of
moreover, however, thus, etc.)
If, in a given occurrence in a string, a word class has more than one of the
string relations listed above, its representation will be the sum (sequence) of
all the string relations which it has in that occurrence. For example, if in a
given string, N has two adjuncts, A on its left and E on its right (as in young
men here), its representation there should be 'ane'\ a.'ane'.e would cancel
to n.
There are certain conditions which the representations must meet. No two
different classes should have the same representation (unless the string
relations of one are a subset of the string relations of the other), for then we
could cancel sentence forms that have one of these classes instead of the
other. No proper part of a whole string should cancel out by itself, for then if
only that part (or its residue in the string) occurred instead of the whole
string, it would cancel as though it were well formed. If a proper part of a
string is itself a string (and so should cancel), then it should be considered a
distinct string; otherwise the extra material in the longer string has the
properties of a string adjoined to the shorter string. For example, in NtVN
(He reads books) and NtV (He reads) we have two distinct strings, with
appropriate representations for their parts.
Most word classes occur in various positions of various strings. For each
of these occurrences there would be a separate representation, and the
occurrence would be well formed in a particular sentence if any one of these
representations cancelled with its neighbors. As is seen in the next paragraph,
for example, N has not one but several representations. We do not know
which, if any, of these is the appropriate one for a given occurrence of A in a
particular sentence until we see which, if any, cancels one of the representa¬
tions of the neighboring words in that sentence.
Generally, host words carry the inverses of their adjuncts, as in ’ane'
above for N; heads (as markers) of strings carry the inverses of the whole
string (so that the members of such strings do not have to carry the inverses
of their next string member): e.g. in which he will take we can represent
which by wv'fn', he by n, will by/, and take by v, so that after cancellation all
that will be left is w to indicate that a well-formed which string occurred
here. Order of adjuncts can be expressed as follows: in respect to noun
adjuncts, N, TN,AN, 7A/V all occur, but not A TN (e.g. the star, green star,
the green star, but not green the star). We have to give these classes,
therefore, the following representations (in addition to others):
TAN
t a n
’ta ’tn
'an
Then the above examples would cancel down to n in the following sequences
of representations: n, t. 'tn (the star), a. 'an (green star), t. 'ta. 'an (the green
star)- but no representations could cancel green the star (ATN).
Permuted elements require separate representation. Thus saw as a verb
requiring N object is represented vn'; but since the object can be permuted
(as in the man he saw) the verb also has the representation ’nv.
Some consequences of language structure require special adjustment of
the inverse representation:
Delays. Sequences of the form x'x'xx will cancel only if scanned from the
right. If the scanning is to be from the left, or if the language also yields
sequences of the form yy'y'y which can only be cancelled from the left, then
it is necessary to insert a delay, i'i, to the right of every left inverse (and to
the left of every right inverse) which can enter linguistically into such
combinations. We obtain in such cases x'i'ix’xx and yy'yi'i’y, which cancel
from either direction. This occurs in English when a verb which requires a
noun as object (and the representation of which is vn'), or certain string
heads like which (in a representation ending in n') meet a compound noun
(the representation of which is n’.nn). The representation vn', for example,
is therefore corrected to vn'i'i, so that, e.g.
take book shelves
vn'i'i n 'nn
cancels from the left (as well as from the right) to v.
Conjugates. There are also certain rare linguistic situations (including
intercalation) which yield a noncancellable sequence of the form z'x'zx. An
x' which can enter linguistically into such an encapsulated situation has to be
representable by its Z conjugate zx'z', which enables the x' and z to permute
and the sequence to cancel. If a string AB, represented by a’a, encircles X
(i.e. X is embedded within AB), once or repeatedly, then the relation of
encirclement requires each A to be represented by xax': then, e.g., AAXBB
yields xax'.xax'.'a.'a, cancelling to x. (Dots are placed to separate the
representations of each capital letter.) In some cases, a particular word X
(such as markers discussed below) requires the presence of particular other
words or classes Y at a distance from it, i.e. with intervening Z, after all
cancellable material has been cancelled: YZXoccurs, but not ZXwithout Y.
The representation of X then has to include 'z'yz so as to reach over the Z
and sense the presence of Y. Thus the representation for than will check for a
preceding -er, more, less', the representations for neither, nor will check for a
preceding not (or negative adverb like hardly) or a following nor: More
people came than I had called-. He can hardly walk, nor can he talk.
Exocentrics. Another problem is that of a string XY which occurs in the
position of a class Z, i.e. which occurs where a Z was linguistically expected:
here the X would be represented by zy', so that zy'.y would cancel to z.
(This XY includes grammatical idioms.)
£ <U Q
(A
^ <
FIG. 7.3. Cycling automaton for an English sentence.

Q
§ *
<U
JZ
c
CT3 <3
JZ -5!
<L>c_
>
Markers. If, in any linguistic situation requiring a special representation,

one of the words occurs only in that situation or in few others, then that word
can be treated as a marker for that situation, and all the special representa¬
tions required to make the sequence regular can be assigned to the marker.
This applies to the heads (initial words) of many right adjuncts, e.g. that,
whether (that he came, etc.). This makes uncouth representations for the
markers (e.g. for the wh- words), but it saves us from having alternative
representations for more widely occurring word categories. Similarly,
certain classes of verbs occur as (adverbial) interruptions in sentences: e.g.
Celsius, 1 think, is Centrigrade. Rather than put the inverse of these verbs as
an alternative representation of every noun (so that the noun plus verb
interruption should cancel), we put the inverse of the noun onto the
representation of these particular verbs: think is not only V that' (to cancel in
I think that . . .) but also ’nd (so that / think is n.’nd which cancels into an
adverb, as proposed under Exocentrics).
Among the linguistic phenomena, in addition to the above, which were
inconvenient for the inverse representation was:
Dependent repetition. The co-ordinate conjunctions permit recurrence of
the preceding word class or sequence but not of most others (N and N, V and
V, but not N and V). One or another of the participating words has to carry
inverses that restrict the class (or class-sequence) after and to being the same
as the one before and-, and we must allow for sequences which may intervene
between the class and and. For example, we may give each class X a
representation xx' + '.18
Figure 7.3 is an example of one representation of a rather complex English
sentence.
The hierarchy of rewrite equations for sentence components, and the
cycling automaton for sentence well-formedness, may seem too laborious
for use in sentence description. Their chief utility may be in what they show
to be possible for linguistic theory, rather than in actual applications.
Thus the hierarchical equations show that each sentence has an underlying
minimal-sentence structure and a transformational relation to other sentences.
And the cycling automaton shows that sentence well-formedness can be
formulated computably, independently of the transformational history of
the sentence. Nevertheless, the hierarchical equations provide overviews of
the syntactic characteristics and differences of various languages. And such
devices as the cycling automaton can be developed for special-purpose
analyses of sentences and texts.
lx In XY and XY, it suffices to provide for Xand X, since each Xcancels its Y. 4- represents
and.
8
Analysis of the Set of Sentences
8.0. Introduction
Chapter 7 presented the analysis of occurrences of language as far as it can

be done in what is called structural linguistics. The analysis ended up with
two general limitations. First, the co-occurrence of individual words were so
numerously varied that regularities could be stated only as combinings of
classes of words, with the individual word-choice as between the class
combinations left unstated. The second limitation was that these regularities
held only within sentences but not over whole discourses, i.e. not between
the successive sentences of a longer utterance.
Two procedures of analysis point to resolutions of these limitations. The
first is the hierarchy of expansion-equations (7.8), which isolates within each
larger sentence a central elementary sentence plus a hierarchy of expansions
on its parts. This constituted a first step in transformational decomposition
of sentences (8.1), by showing that inside every sentence there exists a
minimal sentence formed by a short sequence of words of the given sentence.
The decomposition into component sentences was completed when it was
shown that each of the expansions on the parts of the central sentence is itself
a transform of a sentence, ultimately of a minimal sentence (7.6). The
complete analysis yielded a decomposition of each sentence into minimal
(elementary) ones (8.1.5), the word-choices in the larger sentences being
merely those of the component elementary sentences.
The second procedure was the application of the co-occurrence analysis of
Chapter 7 within the confines of one discourse at a time, in order to see what
restrictions or regularities appear in sentence sequences (ch. 9). It was found
that the words which appear in some grammatical relation to each other, in a
given sentence, have high likelihood of appearing in later sentences also in a
grammatical relation, but not always in the same relation.
Thus, after The lymphocytes contained antibody we may see, at one point
of the discourse or another. Antibody was found in lymphocytes, the
antibody content of lymphocytes, and the antibody in the lymphocytes. It was
then found that the set of grammatical relations in which these words were
repeated (i.e. the set of linguistic transformations) was not arbitrary (e.g. we
would not find here Antibody contained lymphocytes). Rather, it had a
structural characterization (as preserving the likelihood-grading of word
Analysis of the Set of Sentences 197
combinations); and it had the semantic property of being paraphrastic. The

transformations change the shape and arrangement of the words, but not
their choice, so that given a base sentence (4.1) the word-choices of all its
transforms are the same and need not be stated separately for each transform.
When completed, the transformational relation provided also a structure for
the set of sentences (6.3, 8.15).
The process of paraphrastic replacement of grammatical relations is
presented in 8.1. In 8.2 it is shown that this process is the result of
fundamental operations in sentence formation, primarily reduction. Further¬
more, it is found that when we undo this process, we reach the base structure
that underlies language.
We begin with the question of what causes the paraphrastic variation of
grammatical relations between the same word sets in successive sentences.
In each discourse particular small sets of words are involved with various but
not random grammatical relations among them. For example, in the article
analyzed in 9.3, one sentence contains optical rotatory properties of proteins,
another rotatory properties of proteins, and a third proteins undergo changes
in specific rotation, all exhibiting proteins and rotate in grammatical relation
to each other, but different relations. Also, we find The rotatory power of
proteins is sensitive to experimental conditions, and we find under which it is
measured (where the which repeats conditions and it repeats rotatory power
making it Their rotatory power is measured under these experimental con¬
ditions), and dependence of the rotatory properties of proteins on wavelength,
and dependence of optical rotation on experimental conditions, all exhibiting
grammatical relations between rotate and conditions.
If we want to investigate this variation, we have to see what is common to
the variations in many discourses, each with its own small set of words held
together by the relations. We find indeed that particular grammatical
constructions participate in these variations, (a) There are variations among
free-standing sentences: e.g. the transitive active as against the passive, the
assertion as against the interrogative, (b) There are also variations between
free and bound (component) sentences: the free sentence form ending in a
period, as against nominalized sentences (such as dependence of optical
rotation on experimental conditions, above), and also as against adjectivized
sentences such as relative (wh-) clauses and subsidiary clauses (e.g. this
being the case). In each of the (a)-forms one can find approximately the
same normally-selected word-choices (i.e. choices with higher-than-average
frequency), and similarly in each of the (b)-forms.1 And just as a word’s
particular selection, of what are its more-frequent-than-average operators
1 A crucial fact here is that the set of all relative clauses is the same as the set of all sentences
(including resultants of ‘fronting’, 3.4.2), except that an initial argument (or one statably near
front position) has been replaced by a pronoun (3.4.4). The same applies, with certain
modifications, to some subsidiary clauses. These can therefore all be best analyzed as bound
transforms of free sentences.
or arguments, is an indication of its meaning, so also the fact that these

different grammatical constructions have the same selection of word-choices
indicates that the complexes of grammatical relations within them have the
same meaning.
In the inter-sentence transformations, (a) above, we have a system of
paraphrastic relations among certain sentential constructions, or among the
sets of sentences in these constructions. These relations were called trans¬
formations because they constitute a set of partial transformations on the set
of sentences, each transformation mapping the subset of sentences which
have the properties of one of the constructions above onto the subset of
sentences which have the properties of another construction, preserving
word-choice and meaning. It is because of this that these transformations
provide a decompositional structuring of the set of sentences (6.3).
In addition, transformations between free and bound sentences, (b)
above, provide a method of decomposing individual sentences into elementary
sentences plus certain expansions. Many of the expansions are ‘adjuncts’,
i.e. transforms of sentences which are being adjoined to another sentence:
e.g. the expansion of article to long article, in He wrote a long article,
transformed from He wrote an article which is long, from He wrote an article;
the article is long. Other expansions involve ‘nominalizations’, i.e. transforms
of a sentence when it is the subject or object of a further verb: e.g.
destruction of Dresden in He saw the destruction of Dresden, from He saw
people’s destruction of Dresden, He saw people's destroying of Dresden, and
somewhat differently He saw people destroying Dresden, He saw that people
destroyed Dresden. (The status of these further verbs, e.g. saw here,
presents certain problems to string analysis and to transformational analysis;
the difficulty is resolved in operator analysis, 8.2.)
8.1. Sentence transformations
8.1.0. Word-class grammars and word-choices
In the sentence analyses reached in Chapter 7, the sentence structure was

composed of word-class sequences. In constituent analysis, the components
for each sentence were word-class sequences that were not themselves
sentence structures. In string analysis, as also in the expansion equations of
7.8, for each sentence there was a central component which was a sentence
structure, while the other components were not. As between these two
analyses, string analysis (7.6) not only brought in the advantages of contiguity,
but also reduced the environment for syntactic events from the whole
sequence in a sentence to the internal structure of each string plus the
adjunction relation between strings. This simpler framework is a real
property of sentences, as can be seen not only from its relation to sentence
boundaries (7.7), but also from the fact that transformations, which are a
further property of sentences, can be located in respect to the string
structure of sentences (as is seen in 8.1.4 in the case of English).
However, all word-class components of sentences fall short in certain
respects, as noted in 8.0. For one thing, they do not recognize important
similarities in word class, word-choice, and meaning between certain word-
class sequences. For example, neither string nor constituent analysis provides
in a principled way for the similarities between certain adjunct (modifier)
strings and center strings: e.g between AN and N is A (the long article and
The article is long), or between a relative clause and a sentence minus a noun
(as in which he gave to me yesterday and This he gave to me yesterday). For
another, word-class sequences as components produce not sentences but
only sentence structures: to produce the actual sentence would require
characterization of the specific word-choices, which has always been con¬
sidered a body of data so vast and unstable as to be beyond the powers of
grammar. Finally, many problems concerning the meaning of sentences
were touched upon only tangentially: for example, paraphrase and ambiguity.
8.1.1. The transformational method
Most of these inadequacies are resolved by a further method of analysis, that

of transformations. This method utilizes the word-choice similarities between
the sentence expansions of 7.8 and corresponding independent sentences.
For example, in the expansion equation ANl=N' (as part of some sentence)
the normal choice of particular A for each N\ was the same as in Nlj is A. In
the great majority of cases, one can say that the word-choice in the
independent sentence form N is A is preserved in N which is A and in AN,
which are its transforms. The which is A and the modifier A are transforms of
Nj is A when the latter is adjoined to another sentence containing Nh
The search for preservation of word-choices was a matter of finding pairs
or sets of sentence structures (‘frames’), in which the corresponding three or
more word-class positions were normally (selectionally) filled by the same n-
tuples (pairs, triples, etc.) of actual words. Roughly, one sees this in the
active (N\tVN2 structure, with / for tense) and passive (N2t be Ven by N\)
pair: John will see the picture and The picture will be seen by John, but not
The picture will see John nor John will be seen by the picture. Similarly,
relative clauses have the same word-choices as sentences consisting of
the same word-sequences with the relative pronoun being replaced by a
preceding noun: I read a novel which he wrote and He wrote a novel, but not /
read a lamp which he wrote and correspondingly not He wrote a lamp.
There are two problems in making a precise formulation of word-choice
preservation. One is that in many sentence structures there seem to be
special domains of word u-tuples over which the preservation held. This was
resolved by showing that the different domains were due to preservation in

different sentence structures. Thus in the passive N2 be Ven by Nx we find an
active NltVN2 for The picture will be seen by John but not for (1) The
picture will be seen by candlelight or... by midnight where there is no Candle
light will see the picture; however, we can show that (1) presents the word-
choices of (2) People will see the picture (or Nx will see the picture) plus (3)
People’s seeing the picture will be by candlelight or ... by midnight. By the
side of (2) we have (2') The picture will be seen by people, where by people
can be zeroed—this being itself one of the crucial transformations, discussed
below. Then (2') plus (3), i.e. The picture will be seen by people by
candlelight, can be reduced to (1). In this way we do not have to define
subclasses of N, e.g. of place or time or such ad hoc categories as (in this
case) light, for which we would say that the passive by N is not preserved
from the active subject. Instead, we show that even in these cases the passive
preserves the word-choices of the active, but that the second by N above
comes from a different source.
The other problem in formulating the preservation of word-choice arises
from the difficulty of working with word-choice as data. We cannot list all
the word triples that occur in the NtVN sentence structure, not only because
they are too many but even more because not all speakers agree about all
triples and many speakers are uncertain about some triples (as for example
in (4) John wrote a wall). Furthermore, the list of triples varies with the
context—e.g. (4) might be said about wall newspapers or graffiti; and it
varies with time as subject matters and customary locutions change. In order
to deal with the preservation of word-choice we therefore have to introduce,
as a research procedure, a new degree of freedom on grammars: the varied
acceptability of sentences. Instead of certain sentence structures having the
same word-choices—or not having the same—we will speak of the word-
choices having the same acceptability grading in both sentence structures.
We can reach this grading if we merely recognize that the n-tuples of words
have varied inequalities of unacceptability in each sentence structure of n
word-classes.
8.1.2. Sentence acceptability grading2
To introduce this system of inequalities, we note first that not every word-
sequence is either definitely a sentence of the language or definitely not one.
Many word-sequences are definitely sentences, e.g. / am here, and others
are definitely not, e.g. ate go the. But there are also many word-sequences
which are recognized as metaphoric, or as said only for the nonce, i.e, as
2 Large sections of 8.1.2-3 and 8.1.5-6 are revisions from Z. f tarris. Mathematical Structures of
Language, Intcrscience Tracts in Pure and Applied Mathematics 21, Wiley-Interscience, New
York, 1968.
extended by the present speaker on the basis of ‘real’ sentences, e.g. Pine
trees paint well, perhaps analogized from such sentences as Spy stories sell
well. Other word-sequences are recognized as containing some grammatical
play; and some are pure nonsense sentences (but may be accepted): The cow
jumped over the moon.
It is also easy to find a word-sequence for which speakers of the language
disagree, or readily alter their judgement, or cannot decide, as to whether it
is a grammatical sentence of the language or not: e.g. He is very prepared to
go; He had a hard crawl over the cliff. There can be various degrees and types
of marginality for various word-choices in a form: e.g. the gradation from He
gave a jump. He gave a step to dubious He gave a crawl, He gave a walk to
objectionable He gave an escape.
Alternatively, instead of speaking of different acceptabilities for word-
sequences as sentences, we can speak of different sentential neighborhoods.
For each one of the marginal sentences can be an acceptable sentence in a
suitable discourse, or under not; whereas the completely nonsentential
word-sequence, i.e. those which do not belong to sentence forms (such as go
wood the), appear without acceptability ordering, and only as subjects of
special predicates such as the metalinguistic ones (e.g. in this case,. . . is not
a sentence). For extending the grammar, the neighboring sequences in the
discourse can supply the source of the extension: While you had a walk along
the valley, he had a crawl over the cliff. For nonsense, the neighborhood
might be nonsensical in the same way, as in a fairy-tale or nursery rhyme.
Aside from the obvious cases of marginal sentencehood, many sentences are
to be found only in particular types of discourse, i.e. with particular
environing sentences; and many of these would indeed be but dubiously
acceptable sentences outside of such discourses. Thus The values approach
infinity is sayable in mathematical discourse but rather nonsensical outside
it.
Acceptability does not depend on the truth of the sentence: The book is
here is a grammatically acceptable sentence even if the book is in fact not
here. It does not even depend in any simple way on meaning (as in The cow
jumped over the moon). While there is a certain connection between
meaningfulness, in the colloquial sense of the term, and acceptability, the
connection is merely that a sentence which is acceptable in a given neighbor¬
hood has meaning in that neighborhood. We cannot assert that it is acceptable
because it has meaning (i.e. because the combination of its word meanings
makes sense), for in many cases a sentence is acceptable not on the basis of
the meanings that its words have elsewhere. If Meteorites flew down all
around us is acceptable, it is not because this is a meaningful word com¬
bination, since flew means here not ‘flew’ in the usual sense but some
movement distantly similar to it. Nor can we say that acceptability exists for
every word combination which makes sense in any way, for there are word
combinations which can make sense but are not acceptable sentences: Man
sleep; Took man book. The air swam around me may have low acceptability,
although it can have meaning.
The standing of a word-sequence (carrying the required intonation) in
respect to membership in the set of sentences is thus expressible not by two
values, yes and no, but by a spectrum of values expressing the degree and
qualification of acceptability as a sentence, or alternatively by the type of
discourse or neighborhood in which the word-sequences would have normal
acceptance as a sentence.
The complicated and in part unstable data about acceptability can be
made available for grammatical treatment by means of the following pair
test for acceptability. Starting for convenience with very short sentence
forms, indicated by ABC, we make a particular word choice in each word
class except one, in this case A: say p in B and q in C, written as Bp, Cq. Then
for every pair of members An Aj of A we ask how the sentence formed with
one of the members, At in (1) A,BpCq, compares as to acceptability with the
sentence formed with the other member, Aj in (2) A]BpCq. If we can obtain
comparative judgements of difference of acceptability, we have for each pair
Ah Aj either (1)>(2) or (1)<(2) or roughly (1) = (2).3 In this case the relation
is transitive, so that if A,BpCq>AjBpCq and AJBpCq>AkBpCq it will always
be the case that AlBpCq>AkBpCq. We would then have an ordering of the
members of A in respect to acceptability in ABpCq. If the judgements which
we can obtain in this test are not quantitative but rather in terms of
grammatical subsets (metaphor, joke, etc.) or language subsets (fairy tales,
mathematics, etc.), then some reasonable classification of the subsets could
be devised so as to express the results in terms of inequalities (At being
acceptable in the same subset as Ar or in a subset which has been placed
higher or lower). The results of the test might also take the form of some
combination of the above. It is possible that we would obtain a relation in the
set A such that for each pair either equality or inequality holds in respect to
ABpCq, but without the relation being transitive. The system of inequalities,
obtained from such a test, among sentences of the set ABpCq will be called
here a grading on the set ABpCq\ each sentence A,BpCq which has a grading
in respect to the other sentences of the set will be called a (sentential)
proposition. As seen in note 3, there may be more than one grading over the
A in ABpCq, if there is more than one transformational source for ABC.
Other gradings can be obtained for all members of A in respect to every
We may obtain more than one answer as to the acceptability difference between A,BltCcl
and AjBpCq. For example. The clock measured two hours may be judged more acceptable than
The surveyor measured two hours, in the sense of measuring off, i.e. ticking off, a period of
time, but less acceptable than the latter in the sense of measuring something over a period of
two hours. In all such cases it will be found that there are two or more gradings (in the sense
defined immediately below ) over all members of A in ABpCq\each grading is associated with a
sharply different meaning of ABrCtl and, as will be seen later, a different source and
transformational analysis of ABpCq. Thus the sense in which surveyor is more acceptable than
clock here is the case where measured two hours is a transform of measured for two hours.
word-choice in BC. In the same sentence form ABC we can also obtain a
grading over all members of B for each word-choice in A and C, and so for all
C for each word-choice in A B. We may therefore speak of a grading over the
n-tuples of word-choices in an n-class sentence form as the collection of all
inequalities over the members of each class in respect to each word-choice in
the remaining n—1 classes.
Given the graded 77-tuples of words for a particular sentence form, we can
find other sentence forms of the same word classes in which the same /?-
tuples of words produce the same grading of sentences. For example, in the
passive sentence form N2t be Ven by (where the subscripts indicate that
the N are permuted, with respect to N]tVN1), whatever grading was found
for A dog bit a cat, A dog chewed a bone, A book chewed a cup is found also
for A cat was bitten by a dog, A bone was chewed by a dog, A cup was chewed
by a book. And if The document satisfied the consul may be found in ordinary
writing, but The set satisfies the minimal condition only in mathematical
writing, this difference in neighborhood would hold also for The consul was
satisfied by the document and The minimal condition is satisfied by the set.
This need not mean that an 7?-tuple of words yields the same degree of
acceptance in each of these sentence forms. The sentences of one of the
forms may have less acceptance than the corresponding ones of the other,
for all 72-tuples of values or for a subset of them. For example, Infinity is
approached by this sequence may be less acceptable than This sequence
approaches infinity; but if the second is in mathematical rather than colloquial
discourse, so would be the first.
In contrast, note that this grading is not preserved for these same 72-tuples
in N2tVNi (i.e. a sentence structure obtained by interchanging the two N): A
cat bit a dog is normal no less than a dog bit a cat; A cup chewed a book is
nonsensical no less than A book chewed a cup; but A bone chewed a dog is
not normal, whereas A dog chewed a bone is.
To summarize:
1. Inequalities. Using any reasonable sense of acceptability as sentence,
and any test for determining this acceptability, we will find, in any sentence
form A(X, . . . Xn) of the variables (word classes) X\ . . . Xn, some (and
indeed many) 77-tuples of values for which the sentences of A do not have the
same acceptability. Alternatively, given this A, we will find some 77-tuples of
values for which the sentences of A appear in stated different types of
discourse.
2. Preservation of inequalities. Often, we can find at least one other
sentence form B(Xt . . . X„) of the same variables, in which the same
/i-tuples yield the same inequalities; or alternatively in which the same
n-tuples show the same difference as to types of discourse.
The degree of acceptance, and sometimes even the kind of acceptance or
neighborhood, is not a stable datum. Different speakers of the language, or
different checks through discourse, may for some sentences give different
evaluations; and the evaluations may change within relatively short periods
of time. Even the relative order of certain /7-tuples in the grading may vary in
different investigations.4 However, if a set of //-tuples yields sentences of the
same acceptability grading in A and in B, then it is generally found that
certain of the instabilities and uncertainties of acceptance which these
/7-tuples have in A they also have in B. The equality of A and B in respect to
the grading of these /7-tuples of values is definite and stable.
In this way, the problem of unstable word-choices is replaced by the stable
fact that the same (perhaps unstable) word-choices which order the accept¬
ability of sentences in one sentence form A do so also in some other one B.
A sentence may be a member of two different gradings, and may indeed
have a different acceptability in each (e.g. The clock measured two hours,
below and n. 3 above). It will then have a distinctly different meaning in
each. Some sentences have two or more different accepted meanings;
certain ones of these have normal acceptance as members of two or more
different graded subsets, in which case they will be called transformationally
ambiguous. Consider the ambiguous Frost reads smoothly (in the form
NtVD). We have one grading of a particular subset of sentences, one in
which Frost reads smoothly and Six-year-olds read smoothly are normal,
while This novel reads smoothly is nonsense; this grading is preserved in the
form N tried to VD: Frost tried to read smoothly, Six-year-olds tried to read
smoothly, but not This novel tried to read smoothly. In another grading, of
another subset of sentences. Frost reads smoothly shares normalcy with This
novel reads smoothly, while Six-year-olds read smoothly is dubious; this
grading is preserved in One can read Frost smoothly or Reading Frost goes
smoothly, One can read this novel smoothly, and only dubiously One can
read six-year-olds smoothly.
The relation between two sentence forms in respect to preservation of
acceptability grading of/7-tuples is not itself graded. We do not find cases in
which various pairs of forms show various degrees of preservation of
acceptability grading. The preservation of the gradings is a consequence of
this. In most cases, over a set of //-tuples of values, two sentence forms either
preserve the acceptability grading almost perfectly or else completely do
not. The cases where grading is only partially preserved can be reasonably
described in terms of language change, or in terms of different relations
between the two forms for two different subsets of /7-tuples.
When we attempt to establish the preservation of the grading between two
sentence forms, we find in some cases that it does not hold for all the /7-tuples
of values in both forms, but only for a subset of them. For example, the
4 There are special cases involving language change or style and subject-matter difference.
For example, people may accept equally The students laughed at him. The students talked about
war, but may grade He was laughed at by the students more acceptable than The war was talked
about by the students.
NitVN2/N2t be Ven by N\ relation holds for The clock measured two hours/
Two hours were measured by the clock but not for The man ran two hours (no
Two hours were run by the man). We must then define the relation between
the two forms over a particular domain of values of the variables. This raises
immediately the danger of trivializing the relation: there are many more
word n-tuples than sentence forms, hence for any two sentence forms of the
same word classes it is likely that we can find two or more n-tuples of words
whose relative grading is the same in the two forms. For example, given only
A dog bit a cat, A hat bit a coat, and A cat bit a dog, A coat bit a hat, we could
say that the grading preservation holds between N\tVN2 and N2tVN\ for this
domain. Interest therefore attaches only to the cases where a domain over
which two forms have the same grading is otherwise syntactically recogniz¬
able in the grammar: e.g. because the domain covers all of a given morpho¬
logical class of words (all adjectives, or all adjectives which can add -ly); or
because the domain, or the set of n-tuples excluded from the domain,
appears with the same grading also in other sentence forms and is too large a
domain for this property to be a result of chance.
One can specify the domain of one grading preservation in terms of the
domain of another. For example, the passive is not found for triples which
appear with the same relative grading both in N{tVN2 and in N/tV for N2
(checking all N2, with a given V):
The man ran two hours, and
The man ran for two hours, whereas
Two hours were run by the man is rare or lacking;
as against:
The clock measured two hours, and
Two hours were measured by the clock, whereas
The clock measured for two hours is rare or lacking.
One can also specify the domain in terms of stated subsets of words. Some
of the sets of n-tuples whose grading is preserved as above have some
particular word, or an inextendable subset of words, or the complement set
of these, in one of the n positions. For example, the passive is never formed
from /r-tuples whose V is be, become: no An author was become by himT
In other cases, a subset of words in one of the n positions is excluded from
the domain of the grading preservation, but only in its normal use; the same
words may occur in that position in a different use. In the preceding
example, we can say that the passive does not occur if N2 names a unit of
duration or measurement (mile, hour) unless V is one of certain verbs
(measure, spend, describe, etc.). But in many, not all, of these cases, it is not
5 It will be seen later that most of the problems of restricted domains apply to an aberrant
case: the analogic transformations. These domains are partly due to the compounding trans¬
formation. For the base transformations, the domains are unrestricted, or restricted in simple
ways.
satisfactory to list a subset of words which is excluded from the domain of a

certain grading preservation between two forms, because these words may
be nevertheless found in the domain, but with very low acceptance. For
example, The clock measured for two hours, excluded in the preceding
example, may indeed occur, in the sense of The clock measured something,
over a period of two hours. And for this sense. The clock measured two hours
has indeed no passive. So The clock measured two hours has two distinct
meanings, one (barely acceptable) which appears also in The clock measured
for two hours (also barely acceptable), and one normal one which appears
also in the normal Two hours were measured by the clock.
Finally, since the use of words can (with some difficulty) be extended by
analogy or by definition, it is in many cases impossible to exclude a word
from the domain of a grading preservation (which will be stated in 8.13 to
constitute a transformation). Thus if someone were to use run in the sense of
run up (as to run up a bill, or to run up a certain number of hours out of a time
allotment), he might say They ran (up) two hours and this could have a
passive Two hours were run (up) by them. But the passive would be only for
this sense of They ran two hours, and not for the one which is a grading
preservation (transform) of They ran for two hours. In general, then, the
domain of a grading preservation is a set of /?-tuples defined not by a list of
words in a particular position, but by the fact that the set of /7-tuples
participates in stated other transformations.6
The set of word //-tuples which preserve grading in two sentence forms is
therefore characterized in most cases by syntactic properties, as above. At
worst, the justification for distinguishing such a set of /7-tuples is residually
syntactic (e.g. to complete a set of grading preserving form-pairs). Each set
of /7-tuples, or rather the word class (in one of the n positions) which
characterizes the //-tuple set, will almost always be found to have a semantic
property (e.g. the be verbs, the unit names); but semantic properties can
often be found also for other word sets which do not play a role here.
We therefore require that the grading preservation hold not necessarily
over all word /7-tuples of a sentence form, but over a nontrivial syntactically
characterizable subset of these. This reduction of our requirement is especially
useful in the case where we can partition the //-tuples of a sentence form A
into subsets such that in each subset the grading is preserved as between A
and another sentence form B\ but the grading in the set of all //-tuples of A is
not preserved in B because the order of the subsets of A (in respect to the
gradings) may not be fully preserved in B.
The existence of restricted domains makes a general definition of trans¬
formations difficult; and there may be difficulties in discovering the precise
domain of some particular transformations. However, the fact that many
h Cf. Henry Hiz, ‘Congrammaticality, Batteries of Transformations, and Grammatical

Categories’, Proceedings, Symposium in Applied Mathematics 12, American Mathematical
Society, 1961, pp. 43-50.
transformations have no restriction as to word subsets gives an initial stock

of well-established transformations for a language which helps in determining
various remaining transformations.
Hitherto we have considered sentences the grading of whose word-
choices is preserved in some other set of sentences. However, for all
sentence forms A, except the relatively short ones, the grading-preserving
relation will be found, not only as between A and some other sentence form
B, but also as between A and certain sets (pairs, triples, etc.) of sentence
forms. Thus the form N3 N2 t be Ven by Nx (e.g. Wall posters were read by
soldiers) has this relation not only to Nx t V N3 N2 (Soldiers read wall posters)
but also to the pair N2 t be Ven by Ni, N2t be P N3 (Posters were read by
soldiers, Posters were on a wall), or the pair Nxt VN2, N2tbeP /V3 (Soldiers
read posters, Posters were on a wall). Similarly, the sentence form Nfs Vn
Vsst N2 's An (as in His arrival caused her lateness) has this relation not only to
N2’s An t be Pss Nx's Vn (Her lateness was because of his arrival) but also to
the pair Nx t V, N2 t be A (He arrived, She was late).1
What makes this manageable is that it will be found that there is a small
number of elementary sentence forms, and that a small number of (trans¬
formational) physical differences suffices to connect every sentence form to
one or more of these elementary ones. All sentence portions which appear
as substructures in pre-transformational linguistics (e.g. the compound
noun wall posters) are relatable to sentences (e.g. Posters are on a wall) in a
regular way and with preservation of grading for the word-choices involved.
8.1.3. Transformational Analysis8
We now take two sentence forms A and B, of a given n word classes or

morpheme classes or subclasses (as variables), where A and B differ in some
7 Vn, An represent nominalized verb and adjective (arrival, lateness)', t/SI is a verb whose
subject and object are nominalized sentences.
8 The system of partially ordered expansional equations, which provide an apparatus for
deriving a sentence from an elementary sentence, is presented in 7.8 above, with bibliography.
The possibility of deriving or synthesizing sentences from axiomatic (elementary) ones, in a
grammar which is seen as an axiomatic theory of the language, is outlined in Z. Harris,
[Methods in] Structural Linguistics, Univ. of Chicago Press, Chicago, p. 373. Transformations
proper were developed in the service of discourse analysis, and a first list of English trans¬
formations appears in Z. Harris, ‘Discourse Analysis’, (1952) Language 28, 1-30. How
transformations are a necessary development out of problems of traditional grammar, and the
sentence algebra that results therefrom, is presented in Z. Harris, ‘Co-occurrence and
Transformation in Linguistic Structure’, Language (1957) 283-340, reprinted in Papers in
Structural and Transformational Linguistics, Reidel, Dordrecht, pp. 390-457. Linguistic and
logical properties of transformations have been studied by Henry Hiz in many papers funda¬
mental to the subject, including the works cited in n. 6 above and in ch. 11 n. 12andch. 12 n. 15.
Transformations as well as constituent analysis (‘phrase structure’) have been adapted to a
generative system of grammar for creating sentential word-sequences, in a major series of
publications by Noam Chomsky, beginning with his Syntactic Structures, Mouton, The Hague,
1957; ‘A Transformational Approach to Syntax’, in A. A. Hill, ed., Proceedings of the Third
Conference on Problems of Linguistic Analysis in English, Univ. of Texas, Austin, 1962;
Aspects of the Theory of Syntax, MIT Press, Cambridge, Mass., 1965.
fixed morphemes or small sets of morphemes (as constants of the forms), or

in the order or omission of certain classes, and where in each form the
sentences produced by each set of n values, one for each variable of the
form, are graded as to acceptability or discourse environment. Since a
particular sentence form may have two or more gradings for the /z-tuples of
values of its variables (i.e. may be syntactically ambiguous, as in Frost reads
smoothly), we will use the term propositional form for a sentence equipped
with a particular grading of the /z-tuples (and having only one of the
ambiguous readings, if there are such). We define a transformation A B
between the two propositional forms A, B, over all /z-tuples of their values or
a syntactically characterizable subdomain of them, if the grading of the
/r-tuples is approximately identical for the two. When we define a set of
propositional sentences (propositions), where each proposition is a sentence
with a position in a grading, and is identified by an /z-tuple of values in a
particular propositional form, then each transformation take1 each proposition
Aj of one propositional form A into the corresponding proposition B, of
another form B. Each grading preservation gives thus a partial trans¬
formation in the set of propositions.9
A modification of this definition can be made for the case where the
relation holds between A and a whole set B of sentence forms (end of 8.1.2).
Transformational analysis is thus not primarily an indicator of the structure
of each sentence separately, but rather a pairing of sets of sentences, and so
of the corresponding sentences in each set, preserving sentencehood (which
includes preserving approximate grading of acceptability as sentence).
Transformations generate an equivalence relation in the set of propositional
forms (and in the set of propositions), and impose a partition on them. Each
proposition obtained by a given /z-tuple in one form may be called a
transform of the corresponding proposition obtained by the same /z-tuple in
the other form. If one proposition is a transform of another, it occupies in
the grading of its set the same position as the other does in the grading of the
other set. Transformations are thus defined not directly on sentences, but on
propositions, i.e. on graded sentences (sentences as members of a grading).
Given a transformation A i<-*A2, we now consider the difference between
the morpheme-sequences in A, and in A2. Since the various propositional
forms which are transforms of each other contain the same word classes, the
forms, as word-class sequences, cannot be arbitrarily different from each
other. For two transformationally related forms, the difference in sequence
of words (or morphemes) and word classes is, in general, any of the
following: a permutation of word classes or constants; the addition or
The term 'propositional sentence- or 'proposition- as defined here differs from 'proposition' in
logic, where it represents the set of all sentences that are paraphrases of each other. However, it
is an approach in natural language to the 'proposition- of logic, for paraphrases in language can
be defined only on the basis of propositional sentences. Propositions will later be defined as
particular sentence pairs (8.1.6.2).
omission of a constant; and only in limited ways, to be discussed below, the

addition or omission of a class. And since the individual sentences which are
transforms of each other contain the same /7-tuple of word values, they add
no further difference to the above differences between the propositional
forms.10
Given a transformation A \<-+A2 the permutations, additions, and omissions
which differentiate a morpheme-sequence (a proposition) in A2 from its
inverse image in A, are the same for all the propositions in A2. This constant
difference in A2, which is associated with each transformation, will be called
the trace of the transformation, and may be looked upon as an operation on
the morpheme-sequences of A, yielding those of A2. The trace may be, for
example: in the absence of an expected operator (as in and Mary piano after
John plays violin)-, or in the absence of a required argument (as in He reads
all day from He reads things all day)-, or in an extra argument (as in He stood
the box on end against The box stood on end, from He caused the box to stand
on end)-, or from the fact that a particular subset of words admits the zeroing
of a particular operator (e.g. He spoke two hours is zeroed from He spoke for
two hours, above, whereas He spoke two words does not result from such a
zeroing).
The importance of the trace is that it is a physical deposit in one member of
the transformationally related pair of propositions. Each proposition can be
covered disjointly by segments, each of which is a transformational trace or a
residual elementary sentence (which can be considered the identity trans¬
formational trace). For this purpose, however, we have to accept as segments
certain zero arguments (traces of zeroing, which erase the phonemic content
of certain morpheme occurrences), and certain change-indicators (traces of
permutations and morphophonemic change).
We thus find transformations to be not only a relation among propositions
but also a segmental decomposition of propositions.
For convenience in investigating the set of sentences under these trans¬
formations, we may define the positive direction of the operation in a
consistent way for all Ap^-A, pairs. Here, two criteria are used, which apply
equivalently for some pairs of related propositional forms, and complement
each other in the remaining pairs:
10 The one difference between sentences which goes beyond the difference between their
corresponding propositional forms obtains if a word of an /i-tuple has one form in one sentence
and another in its transform, i.e. if the shape of a word changes under transformation (e.g. -ing,
below). This is covered by the morphophonemic definition of the word; if the change is regular
over a whole grammatical subclass of words, it is included in the morphophonemic trans¬
formations (below). A similar situation is seen when the regularity of morphemic difference
applies in certain cases to families of syntactically equivalent morphemes rather than to
individual morphemes. Thus is some cases -ing and zero are equivalent nominalizing suffixes on
verbs (imposed by certain operators): e.g. He felt distrust or trust, in contrast to a dislike but a
liking. At the same time -ing occurs for all these verbs as a different ‘verbal noun’ suffix
(imposed by other operators): He kept up his distrusting or trusting or disliking or liking.
1. We take the arrow in the direction Ax—»A2 if the number of sentence

forms which include the sentence form or trace present in A2 is smaller than
the number of sentence forms which include that present in Ap. for example,
we write I say this-^This I say rather than This l say—>l say this because the
sentence nominalizations Sn do not exist for the This I say form (My saying
this, no This my saying). If we took This I say as source, then nominalization
would not take place on the source, but would on the permutation to I say
this and again not on the permutation to This say I. However, if we take Isay
this as source, then nominalization takes place on the source but not on the
permutations. This is clearly simpler than when This I say is taken as source.
2. We take the arrow in the direction A \-*A2 if the number of morphemes
(including reconstructible morphemes) in the trace in A2 is not less than in
A\. In many cases the number of morphemes is observably greater: It is
old—■>It is very old\ It is old—>I think that it is old. In other cases the number of
morphemes is the same, and the transformation consists in some other
regular change to a more restricted form: He will come only now—* Only now
will he come. Finally there are transformations in which the number of
morphemes seems to decrease; however, the formulation of zeroing and
pronouning shows that the apparently lost morphemes are still present in the
derived sentence, but in zero phonemic shape (end of 8.1.4, item 1).
There are also other criteria which agree with these. For example, we take
the direction as A \-*A2 if the discourse environments of A, are much more
varied than those of A2.
8.1.4. Survey of English elementary transformations
The specific transformations found in English fall into a few elementary

types or are successions of such elementaries. A check-list is given here as an
example of the range of the transformations in a language.
1 Zeroing of words reconstructible from the environment. This applies to

recurrence of words in specifiable situations (G136), as in John will come but
not Mary from John will come but Mary will not come, and in John advised
me to go from John advised me for me to go. The zeroing is in most cases
optional. When the repeated word is a noun it can also in many situations be
reduced to pronoun rather than to zero, as John wrote and he also telephoned
or John wrote and also telephoned from John wrote and John also telephoned.
In some positions a repeated word is always or almost always reduced, as in
John washed himself from the rare John washed John (for the same person).
There is also optional zeroing of indefinite nouns or pronouns, in many
positions (G153), as in He likes to read from He likes to read things, and in
She likes to watch dancing from She likes to watch people's dancing. The
grammar is made more regular and compact if we assume in these cases that
there was an indefinite noun in the given position. If we do not, then the
many verbs like read have to be both transitive and intransitive; and verbs
like watch have to have not only a nominalized sentence (people’s dancing)
as object but also a verbal noun by itself (dancing).
Finally, in particular environments, there is optional zeroing of words
which are unique or highly favored there (G158). For example, the relative
pronouns which, whom can be zeroed in the book which I needed (to the
book l needed), the man whom I know (to the man l know). It can also be
argued that / ask has been zeroed in obtaining Will he go? from I ask: Will he
go? from / ask whether he will go, and that / request you has been zeroed in
Come! from / request you that you come (this being the only way to explain
the yourself in Wash yourself!). Also it can be suggested that to come or to be
here is zeroed in I expect him from I expect him to be here-, otherwise we
would have to say that expect can have as object not only a sentence (I expect
John to forget, I expect John to come), but also a noun alone (with possible
modifiers), as in 1 expect John. Furthermore, as has been noted, the
gradation of likelihoods of nouns as objects of expect is roughly the same as
that of the same nouns plus to come or to be here as objects of expect. And the
meaning of expect with a noun object is the same as the meaning of expect
with the same noun plus to come or to be here as object.
2. Permuting word classes (3.4.2-4). This is limited in English to a few

positions or to particular words. One type is the ‘fronting’ of the object of a
verb to before the verb or its subject, and the rarer moving of subject and
object: The teachers this deny, more commonly This the teachers deny, and
rarely This deny the teachers, against the normal English The teachers deny
this. This is an alternative to the normal linearization (3.4.1), as is the
interruption of a sentence by a secondary sentence (John—this will surprise
you—is coming back, 3.4.4). The length-permutation and the moving of a
modifier to before its host are apparently true permutations, i.e. rearranging
of previously linearized words. So seems to be also the permuting of subject
and tense after certain adverb-like words, as in Little will he expect this from
He will little expect this, or in Only then will he speak from He will speak only
then. This permutation is accompanied by other changes in such sentences as
Will he go? from I ask whether he will go, and Did he stop? from I ask whether
he stopped.
In contrast, some apparent permutations are not due to any process of
moving words, but rather to a complicated zeroing of repetitions. Such are,
for example. These books sell easily as against They sell these books easily
(G368), and also the passive (1.1.2).
In these transformations the information is preserved, but there are
differences in nuance, especially as to what is the topic and as to the aspect
(completive, habituative, etc.) of the predication G304, 364).
3. Single-word adjuncts. There are a few words in English about which we

can say transformationally no more than that they can be inserted next to a
particular word class in a sentence to produce an expanded sentence: e.g.
very in He is very slow from He is slow. There is no natural way of deriving
the occurrences of very from a more general kind of adverb on sentences (as
in transformation type 4, below), even though such derivation is possible in
French, whence very was borrowed.
4. Sentence nominalization (G40). This refers to a sentence appearing, in

modified shape, as the subject or object of a further verb (G77). As object:
e.g. John draws graffiti in I know that John draws graffiti, I asked whether
John draws graffiti, I prefer for John to draw graffiti, I saw John draw graffiti,
1 saw John drawing graffiti, / know of John's drawing graffiti. As subject: e.g.
That John draws graffiti is a secret, John's drawing graffiti has ended, or That
John draws graffiti is unimportant, or That John draws graffiti surprised me.
When the nominalized verb can take more forms than just -ing, we find for
example The growing of marijuana is prohibited, The growth of profits
doubled. The growth was a dark brown. When the predicate of the nominalized
sentence is an adjective or noun we find also such forms as in / know of
John's sadness from John is sad, and John's anger surprised me from John is
angry.
There are also cases where the nominalized subject or object is not a
whole sentence but only the verb with object: / can read it, I tried to read it, I
tried reading it. He is reading it; similarly He is receptive to it. He is a recipient
of it. These can be treated as a case of the nominalized full sentence, above,
with zeroing of the subject in the nominalized sentence on grounds of being
the same as the preceding subject. Such an analysis fits the meaning, and in
some cases also the history of the forms (G289).
There are also bi-sentential predicates, under which both the subject and
object are nominalized sentences (G85): That he seemed angry caused my
departure. His coming here proves that he is innocent. The predicates can be
(chiefly): verb V, preposition P, or adjective plus P, or noun (N) plus P, or
PNP, all with be to carry the tense (e.g. above, and His departure is with my
approval. His departure is due to my suggesting it. His departure is the result
of my suggestion. That he wrote at all was only with the hope of her
answering). The meaning of the original sentence is preserved, with the
meaning of the added material being added thereto.
5. Conjoined sentences (G85). Finally, there are sentences S' composed of

two sentences Sf, S2, with a conjunction C between them. S, and S2 are not
nominalized, unless the whole of S is the subject or object of a further verb
(as in His writing and her not answering caused a stir). Many zeroings can
take place in S2, rarely in S|, in situations that depend on the type of C
(below). The sequence CS2 can move to many positions within S, and even
before 5): The concert was given inside because it was raining; The concert,
because it was raining, was given inside; Because it was raining, the concert
was given inside. If C is and, or, but, or certain conjunction-like words such
as so, then CS2 cannot move to before S\. The conjunction can also be be P
(or be AP, e.g. is due to): His departure was with (or: due to) my approval.
He departed with (or: due to) my approval. The PS2, APS2 can then move to
in or before 5^ Due to my approval, he departed.
There is also an important conjunction wh-, though it is not generally
recognized as such (3.4.4). Wh- can join S2 provided there is some
segment occurring in both. This segment (a noun with its modifiers, or PN,
or a nominalized sentence) in S2 is then pronouned into -ich, -o, -en, -ere,
etc., adjoined to the wh-, and the CS2 thus transformed usually moves to
immediately after the segment's first occurrence, in 5) (and to before it when
which is, who is is zeroed): given A man telephoned and The same man would
not leave his name, we obtain A man telephoned, who would not leave his
name and more commonly A man who would not leave his name telephoned.
A necessary condition is that the repeated segment have the same referent,
as above. The wh-S2 is called a relative clause.
Types 1-3 above transform free sentences into free sentences, with type 3
and some cases of type 2 adding material to the original sentence. Types 4, 5
transform free sentences mostly into bound sentences (i.e. component
sentences), although conjunctions can be looked upon as transforming two
free sentences into one free sentence. In any case they add material to the
original.
There are transformations which can be analyzed as products of elementary
transformations of the above types. For example, adjectives before nouns
are the product of a wh- conjunction (type 5) acted on by zeroing of wh-
(type 1) whereupon the residual adjective is moved (type 2, noted under
type 5):
A book appeared; the same book is new

—» A book appeared which is new
—» A new book appeared.
Another example is the question, which is the product of the following

successive elementary transformations:
I ask whether John will go or John will not go

repetitional zeroing: / ask whether John will go or not
uniqueness zeroing: I ask whether John will go
permutation with changes: l ask: Will John go?
uniqueness zeroing: Will John go?
This derivation of the question makes it possible to obtain both the yes-no
question (Will John go?) and the wh- question (as in Who will go?) from the
same transformational sequence on two slightly different kinds of source
(G331). Many other transformational relations among sentences have partial
similarities which suggest that they are composed in part of the same
elementary processes. Such is the passive (1.1.2, G362), which contains the
same -en as the perfect tense (The book was written, John has written it) and
the same by as in nominalization (The book was written by John, the writing
of the book by John). Such also is the set of nominalizations based on -ing
listed in 4 above. The composition of these transformations is resolved by
the more powerful methods of 8.2 below.
8.1.5. Algebra of the set of sentences under transformations
In transformational analysis of sentences we have two sets whose structure

can be studied: the set of transformations and the set of sentences. Each
transformation has a domain of sentences on which it can act and a range of
sentences which result from its action. A transformation can be identified
with the range of sentences it produces: e.g. the passive transformation is
characterizable by the set of passive sentences. Then if the domain of some
transformation, say the question, includes the passive sentences, we can say
that that transformation acts on (the results of) the passive transformation
(as well as others). Some transformations (roughly types 1-5 in 8.1.4) are
elementary, i.e. not composed out of any other transformations; the remain¬
ing ones are the products of successive transformations (acting on the same
sentences). A transformation can act on the resultant of another only if its
domain is included in the range of that other. Hence certain transformations
in the language are partially ordered in respect to the monotonically
decreasing domains for which they are defined. And when the analysis of a
sentence has more than one transformation operating at a single point in it,
either they are simultaneous (on the same operand) or the domains of the
successive transformations are monotonically decreasing (G364).
Because transformations are defined as going from a particular domain to
a particular range, they offer a simplification in the characterizing of
sentences: the process of composing sentences out of elementary sentences
through transformations is a finite state process. This is due to the fact that
the domain of a transformation is defined in terms of elementary sentences
and of the transformations which produced the non-elementary sentences in
that domain.
The succession of transformations in forming a sentence can be considered a
multiplication, and it is possible to define an associative multiplication on
the set of transformations of a language. However, the types of elementary
transformations pnd of the non-elementary ones which are their products
are so limited that the devices needed for making the set of transformations
closed under this multiplication are ad hoc. The structure of the set of
transformations would be of interest only if languages contained large
numbers of different transformations, actualizing many different combinations
of the elementary ones.
In any case, the set of transformations provides interesting structures in

the set of sentences, S. In the set S as a whole there is perhaps no
transformation (other than the identity) which can operate on all members
of the set (which includes all sentences that are already the resultants of
various transformations). However, for every member of S there are some
transformations which operate on it to produce a member of S. Every
member of S has a decomposition by members of the set of transformations
into a particular (partially ordered) subset of elementary sentences (or
rather, of occurrences of these, since some may appear more than once in
the decompositions). A transformation’s acting on a sentence does not alter
the composition of the sentence as it had been hitherto.
In terms of 8.2, we can define Oao, Onoo, Ooon (3.1) as binary operations;
and we define all the other second-level operators as well as all trans¬
formations as unary operations: S^>S. Any unary transformation/which is
defined on the whole set 5 as domain,/: S-+S, is a 1-1 mapping of the set of
propositions into itself. The other unaries are isomorphic partial trans¬
formations in the set of S, each / sending one subset (which satisfies the
conditions for the operand of that/) into another subset. The nominalization of
a sentence when it is under a second-level operator is thus not an independent
transformation but an argument-indicator (G40).
As to the binaries: a subset of them acts as a binary composition in the set
S, mapping SxS onto S. For example, and, and to a lesser extent or, may be
found (with somewhat different effects in some cases) between any two S,
even between question and assertion, etc. {Will you go, or else I will go; / will
go, and will you go?). The rest are partial transformations from a subset of
SxS to a subset of S.
Given a suitable list of elementary transformations in a language, each
graded sentence (proposition) in the language can be characterized by a
unique ordered product of elementary sentences and elementary trans¬
formations, aside from commutativity of certain transformational products.
If two transformational products which differ in more than this local
commutativity produce the same word-sequence, then the word-sequence is
an ambiguous sentence, i.e. it represents two graded sentences (and their
meanings).
The special provision for commutative sub-sequences of a transformation-
sequence that characterizes a graded sentence can be obviated if we say that
what characterizes a graded sentence is not a sequence of transformations
but a partially ordered set of transformations, which is identical with the
sequence except that each commutative sub-sequence is an unordered set at
the point in the sentence that was occupied by the commutative sub¬
sequence. In the semilattice decomposition below, for instance, the trans¬
formations at 12 and 13 are unordered among themselves, but the set of
them is linearly ordered in respect to those at 11 and 14.
While the trace of a transformation is the same wherever it occurs in the
partially ordered characterization of a sentence, the operand of the trans¬

formational operation is different for each occurrence of it, and can be
stated in terms of its position in the partial ordering.
The transformations impose two different and important partitions on the
set of propositions (graded sentences).
One partition is according to the trace which each proposition contains:
two propositions are (transformationally) equivalent if they contain the
same ordered traces. The relation R of having the same transformation-
product trace (without regard to the elementary sentences on which the
transformations operate) is an equivalence relation on the set P of pro¬
positions. The factor set P/R is the set of transformation applications. The
mapping of the set Tonto its factor set P/R, assigning to every proposition in
P the product in P/R whose trace that proposition contains, is the natural
mapping of P onto P/R. The propositions which contain the traces of no
transformation are sent into the identity of the set of transformation-
sequences. These sentences are the kernel of the natural mapping, and will
be referred to as sentences of the kernel, or kernel-sentences, K. They
are important because, for each graded sentence, they are the residual
(elementary) sentences Se under transformations. (This requires that sen¬
tence expansion be a transformation, as at the end of 8.1.4.1.).
The other partition is according to the residual sentences which are
contained in each proposition under transformations: two propositions are
in the same equivalence class if they contain the same kernel-sentences and
ordered operations on these. All sentences which are transforms (not only
paraphrastic) of each other contain the same residual sentences. The
statement that sentences Ah B, are corresponding members (with same
word choice) of sentence sets A, B which are transforms of each other is
expressed in transformational theory by saying that A^Bj if and only if
there exists a succession of elementary transformations (including the
identity transformation) which sends one of these into the other. Two
occurrences of e.g. They have reduced wages are not transforms of each
other if one is derived from They reduced wages (via X has the situation of the
same X reducing wages) and the other from They have wages which are
reduced (via X has the situation of Y reducing X’s wages). As these trans¬
formations have been defined, sentences differing only in transformations
must have the same residual sentences. These residual sentences of each
sentence contain no transformational trace, and are therefore in the kernel
of the natural mapping above. In the partition of the set of propositions into
subsets whose members are transforms of each other, each subset contains
one or more kernel-sentences, which are contained in each sentence of the
subset. These kernel-sentences generate the subset by means of the trans¬
formations.
This transformational structure of sentences makes it possible to define
sentences as those objects among which the transformational relation
obtains. A word-sequence is a sentence if it has a transformational relation

to a sentence.
We have seen that each sentence has a partially ordered decomposition
into elements. The decomposition is unique for each proposition, if analogic
transformations are taken as single elements. There are certain restrictions
on the combinations that occur in a decomposition; certain combinations
occur in no decomposition.
With the elements taken to be kernel-sentences and unary and binary
elementary transformations, each proposition of the language can be written
uniquely as a sequence of element symbols requiring no parentheses, for
example in the manner of Polish notation in logic. Certain sub-sequences are
commutative (representing elements unordered in respect to each other).
Because both the words and the transformations are partially ordered,
and because a transformation is carried out precisely when the conditions for
it are satisfied in the sentence (3.3.4), all this structure forms a lattice. For
sentences which contain only one kernel-sentence, the decomposition can
be represented as a lattice with the kernel-sentence as null element and the
given sentence as universal element. The points represent the kernels and
transformations: c is the least upper bound of a, b, if c is the first that can be
applied after both a and b (i.e. on the resultant of a, b), with co-arguments
linearly ordered, on the way from the kernel-sentence to the given sentence,
in this decomposition; and correspondingly for the greatest lower bound.
The transformation at the universal element can be taken as the final
sentence intonation. The decomposition of a sentence containing more than
one kernel is a semilattice; in this case, each conjunction Cis oriented, with a
distinction of right-hand and left-hand, because S{CS2 (which would be the
resultant of ordered 5), S2) is a sentence related, but not generally equivalent,
to S2CSi (which would be the resultant of ordered S2, 5)).
The set of all lattices and semilattices, one for each proposition, has
certain properties, i.e. certain dependencies among the occurrences of
particular transformations in particular relative positions within a lattice.
The most obvious one is that for each lattice, whose universal element is an
arbitrary sentence, the null elements are always a kernel-sentence. For each
kernel-sentence beyond 1 there is precisely one Ooa point which is its l.u.b.
in respect to one of the other kernel-sentences in the semilattice.
As an example of the decomposition of a sentence we take the following:
The adrenal appears more and more as a prime endocrine gland: its
importance grows in the animal scale: among the mammals, it has
become indispensable for life, its removal leads rapidly to death: its
functions are multiple.
This sentence is analyzed in Figure 8.1.
A characteristic of transformational analysis is that languages are rather

similar in their transformational structure, and that given a sentence in one
FIG. 8.1. Semilattice of an English sentence.
1. K: Adrenal is a gland. 19. Ooa: wh-.

2. m000: the. 20. mOoa: the.
3. K: A gland is endocrine. 21. A: A mammal lives.
4. O00: wh-. 22. p: A mammal has life.
5. K: A gland is prime. 23. O0: Adrenal is indispensable
6. K: A gland is endocrine. for a mammal’s life.
7. Oon: wh-. 24. Pronoun:: mammal—fit.
8. Oon: wh-. 25. p: has become.
9. p: appears as. 26. A: Nj removes adrenal.
10. 00: more and more. 27. p: Nj effects removal of
11. A: Adrenal is important. adrenal.
12. Pronoun : Adrenal—* it. 28. Pronoun : adrenal—*it.
13. p: Adrenal has importance. 29. A: A mammal dies.
14. 00: Adrenal's importance 30. p: A mammal suffers death.
grows. 31. 0(>0: Nj’s removal of it leads to a
15. p: Adrenal's importance has mammal’s death.
growth. 32. z: Nj’s^> zero.
16. 00: Growth of adrenal's im¬ 33. mOQ: Nfs removal of it leads
portance is along a scale. rapidly to a mammal's death
17. nr Adrenal’s importance 34. Oao\ It has become indispens¬
grows along (or: in) a scale able for a mammal's life,
18. A: Animals are (ranged) in a removal of it leads rapidly
(
’ scale. to a mammal's death.
language and its translation in another, the decomposition of each sentence

in terms of the transformations of its language will be quite similar. An
example is the analysis in Figure 8.2 of the Korean translation of the first part
(itself a whole sentential structure) of the sample sentence above.11
8.1.6. Structures in respect to transformations
8.1.6.0. Introduction
Transformational analysis creates in a language many subsets of sentences,

among which various structural and semantic relations hold. A few examples
follow.
8.1.6.1. Unambiguous subsets of sentences and decompositions
If the intersection of word subclasses were empty, and if there were no

degenerate results from the paraphrastic transformations, there would be no
grammatical ambiguity; the only ambiguity in language would be due to the
spread of meanings of the words in the elementary entities (kernel-sentences
and second-level operators). However, in natural language, it is frequently
the case that different classes or subclasses of words, distinguished by their
being part of different elementary entities, have some members in common,
which reduces the vocabulary needed for a language. This is a source of
degeneracies in the set of produced sentences. It can result in the same
sequence of words being produced by different transformations on different
kernel-sentences.
If we wish to form a language without grammatical ambiguity, we can
associate each ambiguous sentence with its transformational decompositions
(below), or with a distinguished partial sentence of it (8.1.6.2); in the latter
case we have an unambiguous language which consists only of sentences
(more precisely, sentence-pairs) and not of analyses. Given a word-sequence
which is a grammatically ambiguous sentence, the two or more propositions
11 The translation and analysis is the work of Maeng-Sung Lee.
Fig. 8.1 cont.
35. O,OCT . . . among mammals, 40. K: A gland functions.

36. mO(> the (on second mammal's), 41. Pronoun: gland—* it.
37. z: a mammal's-^zero; the 42. p: A gland has functions.
mammal's-^ zero, 43. Oa: Its functions are multiple.
38. Oo semicolon, 44. Ooa: Semicolon.
39. on semicolon. 45. On: Sentence intonation.
Note: K, kernel-sentence; m, morphophonemic change (7.2;; p, preverb (G289); z,
zeroing.
taukta, pusin un curjyoha n noepunpi san ulo poi a ciko iss ta

more and more adrenal principal endocrine gland seem/appear
More and more the adrenal is coming to appear to be a principal

endocrine gland.
FIG. 8.2. Semilattice of a Korean sentence
1. K. pusin-i san-i-ta ‘Adrenal is a gland.’

2. O,,: u n on pusin (marks the topic-word of the sentence).
3. K: san-i ncepunpi-lul ha-n-ta ‘A gland does internal secreting.'
4. Oao: E-nominalization (equivalent to wh-): —> pusini ncepunpilul hanun sanita
‘Adrenal is a gland which does internal secreting.’
5. z: on the appropriate verb hata: —» pusini ncepunpisanita ‘Adrenal is an
endocrine gland.’
6. k: san-i cuyoha-ta ‘A gland is principal.’
7. K: same as 3.
8. Oot): E-nominalization: —> ncepunpilul hanun sani cuyohata ‘The gland which
does internal secreting is principal.’
9. z: on hata: —> ncepunpisani cuyohata ‘The endocrine gland is principal.’
10. O00: E-nominalization: —» pusini cuyohan ncepunpisdnita ‘Adrenal is a principal
endocrine gland.’
11. p: poi-ta ‘appears’: —» pusini cuyohan ncepunpisanulopointa ‘Adrenal appears
to be a principal endocrine gland.’
which are expressed by the word sequence differ in their decomposition

semilattices. These semilattices differ at two or more linearly ordered
points, the later-operating point obliterating the difference brought in at the
earlier point. For if two lattices have different operations—second-level
operators, sameness statements (G96), reductions—at only one point, and
yet produce the same word-sequence, it would have to be the case that the
different operations at that point in the two lattices introduce the same
words in the same positions of the sentences under construction; but these
words would be appearing, in the two sentences, as parts of different
operations which act in the same way in the order of operations. Whether
this is possible, i.e. whether a given language contains two identically
operating transformations such that a part of one consists of the same trace
as some part of the other, can be seen from the list of transformations for the
language. In English, this does not seem to happen. If, then, we disregard
this possibility, there will have to be two or more differences between two
decomposition semilattices of an ambiguous word-sequence.
Any two decompositions of an ambiguous sentence must be similar to
each other (except for the differences specified below); otherwise they
would not yield the same word-sequence. In most cases, they differ in a
kernel-sentence and also in one or more transformations (which obliterate
the difference in the kernel-sentence). For example, in the ambiguous Frost
reads smoothly, one decomposition (omitting some details) is shown in
Figure 8.3.1. N is used here for a disjunction of all N which might occur in
this position of the kernel-sentence; the disjunction could be pronouned as
anything, things, etc. After point 2 the resultant is:
Frost’s reading of N2 is smooth.
After 3, with zeroing of the indefinite N2,
Frost’s reading is smooth.
Given 4 without 3, we have Frost reads things smoothly; but the combined
effect of 3 and 4 is:
Frost reads smoothly.
The other decomposition is shown in Figure 8.3.2. N is used here as in Figure
8.3.1, but it covers a different subset of nouns, and the corresponding
pronoun would be one. As in Figure 8.3.1, z is the zeroing of an indefinite,
but it applies to an indefinite in a different position.
Fig. 8.2 cont.

12. p: ci-ta ‘comes to: —»pusinicuyohan ncepunpisonulopoiocinta ‘Adrenal comes
to appear to be a principal endocrine gland.’
13. p: iss-ta ‘is . . . ing’: —* pusini cuyohan ncepunpisonulo poidciko issta ‘Adrenal
is coming to appear to be a principal endocrine gland.’
14. Oa: adverb taukto ‘more and more’.
15. Oa: sentence intonation.
5" Oa: sentence intonation: —» Frost reads smoothly.
4" m: Sn is A-+SAly (denominalization).
3" z: of N2 —» zero.
2" Oa: is smooth.
1"
K: Frost reads N2.
FIG. 8.3.1. Decomposition lattice 1 of an ambiguous sentence.
5. Oa: sentence intonation: Frost reads smoothly.
m: Sn is A -*SAly.
z: N{s —» zero.
Oa: is smooth.
1A K: Ni reads Frost.
FIG. 8.3.2. Decomposition lattice 2 of an ambiguous sentence.
After 2, the resultant is:

One's reading of Frost is smooth (or: goes smoothly).
After 3, it is:
The reading of Frost is smooth (or: goes smoothly).
And after 4, it is:
Frost reads smoothly. (If the denominalization were made without 3, it
would yield One reads Frost smoothly.)
In some cases, the kernel-sentences of both decompositions are the same,
and all differences are in the transformations. This requires that the import¬
ation or change of words by some succession of operations in one decom¬
position be the same as that due to some other succession in the other
decomposition. This is possible, since there are several cases in which a
particular word or morpheme (e.g. be, -ing) appears in more than one
transformation. Indeed, one of the motivations in defining the elementary

transformations was to have each segment of a transformation occur in only
one transformation. This result could not be completely achieved in the case
of English. Elence it is possible that identical traces may be produced by
different successions of transformations, especially if some of them are
zeroing, which may zero parts of the trace that is due to preceding trans¬
formations. Any such situations which can arise can be determined from an
inspection of the list and product table of transformations for the given
language.
11. O0: sentence intonation.
10. z: on K2, 3.
9. Ooa: and.
7. Ooa\ and.
3. m: slowly.
2. Oa: is slow.
1. Kl: A man walked.
FIG. 8.4. Decomposition 1 of an ambiguous sentence.
For a simple example, we consider A man walked and talked slowly. One
decomposition is given in Figure 8.4. The resultant sentences after each
point in the semilattice are:
After point 2: 1, A man’s walking was slow.
After point 3: 7, A man walked slowly.
After point 5: 2, A man’s talking was slow.
After point 6: 2, A man talked slowly.
After point 7: 7, A man walked slowly and 2, a man talked slowly.
After point 9: 1, A man walked slowly and 2, a man talked slowly, and 3,
man in 2 refers to the same individual as man in 7.
After point 10, where z replaces K3 by zeroing those words in K2 which
are identical with corresponding words in Kl (in this case a man, slowly)
and the permuting of the residue of K2: A man walked and talked
slowly.12
The other decomposition is given in Figure 8.5. The resultant sentences after
each point are:
After point 3: 2, A man's talking was slow.
After point 4: 2, A man talked slowly.
12 The kernel-sentences are numbered in order of appearance in the lattice.
A man walked and talked slowly. = 9. Oa: sentence intonation.
8. z: on K2,3.
7. 000: and.
5. 000: and.
6. K3: Man of 2. K2: A man talked. 1. Kl : A man walked

K3— 1 is same
individual as man
of K3—2.
FIG. 8.5. Decomposition 2 of an ambiguous sentence.
After point 5: 1, A man walked and 2, a man talked slowly.

After point 7: 7, A man walked and 2, a man talked slowly and 3, man in 2
refers to the same individual as man in 7.
After point 8, where the zeroable words are in this case only a man: A man
walked and talked slowly.
The differences between the two lattices are in point 2 (and 3) of the first
one, which are lacking in the second, and in the differences in repeated
arguments as between point 10 of the first lattice and point 8 of the second.
8.1.6.2. The distinguishing partial sentence
Instead of differentiating the two decompositions of an ambiguous sentence

by the two or more points of difference in each, we can, alternatively,
differentiate them by a distinguished partial sentence in each. In a lattice
presenting a decomposition of a sentence, the resultant at a given point
(i.e. the sentence formed at that point, by the kernel-sentences and the
operations at all lower points) will be called a partial sentence of, or
grammatically included in, the resultant of each higher lattice point.
Two semilattices of an ambiguous sentence have different partial sentences
over a statable section: from the earlier-operating point of difference to the
later-operating point of difference (which cancels the first). Each of the two
semilattices, and hence each of the two propositions which they decompose,
can be distinguished by the least (most included) partial sentence which
does not appear in the other semilattice. If the semilattices differ in kernel-
sentences, the different kernel-sentences are the distinguishing partials.
If not, and one of the semilattices contains a zeroing of a previous operation,
then it is distinguished by the previous zeroable material, while the other is
distinguished by the first partial sentence in which the lack of the zeroable
material is recognizable. Hence for each n-fold ambiguous sentence we can
form n subsets of sentences, each subset consisting of that sentence and the
relevant distinguishing partial sentences. For the examples above, the
sentence-pairs are:
1. (Frost reads smoothly; Frost reads N),
2. (Frost reads smoothly; N reads Frost);
1. (A man walked and talked slowly; A man walked slowly),
2. (A man walked and talked slowly; A man walked).
We can form a set of subsets of sentences as follows: For each ambiguous
sentence, a subset is formed for each different decomposition of the sentence
and the differentiating partial sentences of that decomposition, these being
in some cases simply the kernel-sentences and in other cases the kernel-
sentences carrying a particular transformation. For all non-ambiguous
sentences, the subset is simply the given sentence itself. This set of subsets
of sentences contains no ambiguities, and maps isomorphically onto the set
of propositions of the language.13 This set of subsets can be obtained from
the set of sentences by means of the elementary transformations, and
provides us with a set of sentence-pairs (or more generally sentence subsets)
containing no ambiguities, which is expressed purely in terms of sentences
(subsets of sentences) without explicit reference to the analyses or gradings
of the sentences.14
8.1.6.3. Graphs of the set of sentences
The existence of a unique decomposition for graded sentences makes it

possible to order the whole set of graded sentences in respect to their
decompositions, and to investigate various relations among sentences and
certain properties of the whole set of sentences.
There are various types of graphs which represent the relations among
sentences in such a way that properties of transformations or of subsets of
sentences can be obtained from properties of the graph. First we describe a
graph which gives the grammatical inclusion relation among all propositions
(i.e. graded sentences) of the language.
13 The isomorphism with the set of propositions depends on the fact that no word-sequence
can appear twice in an acceptability-graded subset of sentences. If a sentence is n-way
ambiguous, it appears in n different acceptability-graded subsets of sentences, i.e. in n
partitions of the set of propositions.
14 Since ambiguous sentences have various transformational decompositions corresponding
to the various grammatical meanings, and since the traces of these various transformations
affect the word-repetition which is required in all regular SCS . . . CS and discourse
neighborhoods in which the ambiguous sentence appears (9.1), it follows that each n-way
ambiguous sentence appears in n different sets of SCS . . . CS or discourse environments. The
neighborhood therefore differentiates the different meanings of an ambiguous sentence (i.e. in
most cases a sentence is not ambiguous in its neighborhood, and can be disambiguated by
appeal to its neighborhood), unless the differentiating parts of the neighborhood have been
zeroed.
Consider the decomposition lattice of an arbitrary proposition 5). Here

the points are operators (including the null points which select the kernel-
sentences, and the universal point which introduces the sentence intonation);
the lines connecting the points can be considered to represent a minimal set
of independent partial sentences of 5) sufficient to characterize 5). Because
of the partial ordering of the transformations, some of the partial sentences
of A] do not appear in the semilattice.
We now construct a graph representing all the partial sentences of St, i.e.
all the sentences which can be obtained in the course of any order of
decomposition of 5), as follows: each operator point of the lattice is replaced
by an element consisting of a directed edge (representing the second-level
operator or the transformation) from the end point of the preceding
element, to a terminal vertex (which represents the resultant partial sentence);
except that the null point is replaced by a vertex alone, and the universal
point is replaced by nothing. In addition, each set of n lattice points which
are unordered among themselves is replaced by 2"—2 vertices representing
all the sentences that can be formed by combinations of the n operations
from 1 up to n— 1 at a time (we draw connecting edges first for each pair, then
each triple, and so on, up to the n-tuple of the n points; each set of
connecting edges ends in a vertex representing the partial sentence).
In this directed graph of 5,, whose edges are operations, all the vertices of
the graph are all the partial sentences of 5). The graph for each partial
sentence Sq of 5) gives all the partial sentences of Sq, and is a subgraph of the
5) graph. A particular partial sentence Sp of 5) may also be a partial sentence
of some other sentence S2. In that case the graph of Sp is a subgraph of the S2
graph as well.
We now consider the fact that, in the unbounded set of sentences, every
sentence is a partial sentence of certain other sentences. It is therefore
possible in principle to construct a single such graph for the set of all the
propositions of the language. Each proposition is a vertex, and appears only
once. It is the terminal vertex for all the transformations (edges) that
produce it from its partial sentences, in all the ways in which this can be
done, and it is the initial vertex for all the transformations which produce
another sentence from it. Because the vocabulary and kernel-forms and
transformations are each only finitely many, there is a finite number of edge-
sequences eu . . ., e„, such that for arbitrary eh 1 </'</?, we have that e,
reaches a particular vertex A, after which all that can happen are repetitions
of portions of the edge-sequence entering A,, starting with various preceding
vertices; these can obviously be enumerated. We can then connect (by the
and of logic) all the vertices Ah i=l,. . ., n, into a vertex U\ the and of logic
occurs between arbitrary S’, and is available in natural language. The
sentence represented by U is the informationally maximal sentence of the
language, because it has the following properties: all further sentences
which can be formed transformationally from it contain repetitions, to any
number, of one or more distinguished parts of U; or a referential^ un¬

bounded regress of metalinguistic sentences about metalinguistic sentences
about U, its parts, and its extensions. Since the information contained in any
repetition of a sentence part, or a metalinguistic statement about it, is
statable in terms of that sentence part, we can state the informational
content of any further sentence in terms of the parts of U. Since U is finite,
the number of its parts is finite. Except for these foreseeable and enumerable
further sentences built out of U, each proposition of the language appears
once in the graph of U as a distinguished partial sentence of U and its graph is
a subgraph of the U graph. The construction for U does not necessarily hold
for those sentences in mathematical and other symbolisms which do not
have a structurally equivalent translation in natural language.
Put differently: for a particular language at a particular time, it is possible
to construct a graph of the longest (finite) sentence which does not contain
iterations of distinguished parts of itself. It is then possible to describe in a
(finite) adjoined sentence what are these distinguished iterable parts and
what is the informational effect of iterating each of them (to any number of
iterations).15
We thus have a directed graph of all the sentences of the language, and
for informational interpretation a finite graph. Any sequence of edges
connecting two sentences is a transformational path which takes us from one
sentence to the other, where a transformation is taken positively when going
in the direction of the edge, and inversely when going against the direction of
the edge.
8.1.6.4. Language-pair grammars
The grammars of all languages are not arbitrarily different from each other.
The similarities in basic structure are sufficiently great as to make it possible
to take the similarly formulated grammars of any two languages and
combine them into a single grammar which exhibits that which is common to
the two languages and relates it to the features in which the two grammars
differ. Explicit and rather obvious methods for doing this can be formulated,16
and can be so devised as to exhibit the structural differences between any
two grammars.
In particular, the types of elementary transformations of various languages
are quite similar one to the other. The elementary transformations of a
particular type in one language have approximately the same semantic effect
15 For example, one can state finitely the meaning of any number of iterations of very at any
point in which it occurs; one can state in a (finite) sentence the meaning of n iterations of, say,
son’s or of son’s daughter’s as a function of n. The sentences which describe the information of
each iterable part can then be conjoined by logical and to make one sentence giving the
information that is contained in all iterations of all parts of U.
16 Z. Harris, ‘Transfer Grammar’, International Journal of American Linguistics 20 (1954)
259-70.
as the corresponding ones of the same type in another language, to the

extent that appropriate correspondences can be determined. Furthermore,
as was seen at the end of 8.1.5, if a sentence in one language is trans¬
formationally decomposed and is also translated into another language, the
transformational decomposition of the translation will consist largely of a
similar ordering of corresponding kernel-sentences and operations.
This means that in principle language translation can be effected by
decomposing sentences in one language, translating the components (ele¬
mentary sentences, second-level operators, and transformational instruc¬
tions) into another language, and recomposing the sentence in the other
language. Word meaning-ranges correspond (hence translate) better in
these components than in the given sentence.
8.1.7. Transformations as a theory of syntax
Transformations present a single relation (as method of analysis) which

provides a decomposition of each sentence (8.1.4), and also a decomposition of
the set of sentences (8.1.5). Given the existence of elementary sentences
with their word-choices, it can thus provide a theory of the non-elementary
sentences. As a grammatical system, transformations go beyond traditional
word-class grammar, providing a grammar that characterizes the actual
sentences of a language. They do this not by specifying the word-choices of
every sentence but by accepting as given the word-choices in elementary
sentences, and then indicating in which sentence structures these choices are
preserved. In so doing they reduce word-choice from the unbounded set of
possible sentences to the very large but finite set of all possible elementary
sentences. This is done within the universe of traditional grammar, i.e. not
by investigating the imprecise and unstable word-choices but by showing
stable, grammatically regular construction-differences which preserve word-
choice. This has proved a large step toward a least grammar (2.2).
The transformational grammar, though more compact than the traditional
ones, is more powerful in that it makes many relevant distinctions in
grammatical form. Most importantly, transformational grammar shows that
a word-sequence is not enough to characterize a sentence, because the same
sequence may have resulted from two or more different source sentences,
each with a different transformation, thus creating two or more sentences,
each with its meaning. In contrast, the word-sequence of an elementary
sentence suffices to characterize it, wherefore sentences are best analyzed as
transforms of their elementary sentences. Transformational analysis also
brings together grammatical structures that are indeed related. For example,
This is proof of what I said comes out as a transform of This proves what l
said, instead of the two being unrelated sentences that merely contain
etymologically related words. Also, a great amount of multiple classification
is avoided. Words which have several different types of object are shown to
have a single object-type with transformations on it (e.g. by the side of /
expect John to be here, the form I expect John contains not a different object-
type for expect but merely a zeroing of to come). Nominalized verbs and
adjectives (e.g. absence, and so most abstract nouns) are verbs and adjectives
(absent) under a further operator. And much of morphology is treated as
residues of transformations (e.g. His boyhood was harsh is a transform of
some such sentence as His situation of being a boy was harsh—as indeed it
was historically). The success in doing all this shows that transformations are
not an artifice or short-cut of structural description but a real relation among
sentences and a real property of language.
The inadequacy of word-sequence as sole characterizer of sentences has
led all grammatical models to characterize sentences as (regular) sequences
of (regular) sub-sequences of words, these sub-sequences being in general
not themselves sentences (e.g. constituents, strings). In transformational
grammar these entities which are in general larger than words but in general
smaller than sentences are shown to be simply the elementary sentences of
the language.
With the preservation of word-choice comes meaning-preservation.
Somewhat as two words which have the same selection (co-occurrences) are
synonyms, so are two grammatical constructions which support the same
selection; and indeed the transformations are found to be paraphrastic,
aside from considerations of style and nuance. Transformational grammar
also states grounds for a sentence being ambiguous (if it results from
different transformations on different sources)—this is in contrast to lexical
synonymity and lexical ambiguity.17 Furthermore, because the domains of
many transformations are fuzzy, transformations create various types of
marginal sentences, intermediate between the word-sequences which are in
the language and those which are not. In these ways transformational
analysis brings some meaning-distinctions into grammar, and—what is
more important—separates the meaning-bearing arrangements of words
(e.g. in the structure of each elementary sentence) from the non-meaning¬
bearing arrangements.
In respect to the structure of each sentence, any transformations that take
place in it are located in the partial order of the words that compose the
sentence, since each transformation takes place precisely at the point at
which all the conditions for which it is defined have been satisfied.
In respect to the structure of the set of sentences, sentences can be defined
as objects which are reached from other sentences by a directed path,
which is a transformation defined as being able to operate at the initial
point of that path. In some situations, it may be possible to define a
grid of transformational differences rather than directed paths; in that case,
17 Differently from lexical ambiguity, this grammatical ambiguity is explicit: an alternative
source.
something is a sentence if it differs from other sentences only by some

combination of such elementary differences. In this way sentences can be
characterized not only by their compositions but also by their relations to
other sentences and by their membership in subsets of the set of sentences—
so to speak by their behavior in the set of sentences.
Nevertheless, transformations do not constitute a convenient and complete
system of inter-sentential relations. They are too limited: they are not the set
of all inter-sentence structural differences, because only certain differences
between sentence constructions (8.1.0) preserve word-choice. And although
each transformation is a mapping from one subset of the set of sentences
(the sentences of one construction) onto another (the same-word-choice
sentences of another construction), nevertheless the set of transformations
is far from being the set of all subset-mappings in the set of sentences.18 Only
for a few statable differences between free or bound-sentence constructions
is it the case that (all or a distinguished subset of) the sentences in one
construction preserve the word-choice likelihood gradation of (all or a
distinguished subset of) the sentences in the other construction. We are thus
dealing with a constraint on construction-difference, that is an equivalence
(in respect to word-choice likelihoods) among particular constructions,
rather than a general property of transformability among sentence construc¬
tions.
On the other hand, there are too many transformations for each to be a
separate primitive constraint in the language. Once this fact is taken into
consideration, it is found that one can factor out from the transformations a
small set of elementary transformations: expansions of sentences, such as /
know that John left from John left, and deformations (altered grammatical
relations, usually with addition or replacement of certain fixed morphemes),
as in A lynx is what I saw or my seeing a lynx, both of them from / saw a lynx.
The transformations that are found in a language are either elementary ones
or else products (successive applications) of elementary ones. The set of
transformations thus has a certain structure in terms of elementary trans¬
formations. However, by no means all possible products of elementary
transformations are found as transformations in use in a given language.
Hence we do not have a composition in the set of elementary transforma¬
tions, although we have a decomposition of the existing transformations of a
language (6.3).
With all the systemic contributions of transformations, they fall short not
only in the respects noted above, but also in some not unimportant details.
For example, the relation of ‘deriving’ one sentence from another is too
ls Thus He painted two hours and He painted for two hours are transforms of each other, but
He painted two boys and He painted for two boys are not. (Hours here is in a different subclass of
N. namely that of time words, than is boys.) The artist's painting was of two boys and Two boys
the artist painted are transforms of The artist painted two boys; but Two boys painted the artist
and The boy painted two artists are not.
general. The sentences which are the derivational sources of a transformed

sentence are actually present in it as components: derived sentences are
simply abbreviations of, additions to, and rearrangements of, their source
sentences. In addition, a few issues are left unresolved. One is: what
determines the structure and word-choices of the elementary sentences?
Another: what determines the particular elementary transformations, and
what gives them the power to be paraphrastic to the grammatical structure of
their operand (source) sentence (i.e. to sustain the same word-choices)?
A third: why can transformations operate in the same way on derived
sentences as on elementary ones? That is, what gives the resultant of many
transformations (not all, e.g. not the interrogative) the same overt structure
as operand for further transformations? These problems are resolved in 8.2.
8.2. Operator theory
8.2.0. Motivation
Transformations, as we have seen, are word-choice-preserving differences

among observed sentences, which can derive one sentence from another,
and ultimately from distinguished elementary sentences. But it has been
seen that these differences are not sufficiently varied, nor do they combine
sufficiently freely, to constitute an overall system of derivations in the set
of sentences. We therefore consider an alternative (operator-argument)
theory in which transformations are taken to be the effect of something
else.1’' In particular, we utilize the limitedness of the set of transformations
in such a way as to obtain them as a further step on a more general process
that forms elementary sentences directly out of words. To this end, we
observe that preserving word-choice is equivalent to saying that any two
sentences which are transformationally related to each other are carrying
the same base sentence (4.1): not two sentences with the same word-choice
but two forms of one sentence with its word-choice. The transformations can
then be seen as (a) simply further word-choices adjoined to a sentence
already formed, and (b) shape-changes in words that have entered the
sentence.
In type (a) we have increments. These increments are found to have
approximately the same kind of structure as the sentence on which they
act. In English, this structure is roughly a tensed word or phrase (called
operator) after a noun phrase or nominalized sentence (called subject or first
argument) and possibly before a second and third argument (object). Thus
in the underlying sentence Sheep eat grass we can have the increment / know
19 Cf. M. Gross, Methodes en syntaxe, Hermann, Paris, 1975, p. 9: ‘Les transformations

pourraient n’etre qu’un dispositif experimental qui permet de decouvrir et de localiser des
contraintes syntaxiques.’
. . ., producing (as a sentence in its own right) / know sheep eat grass, where
the status of know between its two arguments (/ and eat) is like the status of
eat between its two arguments (sheep and grow). (These are 3, 4, 5 in 8.1.4.)
As to the type (b) transformations, which delete words, or insert a fixed
word or affix, or permute words, these are found to be the resultants of a few
kinds of change in form: primarily reduction of the phonemic shape of words
in the underlying sentence, and the moving of certain words toward the
beginning of the sentence. The underlying sentence, with its word-choices,
remains in the transformed sentence. (These are 1, 2 in 8.1.4.)
The particular development from a transformational to an operator-
argument theory is motivated as follows. We note that each transformation
from simpler source sentence-forms as domain to more complex resultants
as range has more restrictions in the range than in the domain, because each
transformation acts on all or a proper (restricted) part of its domain. Thus in
the resultants of the passive, America was left by me last year is dubious, i.e.
somewhat restricted, in contrast to the normal America was left by the best
writers of the generation and the uncertain The office was left by me at 5,
whereas in the active source I left America last year. The best writers of the
generation left America, I left the office at 5 are all equally normal. In the
English interrogative, if we compare Who met her?, Whom did he meet?, Is
he coming?, Did he come? with the sentences of the source domain He met
her, He is coming. He came, we see in the interrogative form restrictions in
the position of whom (compared to that of her) and in is and the tenses.
This greater restrictedness in the resultants of transformations leads us to
seek other pairs of sentence-forms which satisfy the conditions for trans¬
formation and in which one form is less restricted than the other and is
therefore to be considered the source. One such case is the many affixes for
which an affixless source can be found. A trivial example is pro- before
nouns, derivable from for. He is pro-war from He is for war or He is in favor
of war. Here, ‘X (e.g. pro- above) is derivable from a source Y (e.g. for)'
means, in the observational terms of hearer acceptability, that for all
sentences of the given form containing X, the sentences with Y in place of X
are ordered (as to acceptability) roughly like those containing A: Women are
less pro-war. Children are pro-air, Rocks are pro-universe have among them
much the same acceptability ordering as do Women are less for war. Children
are for air. Rocks are for the universe. In most cases, the sentences with Y
replacing X are a subset of the sentences containing X. Another example is
tin- before adjectives. Under this condition—before adjectives—im-has
the same combinations as less than rather than as not: untrue (but not
unfalse), unhealthy, (but not unsick), unlike (but not undifferent; hence also
unscholarly for unlike a scholar), are replaceable by less than true (but no less
than false), etc.; hence we would consider these occurrences of un- to be
reductions of less than. A suffix example is -less, e.g. in hopeless, derivable
from a free word less, ‘without' (e.g. in a presumed source less hope), which
is still visible in a year less a day. Less obvious is the derivation of -hood from
state of (being) or the like, asm boyhood from state of being a boy, where it is
known that -hood is a suffixal form of a Middle English free word had which
in free position must have been replaced historically by some such word as
state (G 175-6).
In establishing such derivations, we consider the current affixes, restricted
in their positions and in the words to which they are attached, as reductions
of currently existing free words lacking these restrictions. As another
example, note that when we establish adequate derivations for the inter¬
rogative forms, e.g. I ask whether he came as source of Did he cornel, we find
that the source is free of the special restrictions of the interrogative (G 331).
As we establish least restricted sources from which the various more
restricted forms can be derived, we note that these least restricted forms
have a common structure. The structure consists of certain words, chiefly (1)
verbs, (2) adjectives, (3) prepositions, and (4) con junctions, all of which will
be called operators, appearing after or between (a) nouns or (b) sentences
(including nominalized sentences), which will be called their arguments;
(la) John coughed, John poured water, (lb) John's coughing stopped, John
knows she left, John's coughing caused her to leave, (2a) John is tall, (3a)
John is for war, (3ba) John left for Europe, (4b) Her leaving is because of
John's coughing. She left because John coughed.
This structure is so widespread that it becomes of interest to take every
sentence form that does not fit this structure and seek to derive it from one
that does, even if no restriction is circumvented by the derivation. An
example is the deriving of adjectival adjuncts and relative clauses from
secondary sentences: e.g. / saw a tall man from I saw a man who was tall from
I saw a man; the man was tall, where the secondary-sentence intonation,
written as a semicolon, is the conjunction between the two sentences (type
4b above).
To see an extreme example of a word that does not fit this operator-
argument structure, consider yes. This word seems to be outside the
grammatical classes and their co-occurrence conditions. However, note the
following: to the question Did you close the door? a person who has done so
can truthfully answer Yes, or its equivalent here Yes, l closed the door. But
if, upon hearing the question, he closes the door and then says Yes, or Yes, /
closed the door, he is dissembling and not quite truthful. This means that the
Yes is not related to the tense of the answer (/ closed the door, which he had
indeed done before answering), but to the tense of the question: Yes could
be said truthfully only if the door had been closed before the question was
asked, since the question was in the past tense. Hence yes is a referential
word (5.3), containing a repetition of the question (and not a newly made /
closed the door). Yes is thus an O,, (information-giving) operator on the
question-sentence, something like / say yes to your saying ‘Did you close the
door?’. It is thus not unrelated to the operator-argument structure.
It turns out that we can find operator-argument sources for almost all
words and sentences of the language, with few definite restrictions on word-
combination other than those of the operator-argument structure. Further¬
more, the transformations that produce the derived sentences out of the
sources are found to consist almost entirely of reductions in the phonemic
composition of words, specifically of those words which add little or no
information to their operator or argument in the sentence. (Nevertheless,
various languages have important and common transformations (deform¬
ations) which are not readily analyzable as reductions, such as are seen in
The organizing of the event was his, and Her cooking is mostly of vegetables,
or in pairs such as John is the author, The author is John.) We thus approach
a situation in which, first, the only restriction on words in the source
sentences is the operator-argument relation—i.e. this relation creates
sentences—and, second, virtually all other sentences can be considered to
be merely reduced (or otherwise transformed) forms of their source sentences.
This is the system described in Chapters 3-5.
8.2.1. Operator-argument analysis
Given these considerations about sentences, we can state a partial-order

structure which is common to all S. The dependence relation among the
words of a language creates a partial ordering on the words in a sentence. In
the unreduced sentences, each occurrence of a word must be accompanied
by an occurrence of words which satisfy the dependence: in Mussolini
thought Germany would defeat England, the occurrence of thought depends
on two words: one which depends on nothing further (here Mussolini) and
one (defeat) which depends on further word-occurrences; and the occurrence
of defeat depends on the occurrence of two words which depend on nothing
further (Germany, England).
Furthermore, a single occurrence of a word (or argument) cannot satisfy
the dependence requirements of two different word-occurrences (as oper¬
ators): the argument has to be free for the operator, i.e. not an argument of
any other operator-occurrence. This is seen when we consider the com¬
ponents of a sentence. If in a given sentence a depends on m{, m2 and b
depends on n i, n2, and c depends on a.b the question is whether one of the m
might be the same word-occurrence as one of the n. The occurrences of a and
b had to be accompanied by their arguments before c operated on them. But
before c did so, each of a and b with their arguments constituted independent
sentences m\am2 and nxbn2. No word-occurrence could have been common
to these two sentences before c acted on them. If, in the extended sentence
formed by c, one of the n is the same word as one of the m, that n may be
zeroed but its zeroed occurrence would still be satisfying one of the depend¬
ence requirements of b, while the m, unzeroed, satisfied only a and not b.
Compare (1) / locked the door after washing with (2) I locked the door after
washing myself. In (1) it might seem that I serves as first argument (subject)
for both lock and wash; but in (2) we see from -self that wash had its own first
argument, which was zeroable as being a repetition of the one for lock. Thus
auxiliary-like verbs (‘preverbs’, G289) such as take in I'll take a look, and
even can in I can go, and have in I have gone are not taken here as some
special expansion of the following verb, but are analyzed as being operators
with their own subject, which is however the same as the zeroed subject of
the following verb (which is the second argument of that take, can, have).
We thus arrive at a theory not merely of sentence transformation, or of
derived sentences, but of sentence construction ab initio: it deals not with
relations among already-extant elementary sentences, but with the formation
of word-sequences, first into elementary sentences and then in the same way
(i.e. by satisfying the dependence) into all base (unreduced) sentences, and
finally with the reduction or deformation of these sentences to yield the
other sentences. This theory represents first a dependence process of word
juxtaposition (operator on argument), with inequalities of the likelihood
(from zero up) for individual words to occur with particular other ones. On
these word juxtapositions there acts the shape-changing process (kindred to
morphophonemes, 7.4), which consists primarily in reducing words whose
likelihood in the given operator-argument situation is very high. These
shape-changes on words, when suitably chosen, can produce the same
'derived’ sentences as can the transformations on sentences. The 'derived’
sentences are thus the ‘same’ sentence, with the same meaning.
When this is done, it is seen that the cumulative construction of a
sentence, and the likelihoods in word-choices, and the word changes, are all
defined in terms of the finite (though large) set of operator-argument word-
pairs (or triples) rather than on the unbounded set of sentences. It is seen
that the high likelihood on which word-reducibility depends is apparently
definable, in a complex way, in terms of word-pair frequency in specified sets
of discourses, whereas the transformations and the well-definedness of the
set of sentences depended on subjective evaluations of sentence acceptability.
And since the changes are only in the shape of a word (even if to zero) and its
position, but not in its having been chosen, or in its operator-argument
relation to other words in the sentence, it follows that the choices and
operator relations of the words in the sentence—on which the meaning of
the sentence depends—are not altered by the changes, something that does
not clearly follow in the case of transformations.
The two characteristic processes, of operator-argument dependence and
reduction in the shape of high-likelihood words, leave room for additional
processes or entities which may appear in one language or another. For one
thing, the operator-argument dependence creates a partial ordering of
words. Since speech is in essential respects linear, mapping the operator-
argument order onto a linear one leaves room for cases of alternative
linearization; some of the apparent permutations of words are due to this.

For another thing, the high likelihood is only a necessary condition for
reduction. The absence of an a priori sufficient condition leaves room for
various languages to favor or disfavor reductions in particular circumstances:
for example, English does not normally zero the subject of an independent
sentence, even if it is indefinite or repetitional (no Is quiet or Are coming?).
The fact that all vocabulary choices are either operators or arguments, even
if in reduced form, does not preclude the existence, in many languages, of
additional phoneme-sequences that have a grammatical status, especially
such as serve to indicate the operator-argument status in the given sentence,
and such as are determined by the course of constructing the sentence (e.g.,
the -ing which shows that a verb is appearing as argument of a further verb).
In addition, the phonemic distinctions and phonetic composition of words
are arbitrary in respect to the operator-argument relation, leaving room for
minor structures and processes (e.g. assimilation) involving these.
There are also processes that are independent of the two characteristic
ones, but whose effects can be reinterpreted mostly or entirely as though
they were due to the two characteristic processes above. One such process
is morphology, found in many languages, whereby words have affixes
indicating plural, tense, feminine, diminutive, or the like. Such affixes may
go back to pre-syntactic times, but once they are embedded in sentences that
are constructed by the operator-argument relation they can be analyzed in
the latter terms, i.e., for example, ten stones, as though from stone and stone
up to ten, or tigress from female tiger. Another at least partly independent
process is analogy, whereby for example a derived form (affix or word
combination) may be used with a given word A because it is commonly used
with some similar word B, even though the underlying operator-argument
form existed for B and never did for A. Again, we can analyze the case for A
as though it came from the required operator-argument form on A. The
justification is that the underlying operator-argument form in B is what gave
the derived form in B its grammatical ability to carry its meaning, and if the
latter exists also for A it is by reference to a similar operator-argument
relation posited for A.
It is seen from this survey that while the two characteristic processes
suffice to produce and characterize all sentences of a language, a few other
processes may be in effect, but in a way that is limited and that leaves the
basic processes as framework.
The methodological novelty in the operator-argument theory is to speak
not of the restrictions in word combination but of the constraints that
produce these combinations—constraints both in the building process and
in the reduction. A single constraint, repeated, builds both the elementary
sentences and the ones that these expand into. The applications of the
constraint in the building of an utterance are partially ordered, and the
reductions in the utterance can be located in this partial ordering, so that
the utterance is built and reshaped stepwise. Given the operator-argument

word-pairings with normal likelihood, we obtain the actual sentences of the
language.
A further step here beyond transformational theory is that the reshapings
are not an arbitrary catalog of whatever is needed to derive all observed
sentences from their sources, but a particular kind of reduction and position-
change. Then for each given sentence that does not have the directly visible
operator-argument (predicational) form there can be found a paraphrastic
unreduced sentence having that form and differing from the given sentence
only by certain ordered applications of the stated reshapings. The unreduced
(base) sentences are less constrained than the reduced forms, since reduc¬
tions are additional constraints in shape and often also in the conditions they
require for taking place. Hence the grammar of the unreduced sentences is
simpler.
The relation of the paraphrastic sentences to the base ones is that of
reshapings of an original shape rather than the general relation of a derived
sentence to a source, since the original unreduced sentence remains present
in all reshapings, only with stated phonemic reductions or limited position
changes: the source of the sentence is inside that sentence itself. In respect to
structure, the base sentences and the reduced ones are not quite comparable.
For the base sentences, there is a single structure which fully accounts
for each of them. In the reduced sentences, however, there are various
unrelated local constructions affecting various parts—plural, tense, paradigms,
clauses—with certain subsets of sentences having constructions that differ
from anything else in the grammar, as in the interrogative form, or in
sentences of the type of It is time that he left.
Thus traditional grammar does not offer a single structural system adequate
for all sentences. As to the theories of strings and of transformations, these
center on important features of all sentences—contiguity and derivability
respectively—but they do not systematize all that is formal about the
utterances of a language. In contrast, the operator-argument theory presents a
single small set of constraints which produce all sentences from the words of
the language, with provision for distinguishing the normal ones from the
rare and the marginal, and for having such other factors as contiguity,
change, and analogic formation. The theory deals entirely with the data of
sentences, as objects or events. It is based on sentences as said: and it is
extended to those that ‘can’ be said, by potentialities arising from construct¬
ing a least grammar for those that are indeed said, normally or rarely. This
means that the structure of language which expresses most generally the
relations—or constraints—of the sentence-parts is not necessarily quite the
same as the composite structural guides which determine how a speaker
makes a sentence or how a hearer understands it.
In certain respects, operator theory is an end-point of least-grammar
theories of syntax, since it shows that the meaning-bearing arrangements of
words have the regularity of a mathematical object, while the reductional-

transformational rearrangements are not meaning-bearing and do not have
such regularities.
8.2.2. Adequacy of the analysis
We consider now the adequacy of the operator-argument theory. There are

certain data situations which would falsify the theory if they were found in
language, but they seem to be always avoidable. One such is if word
combinations are entirely random, or could not be described by any regu¬
larities. A less crucial, and encapsulatable, problem would arise if there
were operators on N— the lowest arguments—which did not create a
sentence. This would be the case if it were impossible to derive the plural
from an operator on a sentence instead of from some operator on a noun, or
if it were impossible to derive and between nouns from and between
sentences. Yet another (perhaps circumventable) situation that would falsify
the theory would be if an operator on an operator could be different from an
operator on a sentence: that is, if an operator could act differently on a verb
without the presence of that verb’s arguments than with them (in the former
case the verb would be appearing without satisfying its argument-require¬
ment). Even more problematic would be the ability of an operator A under a
further operator B to take an argument X which A could not otherwise take,
since A had to have its argument before B operated on A. In contrast, no
problem arises from A refusing an argument X when A is under B\ for then
we would simply say that when A is carrying X, then B does not select A (i.e.
B does not operate on any occurrence of A that carries Y).
A serious but circumventable difficulty would arise if there were words
which could take, after being reduced, additional co-occurrents (selection
within its operator or argument classes) that they could not take before
being reduced. This would not accord with some statements in the theory. It
would also mean that the base does not carry all the information in the
language. If such an apparent situation is found, one would try to show that
the reduction contains some additional word, in zeroed form, which is
responsible for the added co-occurrents.20 In contrast, no problem arises if
a word X, after being reduced, loses the ability to take certain co-occurrents
A that it could take before being reduced. For this effect can be obtained
simply if the domain for the reduction does not admit A, i.e. if the reduction
takes place in X only when X does not occur with A.
The present theory arrives at a mathematical formulation for the partial-
ordering relation, though not for the reductions. There are certain situations
20 This analysis is needed because some words when under particular operators change their
further co-occurrents (3.2.2). Hence one would seek a zeroed intervening operator in the
reduction.
which would obstruct the mathematical formulation—unless they can be

shifted into the reductional description. One of these is the case of operators
rejecting certain words in their argument class. The only serious example of
this in English is that of the auxiliaries, can, may, etc., which reject any
second argument whose subject (first argument) is not the same as their own
subject. In John can drive we say that the second argument of can is John
drives, with John requiredly zeroed as a repetition (the argument cannot be
drive alone, since drive must have its subject with it). But the second
argument after John can must not be Tom drives. Such an exclusion violates
the rule that an operator can take any member of its argument class, even if
very rarely. One can only seek some other operator (e.g. able) which is not
limited in this way (even though I am able for Tom to drive may never be
said, on semantic rather than grammatical grounds), with can being an
optional variant when the two subjects are the same. (This is not the history.
Some auxiliaries had been verbs which fairly recently accepted a different
subject in their second argument. However, we can say that when they
ceased to accept this they ceased to be operators in their own right and
became variants of the semantically closest verb which was not thus restricted.
Such analyses can be made with good reason in some other cases of
restricted words.)
In many languages we also meet certain situations which detract from the
advantages of the theory, but which present a problem only if they are
widespread. Even so, the problems can be circumvented. The chief example
is of sentences which cannot be obtained by known reductions from an
operator-argument structure: e.g. certain borrowed foreign phrases, and
partially frozen expressions (such as the more the merrier), and linguistic
jokes (such as let’s you and him fight). One might say that these constitute a
small set of locally asyntactic word-sequences which function like an addition to
the vocabulary. Or one can create ad hoc transformations to derive them
from related regular sentences (perhaps by the actual historical route).
Somewhat similar is the case of any morphology which we cannot derive
from an operator-argument structure, or which one might consider to
involve too costly a derivation, e.g. the plural suffix. Here one might simply
say that the language contains pre-syntactic N pairs with added singular-
plural meaning (and with the plural-agreement restriction), e.g. stone-
stones, with one or the other occurring in a given N position according to the
meaning that is intended.
Languages whose vocabulary elements are not in general free words do
not interfere seriously with the operator theory. In Semitic, most words can
be constructed in a regular way from consonantal roots possessing dictionary
meanings, intercalated with vowel 'patterns’ possessing grammatical meanings.
This construction does not preclude the operator-argument analysis if we
can reasonably choose one root plus vowel combination as the base word,
and consider the various shifts to other vowel patterns as various operators
on the base; it is indeed reasonable to suppose that such may have been the
distant history. Such a solution is clearly appropriate in Semitic, where the
simplest verb form can be the source from which other verb forms and
related nouns are obtained; in addition, some simple nouns would be taken
as sources for verbs.
More generally, the operator theory does not intrinsically require that the
source form of operators and arguments be free words. The fact that to some
extent affixes can be derived from free words makes it possible to see the
similarity in operator-argument structure between languages which have
morphology and languages which do not. But it is not essential for the
present theory.
The problem is different if we wish to construct words like the English
conceive, perceive, contract, etc., let alone dazzle, frazzle, etc., out of two
free words, and more generally out of an operator-argument relation.
There is not enough regularity to permit this in English. Furthermore, the
small number and encapsulated nature of these composite words allow us to
consider them as indivisible entities in the syntax; their composition is then
a semantically relevant pre-syntactic structure in the vocabulary.
Similarly, if nouns consist of a stem plus a required case-ending, we would
take the ending, at least in the nominative, as a required indicator of
argument position. Then for the operator-argument analysis the stem
would function as a free word, though in being said it would carry the
indicator. Likewise, if nouns have gender affixes with no consistent male-
female pairing, the affixes could be considered argument indicators; but if
they have paired male-female affixes, one of those affixes (depending on
morphophonemic convenience) would be considered as (male or female)
operator on the base noun (which, in the absence of this operator, would
mean the other gender and would contain the other affix as noun indicator
rather than as gender indicator).
Another aberrant situation is that of reductions which are required rather
than optional. When the required change a follows on an optional one b, we
can say that the ordered pair ba together is optional. In (1) the glass which is
broken, the zeroing of which is is optional, since (1) exists; but once which is
is zeroed certain residues must move to before the host, yielding (2) the
broken glass', the two changes together change (1) optionally into (2). When
the required change acts directly on the source sentence, we cannot make it
part of an optional event .John thinks that he is right must be derived from (3)
John thinks that (the same) John is right, but (3) is not normally said. The
ultimate sources, the elementary base sentences—what were called the
kernel-sentences in transformational theory—indeed do not have required
reductions, but the extended sentences of the base, such as (3), may. This
means that not all base sentences are actual sentences of the language, even
though complicated ad hoc restrictions would be needed if we were to
formulate a grammar that would routinely exclude (3). As long as required
reductions hold for only a few grammatical situations, such as in (3), we can
still claim that the base is a subset of the sentences of the language (i.e. that
all source sentences exist in the base)—aside from the few required reductions
—and that all information statable in language is statable in its base subset.
If, however, required reductions in base sentences were many and diverse,
this claim would have to be given up; the base would then become a
grammatically regular set of forms that underlie the sentences of the
language rather than being a subset of it.
Another problem can arise in the case of analogy. In this process, given a
form Ah together with Af differing from it by certain reductions (or additions
or replacements of words), and given 5, (but no 5y), there is formed a new BJ
having the same overt difference from B,. In most cases, Bf is then under¬
stood as though it had the same reductions or other grammatical differences
from B, as Ay has from A,. In some cases, however, the words of B do not
admit of the changes made in A, or By may have been made on the model of
Aj without Bi existing in the language. If these latter difficulties are very
numerous, the explicatory value of the present theory for the individual
sentences of the language is decreased.
There are also certain properties which are common in languages, but
which are not essential to the adequacy of the theory, although some of them
increase its utility or simplify its recognition of sentence analysis. One is the
fact that reductions are ordered in respect to the operator-argument
ordering. Another is the availability in many languages of operator indicators
and argument indicators; e.g. case-endings are argument indicators for the
zero-level arguments N. A third is having different words in the different
dependence classes N, On, Oa, etc., with only few words multiply classified.
If a language had a great many words which are members of several classes,
then the classes, which in any case are defined as relations among word-
occurrences (‘word tokens’), lose the support of being distinguished by their
word members (‘word types’).
Some features of traditional grammar are readily describable in operator-
argument terms. Such are pronouns (as repetitions); the infinitive (as an
argument indicator making operators into arguments); the status of adjectives,
prepositions, and conjunctions (as operators); the interrogative and im¬
perative (as metasentential operators); the absence of to under auxiliaries
such as can (as zero variant of the infinitive). Other features are more
difficult. Such are, in English: the various nominalizations; the borderline
beween ‘indirect’ objects and adverbial modifiers as in rely on something
(object), sit on something (object or adverbial phrase?), stand on something
(adverbial phrase); the requirement for a before ‘count’ nouns in the
singular. There are also problems with many individual words whose use is
restricted or specialized in various irregular ways: e.g. afford (He can afford
to buy it; no He affords to buy it), supposed to (which differs in selection and
meaning from suppose: I’m supposed to be there tonight is not paraphrastic
to People suppose me to be there tonight). In some cases the difficulty of a

regular description is due to a construction being in flux, historically.
The regularities which can be stated, secondarily, in terms of the operator
-argument structure may fail to hold in some cases. For example, each
operator when combined with its arguments forms a sentence. In almost all
cases the resultant sentence can stand by itself, in suitable linguistic or
situational context. However, there are in English certain operators which
produce sentences that are not comfortable unless some further operator
acts upon them. This happens in English for many intransitive verbs. Thus
enter operating on John yields John enters; but this sentence would not
normally be said by itself (except in stage directions). Instead we have John
is entering, John enters before everyone else, and the like.
For some features of grammar, either the operator-argument source
or the derivation from it is felt to be artificial, though technically
grammatical; but precisely these sources have grammatically explanatory
advantages. In English this is the case for the tenses (G265), for the
plural and quantifiers (G244), for the passive (G362), for not (G221), and
to a lesser degree for affixes, referential, possessive 's, the, and the
demonstrative pronouns.
All of these difficulties are treated either by admitting special forms or
special domains of known types of reduction, or by creating certain of the
base (source) sentences or intermediate sentences of the derivation as
reconstructed sentences which no one would ever actually use in speaking
(and never or hardly ever in writing); in extreme cases the reconstruction
may be marked by a dagger (t). For example, for John drives slowly, which
is not in operator-argument form, we reconstruct John drives; said John's
driving is in slow form (or even ... is in a form; said form is slow); to take an
extreme example, for John entered we reconstruct I say that John enters;
John's entering is before my saying that John enters. These reconstructed
sentences are not simply operator-form paraphrases selected ad hoc\ they
are operator-argument constructions which differ from the given (attested)
sentence only by established reductions and linearizations, applied in the
conditions which these reductions and linearizations require. Thus they
differ from attested sentences in the same way that attested sentences differ
among each other. No line can be drawn between the reconstructed sentences
and the normal ones, and only with great difficulty can grammatical rules be
stated that would exclude the reconstructed forms. Indeed, any grammar
based only on sentences which would be generally accepted as normal will
have odd lacunae and oddly limited rules, the more so in that the set of
sentences is not well- defined. But if we state a system of constructions and
relations—let alone one with reasonable mathematical characterization—
which is just sufficient for all ‘normal’ sentences, we will unavoidably admit
many sentences which no one would ever say, either because they are far too
labored, or because the availability of reduced forms has overshadowed the
unreduced, or because the vocabulary needed for the unreduced form is not
(now) available.
In some cases historical replacement has led to this vocabulary lack, as in
the reconstructed base state of (being a) child for childhood, where -hood is
the modern form of a once-free word meaning ‘state’. But in many cases it is
not a question of historical loss. Rather, it is a matter of the partial
independence of syntax and vocabulary. True, the syntactic structure is
arrived at as a relation among words in utterances. But when that structure is
stated, it is not in respect to specific words—which in any case can drop out
of the language or enter into it—but in respect to the ability of variously
related words to fill that structure. In particular, the N, O,,, etc., classes of
the present theory are defined by certain dependence properties of words,
not by a particular list of available words. Given a particular sentence,
and a reconstructed one from which the given can be reached by known
reductions, the reconstructed form may be peculiar because a particular
word in the given sentence does not normally occur (or no longer occurs) in
the position it is supposed to occupy in the reconstructed sentence, or does
not have normal likelihood in respect to the other words in the reconstruction.
This is easiest to see when the word had indeed occurred there and had then
been replaced. For example, at the time when the percursor of -hood;
(above) had still existed as a free word, we would have had by the side of the
compound noun child-hood (the vowel in it was then a) also the had of (being
a) child, just as by the side of birth-rate we have the rate of birth', the stressed
had was the free variant of the unstressed -had. When, however, had ceased
to be used in free-word position, the word that replaced it there—state or
the like—became the free-word suppletive variant of the unreplaced suffix
-hood, yielding the reconstruction the state of (being a) child.
Because the operator-argument theory refers everything in sentence
structure to just a few principles—and ones which can be considered
reasonable, given the use made of language—the theory has an explanatory
value, showing how the properties of the grammar and those of particular
forms—even of some peculiar ones—arise from these first principles. For
example, we see why derived sentences have the same general structure as
do base sentences (8.2.1, G77); how cross-reference from one word-position
to another is managed without a system of word-addressing (5.3.2); why
grammatical constructions are nested one in another (2.5.4) and why
permutations are limited as they are (3.4.3); the distinction between the zero
reduction of a word and non-occurrence of that word (G136, 150); why
certain transformed (reduced) sentences, including metaphors, do not
support further transformations which their apparent form would admit
(G405); why some of these require modifiers (e.g. not He peppered his
speech but only He peppered his speech with asides)', and why many word
subclasses are fuzzy as to membership (e.g. the auxiliaries, the reciprocal
verbs). In respect to particular constructions, we see, for example, how
tense (G277) and plural (G244) can be fitted into the rest of the grammar,
and why certain peculiar uses occur in them (eh.5 n.5); how compound
tenses fit in grammar (G290); how ‘topic’ permutation takes place (3.4.2,
G109); how the relative clause is formed, and modifiers in general (G227);
what in a sentence can be asked about and what can carry a relative clause
(G121); how yes-no questions and wh- questions have a common derivation
(G331); the grammatical basis for the distinction between use and mention
of words (5.4); how we obtain the peculiar grammatical properties of the
comparative (G380); how the special limitations of the passive arise (G364);
how peculiar sentences arise such as It is true that John left, It is John who
came (G359); how certain words such as any, but, only come to have uses
that are almost opposites (G327, 401); why one can say He truly came on
time or He will undoubtedly come but not He falsely came on time or He will
doubtfully come (G306).
As an analysis of language, the operator-argument theory has certain
overall advantages:
First, it covers a language more fully than has been expected of grammatical
theories. It can in principle yield not just the sentence forms but also the
actual sentences, and this with the aid of a list not of elementary sentences
but just of normally likely operator-argument pairs; to prepare such a pair
list, with the aid of many fuzzy subclasses of words, would be a tremendous
task which no one would undertake, but it can in principle be done. A word-
sequence which is fj-ways ambiguous on grammatical grounds (e.g. G143) is
obtained not once, as a single sentence, but n times, as n sentences
originating from different sources via different reductions which degenerately
yield the same word-sequence. The theory can also account for various kinds
of nonce-forms, daggered reconstructions (as above), joke forms, and
fragments, as particular kinds of departures (chiefly by extending the
domain of a reduction) from the main, but not well-defined, body of attested
sentences. At the same time, the theory does not overproduce, does not
yield word-sequences which are not sentences. The product of the theory
includes every contribution to the substantive information of a sentence, as a
fixed interpretation of each step in producing the sentence. And it does this
entirely in accord with the apparatus of the grammatical theory (especially
selection), without appeal to any mental structures.
Second, the operator-argument theory is intended to be maximally
unredundant in that everything that is constructed is then preserved under
further construction. In particular, the existing vocabulary is utilized more
fully, as when affixes are obtained as positionally restricted reductions of
free words (which are already attested otherwise), or when pronouns are
obtained as reduced repetitions of a neighboring word. The existing con¬
structions are utilized maximally, for example in keeping multiple classi¬
fications to a minimum, or when the intransitive read is obtained by second-
argument zeroing from the otherwise existing transitive read (He reads all
the time from He reads things all the time). The meaning-interpretation is
simplified by showing the reductions and alternative linearizations to be
substantively paraphrastic. All this means that fewest constraints have to be
stated in order to obtain the sentences of the language and to relate them
relevantly to each other.
Third, the theory provides an effective procedure for characterizing
utterances. It does so by locating all its events in the finite though large set of
operator-argument constructions rather than in the unbounded set of
sentences (let alone whole utterances). And it retains its effective property
by having only stated types of deformations—primarily reductions and
alternative linearizations—and by having one necessary though not sufficient
condition for determining the domain of almost all reductions, namely low
information (high expectancy) in respect to the immediate operator or
argument. When grammatical events are formulated in such a way that the
relevant environment for them is the operator-argument pair rather than
the sentence, then the reduction of a word can be based on its low
informational contribution, ultimately referable to its relative frequency in
the given position, rather than on the acceptability of the sentence as
reduced.
Lastly, because the events of construction and of deformation are uniform
in their basis—and because nothing extra-linguistic is presupposed—it
becomes possible to compare structurally, and to measure in respect to
complexity (and the cost in departures from equiprobability), each analysis
of a sentence (especially if more than one analysis is possible), and also the
sentence structure itself, and even the whole language.
Most generally, one reaches a fundamental pervasive structure consisting
of a few conditions (constraints) on word combination, with the successive
or simultaneous satisfaction of these conditions producing the specific con¬
structions of language occurrences. What is gained here is not only the
simplicity of the structure, but also the finding that each of these elementary
constraints makes a fixed contribution to the meaning of every sentence in
which it appears.
The major cost of reaching a simple and regular syntactic system has been
found to be moving many form-differences from syntax into morpho¬
phonemics, i.e. into the phonemic shapes of morphemes (words). However,
the phonemic shapes of words are ‘arbitrary’, i.e. they can only be listed
since there is no regularity which would enable them to be derived. And
various morphemes in various languages have more than one phonemic
shape (7.4). Therefore when we replace morpheme-combination differences
by phonemic-shape differences (e.g. by saying that -hood is a suffixal
phonemic shape of state, or that different declensions are just matched sets
of different phonemic shapes of the case-endings), we are only adding to
existing lists of the phonemic shapes of words; we are not bringing a new
degree of freedom into the grammar. Furthermore, the total phonemic
content of words is neutral to syntax: it does not correlate with the meanings
or the combinability of the words. Hence moving a form-difference into the
stating of phonemic shapes of words is moving it out of the stating of
syntax—so long as this moving preserves the selection (combinability) of the
words affected, e.g. so long as the selection of -hood as an independent suffix
is included in the selection of state of which -hood is to be considered a
variant.
In evaluating any theory of syntax, certain limitations must be kept in
sight. There exist additional processes of considerable importance in the
formation or availability of sentences in a language, or in the way speakers
understand the sentences (i.e. relate them to others). Such processes are:
the phonemic similarity among syntactically related words, morphology,
borrowing, historical change in phonemes and in use-competition among
words, and above all analogy. These processes are presently external to any
systemic syntactic theory, but they affect considerably the form of many
sentences which the syntactic theory has to account for. Since the present
book offers not a survey of syntactic description but rather a theoretical
framework, it suffices here to note that given the forms which result from
these other processes, we can reconstruct for them operator-argument
‘sources’ in the same general way we use for attested sentences in the
language. The forms are as though derived by the known types of deformation
from the known types of sources. This is the case for all established
sentences except some listable, usually frozen, expressions.
Some features of language remain effectively outside the syntactic frame¬
work: for example, the phonemic composition of words, and the quasi-
morphemic composition of certain words (e.g. the si-, gl- words, the -le
words, and the con-, per-, -sist, -ceive, etc., set). Other features are only
tangentially disturbing to the operator-argument structure: for example,
popular etymological perception of particular words, morphological paradigms
over classes of words, and the intentional departures in vocabulary and
syntax seen in argot, linguistic jokes, and in some literature.
There remains one analysis which is outside current syntax but which
would simplify the application of the operator-argument construction. This
is factoring words into syntactically more elementary word-components. In
general, syntax starts from the existing vocabulary. Any detailed syntactic
theory will have to admit some regularization of the vocabulary, such as
combining restricted suppletive words into a single word (is, am, are into be\
-s, -z, -en, etc. into a plural morpheme), recognizing some suffixes as
variants of free Morels, etc. The present theory goes beyond this in order to
give operator-argument sources for words that do not seem to accord with
operator-argument structure (e.g. the from the phrase that which is, or
which, who, etc. from wh- plus a pronoun; also able or the like as source for
can). Such analyses are not based on any semantic criterion, which is
not controllable, but on their usefulness in reconstructing appropriate
operator-argument sources by means of established types of reduction.

There are also considerations of systemic simplicity. For example, English
has very few Omm operators, whose argument-requirement is a triple
N, N,N. One such is put: we have He put the book on the table, He put the
book down, He put the book (in a place) where they could see it, but not an
independent He put the book. One could decompose put into the operator-
argument construction place . . . so that, with He put the book on the table
reconstructed into He placed the book so that the book was on the table. If
such decomposition were reasonable for all Onnn, we would eliminate this
little-used operator class.
Replacing a word by a phrase or operator-argument construction, as
above, can be carried out just on the exceptionally irregular or restricted
words. One might also think of carrying it out in a systematic way, in order to
modify both the vocabulary and the details of syntax into a maximally simple
joint system. This last may be virtually undoable and perhaps of little value
for the whole language, but it becomes a matter of great interest in the
sublanguage of science (10.3).
8.3. Sentence semilattices
The considerations of 8.2 show that the words of a sentence have the relation
of a branching process in which the dependence of a on b, . . ., d is
represented by each of b,. . ., d branching out of a, with b,. . ., d linearly
ordered (because likelihood is not preserved among the different orderings
of co-arguments, as in John ate cheese and Cheese ate John). More specifically,
for all word-occurrences a, b in a sentence S, we set a^b for the case where
either b is the same as a, or where b is reachable from a by a sequence of
dependings. When the length of this sequence is not zero, a>b: and when it
is zero, b is identical with a. Here a is an upper bound of a subset Z of word-
occurrences in S if a^c for all c in Z, and it is the least upper bound (l.u.b.) of
Z if a’â for every upper bound a' of Z. The word-occurrences of a base
sentence have the join-semilattice property, namely that every two occur¬
rences have a l.u.b. in the sentence. If a>b and there is no c such that
a>c>b, the a is said to cover b, or a is the (immediate) operator on b, and b
is an argument of a. Here there is the added property that the word-
occurrences covered by a are linearly ordered; they are called co-arguments
of each other: e.g. John and cheese in John ate cheese. The cover relation is
the previously described dependence relation among word-occurrences in a
sentence, and is the one elementary relation of sentence construction. Also,
every subset of S which contains its l.u.b. is a component sentence of S. A
join-semilattice property of word-occurrences is that no two occurrences a,
b have a greatest lower bound other than a or b themselves. This semilattice
differs in various respects from the transformational lattice of 8.1.5.
The actual sentences are a fixed linearization of the semilattice; in certain

situations, there are also alternative linearizations which differ from the
normal one in stated simple ways.
A major problem arises when one sentence (the ‘primary’) is interrupted
by another (3.4.4). We see this first of all when the interruption is an
independent sentence, as in John—it was late—left hurriedly. Here we can
say that the interruption is represented by a separate lattice, and is located at
a particular point in the linearization of the sentence. If the interruption
starts with a conjunction (an Ooa), as in John—only because I suspect him—
began accusing me, we can only say that it is a conjoined sentence that had
been permuted to this position. (This can also be said for the preceding case,
if we take the dash as a conjunction.)
The problem becomes more complicated when the interruption is a
relative clause. It was noted in 3.4.4 that the relative clause is a separate
secondary sentence which begins (in its linearization) with an argument that
is the same as a word (‘the antecedent’) in the primary sentence. If the
relative clause were taken as a conjoined sentence permuted to inside
the primary, we would need a complex addressing system to indicate the
location of the antecedent, to after which the relative clause is permuted.
This addressing capacity is needed, even without the permuting, to tell us
which two words (one in the secondary and one—the antecedent—in the
primary) are the same (usually with the same referent). Languages do not
use such complex addressing systems, and have obviated the need for them
by having the two ‘same’ words appear in easily stated fixed positions
(linearly or in the partial order) in respect to each other. Clearly then we
have to take the relative clause and the sameness statement as interruptions
of the primary, at the point in the linearization (or the partial order) at which
they appear. Once we recognize such a structure, we can consider the
reduction-instructions of 5.2 to be also secondary metalinguistic sentences
which interrupt the primary sentence at the point in which the conditions for
the reduction have been satisfied by the sentence.
The various secondary sentences are all separate sentences with their own
lattice structure, but they all take effect on the primary sentence at a
particular point of it. The point in question can be identified in the semilattice of
the primary sentence, but it is transparent to the operator-argument
relation expressed by that lattice.21
8.4. Individual language structures
The operator-argument theory of language fails to provide a complete

description of the grammar of each language. This failure results from three
:l The solution to this problem, namely the transparency of the metalinguistic material in
respect to the primary-sentence lattice, comes from Danuta Hiz, and in a somewhat different
way from Henry Hiz,
limitations. One is that the theory contains only a necessary but not
a sufficient condition for reduction, so that the specific reductions and
permutations in each language have to be listed. The second is that indi¬
vidual languages may have many and prominent regularities which are not
germane to the operator-argument structure, such as phonemic similarities
within or among particular subsets of words, or morphological paradigms.
The third limitation is that some languages have created local restrictions
on the general system, by letting certain sentence-making steps become
requited rather than optional: in a given situation, a particular word may be
required (e.g. repetition of subject under can, may), or a reduction may
be required (e.g. zeroing of this repeated subject).
Differences in these respects give languages different grammatical features
(aside from the differences in vocabulary—in phonemic content and com¬
binatorial ranges). Some ('inflectional') languages have interrelated systems
of affixes, others have non-systematized affixes strung onto host words or
onto stems, and yet others have no explicit affixes. For the operator-
argument theory the affixes are of little moment, since they are equivalents
of free-word operators on the host words, but for the grammar of the
individual language these differences are important. Some languages also
have for all or many words special forms which are foreign to the operator-
argument relation, but can be shown not to violate it. Thus in languages
which lack enough free words, because most words have stems or roots
which do not occur without one or another of a set of attached morphemes,
one can (as noted above) consider the combination of stem with one
attached morpheme to constitute the base free word, with the other forms
obtainable from it by known operations of the grammar. All such features,
including the fixed phonemic (and syllabic) properties which words have in
some languages, are important regularities in the description of the language,
but secondary here to its operator-argument and reduction structures.
In addition to these word-structure properties, many languages have word
subclasses with important regularities which again do not violate the operator-
argument and reduction systems but rather are in many cases by-products of
them. Many of these subclasses are fuzzy in their properties (e.g. the
'human' class in English) or in their membership (e.g. the English auxiliaries).
In certain cases, members of a subclass constitute a required or virtually
required participant in sentence composition, above all in the case of
grammatical paradigms: tense and person, with aspect in some languages;
also noun-case, gender, plural, and agreement. Many of these overtly
required constructions create grammatical categories for the language. Such
categories may apply semantically even to simple words which have no
explicit construction, if the absence of the construction means a particular
value of the category (e.g. if a word without plural affix is singular,
in meaning and in agreement, rather than being unspecified as to
number).
Furthermore, languages differ in transformations, i.e. in the way sentences

are combined and changed, and in overtness of the forms used therein.
Some languages have regular affixal forms for making an operator into the
argument of a further operator (infinitive, participle, gerund). Some have
very visible and unchanging forms for certain transformations (e.g. the
Semitic conjugations), or have specialized transformations requiring a
complicated reductional path (e.g. the English construction seen in It was
this that I said, G359). Many languages have regular ways to make almost
any operator or argument in a sentence into the subject of a paraphrastic
sentence, but they differ in that in some languages some of these rearrange¬
ments are particularly frequent, made in particularly conventionalized and
overt ways.
All of these additional and overt regularities, plus the avoidance of certain
constructions (e.g. of zeroing the main subject in English), give particular
languages and families of languages unique structural appearances (1.5,
end). To the extent that they are intertwined in form or in grammatical
function, they may give the language the appearance of an overall gram¬
matical character or strategy (‘genius’). These regularities and intertwinings
may increase through analogy and conventionalization, yielding the dia¬
chronic effect which Edward Sapir called ‘configurational pressure’ or drift
in the language.
What is ultimately, if not always overtly, simple in language is the set of
basic and universal processes, especially the operator-argument process
that makes sentences out of vocabulary. The final description of an individual
language—resulting from all the processes, similarities, and events in it—is
not simple and not universal. What makes the simple structure basic is that
the other, individual, features can be shown to operate as additions to it, or
can be transformed as though they are products of it.
The total regularities and similarities in the language come into play in the
problem of how to analyze each particular sentence, especially by the
a priori and objective procedures of computer programs. To formulate
a regular procedure for discovering (recognizing) the composition of a
sentence, and its grammatical relation to other sentences, is a complex
endeavor, over and above the general sentence-synthesizing grammar. It is
not simply a matter of taking the inverse of the steps that produced the
sentence, both for essential problems of process-inversion, and because not
a few of the synthesizing steps are degenerate, a consideration which goes
unnoticed in the sentence-producing system. That is, two different reductions
on two different operator-argument combinations may yield the same
word-sequence, which is then ambiguous. Faced with the ambiguous sequence,
one has to reconstruct both reductional paths from the two sources, and then
choose between them (disambiguate). In many situations, the degeneracy is
only local: one of the reconstructions does not fit the other (later) material in
the sentence and has to be abandoned. But many globally degenerate
ambiguous sentences remain. In analyzing written rather than spoken

material, additional difficulties appear, because various grammatically-
relevant intonations are not adequately distinguished in writing, and some
punctuation such as commas is used erratically.
For these and other reasons, a procedure for analyzing sentences has to
combine its own logic with the inverse of sentence-production, and with
everything that can distinguish which productional path has been taken in
the given sentence. Such distinction is possible because of various factors:
for one thing, to a large extent operators and arguments are different words
(as against being the same word used in different relative positions), which
facilitates their identification. For another, many languages have affixal or
other indicators that mark the operator or argument status of a word-
occurrence , in some cases even indicating the extent of the sequence (if it is a
sequence) that is thus marked. Thirdly, most or all reductions leave a trace
(which may be degenerate) in the resultant form: either a specific entity
(e.g. the interrogative intonation), or an operator’s having an argument
of a different class than it should have (G19, 108), or an otherwise un¬
grammatical word-sequence (G30, 150). Also, most relations are in situ—
the word simply receives a different phonemic form, often zero; position-
changes (so called permutations) are in many languages few, and limited as
to how far the word can move (3.4.3).
Given these considerations, it is clear that sentence analysis (recognition)
is a major task in itself, and requires its own strategy. For each language
or family of languages, one may need particular detailed strategies and
algorithmic presentations of how the particular reductional traces fit into the
operator-argument reconstruction.22
Partly related to the difference between the theory of sentence production
and the strategy of sentence analysis is a difference which arises in the actual
use of language, namely between the speaker and the hearer. In the first
place, the user’s perception of the language, informally or as a result of
learning traditional grammar, does not have to be the same as the system
of constraints described above. People can remember and hold in view a
much larger stock of relations and structures than the few required in the
operator-argument theory. They also tend to avoid the great number of
repetitive small steps such as are used in this theory’s processing of a
sentence, both in building up a sentence and in changing its form. For these
reasons, it is not surprising that people seem to operate grammatically with a
rather large stock of words and constructions, even though many of these
can be derived theoretically from a much smaller stock. This is especially
common in the case of derived words and derived sentence structures
whose derivation is very complex, or involves delicate distinctions or
obsolete forms (e.g. source-words or intermediate structures, necessary for
22 Cf. ch. 4 n. 1.
a derivation, that have gone out of use). Certainly, speakers of English,

following frequent combinations and overt similarities and regularities, use
such a word as childhood, or the tenses, or the auxiliary verbs (can, may), or
the passive form, all as individual words or structures, without considering
that the complexities that these bring into English grammar can be avoided
by deriving them in particular ways from simpler forms.
In addition to this, the speaker’s relation to what he is saying differs in
important respects from the hearer’s relation to what he hears. For one
thing, the speaker has little reason to note the traces left by the reductions
which take place in his sentences, whereas these traces are important for the
hearer. Indeed, the speaker need not be aware if his words have homonyms
(or, in writing, homographs), and if the reductions in his sentences are
degenerate either locally or globally over the sentence; but the hearer has to
note them when there is danger of ambiguities, and he has to know what in
the environing material can disambiguate them. Failing disambiguation, the
hearer may not know what the speaker meant. There are also differences in
respect to reference. One issue is whether the reference is to the same
individual or not. The speaker can say / saw men enter and, later, leave e ither
if he knows them to be the same men or if he does not know, or if he knows
them to be different. He says / saw men enter and, later, men leave only if he
does not know them to be the same or does not want to impart that
information; if he says I saw men enter and later l saw them leave he is saying
they were the same men. In the first case the hearer does not know which
information the speaker meant. More generally, when the speaker uses a
free pronoun instead of a noun, he knows which previous occurrence he is
referring to, whereas the hearer has only the pronoun to go by and in many
cases cannot be sure which other word is being referred to.
Most generally, the speaker's sentence is both partially and linearly
ordered. The grammatical relations and the likelihoods (hence also meaning-
specifications) and the reductions are all in respect to the partial ordering,
but at the end there is a linearization. This is not to say that the speaker first
formulates the sentence in a partial ordering: he may think in elementary
sentences which are already linearized; unintended Spoonerisms may show
if the grammatical partial ordering affects the order in which words have
been chosen by the speaker. In any case, the hearer meets only the linear
form; if the partial ordering has reality for him (as grammatical relations), it
is only at the point where he has completed his understanding of the
sentence. It has been shown (by Lashley23) that hearers try to process the
grammatical relafions in the sentence as they hear it, in the order in which
the words appear; they may have to reanalyze the beginning of the sentence
if later material in it does not fit their analysis up to that point.
23
Cf. ch. 3 n. 4.
IV
Subsets of Sentences
9
Sentence Sequences
9.0. Motivation: constraints between sentences
Chapter 7 established the sentence as an utterance-segment within which are

contained the major constraints, those of operator-argument relations.
Chapter 8 showed that sentence structures form a system of relations among
sentences. Chapter 9 introduces the first globally new constraint in language,
beyond the constraints that created sentence boundary. This is a relation
that holds only over domains larger than a sentence. In Chapter 9 the
domains are continuous speech segments larger than one sentence. (In
respect to this relation, single sentences are not merely short discourses, for
the discourse introduces a new relation.) In Chapter 10 the domains are sets
of sentences that have in common a speaking-situation: these domains can
be defined (a) externally (e.g. articles in a particular research line, or
conversations when people are involved in a particular joint interest), or (b)
structurally (if the sentences in a corpus of discourses show the same kind of
structural relations that are found in discourses that satisfy (a)). As to
meaning: much as the operator-argument structure of the one-sentence
domain means predication (saying one word about another), so the pluri-
sentence structure means subject matter (saying more than one thing about
the same referent or the same class of referents).
Thus further, pluri-sentence, relation is reached by considering two new
problems: what about constraints that are not contained within the limits of
a sentence? That is, are there constraints that reach among successive
sentences? And what relations hold among sentences not throughout the
language but specifically in certain subsets of sentences: when they are in the
sentence succession that constitutes a spoken or written discourse; and when
they are all in one subject matter (sublanguage)? We will see that in all these
cases there is indeed a further constraint, but not on the major word-class
combinations. Rather, it is on word-choices within these.1 Specifically there
is a demand to repeat word-choice in conjoined sentences (9.1); and a
demand to repeat the operator-argument relation among words of particular
subsets in discourse (9.3) and in sublanguages (ch. 10). In the case of
' There is no further constraint on word classes because sentence boundaries were reached
and defined precisely as the point beyond which no constraint was carried over stochastically
from the preceding word classes (7.7)
256 Subsets of Sentences
discourse, this yields a double-array structure, as against the linearized

partial-order structure of single sentences.
9.1. Conjoined sentences
A sequence of even just two sentences, if they are conjoined by connectives

(a subset of O00), reveals a constraint. This constraint can be considered as
the likelihood constraint imposed by each of these Oao on its pair of
sentential arguments. To set the stage, we note that certain conjunctions
demand certain minimal differences between their two argument sentences:
and may be found between two identical sentences (e.g. I welcome you and /
welcome you, only slightly unusual); or expects at least one difference (John
will write it or Mary will write it. We will write it or we will type it, with John
will write it or John will write it being quite peculiar outside of logic
examples); but expects at least two differences (John will write it but Mary
will sign it or . . . but Mary will write it too, whereas there is no John will write
it but Mary will write it except in special cases where differences of expected
kinds have been hidden by reductions, e.g. in John wrote it, but John wrote it
sloppily, or in semantically opposite operators (which conceal a zeroed
contrastive sentence) such as I studied it but I forgot it.
We now consider the general question of constraints on sentences under
various conjunctions and other binary second-level operators (0GO). No
clear constraints have been noted in respect to specific vocabulary; but
under most binaries it is expected that there be at least one difference
between the arguments (except e.g. We will arrive when we will arrive). The
major constraint here is that there be at least one word that occurs in both
sentences, except in so far as one occurrence was removed by reductions. To
see this, consider Sx0ooS2, where Sx, S2 are the arguments of O00. For a
given SxOaoS2 sentence, its acceptance as something unexceptional rather
than surprising or peculiar is either roughly equal to, or else lower than, the
lesser of the acceptances of Sx and S2 separately. Thus in (1) We stayed at
home because we were tired, we find that Sx, S2, and Sx0ooS2 are all equally
unexceptionable. In (2) We stayed at home because today is Tuesday, S, and
S2 are unexceptionable, but SxOoaS2 is odd, or leaves the hearer waiting for
something more unless that explanatory something is already known to the
hearer. And in (3) We stayed at home when vacuum consumed the chair, both
S2 and SxOaoS2 are peculiar while Sx is unexceptionable. When SxOoaS2 has
lower acceptance than all its components, as in (2), we note that its
acceptance can be raised by conjoining an unexceptionable S3 which contains a
word of and a word of S2—for example, and Tuesday is the college's at-
home day in the case of (2).
This suggests that to retain in a resultant sentence the level of acceptability
of its two components (i.e. the degree to which they are unexceptionable)
Sentence Sequences 257
we need word-repetition among the components. There are cases where the
unexceptionability is retained without word-repetition; but then we can
always find zeroable sentences which would provide the word-repetition and
which we can reconstruct as having been present, but well known and
therefore zeroed. 1 hus given (4) We stayed at home because John was tired,
where S]0ooS2 is as unexceptionable as 5), S2 but without repetition, we can
assume some explanatory zeroed sentence connecting John with we (e.g.
John is one of us, or John is our son).
The zeroed adjoined sentences which provide the missing repetition are
in many cases matters of common knowledge, or definitions, which are
zeroable as adding nothing to the information in S\OoaS2. For, example, in
(5) I wrote it on Tuesday because l had to hand it in Wednesday morning, the
conjoined repetition-supplying sentence would have been and Wednesday is
the day after Tuesday, or the like (2.2.3, 2.3.2). When our supplying the
presumed zeroed sentence OaoS3 raises the acceptance of S{000S2, as in (2),
it is because the 0ooS3 of (2) was not known to the hearer, whereas in (4) and
(5) it is known. Thus a sequence S\000S2000 . . . 0ooSn is generally as
acceptable as the least acceptable S, in it if each component St contains at
least one argument or operator which is also present in some other com¬
ponent Sj (before reduction). The above sequence can be shortened without
loss of acceptability by zeroing any S in it which is common knowledge, at
least to the speaker and hearer. Omission of any other S may make the
sequence less acceptable than any of its components. The fact that sentence¬
connecting has this property leads to important results about the structure of
discourses and the nature of argumentation, as will be seen below.
9.2. Metasentences
In the reconstructed sentences that provide word-repetition under Oao we

see that to a sentence 5) we can adjoin (by an additional O00) sentences
which talk about this Sh and which do not appear because they are known
and have been zeroed. This is possible because metalinguistic sentences,
which talk about sentences and their parts, exist (5.2), and secondly because
the ability of a metalinguistic sentence to refer to one preceding it in the
Sx0ooS2 sequence is supplied by the addressing (referral) power which is
contained in the relative-clause procedure that adjoins one sentence to
another (3.4.4, 5.3). Metalinguistic sentences which refer to the ‘host’
sentence to which they are adjoined will be called metasentences. When
they speak about a word, they are referring to an occurrence of that word in
the host sentence, with the sense and selection which it has in that occurrence.
For example. The kite flew high; in this sentence the word kite refers to a bird,
not a toy; also Everybody wanted the funny hats, but the dolls (stressed)
wanted only the girls, where the stressed word dolls is the object and the word
girls the subject of the word wanted; and He asked for cardizem; the word
cardizem here is the name of a medication (usually reduced by ‘abus de
langage’ to He asked for cardizem, which is the name of a medication).
The metasentence adjunctions can fill many functions for the host sentence.
They can state that a word-occurrence X in one position is the same word
(with the same selection, i.e. in the same sense), or that it refers to the same
individual, as a word-occurrence Yelsewhere; in this case the metasentence
can be replaced by reducing A" to a pronoun (possibly zero) of Y. For
example, I saw some high-flying kites, which are hard to keep up from / saw
some high-flying kites; kites—the preceding word refers to the same as the
prior word—are hard to keep up; note that neither the which nor the
sameness statement that is its source would be said if the prior occurrence of
kites referred to the bird ‘kite’, since the predicate in the second sentence is
in the selection only of the ‘toy’ occurrence of kite (5.3.3).
The metasentences can state that the speaker is talking about something,
in which case the word for that thing in the host sentence can be pronouned
as a repetition from the metasentence, which is then zeroed; this gives a
demonstrative (deictic) meaning to what is constructionally just a repetitional
pronoun, as in This coffee is cold from, say, / am speaking about some coffee;
this coffee is cold (5.3.4). The metasentences can also supply grammatical
and dictionary information about all parts of the host sentence. When they
do, the sentence—host plus its metasentences—becomes a self-sufficient
universe requiring understanding only of the given metalinguistic material
(grammar and dictionary) but not of the host sentence itself, since the latter
is.explained by the metasentences.
That this is indeed the case has been remarked above (5.2), namely: we
can take an arbitrary sequence of English phonemes that satisfies English
syllabic structure, and can make it an English sentence if we adjoin to it
appropriate metasentences, for example ones stating that an initial sub¬
sequence in it is the name of a scientific institute, and that the middle sub¬
sequence ending in a (spelled) 5 is a term for some laboratory technique, and
that the final segment is some science object (e.g. a chemical compound).
This means that when we hear a sentence which we accept and understand as
English, it is because the metasentences which performed the above kind of
task for it had been zeroed as being common knowledge. Thus a sentence of
a language is simply a phonemic sequence of it with known or guessable,
usually zeroed, metasentences identifying its successive segments as entities
permitted there by the metalanguage and usually known by the hearer.
Since different grammatical and dictionary metasentences may apply to
the same sentential phoneme-sequence, the zeroing of them may leave the
phoneme-sequence ambiguous. This takes place when the metasentences
state different reductions on different operators and arguments, which
happen to be degenerate as to their resultant phoneme-sequence. Another
kind of degeneracy results when metasentences are erroneously zeroed; for
example, if we zero a metasentence which identified a word in the host as the
name of an object or relation rather than as referring to the object or
relation, we would obtain the effect of ‘abus de langage’ as above.
9.3. Discourse analysis
If we now move from SOaoS sequences to the general case of connected

discourse, i.e. of long utterances, we find a property that goes importantly
beyond the word-repetition seen in 9.1. In each discourse we can find two or
more particular subclasses of words appearing in many sentences, in operator-
argument relation to each other, or in certain (not all) other grammatical
relations. The subclasses are not necessarily known in advance; they are
established for each discourse from the word-pair repetitions that are found
in it. This is a property that seems to be satisfied by all discourses, even
rather aimless ones, but is not in general satisfied by arbitrary collections of
sentences (for example, one composed of the first sentence on every fiftieth
page of an encyclopedia). The various grammatical relations in which the
word-class-pair repetition is found are approximately paraphrastic to each
other, and are found to be transforms of each other (8.1). This property is
different from the word-repetition of 9.1 in that it specifies repetition of two
or more word classes (not specifically of the same word) in certain gram¬
matical relations to each other.
In practice: given a discourse, we look for sentences or sentence-segements
which have a sequence of two or more words in a grammatical relation
(chiefly operator-argument, or host-modifier) that recurs in various sen¬
tences, in the same relation or in transforms of that relation. We may
thereupon reconstruct the sentences or segments containing the transformed
relation, so that a canonical sentential form is selected for all segments
containing that word-sequence. Thus, given Lymphocytes produce antibody
and Antibody is produced by lymphocytes, we might transform the passive
back into the active, to obtain two occurrences of Lymphocytes produce
antibody. We then seek segments in which one of the words in the original
sequence appears in the given relation to other words (not present in the
original sequence): e.g. Lymphocytes contain antibody. This enables us to
suggest that the new words (e.g. contain) are in the same equivalence class as
the words they ‘replace’ (in this case, produce).
To show how this analysis proceeds, we take a complete article, quoted
below. In Figure 9.1 transforms of its sentences are aligned into a table; the
sentences were transformed in a way that would maximize their alignment.2
2 Z. Harris, (a) Discourse Analysis Reprints, Papers on Formal Linguistics 2, Mouton, The
Hague, 1963, pp. 20 ff.; (b) ‘Discourse Analysis’, Language 28 (1952) 1-30 for the general
method; (c) ‘Discourse Analysis: A Sample Text', Language 28 (1952) 474-94.
The Structure of Insulin as Compared to that of Sanger’s A-Chain’
by
K. Linderstrom-Lang and John A. Schellman
1 The optical rotatory power of proteins is very sensitive to the experimental

conditions under which it is measured, particularly the wavelength of the light
2 which is used. Consequently single measurements of optical rotation do not give
an adequate description of rotatory properties, though they have often proved
very useful in the characterization of proteins and the detection of changes in
3 structure in solution. The diversity of the factors which affect optical rotation is in
many ways an advantage, for the variation of each parameter yields a separate
source of experimental information.
4 One of the authors (J. A.S.) has for some time been engaged in a study of the
rotatory properties of several proteins including the effect of temperature,
5 wavelength, pH and the denaturation reaction. One phase of this research, the
dependence of the rotatory properties of proteins on wavelength, is recorded
here because it is of special importance to the problem at hand.
6 In agreement with the observations of HEWITT it was found that all the
polypeptide systems which were investigated obeyed a one term Drude equation
within experimental error (usually less than 1%);
7 where [a] is the specific rotation and A the wavelength of the measurement. A and
Ac have a certain amount of theoretical significance but will be regarded here as
empirically determined quantities.
8 The dependence of the optical rotation on experimental conditions results
from the fact that A is in general a function of temperature, pH, ionic strength,
9 denaturation, etc. Ac, on the other hand, varies only as a result of drastic changes
in the protein system, in particular denaturation by urea or guanidine or titration
10 to a pH between 10.5 and 12.0. The values of Ac and [a]f]' are given in Table I
11 for a number of proteins and related substances. (A2o may be obtained from
12 [a]o by means of the relation A2() = [«]o'(AD2—A2). Ac is not dependent on
temperature within experimental error (approximately ± 50A).
13 The results tor gelatin are taken from CARPENTEk AND LOVELACE.
14 [Table I was inserted here.]
15 [Note to Table I] Except for insulin the specific rotations were independent of
moderate changes in protein concentration.
16 The substances in Table I are listed in the order of descending values of Ac.
17 The table has been divided into two groups to emphasize an obvious pattern,
namely, all the substances in Group A are native globular proteins, all those in
Group B are not.
Reprinted from Biochimica et Biophysica Acta, 15 (1954) 156-7. Table 1 in this paper has
not been reprinted here, because it is not in language form, and only details the information
given in this paper.
18 The entries in Table I are too few to permit any certain generalizations but the
indication is that Ac provides a measure of the presence of secondary structure in
polypeptides, having values above 2400 A for the ordered configurations of
native proteins, and values less than 2300 A for polypeptides in disordered states.
19 We are here subscribing to the view that heat and chemical denaturation result in
20 the unfolding of the hydrogen bonded structure of a protein. Clupein naturally
falls into the latter group because the internal repulsion due to its high positive
charge makes folded molecular forms unstable.
The presence of the oxidized A-chain of insulin in Group B was not expected.
— In tact it was hoped that the A-chain would exist as an a-helix in solution and
would therefore serve as a model substance of known structure in the study of
23 denaturation. Instead it was found that the A-chain possesses rotatory properties
which resemble those of clupein very closely, but do not resemble those of insulin
24 itself. Most striking is the fact that the specific rotations of clupein and the A-
chain are virtually unaffected by strong solutions of urea and guanidine chloride.
25 Ordinary proteins, including insulin, undergo changes in specific rotations of
26 100% to 300% under these conditions. These results suggest that the oxidized A-
chain is largely unfolded in aqueous solution and are in agreement with the recent
finding that the peptide hydrogen atoms of the A-chain exchange readily with
D20, whereas those of insulin do not.
27 A detailed report of this work will appear later.
In the table in Figure 9.1, column 1 gives the sentence number in the original,
column 2 gives the sentence number transform. Each following column contains
an equivalence class; that is, the entries in each column have been shown to be
equivalent to each other by the procedure described above. It turns out that
throughout the whole article, every sentence of the transform consists of the
HRLK equivalence classes (or of two or three out of these four), with intervening
verb equivalence classes.
The interpretation is, of course, not that all members of an equivalence class
are synonyms of each other, but that the difference for the given discourse
between any two members X,, X2 of one equivalence class X corresponds to the
difference between the corresponding members T,, Y2 of the class Y with which
they occur. That is, the discourse compares X,, Y\ with X2, T2, or equivalently,
Xu X2 with Yu Y2.
The word subclasses in question can be determined in the following
equivalence relation (transitive except for subscript, and symmetric and
reflexive) on word-sequences:
4 After the equivalence classes have been set up and their relative occurrence studied, we
may find reason to say that in a particular case a=b„ a: i.e. that a particular occurrence of the
morpheme-sequence a is not discourse-equivalent to the other occurrences of the morpheme-
sequence a. (This differs from the case of two sentential word-sequences being identical in form
but not in derivation, i.e. source: see 8.1.5.) The a±acase will happen if we find that accepting
the equivalence in this particular occurrence forces us to equate two equivalence classes whose
difference of distribution in the double array described below has reasonable interpretation.
Such situations are rare; and in any case the equivalence chain has to start with the hypotheses
that occurrences of the same morpheme-sequences are equivalent to each other in degree zero.
The equivalence relation then states that a is equivalent to b if for some n, n>0, we have a=„ b
by the formula given here.
s P C H V R U
T 1 Insulin has structure

2 Sanger’s A-chain has structure
1 3 Proteins have optical rotatory power

4 wh Proteins have optical rotatory power
5 particularly Proteins have optical rotatory power
2 6 Consequently rotatory properties
7 though
8 and
9 structure
3 10 Proteins have optical rotation

11 In many ways
12 for
4 13 several proteins have rotatory properties
14 several proteins have rotatory properties

5 18 proteins have rotatory properties

19
20 because
6 21
22 the polypeptide syst. obeyed within a— one —
experimental
error
23 the polypeptide syst. obeyed within a— one—
usually less
than 1%
24 the polypeptide syst. obeyed within N 1
usually less the specific rotation j
than 1%
25 all wh the polypeptide syst.
7 26
27 but
FIG. 9.1. Analysis of a text, (Part I).

L w K Metadiscourse S
compared with 2 T
compared with 1
very sensitive to the experimental conditions 1

measured under the experimental conditions
very sensitive to the wavelength of the light
which is used
are not given adequate 2

description by single
measurements of optical
rotation
the latter has often proved
very useful in the
characterization of proteins
changing in solution ... in the detection of 9
affected by a diversity of factors 3

a diversity of factors is an advantage
the variation of each parameter
yields a separate source of
experimental information
One of the authors (J.A.S.) 4

has for some time been
engaged in a study of . . .
including the effect of temperature
including the effect of wavelength
including the effect of pH
including the effect of the denaturation reaction
depending on wavelength 5
One phase, 18, of this
research is recorded here
it is of special importance to
the problem at hand
22-4 was found, in agreement 6

with the observations of
term — Drude- equation Hewitt
term- Drude — equation
A2) |
A +(-Aac + the wavelength of the >
measurement J
were investigated
have a certain amount of 7

A,
theoretical significance
A, Ac will be regarded here as
empirically determined
properties
5 P C H V R U
8 28 (proteins have) optical rotation

29 results from l
30 fact in general (
31
32
9 33 on the other hand
34 in particular
35 or
36 or
10 37 a number of proteins have values of Md and values of

38 and a number of related have values of and values of
Md
substances
11 39 Md from"1
40 by means of
M2d
=
relation
12 41
42
13 43 gelatin
14 44a (The substance . . . has value of -Md and value

ff. Insulin has value and value
— has value — and value
/S-lactoglobulin has value — and value
15 45 (proteins for"1) the specific rotations
46 but insulin for-1 the specific rotations
16 47 the substances have value
17 48
49 1 substances native globular, all are

1 proteins
50 ( substances not native globular, are
l proteins all
FIG. 9.1. Analysis of a text, (Part II).

L w K Metadiscourse .5
dependent on experimental conditions 8
A is a function of temperature
A is a function of pH
A is a function of ionic strength
A is a function of denaturation, etc.
Ac varies only as a result drastic changes in the protein 9
of system
K varies as a result of denaturation by urea
K varies as a result of denaturation by guanidine
Ac varies as a result of titration to a pH between
10.5 and 12.0
Ac given in Table I 10
Ac given in Table 1
Aio may be obtained 11

A20 (—A2C + Ad) ‘
Ac is not dependent within temperature 12

experimental error on
Ac is not dependent within temperature
approximately ± 50A on
has results taken from 13

Carpenter and Lovelace
of Ac under conditions) (is) Table I 14

under pH 3,2%
under —
(Group
1 A
under
under —
1 under —
— under denaturation by 7M urea

— under —
Group
under
B
— under —
- - under —
were independent of moderate changes in protein 15

concentration
were not independent moderate changes in protein
of concentration
of Ac listed in descending order in 16

Table I
The table has been divided 17

into two groups to emphasize
an obvious pattern, 49-50
in Group A
in Group B
5 P C H V R U
18 51
52 but
53 polypeptides have secondary structure whose presence
is provided a
measure by
54 proteins native with ordered have values
configurations
55 polypeptides in disordered states have values
19 56
57 a protein has unfolding of the hydrogen

bonded structure
58 a protein has unfolding of the hydrogen
bonded structure
20 59 clupein naturally
falls into
60 because (clupein has) unstable folded
molecular forms
21 61 the oxidized A-chain is present in

of insulin
62
22 63 In fact
64 the A-chain as an a-helix
65 and therefore the A-chain of known structure
as model substance
23 66 Instead
67 the A-chain possesses rotatory properties
68 clupein possesses rotatory properties
69 but insulin itself possesses rotatory properties
24 70
71 clupein has specific rotations
72 and clupein has specific rotations
73 and the A-chain has specific rotations

74 and the A-chain has specific rotations
25 75 ordinary proteins have specific rotations
76 including insulin has specific rotations
26 77
78 the oxidized A-chain is largely unfolded
79 and
80 the A-chain has peptide hydr. atoms

81 whereas insulin has peptide hydr. atoms
27 82
FIG. 9.1. Analysis of a text, (Part III).

L w K Metadiscourse S
The entries in Table I are too 18
few to permit any certain
generalizations
the indication is 53-5
of Ac
>2400 A
<2300 A
of Ac
We are here subscribing to the 19

view 57-8
!ed-to by heat
led-toby chemical denaturation
Group B 20
made by internal repulsion due to

its high positive charge
Group B 21
61 was not expected
64-5 was hoped 22

would exist in solution
would serve in the study of denaturation
67-9 was found 23

resembling X very closely
resembling X very closely
not resembling X
The fact 71^t is most striking 24

virtually unaffected by strong solution of urea
virtually unaffected by strong solution of guanidine
chloride
virtually unaffected by strong solution of urea
virtually unaffected by strong solution of guanidine
chloride
undergoing changes of these conditions 25

100% to 300% under
undergoing changes of these conditions
100% to 300% under
Results 71-6 suggest 78 26
in aqueous solution
results 71-6 are in agreement
with recent finding 80-1
exchanging readily with d2o
not exchanging readily
with d2o
A detailed report of this work 27
will appear later
a=0 b . — . a is the same word-sequence as b4

a=„ b . = . env env b (read: env a is „_requivalent to env b, or
env a = env b in degree „_i),
where a, b, . . . are word-sequences, and env a is the remainder of the

sentence (or of a transform of the sentence) which contains a; that is, env a is
the sentential environment or neighborhood of a, and is itself a (possibly
broken) sequence of morphemes. Here env a =„_i env b is taken to mean
that at least some part of env a =„_i the corresponding part of env b, and
that all other parts of env a o the corresponding parts of env b. That
is, n—1 is the highest subscript of equivalence between any part of env a and
the corresponding part of env b.
The subscripts n are intended only to show a small ascending chain of
environment equivalences. And the formulas above are not included to
suggest a decision procedure for discovering a structure (let alone the true
structure) of a discourse. Rather, they constitute an orderly record of how
and at what cost a particular structure of that discourse can be obtained.
The equivalence a=0 b is used only when we can find a chain of equivalence
with ascending subscripts. To find such an ascending chain, it is usually
necessary that a and b occur corresponding in grammatical positions within
their respective sentences, or within the transforms of their respective
sentences. Ubiquitous words like the, in, is will usually not satisfy this
condition, and are therefore useless as a base for a chain of equivalences.
The equivalence a=n b between particular word-sequences may be reached
by more than one chain. The degree n of the equivalence between them will be
understood to be the lowest subscript of a=b in any chain in which a—b appears.
The purpose of this chain of equivalences is to create out of the word
combinations—or more simply the word-occurrences—in the discourse
certain word subclasses which will be found to occur repeatedly in a constant
(under transformation) grammatical relation to other equivalence classes in
the sentences of the discourse.
The possibility of achieving this result is enhanced by further grammatical
analysis of the sentences of the discourse. For example, if the discourse
contains Antibody appeared in the lymphocytes and The lymphocytes secrete
antibody, where the procedure above cannot be applied directly, one can
choose the passive transform of the second sentence so as to make it
Antibody was secreted by the lymphocytes. Then antibody and the lymphocytes
are subject and object, respectively, of the two operator-entities appeared in
and was secreted by. Here antibody =Q antibody, and the lymphocytes =Q the
lymphocytes, and appeared in = , was secreted by. The effect here has been
to align antibody in both sentences as being the subject of the sentence or of
a transform of it, and correspondingly for the lymphocytes as object.
Furthermore, decomposing the sentences by breaking them at O00 points
into their component sentences reduces the complexity of the environment
which the procedure has to take into account, thus enhancing the ability to
build up equivalence classes. Since the transformations, and more generally
the reductions and position-changes, do not change the operator-argument
selection or the substantive information in the sentences, any chain of
environment-equivalences of word subclasses established for transformed
sentences will be valid also for the original (untransformed) text. And since
the grammatical relations, and the reductions and other transformations,
have been established for the language as a whole, and are not being
proposed ad hoc for the discourse which is being examined, there is nothing
circular in applying them. One might however go beyond the whole-
language transformations in one case: one could use any synonym or
classifier of particular words as an aid in furthering the building up of their
equivalence classes, but only if it is clear to everyone involved that these are
indeed synonyms and classifiers of those words in the subject matter and
style of language use of which the given discourse is a case.
This method segments the discourse into recurring sequences of certain
equivalence classes. These classes and the sentence types made of them
constitute a special grammar of the particular discourse. Thus, although the
discourse has to satisfy the grammar of its language, it also satisfies a
grammar of its own based on its particular word-choice co-occurrences. If
we take a translation of the given discourse into another language, we find
that the sentence grammar is then that of the other language, but the word-
choice subclasses and their sentence types are much the same as in the
original discourse.
It is possible, in a discourse, to collect all the sentences which have the
same classes, so as to see how the members of a class change correspondingly
to the change in members of the other classes, and so as to compare these
sentences with sentences having in part other classes. In many cases it is
possible to obtain a rough summary of the argument; and it is possible to
attempt critiques of content or argument based upon this tabulation of the
structure of the discourse.
If we consider many discourses of the same kind, we find that their
analyses are partially similar. For example, in scientific articles it is charac¬
teristic that we obtain a set of sentences in the object-language of the
science, and on some of them a set of metascience operators which contain
words about the actions of the scientist, or operators of prior sciences (e.g.
logical relations). In the table the last column (and some of column 3) is of
this nature.
The result of discourse analysis is essentially a double array, each row
being a sentence of the transformed (i.e. aligned) discourse, and each
column a class of sentence segments which are equivalent by the discourse
procedure. All further analysis and critique of the discourse is based on the
relations contained in the double array. The discourse dependence of later
sentences in a paragraph upon the first sentence is different from the
dependence of S2 upon 5, in S\Oc>(>S2 (9.1). In a discourse, the later

sentences are not independent; it may become possible to extract from them
certain factors of modification which might be considered to operate on
segments of the first sentence to produce segments of the later sentences.
These factors are in general independent of the first sentence, and it is these
factors then that constitute the second dimension. It is in this respect that
discourse, as a two-dimensional structure, differs from language, which is a
one-dimensional structure even in its SOoaS word-repetition requirement.
In the case of standardized types of text, and in the case of certain literary
pieces, such as carefully constructed short stories, the double array may
exhibit the compositional structure of the text, and may even locate and
reveal some of the points on which the composition turns.5
In the double array which is created by a repetition of a sentence type
made of particular word subclasses, one can see various differences in the
successive members of a word subclass, i.e. in the successive rows for a
particular column. One may be able to see some regularity, or some
explainable feature, in the way different columns change in respect to each
other, for successive rows. One may also be able to see other grammatical
features such as preferences in the use of pronouns, synonyms, reductions,
and other transformations in respect to the double array; reference, as noted
in 5.3, is always within a discourse, and may be affected by discourse
structure. In cases of extreme regularity it may be possible to describe a
double array as in part the resultant of two axes: first, the top row introducing
the sentence type, and then a generalization of how successive rows differ
from their predecessors. In longer discourses we generally find not one but
several different sentence types, using in part the same word subclasses but
in different grammatical combinations, each repeated to create its own
double array. (In the example above, sentence 61 belongs to a different
type—covering cols. 4, 7, 8—than 64-81, which covered cols. 4-6, 9, 10.)
The existence of the double array, or the position of a particular row in it,
does not confer any value of truth or consistency upon the individual
sentences making up its rows, nor does it empower us to derive for the
discourse further sentences that would fit or extend the inter-row differences
seen in the array. Nevertheless, in some cases the array can suggest educated
guesses as to which unsaid sentences might fit into the discourse, or even why
particular sentences that could fit into the array were not included in the
discourse.6 Going beyond this, there is a possibility of additional conditions
on the double array which give it the power to identify some of the rows as
conclusions from preceding rows, or as having other semantic relations such
as consistency with the preceding rows. This arises from the fact that certain
5 Cf. the analysis of James Thurber’s ‘The Very Proper Gander' in Harris, Discourse
Analysis Reprints, Papers on Formal Linguistics 2, pp. 20 ff.
6 An example of the latter is in Harris, 'Discourse Analysis: A Sample Text', Language 28
(1952) 474-94.
changes within a column (e.g. substituting classifiers for ordinary members

of the word subclass), or correlated changes within different columns in
successive rows, or the action of particular O00 successions between succes¬
sive rows, can separately or together give direction to the double array.
Then, from being just a set of occurrences of a sentence type, it becomes a
progression in the sentence type, consisting often of a first sentence with
various modifications of it in the successive rows, which semantically support
certain later rows as conclusions or suggestions from the earlier rows. This
additional power will be considered more fully in Chapter 10 in conjunction
with the power gained from the sublanguage condition.
We have seen, in considering discourse, that despite the grammatical
importance of the sentence there are structural relations in language which
are not encompassed in the structure of the sentence. Not even a further step
beyond the sentence, namely a relation between sentence structures, suffices to
yield the discourse property. Rather, the latter is a repetition which is indeed
between sentences, or more precisely between operator-argument entities
(as are all grammatical events), but is not a relation of sentence structures.
Indeed, the repetition is seen fully only when we undo transformations in
some sentences and add transformations to others until we maximize the
alignment of the repeated word subclasses. The double array extends the
sentential relations among the words of the sentences into the more complex
relations among the repetitions, thereby extending the ‘aboutness’ of the
sentence—the operator being predicated about its arguments—to the more
complex aboutness of a discourse. The SOoaS repetition and the double
array show that in order to discuss anything more than a single unadorned
one-sentence observation, one cannot merely string various independent
sentences but must have grammatically interrelated repetition of words in
small subclasses. It follows that, as noted above, a discourse is not simply a
longer sentence, or an arbitrary sequence of sentences, nor is a sentence
simply a short discourse.
10
Sublanguages
10.0. Introduction
Certain proper subsets of the sentences of a language may be closed under

some or all of the operations defined for the language. That is, carrying out a
grammatical operation on a sentence of that subset yields another sentence
of the same subset. Such subsets constitute a sublanguage of the language,
because they satisfy the grammar without involving any sentence that is
outside the subset. It will be seen that the structures of these sublanguages
have important new properties of their own, and also throw further light on
the structure of language as a whole.
The sublanguages are found not by using new methods or by asking new
questions with the existing methods, but simply by using the existing
methods on bodies of data whose provenance is restricted in the real world.
The dependence relation of 3.1.1, and the likelihood-based reductions of
3.3, can have different effects in different data conditions. When carried out
on arbitrary samples of utterances of a language, they yield the zero-, first-,
and second-level word classes. When carried out on each discourse (any
extended sequence of sentences arising in the real world) separately, they
yield the discourse word classes and sentence types of 9.3. And when carried
out on the utterances of distinct social or occupational types of situation
(without regard to the sentences’ occurring in a sentence sequence), they
yield sublanguages with specific word classes and sentence type (10.3-5).
The distinct social or occupational situations are in most cases recognizable
as dealing with distinct subject matters, most characteristically in the various
sub-sciences, but also in such situations as law cases, opinion surveys, or
even heraldry.
There is no particular limit to the number of possible sublanguages in the
open-ended whole language, or to the overlappings among them. But this
does not mean that the whole language is merely an envelope of the
sublanguages, or that its grammar is just a washing-out of the differences
among the sublanguages. The language as a whole has properties which do
not hold for any of its sublanguages. In particular, every sublanguage except
the base (10.1) has an external metalanguage w hich is a subset of the whole
language but not of that sublanguage, whereas the metalanguage of the
whole language is always a subset of a whole language (10.2). To this is
Sublanguages 273
related the fact that sublanguages have the dependence of 3.1.1, but not in
general the dependence-on-dependence of 3.1.2.
A sublanguage can differ from the whole language by omitting some
grammatical properties of the language. This is possible because many of the
constraints that create grammar are satisfiable independently of each other.
A sublanguage may happen not to contain the conditions for a given
constraint. The subset of sentences which forms a sublanguage must satisfy
the whole-language grammar, but it can do so vacuously in the case of some
of the constraints, which are thereupon omitted in the sublanguage grammar.
And it can satisfy in addition certain other constraints which are washed out
in the rest of the language.
In Chapter 10 we will see sublanguages based on grammatical conditions
(10.1-2), and those based on subject matter (10.3-4), also formula-sub¬
languages of science which can be considered distinct linguistic systems
(10.5). There will be also comparisons of natural language to other systems
of a linguistic character, specifically mathematics (10.6), and to entirely
different systems of symbols, chiefly music (10.7).
10.1. Grammar-based sublanguages
Some sublanguages are composed of sentences which satisfy certain gram¬

matical conditions that are not satisfied by all other sentences of the
language. One such sublanguage consists of the elementary sentence, i.e.
those formed by N and On . . . all under just the Ooa (conjunctional)
operators of the language (or under and alone): for example, John sleeps,
John sells art, John puts the book on the shelf, all under and, because, etc.
Such a sublanguage may be of interest because it leaves out the O0 operators
and those which mix TV and O in their arguments: the Ono (e.g. think in I
think he left), the Oan (surprise in My entering surprised the dog), and the
Onoo (attribute in I attribute his appointing you to my phoning him). Since
these operators have particular semantic characters, it is possible that in
certain applications language may never need to use them.
Another grammatically conditioned sublanguage is the base language
(4.1), that is, those sentences which do not result from reductions, position-
changes, and other (transformational) shape-changes. This sublanguage is
important because it suffices for all the substantive information that the
whole language carries, yet is (differently from the whole set of sentences)
transparent in its form: the hierarchy of operators and arguments comprising
each base sentence can be read off mechanically from its word-sequence. If
it is desired to make this sublanguage directly readable as English, or
whatever language the base is in, while retaining reasonable transparency
for processing by humans or by computers, we can add to the base sentences
the effects of the required and the most common transformations: e.g. for
English, tense, plural, the/a, referential indication., and possibly the relative
clause in the base sentences. It may be noted that, since the base structure is
more universal than the reductions, the base sentences of one language
translate more easily into those of another than do the transformed sentences,
so that practical considerations are involved here.
There are also various distinguished subsets of sentences which have
grammatical and informational properties but fall short of satisfying the
conditions of being sublanguages. Such are classifier sentences, definitions,
and ‘general sentences’ such as those that express laws of a science. For
example, in classifier sentences it is not always possible to distinguish
formally or semantically between predicates indicating classification and
predicates indicating other properties (e.g. in Cats are mammals. Cats are
hunters), or to specify the ultimate objects of classification (e.g. A horse is a
mammal, A roan is a horse, A stallion is a horse. The stallion is a roan).
In all cases of grammatically based sublanguages, sentences can be
usefully characterized not only by their decomposition but also by the
sublanguages (or more generally the characterized subset of sentences) into
which they fit. To a lesser extent, this holds also for the subject-matter
sublanguages, and for the subsets of sentences which do not quite make the
sublanguage grade, as above.
10.2, The metalanguage
A third sublanguage satisfies both the above property of having special

grammatical additions and omissions, and also the property of dealing with a
closed subject matter. This is the metalanguage. As noted in 5.1-2, the
metalanguage of a natural language is a natural language. More precisely,
one can form in any natural language various sentences that speak about that
language (or about any other), about the combinations of parts in occurrences
of the language, and the like. There is no way to define or describe the
language and its occurrences except in such statements said in that same
language or in another natural language. Even if the grammar of a language
is stated largely in symbols, those symbols will have to be defined ultimately
in a natural language.
The importance of the fact that the metalanguage is in the language has
been noted in 2.1. Here we consider the fact that not merely is it in the
language but it constitutes a sublanguage of that language. It is distinguished
grammatically from the rest of the language by the fact that its zero-level-
argument (A) positions can be occupied not only by words but also by
sentences, and indeed also by otherwise non-occurring phonemic sequences
{The sequence of phonemes, or of words, he went is a sentence-. The phoneme-
sequence pxk is not an English word). However, it also has the subject-
Sublanguages 275
matter property (10.3) of being restricted to a small, relatively closed, set of

words as its own vocabulary of operators. With these it forms its own set of
sentence types, as can be seen in 5.1. Like other subject-matter sublanguages it
need not have grammatically relevant differences, or grading, of likelihood
among members of a subclass (10.3), though it can describe the likelihood
gradations that are found in the colloquial language. The fact that the
metalanguage of a natural language constitutes a sublanguage does not
conflict with the fact that its sentences are at the same time sentences of a
natural language, satisfying its grammar. The metalanguage succeeds in
describing the operator-argument structure that creates sentences only by
stating it in sentences that themselves have that structure.
The metalanguage itself uses only the operator-argument relation and a
subset of words with individual selections (co-occurrence likelihoods). It is
thus a part of the base of the language (4.1). Within the metalanguage one
can not only characterize the whole base of the language, but also define the
reductions for the language. Thereafter, the metalanguage can itself use
those reductions to make the sentences of the metalanguage less cumbersome.
The relation of the metalanguage to language may appear circular: the
metalanguage is part of language, yet can generate the whole language
without being reachable except via language. This circularity is avoided by
the natural history of language acquisition (whether for an individual’s first
language, or second language, or for the species’ development of language
initially). That is, the gross structure of language, in particular the operator-
argument relation and the selections of various words, can be recognized
from observing the gross regularities in occurrences of language (speech or
writing); or it arises by conventialization in the course of language develop¬
ment (12.3). This suffices to establish a basic grammar and vocabulary for
the metalanguage. In turn, the metalanguage then suffices to characterize
(or generate, if we will) the sentences of the language.
The internal metalanguage of a natural language L0 can be obtained if to
each sentence of L0 we attach by some grammatical connective a sequence
of conjoined metalinguistic (5.2) sentences (of L0) which state the gram¬
matical properties (and, if desired, the definitions) of each word in the given
sentence and of all grammatical relations among these words in that sentence.
For example: In the sentence John likes to read, John is a word in N, like is a
word in Ono with ordered arguments John and read, and read is a word in Onn
with ordered arguments John and things, and -s is the operator-indicator on
like, and for... to is the argument-indicator on read, and the words for John
are zeroed qua repetition while the word things is zeroed qua indefinite. As
has been noted, these metalinguistic appendages are implicitly present on
every sentence of L0: if we did not know them we would not understand the
sentence or know that it is a sentence of L0; and because we know them they
are zeroable, leaving the given sentence John likes to read. If we consider
just the metalinguistic sentences on all sentences of L0, we obtain a separate
subset of sentences which specify the grammatical status of the words of L0

and of their grammatical relations.
Taken together these metalinguistic sentences, or rather summaries of
them, constitute the metalanguage Lx of L0. Their primitive arguments (N)
are the elements of L0 or classes of these. Their first-level operators (on N)
are classifiers and relations of these. The structure of the sentences of Lx is
the partial order of dependence among their words; but the dependence-on-
dependence relation is not required since the membership of TV is specified
directly, the other classes being defined in respect to it. Also the relative
likelihood of word combinations plays no role; and reduction is not necessary
—we can choose (inconveniently) to use only unreduced metalinguistic
sentences. One grammatical property that is essential to the metalinguistic
sentences is the capacity to name elements and their relative location
in sentences, i.e. the existence of classifier and locational operators at the
first level. This is what makes it possible for the metalinguistic sentences to
refer to words and their locations in stated sentences or in all sentences of
L0; in L0 this capacity is used only for cross-reference and for reductions.
In addition, those grammatical properties of L0—whether essential to L0
or not—which are not part of the structure of the metalanguage can be
defined by means of sentences of the metalanguage. Thus, dependence-on-
dependence, and likelihoods of word combinations within the dependence
relation, and also reductions, and alternative linearizations, can all be
defined for L0 in metalinguistic sentences which do not themselves exhibit
these structural features. In respect to these secondary features, the meta¬
linguistic sublanguage has a simpler structure than the language which it
describes. Rather than speak of the metalanguage as being in general richer
than the system it describes, we see that in this case it is simpler, because it is
an extract from the whole language, containing just enough structure to
describe the language itself. And rather than the metalanguage being prior
to natural language, the existence of the metalanguage is made possible by
the prior existence of the language itself, because only the dependence-on-
dependence relation, which is unique to natural language, assures the
arising of a dependence system out of word use that lacked dependence.
It is also possible to look upon the sentences of L0 (including those of its
metalanguage) as being generated by discourses in that metalanguage,
rather than to obtain the metalanguage as appendages to the sentences of
L0. Thus John likes to read can be formed from the sequence of conjoined
metalinguistic sentences that follows it in the example above.
The relation of a natural language L0 to its metalanguage, whether
starting from L0 or starting from the metalanguage, recurs in the relation of
the metalanguage to its own metalanguage in turn. There is necessarily an
infinite regress of metalanguages. But it will be seen that this is the case only
for their referents; beyond a certain stage, their formal composition remains
constant. To see this, we take the grammar of L, as a discourse written in the
Sublanguages 277
metalanguage, Li+l, of L,. As noted above Lx has a structure of its own: its
zero-level arguments include all entities in utterances of L0; its first-level
predicates include, for example, differ phonemically, is a phoneme, is a
word, operates on, occurs with; its second-level operators include highly
likely, is more frequent than. The grammar of Lx is written in L2, which is the
metalanguage of Lx; like everything which is not a whole natural language,
L, has an external metalanguage, and L2 is a sublanguage of L0, not of Lx.
L2 in turn has a structure of its own. Its zero-level arguments are not the L0
entities themselves but the classes and relations of them as established in Lx,
hence in effect the predicates and operations of Lx, such as phonemic
distinction, grammatical events, occurrence with, high likelihood; its operators
include is partially ordered, is less redundant, but not, for example, is more
frequent than, because, differently from whole languages, the relative
frequency of words in a sublanguage (here, Lx) has no grammatical relevance
to the sublanguage (10.3-4). Proceeding further, the grammar of L2 is
written in L3, the metalanguage of L2. Here again we find certain differences.
The arguments in L3 include the predicates of L2 such as classification,
redundancy; in addition, certain operators such as is a word, occurs with, will
be found in all grammars including those of the metalanguages. After stating
these elements as a grammar of L3, written in L4, we find that the grammar
of L4 is formally identical with that of L3. Its referents are different, in that
the grammar of L3, in L4, refers to sentences of L3, while the grammar of L4
(written in L5) refers to those of L4. But the kinds of argument and operator
classes needed to describe L4 are not different from those needed to describe
L3. Thus, although there is an infinite regress of referents in the meta¬
languages, the form of their grammars remains essentially unchanged after a
certain point. The reason for this syntactic result is that Lx has the form of a
factual description—the grammar of a (natural) language, while L2 has the
form of a theory, describing the syntactic structure of such factual descriptions,
and L3 has the form of a metatheory, describing the syntactic structure of
theories, which then repeats in L4 and thereafter.1
This means that the language of metatheory, that is, the sublanguage
which presents the structure of a theory, itself has a metalanguage which is
structurally identical with itself, rather than being a proper subset of it as is
the case for a whole language, or external to it as is the case for subject-
matter sublanguages. To generate a whole language it may therefore be
1 The chain of metalanguages has a certain similarity to the chain of translated sentences in
the story about the mathematician who knew no English and asked a friend to translate an
article of his into English. In the published form three lines appeared under the translated title:
I thank X for translating this paper and its title.
I thank X for translating the above line.
I thank X for translating the above line.
The change of referent in successive occurrences of above enabled the author to copy line 3
himself from the translated line 2.
possible to start with L3, defining it in an L4 structurally identical to it, and

then use it to define L2, thence Lx, and finally the language L0.
10.3. Subject-matter sublanguages
Many sublanguages are composed of sentences which deal with a more or

less closed subject matter—one in which a limited vocabulary is used and in
which the occurrence of other words is rare. Whereas the closure of the
grammatical sublanguages above arose from specific grammatical constraints,
the closure here arises from the limitations of word use due to the limitations
of the subject matter, which are sharp enough to constitute constraints on
word selection when one is speaking within the subject matter. The meta¬
language has both of these properties.
The limitations of use are found by much the same method as the
equivalence classes of discourse analysis (9.3), except that here the environ¬
ment needed for the equivalence is specifically the operator and co-argument,
if any, on a word, or the arguments under it, since that is the main scope
within which selection is defined (3.2). The reason for this difference from
9.3 is that here we are discovering the grammar of a language—a sublanguage
in this case—rather than seeking the maximum alignment of words in a text.
In the sublanguage material, what we obtain from the equivalence classes in
respect to the operator-argument environment is various word subclasses
which co-occur in operator-argument relations, i.e. in base sentences, but
also less transparently in reduced sentences. That is, we obtain various
sentence types, each a particular combination of word subclasses. This is the
characteristic property of subject-matter sublanguages; it holds whether the
sentences are in discourses or occur singly. It is sharpest in science, especially in
the 'hard’ sciences, where words and sentences drawn from outside the
science are both irrelevant and absent.
A detailed description of the word subclasses and sentence types in a
sublanguage constitute its grammar. If we now ask how a sentence can
satisfy both the grammar of a sublanguage and the grammar of the whole
language, the reason is that the subclasses of the sublanguage are each inside
a main word class (or a class with modifiers) of the whole language, as they
could not but be. In the sublanguage of biochemistry, for example, the
sentences The polypeptides were washed in hydrochloric acid or The proteins
were treated wi(h acid would be cases of a particular sentence type, say
Nmo,OsolNso, (where Oso, includes is washed in, etc.; mol is for molecule, sol
for solution); in English they are cases of N2ON, (more precisely, the
passive of N3OnnN2PNi‘. We washed the polypeptides in hydrochloric acid;
the subscript numbers distinguish corresponding N). The difference between
the two representations involves the fact that English Onn allows also the
sentence (1) Hydrochloric acid was washed in polypeptides, for whatever
Sublanguages 279
meaning that would have; whereas in terms of the biochemistry grammar,

(1) would be NsolOsolNmoh which is not a sentence type there.
In the grammar of the sublanguage we find particular subclasses of
arguments and operators, which combine in various ways to form particular
sentence types. These latter necessarily satisfy the partial ordering of words
according to their argument-requirements, and so are subclasses of the
whole-language sentence structures (e.g. Nmo,Oso,Nso, a subclass of NOnnN).
In the sublanguage all the operator-argument sequences, with possibly a
few exceptions, satisfy one or another of its sentence types. In the grammar
of the whole language the word subclasses and sentence types of the
sublanguage do not appear, because when the sentences of the sublanguage
are considered not separately, as a separate universe of data, but together
with arbitrary other sentences of the language, the Oso, words are not found
to be restricted to ordered NmoiNsol pairs in the whole-language data, and do
not constitute separate classes. Thus the sublanguage has constraints which
do not apply in the whole language.
In addition, the whole language has constraints which do not apply in the
sublanguage. Since the sentences of the sublanguages are in the language,
and must therefore satisfy its constraints, this additional property is possible, as
has been noted, only if the whole-language constraints are satisfied vacuously
by the sublanguage sentences. This arises if certain operator classes (e.g.
certain kinds of Ono such as know, hope, with their largely human subjects,
or certain repetition-requiring O00) or certain transformations (including
stylistic ones) never occur in the sublanguage sentences.
It follows that the sublanguage grammar contains constraints that the
language sentences violate, and the language grammar contains constraints
that the sublanguage sentences do not satisfy. Hence while the sublanguage
is a proper part of the language, the grammar of the sublanguage intersects
the grammar of the language (although the intersection of the two is
essential to the sublanguage’s being a language) rather than being a sub¬
grammar of it.
The differences between language and sublanguage structure are important
not only as showing what new forms can be attained in sublanguages, but
also as showing the capacities of linguistic systems beyond that which has
developed in natural language.
The greatest difference arises in the different word subclasses and sentence
types to which the sublanguage is restricted, within the operator-argument
sentence structure of the whole language, and in the possibility of new
argument-requirement types among the operators.
In addition, there is the overriding fact that the metalanguage of the
sublanguage is outside the sublanguage, since the word subclasses of the
metalanguage will not in general be among those of the sublanguage.
The externality of its metalanguage releases the grammar of the sublanguage
from being tied to the least description of (sublanguage) constraints, for now
we can always define a class or its environment or other relation not by its
constraints but by an external metalinguistic statement. Thus it is possible to
have multiply classified words—elements which in certain environments
belong to one class and in others to another, the distinction being made in
the metalanguage: in the immunology example below (10.4) it will be seen
that in some environments wh- is a form of the special time-connective in the
bi-sentential sentence type of that science, while elsewhere in the science it is
the general subordinating connective that it is in the whole language.
As in discourse grammar, but not in the whole language grammar, a
sublanguage can have phrases as members of a word class. Even in the whole
language it is possible for certain phrases to be considered as members of a
word class. For example, providing, provided, provided that, with the
provision that, all appear as conjunctions, classifiable under Oao\ l will go,
providing he will go too, and correspondingly for the others. But all of these
also have an analysis consistent with their parts, without which they could
not exist in the sentence. We can see this for with the provision that if we
begin with:
/ will go, with a provision; the provision is that he will go too.

—» I will go, with a provision which is that he will go too.
—» I will go with the provision that he will go too.
Similarly we can see this for the conjunctional provided that, if we begin
with:
I will go if it is provided that he will go too.
And we may see it for the conjunctional providing that, if we begin with
something like:
I will go, if my going provides that he will go too.
In all these forms, the starting-point conforms to the meaning of provide

and the operator structure of the sentences. With various zeroings, which
are not all common but are here supported by the broad selection of these
phrases (i.e. the range of sentences connected by them), we obtain the
shorter forms and the more general meaning of ‘only if rather than specifically
of ‘provide’. We end up with phrases which do not have quite the meaning of
their words, and where the operator-argument relations are not obvious,
but where the whole phrase can be regarded as a new member of O00.
Similarly in the case of sublanguages (but not for science languages, 10.5)
there has to be a ‘correct’ operator-argument structure to account for each
part of the phrase, but the semantic connection of the phrase to its words can
be tenuous, and in effect we simply have various phrases as members of
word subclasses, defined as such in the metalanguage of the sublanguage.
This includes the possibility of considering, as in the methods of 9.3 and
10.3, that both appear in and is secreted by are members of the operator class
Sublanguages 281
on the ordered pair antibody, lymphocyte, disregarding the whole-language

fact that one is passive.2
The external metalanguage is a factor in most of the other differences:
first, a by-product of the external metalanguage is that the sublanguage need
not constitute a mathematical object by itself. Its word subclasses need not
be defined entirely by their interrelations, because it is possible to define at
least one of them, usually the N subclasses, by listing the members of these
subclasses (in the metalanguage statements).
Secondly, related to the above is the fact that the set of subclasses of the
sublanguage is as though closed. Although new subclasses and sentence
types can enter the sublanguage at any time, one can at any time analyze its
sentences in respect to its currently specified word subclasses. The analysis is
considered to fail for any sentence that cannot be analyzed in these terms,
whereupon a new subclass may have to be defined. What is observable
before analysis is that the sentences of the science seem to be in a nearly
closed vocabulary: in fields where this is the case, sublanguage analysis is
easier.
Thirdly, likelihood differences and gradations among members of one
subclass in respect to each member of its operator or argument subclass are
not relevant to fhe sublanguage grammar, though they may correspond to
the meaning-differences among those members.3 This arises from the fact
that such relations as synonymity and classifier status among the words, and
possibilities of factoring the scientific terms into component terms, and also
any useful difference in meanings, can all be stated in the metalanguage, so
that the sublanguage does not need to have a combinatorial basis for them.
Lastly, since synonymities and classifier relations are to a large extent
specifiable in the metalanguage, there are fewer ambiguities in the base
2 This is also justified by the fact that passive constructions such as is secreted by may not have
in the sublanguage the stative environments and meaning that they have in principle in the
language as a whole. Whereas the restriction of selection and meaning that such constructions
as the passive, or provided, providing, should have may be unnoticed or disregarded in various
sentences of the whole language, in sublanguages they may be systematically disregarded so
that there they have to be accepted not as grammatical constructions but as phrasal members of
a word subclass.
3 The sublanguage subclasses are not merely an extreme case of whole-language combinatorial
likelihood differences. Rather, they provide a classification of the sublanguage vocabulary into
subclasses in respect to combinability, in a way that fits the informational framework of the
subject matter. In contrast, the differences in likelihoods of combination that remain among the
words within a subclass (after synonymy has been eliminated) reflect differences in meaning of
the various terms within the current classification knowledge. Thus the sublanguage classes are
specializations of the dependence relation (3.1.1), not of the likelihood relation (3.2). An
opposite analysis was made for the auxiliaries (3.2.4), whose combinatorial peculiarity (lack of
a different second subject) was taken as zero grade of likelihood rather than as dependence.
The justification there was that auxiliaries were a small set of operators whose combinatorial
peculiarity could be readily ‘cured’ by supplying a synonymous source that was regular; they did
not constitute, as in the sublanguage, a part of a global combinatorial subclassification of words
that proved semantically useful.
(unreduced) form of the sublanguage than there are in the base of the whole
language.4
Aside from the differences between sublanguage and whole language, we
can also consider the relations among subject-matter sublanguages, because
the fact that they are all subclass specializations within the operator-
argument structure makes inter-system relations specifiable, much as one
can specify the cost in respect to the operator-argument system of one
another of a given sentence. The transformational analysis of 8.1 has already
introduced the notion that sentences can be identified not only by their
components but also by their relations to other sentences, as transforms of
them. The sublanguage analysis shows that sentences can also be identified
by another inter-sentence relation: by having the same sublanguage properties.
Sublanguages can have various relations to each other. First, some
sublanguages may largely share the same word subclasses, others may have
partly similar structures, somewhat as certain fields of mathematics are
abstractly equivalent. Secondly, sentences of one sublanguage may have
special relations to another sublanguage. For example, certain sentences of
one sublanguage may be arguments in certain sentence types of another: in
Digitalis affects the heart conractility (an example of a common sentence type
in pharmacology), the second argument (The heart contracts) is an example
of a sentence type in physiology. This particular syntactic relation in
sentences of different fields creates a relation of ‘prior science’ between
sublanguages, where one science acts upon events of the other. One may
even think of large parts of language as being covered by various sub¬
languages, possibly overlapping in their sentences. But in any case, the
whole language—colloquial, general, literary, and so on—is not merely an
envelope of sublanguages. There is always a complete, independently
standing language over and above all sublanguages. Informationally, this
whole language may have a relation to the perceived world (11.6) similar to
that of a science language to its science. But there may be little to be gained
from considering the structure of the whole language to be itself a sub¬
language structure.
It is also of interest to consider the similarities between the structures of
subject-matter sublanguages and discourse structures. The main similarity is
that both have special word subclasses and sentence types, but for partly
different semantic reasons. Discourse has these because of the constraints
due to maintaining a topic, to giving information and discussion beyond
separate observations. The sublanguages have these structures because all
4 Pure synonyms, which rarely exist in whole natural languages, are distinguishable in
sublanguages. Within a particular research line for a given two words, if the likelihoods of their
combining with other words become more similar as the corpus of discussion increases, then the
two given words are synonymous for that research line. This does not indicate that they are
synonymous elsewhere. For example, in the immunology material of 10.4, agglutinin and
gamma-globulin are synonyms of antibody, although their specific meanings, and their uses
elsewhere, are not identical.
Sublanguages 283
the sentences deal with an interrelated set of information. Both tend to

maximize the structural similarity within their restricted universes, and in
both cases the external structural status of sentence-segments outweighs the
importance of their internal structure, so that words can freely belong to
different subclasses, and the subclasses can have whole phrases, or transparent
segments such as passive verbs, as members—all this defined in the meta¬
language, which is external not only for sublanguages but also of course for
each discourse.5
The main difference is that a sublanguage does not have the double array
structure of discourse; it is, like the whole language, naturally represented
linearly, whereas discourse is not. Sublanguages have less of the charac¬
teristic repetition than discourse has, and usually more sentence types.
Discourse analysis maximizes the similarity among its sentences via the
equivalence chain of 9.3, more assiduously than does sublanguage analysis.
The sublanguage grammar allows for new sentences that can be made within
its sentence types, whereas the discourse structure allows only for further
sentences that fit the content of the given discourse.
Most sublanguage material comes in discourses, not in isolated sentences.
Combining sublanguage and discourse analyses yields the greatest results,
especially in applications. If scientific texts are analyzed by both methods, it
becomes possible to organize their information, to summarize it, and to
retrieve specific items. Also, in the discussion or argumentation sections of
articles it is possible to seek what conditions on the successive sentences
enable earlier sentences to serve as support for later ones. Since all of these
applications are made possible by the regularities of repetition and inter¬
word relations within a subject matter, it is not excluded that in some subject
matters the regularities may be so complex as to defy any useful utilization.
10.4. Science sublanguages
Among possible subject-matter sublanguages certain ones stand out as by

far the most important. These are the science sublanguages, each consisting
of the sentences and discourses of the practitioners of a sub-science, issued
in the course of research and discussion. They are important because their
subject is important, and because they throw light on the structure of the
science itself, or in any case on the structure of its information. They
5 In both sublanguage and discourse analysis, maximizing similarity is largely a matter of

aligning sentence-segments. In discourse, we seek the maximal alignment of the sentences. In
sublanguage analysis, we try to get segments that have similar combinations into similar
positions. Thus in the immunology material of 10.4, when we find verbs V, that occur between
A and C (Antibody appears in lymphocytes, A V\C) and verbs V2 that occur between C and
A(Lymphocytes contain antibody, CV,A), we form a class V in formulaic order AVC which
represents the V, verbs with argument order A. . . C, and represents V2 verbs with argument
order C . . .A.
constitute large and complex systems, and not merely a collection of

sentences containing scientific terms; and they have room within their
grammars, in distinguished ways, for disagreements and change—necessarily,
since the grammars are made from the sentences of science discussion,
which indeed contain disagreements and changes.
Attempts to investigate the language of science, implicitly as a means of
organizing its information and its conclusion-drawing, were made chiefly by
the Logical Positivists of the Vienna Circle,6 using the tools of mathematical
logic and, later, probabilistic considerations. The difficulties encountered
there can be avoided if we use, instead, the methods of word-combination
analysis. As has been seen for discourses (9.3) and for sublanguages (10.3),
the repetition and classification of word-combination in such subsets of
language create special grammars which yield much of what was sought as
the language properties of science.
In order to see what a science sublanguage is and how it is determined, an
example is given here. The field analyzed is from the early years of
immunology, and covers particularly the research into what cell in the
lymphatic system is the producer of antibody. In this line of research, lasting
from the late 1930s to the late 1960s, there had been a controversy as to
whether the lymphocytes or the plasma cells of the lymphatic system
produced the antibodies. It was finally resolved by definite evidence for both
cells that they produced antibody, and by recognizing that these names were
being used for different stages of the same cell-line. This situation was of
interest to our present purposes, because one can check whether the analysis
of the sentences in the various articles shows their stand in the controversy,
and what each article contributed to it, and whether the analysis shows the
relation of the resolution of the controversy to the earlier controversial
statements. This served as a test of the ability of the analysis to represent the
information, and the differences in information, in the sentences of the
sublanguage.7
We find here several subclasses of zero-level arguments, most of whose
members form a rather small vocabulary of terms of the science, and a few
subclasses of first-level operators (mostly verbs), some of whose members
are also science terms while others are drawn from general English. There
are also various subclasses of first-level operators, in subsidiary sentences as
modifiers (adjectives), some of which are science-specific and others general
English. The latter include classes of quantity and time words. The second-
level operators, such as the conjunctions and the metascience segments, are
drawn from genefal English, and only in certain cases show science-specific
fi Sec in particular R. Carnap, The Logical Syntax of Language, Harcourt, Brace, New York,
1934.
7 The full report of this work is in Z. Harris, M. Gottfried, T. Ryckman, P. Mattick, A.
Daladier, T. N. Harris, S. Harris, The Form of Information in Science: Analysis of an
Immunology Sublanguage, Boston Studies in Philosophy of Science 104, Kluwer Academic
Press, Dordrecht, 1989.
Sublanguages 285
restrictions. All these appear together in a few sentence types specific to the
science.
In broadest terns, the grammar for the immunology material turned out as
follows (meanings are given here for convenience, but were not used):
The words for ‘antigen’, marked G, were either synonyms of it or names

of various antigens.
Words for ‘antibody’ (A) included local synonyms such as gamma globulin,
indicators of antibody presence such as agglutinin, or names of specific
antibodies. It should be noted that most of these synonyms would not
be synonymous with antibody in other research—gamma globulin is a
broader term than antibody—but locally, in these texts and word
environments, there was no difference, and indeed in most cases what is
meant, for example by gamma globulin here, is just antibody.
There is a class of bodies, body parts, and animal names (B): rabbit,
footpad, etc.
Inject (J) is used between G and B; there are also local synonyms for inject.
In this material immunize was used synonymously with inject, and
normal was used to mean not injected.
There is a large class of ‘tissue’ names (T) with many different tissues and
organs which have to be distinguished by subscripts: 7} lymph, Tn
lymph nodes, Ts spleen, and so on. They all have certain common
environments, with differences among them in respect to detailed
environments.
There is a smaller class of cell names (C): Cv for lymphocyte, Cz for plasma
cell, Cb for blast cells, and quite a few other names, some proposed in
the given article just for particular properties that were being reported.
There is a class (V) of the verbs that occur between A and C (or A and T):
Vi for in (synonymously: found in, present in, appear in; also contain in
the inverse order C-A) as in Antibody appeared in the lymphocytes
(AVjCv); Vp for produced in; Vmiox pass through; F* for secreted by, as
in Antibody is secreted by the cell (AVpC).
There is a large class (VF) of words and adjectives after T or C, naming
properties or changes of tissue or cell, with different subscripts: tissue is
inflamed, cell is mature, cells proliferate, etc.
There is a class (U) of verbs which occur between antigen and tissue or cell;
antigen moves to tissue, perishes in tissue.
Finally there is a class (Y) of verbs between cell and cell; is similar to, is
called; there is also Yc changes into, differentiates into; and some more
special subsets such as Yu contaminated with.
While a subscript on X indicates a subclass of X, a superscript on X (seen

below) indicates a modifier attached grammatically to X.
The major sentence types, with some indications of their examples were
identified as follows:
GJB, for sentences of the type Antigen is injected into a body part or an
animal.
Glf'TT, similarly for Antigen moves from some tissue to some tissue (the
superscript indicates from . . . to).
TW and CW, for A tissue (or cell) has some property or undergoes a
change.
A VC for Antibody appears in, is produced in, or is secreted from a cell.
CYC for Some cell is similar to, or is called, some cell.
CYcC for A cell develops into another cell.
In the cell-transfer studies, in which antigen is injected into one animal,
whereafter lymphocytes are injected (transferred) from that animal to
another, with antibodies then being sought in the second animal, an additional
sentence type is found:
CP’BB, for Cells are injected from an animal into another animal. The P’
with two B differs from the J above, even though injected appears in both
subclasses.
There is a special conjunction, internal to a particular sequence of
sentence types, which is seen or is implicit in almost all occurrences of the
pair GJB and AVC. This is thereafter and its synonyms, marked here by a
colon. It often carries a time modifier, e.g. three days, five hours as in
GJB:'AVC, for Antigen is injected into a body part; three days later (or the
like) antibody appears in cells. Also, Antibody appeared three days after
injection of antigen, and the like. This conjunction takes different gram¬
matical forms (e.g. to in The cell contained antibody to the antigen and even
the relative clause wh- (10.3) ); all of these forms synonymously connect
AVC (or CW or TW) to GJB.
Metascience material, giving the scientist’s relation to the information of
the science, can be separated off. Mostly, this is isolated immediately, as the
highest operator in a sentence: O0 and Ono (e.g. as was expected or Xand Y
have shown that). But there are also cases where metascience operators
combine with operators of the science language proper, as in Antibody is
found in lymphocytes, which could be analyzed as either from (roughly) It
was found that antibody is in lymphocytes or else from We found antibody in
lymphocytes.
The relation of this sublanguage structure to the information of the
science can be seen in various ways. We note first how the known change of
information through time is seen in the change of word subclasses and
sentence types in the successive articles of this period. The class C (cell in
contrast to T, tissue) first appears at the very end of the earliest article, and
thereafter becomes very common. The class S for intra-cellular ultrastructures
appears only midway in the series and becomes common as more cellular
structures become discernible. In C, lymphocytes and plasma cells appear
early, and later we find a welter of different names for lymphatic cells as
Sublanguages 287
more and more cell distinctions are made; many of these names appear only
in particular articles and not throughout. In the subclass V(operators on the
pair A, C), the earlier articles have primarily the member Vt (is found in,
etc.) but the later articles increasingly have Vp (isproduced in), as evidence
of actual production becomes obtainable. Finally, the whole subclass Y
(operator on the pair C, C) puts in a relatively late appearance, both in its
member is the same as, is called, which is used to match various new cell
names that appear in different articles, and in its member Yc (changes into,
develops into), when it becomes recognized that some cell names represent
merely later stages of what other cell names represent.
In sentence types, the earlier A VT (Antibody is found in tissue) is later
replaced by AVC (Antibody is found in a cell), where the neighboring
sentences show that this is an abbreviation or conclusion from A VTX; CWTX
(i.e. Antibody is found in a tissue; certain cells abound in that tissue). Later
yet the AVC occurs without a distantly connected AVT: the antibody is
found directly in the cell. Finally, the later articles introduce a new sentence
type CxYC2 (what has been called by cell name Cx is the same as, or develops
into, what has been called by cell name C2); this becomes necessary as the cell
names proliferate, and especially when it is realized that the plasma cell is
not an alternative to the lymphocyte but a transformation-product of it.
Next, we note how the controversy is reflected in the sentence structure.
The disagreement appears first in the fact that while some papers have
exclusively (1) AVpCy (Antibodies are produced in lymphocytes), others
have A VpCz (Antibodies are produced in plasma cells) to the exclusion of (1)
but frequently accompanied by A VpCy (Lymphocytes have a role in production
of antibodies—as against actually producing them). This superscript r (have
a role in) is in itself an indication of the controversy; it has the same position
(on Vp) in the AVpC sentence formulas as the negative superscript ~ which
denies production, and as the absence of ~ which asserts production; and its
meaning—judging from the English words used and from the different
relations of A VpCy, A Vp~Cy, and A VpCy to their neighboring sentences—
is somewhere between denying and marginally admitting production. Finally,
the disagreement appears explicitly when some of the papers that contain
AVpCz also have AVp~Cy (or else AVpCv under negative metascience
operators); e.g. The experiments clearly refute the idea of antibody production
by mature lymphocytes and Antibody is not produced by small lymphocytes.
It is relevant to the ‘which cell’ controversy that several differences in
experimental conditions were common in sentences involving plasma cells
as contrasted with lymphocytes. In the formulas these differences appear as:
P rather than Jx or P (P indicating more massive and repeated injection of
antigen), ,+ rather than (l+ indicating longer time before antibody
finding), and Vp rather than Vf~ (Vp indicating larger amounts of antibody
found). Thus characteristically GPB:'+AVpCz as against GPB:'~AV~Cy.
In addition, we find AVsCy but not AVSCZ: Lymphocytes (but not plasma
cells) secrete antibody, which can account for their containing less antibody.
Al! this is relevant to the issue of whether lymphocytes produce antibody,
and it is visible in the formulas. There is also a relevant consideration that is
not visible in the formulas, but rather requires a combining of these Cz
formulas with other Cz formulas: this is the evidence that the Cz are large
cells with organelles in which antibody could be stored, which could account
for the plasma cells containing more antibody than the lymphocytes. In
contrast, the Cv are small cells with little internal structure which were
widely considered (without adequate evidence) to be the end of their line.
The resolution was that lymphocytes produced and secreted antibody, then
grew into plasma cells which continued to produce, and store, antibody.
One might ask why the sublanguage analysis is needed to show the
disagreement when it is expressed so clearly. The answer is that in the
sublanguage sentence types, such phenomena as disagreement can be found
on grammatical grounds. The immediate purpose here was to show that the
disagreement which we see in the Lymphocytes produce (or: do not produce)
antiboody sentences, explicitly in the use of refute and not and implicitly in
have a role, appears upon inspection in the superscript position on the Vp in
the formulas. Furthermore, in that position, after the sentences of the
articles have been mapped onto AVC and the other sentence types, the
disagreement can be located a priori and even mechanically, without much
knowledge of the meanings of words: if the texts contain a sentence formula
both with the tilde (~) and without it, then there is either a disgreement or a
difference in experimental conditions. Thus when the sentences of scientific
articles are mapped into sublanguage formulas, as can in principle be done
by computer programs, denials or doubts or hedgings about a given statement if
present can be found in a priori statable locations of the formula for that
statement, and so for other items and aspects of information. This makes
possible various human and machine inspections and processings of inform¬
ation, and also a formulation of the structure of information in the
science.8
Lastly, we come to how the resolution of the 'which cell' controversy
appears in the sentence structures. The information structure that turned
out to be involved in the resolution appears early in the series of articles,
when a new sentence type CYC is introduced. It is introduced for two new
kinds of information: Y without subscript is introduced for is called and is the
same as, for the proliferating cell names; and Yc is introduced for changes
into, develops into, noted above. Implicit in the need for these sentence
formulas is the fact that cell names are artefacts of the scientist's work, and
K More generally, considerations of formula structure may suggest questions about the
interpretation of data. For example, the mere fact that we find both GJB.A VC and GJB.CW,
i.e. that the colon can connect GJB to either A VC or CW, could raise the question as to what
precise relation—of time or of cellular activity—holds between AVC and CW, and what
mechanism represented by the colon has the effect of leading, at least initially, to one outcome
rather than the other, or one outcome before the other.
Sublanguages 289
that one named cell changing or developing into another (CYcC) is the same
fact as a cell undergoing certain development (CW). When evidence ac¬
cumulated that both plasma cells and lymphocytes produced antibody, the
controversy was resolved in terms of Yc: namely by CyYcC2 {Lymphocytes
develop into plasma cells). As before, this is not to say that one would have
known in advance that plasma cells are a later stage of lymphocytes: the
need to have introduced the CYC sentence type does not imply the CyYcCz
case of that type. But it shows that the resolution can be recognized
structurally in the sublanguage, and is related to earlier developments in the
sublanguage; perhaps awareness of the sentence type might have made it
easier to consider the possibility of the CvYcCz case.
In the standardized ways of presenting scientific results, the articles have
different sentence types in different sections: e.g. in the Materials and
Methods section as against the Results section, with certain word subclasses
and also sentence types being common to both sections and certain ones
specific to each section. The Discussion section, in contrast, is not formed
out of new sentence types but mostly out of the Result sentences with special
constraints. These constraints involve primarily quantifiers, classifiers, certain
secondary operators (superscripts, in the formulaic representation), con¬
junctions (more generally, O00 operators), and finally metascience segments
(Ono and O0 operators). Much of the vocabulary of all these further
constraints is common to discussion in many sciences, although certain
classifiers and secondary operators (such as the have a role in above) are
specific to the given sub-science. The science sublanguage structure has
specifiable positions in which we find expressions of the disagreements, the
uncertainties, and the change in knowledge. In the present case, the
disagreements are found primarily in the membership of the C subclass as
second argument of Vp. The sharpest uncertainty is in the r superscript (have
a role in), in superscripts such as + and — for imprecise quantifiers {much,
little), and in secondary operators (as superscripts) and metascience operators
that explicitly express uncertainty {likely, it may be that . . .). Change in
knowledge is expressed by new members of existing word subclasses, by new
subclasses, and by whole new sentence types such as CTC; also by changes in
the uncertainty markers above.
A detailed examination of the sublanguage of the immunology articles
shows that here, as also in other sciences, the boundaries of the sublanguage
have yet to be defined. There are several ways in which we can judge the
boundaries of the early-immunology sublanguage, in each case satisfying the
requirement of closure under the set-theoretic conjunctions and, or of
the whole language, and under the reductions and other transformations of
the whole language.
In the first place, the difference in sentence types between the Methods
and the Results sections could be considered grounds for constructing two
sublanguages, but this is undesirable because of their close and complex
interconnections. Differently, there are distinguished portions of many

sentences, occupying distinguished sections of their semilattices, which
could be considered separate sublanguages. These are above all the meta¬
science segments which operate on the science sentences proper (e.g. Rich
demonstrated that in Rich demonstrated that the acute splenic tumor cell was
identical with the lymphoblast), and the quantifier words in all their occur¬
rences. In these cases, each class of segments has its own interrelated
vocabulary, common to its occurrence in all sciences and not specific to the
immunology material; its relation to the specific science material in the
sentences is much the same as it is in other sciences. One could therefore say
that each science sublanguage, in somewhat different ways, draws upon
these more general sublanguage systems in constructing its sentences.
A different situation is seen if we consider closely related research lines.
For example, by the side of the search for which cell produced antibodies,
there was another body of research to see what happens when one animal
(the ‘donor’) is immunized and certain of its cells then injected into another
animal. This second line of research, which threw much light on the first one,
turned out to use largely the same sublanguage as the first but with certain
additions, mainly the CP'BXB2 sentence type (Cells were injected from
the donor animal to another animal). This additional sentence appeared
inside the GJB.AVC sequence of the ‘which-cell’ articles, to form GJB
CP'B\B2:AVCB2. One can see how the differences and relations between
lines of research are reflected in the differences between their formula
structure. A more general case of this situation will be seen in the relations
between neighboring sub-fields (10.5.3).
10.5. Formulaic languages of sciences
10.5.0. Introduction
The way science sublanguages differ from whole natural languages suggests
that it may be of interest to look at them as distinct linguistic systems rather
than as sublanguages. This view is supported by the fact that when in a given
science we analyze articles written in different natural languages, we obtain
essentially the same sublanguage structure.9 Thus the consistent structure
obtained from all the articles of a field, in whatever languages, is not
necessarily a sublanguage of any one natural language. Furthermore, if we
are freed of the need to state the grammar of the sublanguage in a way that is
y Thus in the immunology study of n. 7 above, analysis of articles written in French yielded
substantially the same formulaic representation as in the case of English articles. It is also
relevant that many articles written in one language are read and discussed by scientists who
speak other languages, and that the peer pressure which is exerted on article writers to adhere
to precise and standard formulations can come from foreign colleagues as well as from same-
language colleagues.
Sublanguages 291
maximally similar to the grammar of its whole language, we can maximize

the regularizations in describing the sentence types of the science information
even if this decreases the similarity to the natural languages in which the
articles were written.
We then obtain a representation of the sentences of the science consisting
of word classes and subclasses (the latter indicatable by subscripts), which
combine into particular formulaic sentence types and formulaic sentences,10
with no synonymy in the vocabulary and no stylistic synonymy in the
transformations—that is, only such transformations as are needed for cross-
reference and for aligning the information into tabular structures. The
formulaic sentences then constitute a canonical form for the information in
the science; and the grammar of least constraints which they satisfy con¬
stitutes a new linguistic system, differing somewhat from the set of natural
languages, and similar in some respects to mathematics. We thus obtain a
representation of science sentences by sentential formulas free from the
specific words of language: an example was seen in 10.4.
10.5.1. A distinct linguistic system
When we consider science languages—i.e. the formulaic representation of

the sentences of science (such as the GJB:AVp~C/ of 10.4)—as distinct
linguistic systems, we see certain major differences from natural language.
It is true that the formulaic sentences, like the word-sentences of natural
language, are projections of partially ordered argument-dependence relations,
with the interpretation that the operator is predicated about its argument;
without this they would not constitute a linguistic system at all. But where¬
as the word classes of natural language are formed by dependence-on-
dependence, yielding nothing smaller than the N, Oa, etc. classes, the word
classes of a science language are formed by a best fit of how certain
individual words (or phrases) appear as operators on other individual words.
In this the relation is similar to likelihoods in the operator-argument
construction, but it is not individually graded as in natural language. Rather,
it is a yes-no distinction: a particular word occurs either sometimes or never
as operator on a particular other word. In this it is a dependence, like the
relation that creates the large N, On, etc. classes of the whole language,
although here it creates the subclasses and sentences of a science language.
Since these special classes are (as subclasses) within the big classes of the
whole language—G (antigen), A (antibody) within /V; J (is injected into), Y
(is present in) within Onn—the sentence types formed by them are subtypes
within the basic operator-argument sentence structure of the whole language.
Thus we have in the science language something new: a number of different
sentence types distinguished by their word classes, but all having the same
10 It should be clear that the formulas are simply canonical representations of sentences.
operator-argument, i.e. subject-verb-object, structure. Just as the special

word classes of the science have characteristic meanings (‘cell’ for C, ‘cell or
tissue events’ for VT). whereas the large classes of the whole language have
only a general grammatical meaning (of noun, verb, etc.), so the sentence
types have characteristic meanings (cell-antibody relation for AVC) whereas
the basic structure of language has just the general meaning of predication.
The sentence types may combine into larger sentences, called macro¬
sentences; this itself is a novelty for language. Thus in the immunology
language, the GJB (Antigen was injected into the footpads of rabbits) and
AVC (Antibody was found in lymphocytes), which appear separately, also
appear frequently, especially in the Results section, conjoined by a colon
(,thereafter, etc.) in GJB: AVC. This statement form, which is the main one in
the research line analyzed in 10.4, can accept an inserted GUT or GUTT
(Antigen arrives by the lymph stream), to make a full macro-sentence-type
GJB:GUT:AVC. In the donor research there is a necessary infixed sentence:
GJBx:CIB\B2'. AVCB2. Some sentence types may thus have to be defined as
being expandable in particular ways.
There are additional differences between science languages and natural
language. The vocabulary of natural language is open, and is identified by
each word-occurrence having one or another dependence-on-dependence
relation to other word-occurrences in its sentence. But the vocabulary of a
science language is closed or at least characterizable in any one moment or
corpus; and any new words are constrained to be characterizable in respect
to the existing classes, there being no general dependence-on-dependence
criterion which could determine their status. New members of a class or
subclass, or a new but related word class, have to fit into an existing or new
but related sentence type—each innovation being explicitly either definite
or unsettled, as will be noted below. It is possible that the complex relations,
and hierarchies of relations, that a science has to take into account may lead
to words having different argument-dependences than those that are current
in natural languages, and this may already be the case with such words as
respectively, ratio, rate, proportion, covary, derivative, and integral, in the
way their use is constrained in science languages.11
11 A possible example of new grammatical constructions in sublangauges is the following. In
whole-language grammars it is not clear whether there is a distinction between (1) •restrictive’
and (2) ‘unrestrictive’ adjectives (and relative clauses), as in (1) a white and black cat contrasted
with (2) a white and black panda. In English, the most overt distinction is that the unrestrictive
relative clause, (2), can be separated off by a comma (here / will draw a panda, which is white
and black, meaning any panda, since all are such), while a restrictive one, (1), cannot (/ will
draw a cat which is white and black, meanjng a particular cat, others having other colorations).
In a sublanguage, it may be found that certain adjectives are always restrictive in respect to
certain nouns, while others are not. In that case, the distinction can be made by adjective and
noun subclasses. The difference between restrictive and unrestrictive combinations affects the
analysis of a sentence, in that for (1) further operators are selected by the noun as modified by
its restrictive modifiers, but for (2) they are selected by the noun alone without any effect from
its unrestrictive modifiers. The difference lies in the order of entry of the modifier and the
further operators.
Sublanguages 293
In addition to the special word classes and sentence types, a particular

science language may contain special constraints on certain subclasses of
words, yielding the effect of special grammatical constructions. One case of
this is seen in the metascience segments that operate on the whole science
sentence, as in NO no {We expect that. ..), On(. .. is demonstrable), O00(. . .
indicates that . . .); here there is a subset of second-level operators which
differ in their specific argument-requirements, but have the property of
occurring freely on all science-language sentences, and of carrying meta¬
sentence meaning. Another such subset of operators carries the meaning of
evidentiality, as in may (Antibodies may appear). Another case is that of
special Ooa, both the science-specific ones that create macro-sentences (in
the immunology language: the colon, meaning ‘thereafter’, in GJB.AVC),
and any specially relevant connectives or sequences on the observations or
conclusions of the science (e.g. is carried by, is identical with, is a condition
for).
Finally, an important kind of special construction is the case of those
secondary operators which become modifiers of particular word classes, but
which fit specially important or even necessary informational categories
rather than being just added secondary information. Such is the set of time
modifiers on the colon above (the superscript t in 10.4); this is the only
position in the immunology language where elapsed time has a recognized
standing (but time order is important as modifiers on V,: in what tissue the
antibody appears soonest). Such also is the set of quantifiers in certain
sentence types (in immunology: how much antibody appears in a tissue, how
many times antigen is injected). And there is the special status of r on Vn to
carry the argument that lymphocytes had less than complete ability to
produce antibody (10.4). Such special grammatical constructions differ from
science to science.
Some comment must be made about the reliability of the methods for
determining a sublanguage grammar. There is clearly a need tor more
experience in discovering these structures, and for methods of testing their
adequacy. Evidence exists that within a science field there is clustering of
words in respect to the grammatical relations among them. Experience
suggests that such clustering will hold in very many specialized subject
matters, and perhaps in all sufficiently small subfields of science. Experience
also suggests that this clustering creates for each field a few word subclasses
and a few ‘sentence types’ as sequences of these word subclasses. However,
short of much additional work, we cannot know whether there may not be
fields in which complexities and degeneracies of word-classification and of
word-class sequencing would preclude the applicability or the utility of the
sublanguage analysis described above.
To the extent that the analysis is applicable, it can be carried out in
an objective and reproducible manner. It is true that some knowledge of
the field analyzed (as well as of language structure), and some heuristic
considerations, may be helpful in carrying out a sublanguage analysis; and

these may well differ for different analysts. However, the criteria by which
word subclasses and sentence types are to be judged are objective; and their
adequacy is tested by seeing if discovery of new classes and types falls off as
the corpus of data (texts in the field at the time) increases, and also by seeing
if the structure yields useful categorization and organization of the informa¬
tion of the science. Any difference among analyses reached by different
analysts would be mostly in details of word subclassification and of sentence
type distinction. These should for the most part be encapsulated within parts
of the grammar. In any case, such differences should be visible upon
inspection of the analyses and should be testable as to their equivalence or
difference in what they represent as the informational structure of the field.
10.5.2. Formulas of information
The science languages are a more precise representation of science informa¬

tion than are the sublanguages. In the sublanguages, which draw upon
whole-language vocabulary and grammar but use it under science constraints,
homonymy is inescapable. Thus, in the English immunology sublanguage,
we have Antibody is found in lymphocytes and Many lymphocytes are found
in this tissue. Found is used here for two different relations among objects of
the science; and in the formulaic representation, these are A V,Cy and CvVT,T
respectively. In the sublanguage, found in, and also contained in, are
homonymous members of each of the two subclasses Vt and Wr In the
formulaic language, the vocabulary simply includes two symbolic ‘words’, V,
and W,.
For the same reason, synonymy (in respect to the subscience) is common
in the sublanguage: e.g. antibody, agglutinin, gamma globulin, are all
undifferentiated members of A in respect to the issue in these articles.
However, in the science language, only A itself is a vocabulary element.
Thus the sublanguage isolates the synonyms, but the formulaic language
eliminates them. It should be understood that when we lose the sublanguage
words and retain only the symbol, we lose little of the total information, and
may even lose nothing if the retained symbols express enough variations.
For the various sublanguage synonyms may have different meanings in the
whole language and even in the rest of the given science, but if the methods
of 10.1-4 permit them to be represented by the same symbol, then any
differences among them are not relevant to the information in the given
corpus or subfield at the given time. Indeed, in the example of A above, the
members of A have different meanings in immunology, and their use here as
undifferentiated words for antibody is a stylistic imprecision that is im¬
material to the given research problem. When the various words used are
replaced by a single symbol A, the information is clearer, and the cor-
Sublanguages 295
respondence between form and content is improved. The tighter structure of

science languages in comparison with science sublanguages thus enables us
to tighten the structure even more, for example by removing homonymity
and synonymity from the system.
Another tightening of form-content correspondence arises in pluri-symbol
words. The sentence formulas require that the words of each sentence be
mapped onto one or another particular sequence of symbols. In some cases
we find single words which have to be mapped onto a combination of
symbols: either a symbol and its superscript (i.e. a word class and a modifier)
or else two symbols (e.g. an operator and its argument). For example,
injection is found to be GJ (i.e. equivalent to antigen injected), or just G.
An extreme case of this is seen when one symbol is in the metascience
segment and the other is in the science-language sentence proper, for
example in They found antibody in the lymphocytes, where the formulaic
representation would be MsMnoAV,Cy {My. metascience subject; Mno\
metascience Ono verb; Vy in; Cy lymphocyte), as against Antibody was
found in the lymphocytes, which would be AVjCy. A solution that is
applicable in some cases is to segment the two-symbol word into two
components of the sentence, e.g. find into establish {that) . . .is present {Mno
. . . Vj), as though the sentence were They established that antibody is present
in the lymphocytes. Such factoring of a word into a combination of other
words is done not on semantic grounds (though it should preserve meaning
in the environment), but only to regularize the mapping of the sublanguage
sentences onto the formulas. Such a solution may-not be desirable in another
kind of pluri-symbol word, namely the classifiers of sentences. For example,
in the immunology sublanguage, response is used for antibody presence, but
only following antigen entry, not for the background presence of antibody.
Thus response is a classifier covering the GJ:AVj portion of a formula. Such
classifiers are not directly reflected in the formulas, unless a subscript is
inserted to indicate that a classifier—i.e. a general word—is being used. It
goes without saying that the status of a classifier is not decided semantically
or etymologically but again by the combinations into which it enters. Thus
immunize is found to be a classifier for GJ, not for GJ:A V; use of that word
reports antigen injection (i.e. an attempt to cause immunity) but does not
contain information as to whether antibodies were thereupon produced (i.e.
whether immunity resulted).
Mapping the sentences of the science sublanguage onto an explicit and
more or less closed formula structure has other advantages and disadvantages.
It requires that decisions be made—if possible—in many details of the
information in the original sentence. For example, for the formulaic repres¬
entation of small lymphocyte, we have to decide: is this the name of a
particular cell type, hence to be marked by a subscript on the symbol for cell
(e.g. Cs); or is it a relevant property of the lymphocyte (one of a recognized
set of properties), to be marked by a superscript indicating that the lymphocyte
in question has this property (C/) carried by restrictive wh- (n. 8 above); or
again is it simply a secondary fact that the lymphocyte in question was
unrestrictively small—permanently or at the time in question—in which
cases the modifier small would be reconstructed into a secondary sentence (a
distinct formula attached here by unrestrictive wh-) The lymphocyte was
small (CVWV)? The difference between these formulaic assignments can be
decisive for how well the formulaic sentence fits into the sequence of
sentences that constitute scientific conclusion-drawing (10.5.4(d)). Of course,
we may not know which representation is intended or is best. But even to
raise the question, and to be aware of the alternative representations and
their implications for the sentence-sequences, is a step toward a more aware
andefficacioustreatmentof scientific information in process of accumulation
(10.5.4(d)).
For formulas to be obtained in a controlled procedure directly from the
texts—or for the texts to be mapped in a regular way onto formulas—
assures the relevance of the formulaic representation, and its coverage of the
texts, and also its sensitivity to change in knowledge. An example of the
sensitivity of sentential formulas to informational change: during the early
1970s, when what was indicated (or understood) by the words site in
biochemistry was changing from a location in a cell membrane to a receptor
molecule there, sublanguage analysis had shown, without appeal to bio¬
chemical knowledge, that site, which was at first clearly a member of the
‘cell-portion’ word subclass, was increasingly getting, as predicates on it,
words that were otherwise predicates on molecules rather than being
predicates on cell-portions; the observations made about ‘sites' were becoming
increasingly similar to observations about molecules (later established as
receptors). The gradual shift in experimental observation was visible early in
the details of the formulas. Somewhat as grammatical transformations can in
some cases show the direction of an ongoing grammatical change, so the
changes in operator-argument environments of a given word can in some
cases show the direction of an ongoing change in understanding.
The text-derived formulas also come closer than the text itself to a pure
representation of the information. It is known that both the selection of
scientific problems and also the views and presentations of the information
obtained are affected by socially fostered fashions and expectations, in
addition to the direct pressure of scientific developments. The procedure for
obtaining the formulas cannot correct for the choice of problem, but it can
correct for any language use beyond what is essential for the information,
whether it is teleological or analogical terminology used for its suggestive or
communicative value, or else dramatic presentation or over-reaching con¬
clusions and the like. The formulas, and the constraints on their conclusion-
sequences, also show something about the particular scientific field, and to
some extent about the character of current scientific work in general.
However, there is here no claim that the formulaic representation gives
Sublanguages 297
everything that can be understood from a reading of the natural-language

text. It may be that the additional vocabulary and transformations of the text
suggest things not captured in the formulas, let alone the fact that the
reader’s thinking is done in natural language and not in formulas. The
formulaic representation here is therefore only for information processing,
whether by humans or by machines, whereas in mathematics it can carry
virtually the whole discourse.
It does not follow from the advantage of the formulaic language that
nothing is gained from the availablity of natural language, even though it is
less precise—especially less precise in respect to science information. In
certain situations there is need for imprecision, when one is dealing with
unsettled questions and with areas where concepts are not fixed because the
operations or relations of the science are not adequately understood. The
recourse to imprecision is always available, since in any case the information
of the science will be written in natural languages—more specifically, in the
science sublanguage—with the formulaic representation serving possibly
for control, and for applications such as are noted below (10.5.4(b)). The
parts of the science sublanguage that are well established in a given period of
the science can be mapped onto the formulaic language to which the
sublanguage has a close and regular structural relation. The rest remains in
the vocabulary of the sublanguage; an example of this is the r superscript in
the immunology sublanguage above. The ordinary-vocabulary sublanguage
segments are not a disorderly retreat from precision, because in this situation
they are encapsulated and have a known location in respect to the formulaic
structure, so that one can inspect what is unclear and what is the relation of
that to the precise knowledge. It is in this way, and with the availability of its
corresponding sublanguage, that the formulaic language of a science is both
a closed system and yet open to uncertainty, disagreement, expansion, and
change.
The change in a science language is contained within its remaining
structure, so that it is not easy to say at what point, if any, the amount of
change has become massive. This is not to say that there may not be
situations in which massive changes of knowledge and understanding would
impose a massive reformulation of a science language, but in such cases, the
old science or perception of the world usually disappears from use, often
only gradually, possibly to be followed by the beginnings of a new science
which fills a rather different niche in human thought and in practical life.
Such a situation has the character of cultural change rather than of science-
language development. Even within an ongoing science, many views that
are erroneous or prove unproductive just fade away, rather than being
buried and explicitly replaced by something else.
The formulaic representation also clarifies discourse structure. In a large
part of science, what counts is not only the observations but also how the
observations were made and in what conditions and boundaries, and how
they are juxtaposed and confronted with other observations and considera¬
tions. In science languages, this is reflected in the fact that science sentences
do not appear alone or in small independent sets, but in rather extended
sequences, that is, in discourse. Characteristically, there are regularities and
constraints in the discourses, beginning with the general discourse property
of repeating word-class relations (9.3). For example, in the standardized
styles of scientific articles, the articles in certain experimental fields have
three essential divisions: Materials and Methods, Results, and Discussion. It
is found, as has been noted, that the word classes and sentence types of the
Methods section limit what can appear in the Results section, due of course
to the plan of experiment, but visible in the reporting sentences. Also the
sentences of the Discussion section are not independent of those in Results;
indeed the Discussion consists largely of Results (and other) sentences,
modified in certain ways (e.g. by use of classifiers) and arranged under a
restricted hierarchy of O00 operators which build an argumentation. Thus
the Methods sentences impose constraints on the Results sentences, while
the Discussion sentences are constrained to stay close to the Results
sentences. In addition, the sequence of sentences within the Discussion
follows a special type of discourse constraint (10.5.4).
The informational relations within science texts are thus indicated in the
structure of the sentence types (the rows of the double array), the correlated
changes in the columns, the local constraints on successive rows (10.5.4),
and the more general relations on the sentence types within a discourse (e.g.
immediately above). Within each discourse there may also exist more
complex paths of informational interconnection among sentences; but these
are not reached at present by any a priori method of analysis.
10.5.3. Boundaries of fields
In discussing science languages, we have to ask what are the boundaries of a

science language and how it is distinguished from others. First, the grammar
of the formulaic sentences, and the correspondences for mapping the
original sentences onto them, can identify what material in the original
sentences does not belong to the science language, and in what way that
material relates to the formulaic sentence. This may be useful in fields such
as law where there is formalizable legalistic material intermingled with not-
readily-formalizable general-language segments (e.g. examples from life)
that have at least a content relation to the formalizable material.
Second, the semilattice analysis of the formulaic sentences is a mine of
information about the science and its boundaries, because the different
science statuses of material in a single sentence appear in different sections
of its semilattice. In the semilattices, the metascience operators are upper
bounds of the operators in the science language proper. If there are any
Sublanguages 299
categories such as time or quantity which draw on a general language of

science to modify particular items of the given science, the appropriate
symbols will appear as secondary operators (also upper bounds) on the
lattice items in question, for example in the t (time) on the colon of
GJ B.'AVC. If there is a prior science to the science in question, it will appear
at particular lower points in the semilattice. For example, in the sentences of
pharmacology, there are as noted above Ono pharmacological operators
whose first argument is a drug noun and second is a sentence (or noun) of
physiology: in the pharmacological sentence Digitalis affects the heart con¬
tractility we have as argument a physiology sentence. The heart has the ability
to contract. Physiology is here a prior science to pharmacology.
Third, relations among sub-areas of a science which do not take place
inside of single sentences are marked by statable similarities between the
semilattices involved, as in the similarities between Methods sentences and
some of those in the Results section.
Fourth, neighboring scientific fields have appropriate similarities and
differences in their formulas. This holds for neighboring research lines
within a field (e.g. the ‘donor’ research, 10.4, end), for neighboring sciences,
and also for many successive periods within a single research line. Indeed,
earlier and later articles within a single research endeavor show appropriate
differences in their formulas, differences which mount up when major
advances are made in the field. All this may make it possible to quantify the
differences between fields, or periods of a single field, by weighting the
differences in word classes and in their memberships, in sentence types and
in the special constructions that are reflected by the superscripts in the
formulas, and finally in any differences in respect to boundaries (e.g. to prior
sciences).
10.5.4. Information and argumentation
We can now summarize what kind of information we may obtain from the
structure of science languages about the science and its structure, keeping in
mind that such terms as ‘information’ and ‘structure of science’ are undefined
and can only be used loosely. The operator-argument analysis, and the
mapping of science sentences onto the formulas of a science language,
suggest that the latter can contain almost all the language-borne information
in the reports and discussions of the science. This means a better fit than we
have in natural language between differences in form and differences in
information—loosely, a better fit between form and information. What
there is to learn from the structure of science languages can be considered in
four structural contributions: (a) the internal structure of all sentential
formulas, (b) the different types of sentential formulas, (c) the kind of
information that characterizes the given science, and (d) the way the
formulas are combined in discourses of the science.
(a) The semilattice structure of each sentential formula isolates and
identifies the various sources that contribute to the information of the
science: the metascience, the prior science, the procedures and reliability of
observation, the quantities. It also instigates clarification of the relation
between particular subsets of terms: e.g. in a given research line, as to
whether small, or mature, is (as noted above) a subtype of cell (hence a
subscript on C), or a necessary condition of it for the given experiment
(hence a superscript on C), or an incidental (non-restrictive) property (to be
stated in a separate formula, attached by the wh- relative clause).
(b) Since only a few particular types of formula are found in each line of
research, they constitute a framework for its information: anything not
expressible in the stated types of formula is excluded, except as being a
specifiable addition to the field described thus far. Having a framework
creates new conditions for the representation of information. Since the
information is fitted into the available formulaic structures, we can obtain a
tabulation of the information in a text, and know where to look for each kind
of information if it is present. This makes it possible, given a tabular
representation, to inspect it for any particular kind of information, and to
summarize or otherwise process the contents of the text. All this holds only if
the formulaic framework has been obtained not by fiat—by someone’s
understanding of the field—but by the regularization of the text itself, or of a
whole corpus of similar texts, in such a way that it constitutes an objective
best fit for the information in the texts (as in 10.4). Furthermore, the fact
that this framework is not made by one expert or another, but is reached by
objective procedures carried out on the discourses of the science, assures in
principle that mapping the discourses onto the framework can be carried out
by computer programs. Finally, it becomes possible to recognize differences
in respect to the framework—imprecisions, disagreements, change through
time.
(c) As to the kind of information: the word-subclass relations that appear
in the various sentential formulas express the various types of fact and
conclusions with which the given science deals at the given time. This
specification makes it possible also to identify the kind and amount of
difference in subject matter between neighboring sub-sciences, and the
amount of change in a science over time. It also excludes from the science
any irrelevant material; such material cannot be mapped into any of the
sentential formulas. Thus a science language is protected against nonsense,
as is mathematics; but, like natural language, it is not protected against error
and falsification.
(d) The way the sentential formulas are sequenced in discourse is important
for inspectable presentation of methods and data, and above all for controlled
Sublanguages 301
and inspectable drawing of conclusions. The latter raises the question of

proof and truth in science.
To the extent that this last can be discussed within present knowledge, it is
a property not of individual sentences (formulas) of a science language, but
of constrained sets and sequences of such sentences. The great successes in
this direction—the syllogism, the truth tables of mathematical logic, mathe¬
matical induction, and the inspectable structure of proof in mathematics—
all indicate that the structural issue is how to control some informational
value of the last sentence of a sequence on the basis of the presence of
information in initial sentences, by means of constraints on their structure
and on connectives between them. It has already been seen in the analysis of
discourse (9.3) that any sustained treating of subject matter involves sentence-
sequences in which (after transformational alignment) something of the /7th
sentence is retained in the /7 + lth, aside from intervening sentences of
various kinds. The only overt whole-language structure that connects suc¬
cessive sentences is the set of bi-sentential operators, O00, including the
conjunctions. These impose weak word-repetition constraints on their
arguments (9.1), but no recognizable transitive or long-range constraints. In
the set of all natural language discourses, the only common constraint stated
in 9.3 is retention of operator-argument (or of some transforms of it) among
words.
In mathematics, the constraints of material implication (or truth tables)
and substitution (with no likelihood differences) determine the possibilities
of proof. In formulaic science languages, these mathematical constraints are
unavailable because of the specific combinations to which the various word
classes are restricted, which produce families of sentence types, thus pre¬
cluding such a general relation as material implication. In natural language
this complexity is due directly to the particular likelihood (and, therewith,
meaning) properties of each word. In science languages, this complexity is
reduced to the combinatorial restrictions of word subclasses; but the reduction
is achieved at the cost of having several different sentence types. But this
very limitation may open the way to a different constraint, new to natural
language, which appears in some successions of science sentences: regularities
of change in the successive words of a given word-class (within successive
sentences) in respect to changes in the other word classes of their sentences.
This word-change is syntactically reminiscent of substitution in successive
sentences of a mathematical proof, but is very different in that it is a
constrained and successive substitution, specifically in the double array that
characterizes discourse (9.3). This word-replacement rests upon the fact
that the word classes of science have a common meaning interrelated with
the meaning of their environing word classes, so that in successive sentences
the words of a given position and the words of their environments (the
columns of the double array) change together.
In addition, science languages have another property, lacking in math¬

ematics, which can be used for constraints on sentence-sequences: the fact
that sentences of the science can carry many varied metascience and
evidential operators acting on them—Ono and Oa in contradistinction to the
bi-sentential (and largely causal or temporal) Ooa. The importance of all this
is seen in the fact that the Discussion sections in scientific articles are built
not out of new sentence types of their own but out of previous Methods and
Results sentence types with more-or-less constrained word-replacement,
metascience and evidential operators, and conjunctions.12
To this should be added one important consideration: the choice of initial
sentences. Whereas in the whole language the choice is arbitrary, and in
mathematics it is in principle limited to proved theorems, ultimately to
structurally recognizable axioms, the practice in science writing is a selection
of Result sentences—made according to the wishes of the writer—with a
selection of hypotheses, posed alternatives, and the like. As in mathematics,
all these sentences are not generally used in a sequence obeying a single
constraint. Rather, initial sentences (premises and results) are stated, and
each may be followed by a constrained sequence amplifying or supporting it,
whereafter another initial sentence can be brought in. The whole becomes—
somewhat like a proof—a partially ordered sequence. Nevertheless it
remains a progression, satisfying the essential requirement that from given
initial sentences certain constraints be handed down through successive
sentences to affect the concluding sentence as a ‘consequence’ related to the
initial ones.13 The greatest problem in the case of science discourses is in
establishing criteria for the initial sentences, in relation to the subject under
discussion.14
In science languages, the effect of the type of sequence considered above
on its concluding sentence is not that the truth of the conclusion is guaranteed.
Even in mathematics the effect of proof is only to guarantee that there is no
loss in truth value, in respect to that of the initial sentences. Since, in
language and in science, we lack a criterion to assure the truth of initial
sentences (at least to the extent that axioms are taken as true), and lack the
truth-retention of the succession-constraints of mathematical proof, we
12 More generally, the set of sentences which go beyond data-description to constitute a

theory of a subject do not merely have additional words, for constructs and concepts that go
beyond the words of the data-description. Rather the new ‘abstract' words have positions in the
theory sentences related to the positions which corresponding ‘concrete’ words have in
sentences of the data-description. The sentences of the theory are thus structurally related to
the sentences of the data, or may contain (as their arguments) classifiers of the latter sentences.
11 For what little it is worth, it may be noted that in English both sequence and to conclude,
which are contained in consequence and conclusion, suggest that the latter refer to reaching a
point at the end of a sequence.
14 The structure of argumentation, and of the data-presentation prior to it, is developed not
only by the individual scientist’s response to the science material, but also by the views and
understandings of the science community in general. It thus has something of the public
condition of natural language (2.5.1, 12.3).
Sublanguages 303
could not even say how to recognize by its structure the truth of an arbitrary
scientific sentence—or of any sentence. But lesser goals may be available,
given the constraints in science languages: for example, consistency, retention
of possibility or plausibility, even the certainty that something follows from
explicitly given initial sentences.
It is of interest here that the syntactic and discourse properties of proof¬
like argumentation in science are also found in sentences expressing causation.
The term ‘causality covers a welter of situations: an agent or a precursor
causes an event; an earlier stage of a process or chain ‘causes’ a later stage;
even membership in a set ‘causes’ a member to have the properties of the set.
But in many cases the reporting of causality, i.e. stating that a causal relation
holds between events, has the structure of a sequence or partial order of
statements culminating in the statement of the caused event, as in the QED
of a proof.
In any case, under the constraints, the final sentence is a consequence of
the initial ones. Any attempt to specify the constraints on succession more
fully, and to formalize scientific argumentation or make it subject to
inspection and control, will meet with great difficulties. Although the
development of science is certainly affected by unpredictable advances in
technology and in methods of observation on the one hand and by social
controls and personal interests on the other, it is also affected by its own
results and conclusions which in themselves are not arbitrary or faddist or
competitive, but rather follow the cumulative direction not only of its
increasing data but also of its own argumentational constraints.
10.5.5. Notation
The structure of the sentential formulas of a science constitutes a notation

for it: that is, as a set of symbols with a syntax (combination-constraints) on
them. In mathematics, notation began as shorthand for the objects and
operations in current mathematical statements; but today, within the limits
of meta-mathematical knowledge, a notation is determined a priori by how a
set is closed with respect to operations and properties. In the various
sciences, no meta-science knowledge suffices to fix a priori a notation for a
science. The alternative proposed here is to derive the notation from what is
actually said in the science, by procedures that assure its objective para¬
phrastic relation to the original, its objective reproducibility, and its coverage
of the field. The latter requires that the notation contain inherently the
capacity to change with any change in what is written—change of knowledge,
of theory, and of attention (i.e. of what is topical) in the science. In 10.4, it
was seen that procedures of sublanguage analysis can serve for that purpose.
An adequate notation does much more (as well as less) than express what
is present in the science. It also limits what can be expressed, which in the
case of always-changing science requires an inherent capacity for notational

change. This capacity is in principle possible since change in science is not
arbitrary even in the case of unexpected discoveries, but almost always
includes continuities in knowledge and even in problematics. The notation
also provides certain kinds of critique of the expression of knowledge. In
part it does so by reducing into precise and prosaic formulas any metaphoric,
rhetorical, or sloganizing presentations which may have been made for
communicational effect, as well as any analogic argumentations, and any
teleological terminology.
Consider, for example, the case of teleological terms, as when a process is
named in terms of its outcome, even though the process must have been
occasioned by other factors before its outcome has come into existence. It is
not always easy to decide semantically whether a term for an entity or event
in a given field is teleological: does the meaning of the word, or its uses
outside the field, relate to its end-product rather than to its substrate or
composition or mechanism? If it does, it is teleological, as an inverse of
reductionism. As an example of the need to avoid end-product teleology: in
evolution, every step has to have had a survival value for the species, without
taking into consideration the value of what it has led to (e.g. in the
development of feathers, before flying with their aid). Another example is
the use of the term ‘communication' for certain intra-species correlated
animal behavior (12.4.1). Thus to use end-product terminology such as
information, messenger in the DNA-RNA literature is somewhat like the
inverse of what we would have if we named information-carrying messages
in the macroscopic world by a term taken from the biochemical action of
‘messenger’ RNA. In notation, such problems as teleological and meta¬
phorical terminology are more manageable: to the extent possible, each
zero-level term of the sublanguage should not have other meanings or
connotations than are required for the selection of first-level operators with
which it occurs in the sublanguage, except for reasonable generalizations or
the like; and similarly for first-level operators in respect to their zero-level
selection. In mathematical research, one finds little rhetoric, but even there
discussion in a natural language shows departures from the limited universe
defined in the notation. In various sciences, there are various degrees of
loosely stated or uncertain data, observations, and argumentation, as also of
metaphoric or end-product talk. This is all the more reason for having by the
side of the natural-language presentations also a procedurally obtainable
mapping into a formulaic language.
10.6. Mathematics as a linguistic system
The question of how language can be described as a mathematical system

was discussed in Chapter 6. Here we will consider how mathematics can be
Sublanguages 305
described as a linguistic system. Only a sketchy account will be given here,

because anything more fundamental would involve broad considerations of
the syntax of mathematics, as well as meticulous grammars of the sentence
structures and discourse (proof) structures in mathematics and logic. The
gross syntax of mathematics and logic is clearly similar to that of language. In
all of these, the texts consist of sentences (propositions, formulas), each of
which is a sequence of symbols that are readable as words, the occurrence
of some types of symbols being dependent on the occurrence next to them of
other types (their arguments). Roughly, in the case of mathematics and
logic, the tilde of negation (~) requires one argument, plus and minus signs
require two, and the signs for equals and implies and contains also require
two arguments, but in different conditions.15
The linguistic status of mathematics is seen in the fact that the sentences of
logic and mathematics are within natural language at least under translation.16
That is, it is possible to read mathematical and logical formulas in a natural
language, although in some cases only by language constructions which
differ structurally from the mathematical notations. (Hence not all math¬
ematical notation satisfies the conditions for natural-language syntax.) The
converse does not hold: not all sentences of a natural language can be
written in mathematical or logical formulas. The symbol-lists (vocabularies)
and grammars (syntax) of mathematics and logic do not provide stable
differences in likelihood of combination, such as language has in its operator
-argument selections; nor do they have the ability to borrow this capacity
from language. Hence, they cannot capture the meanings expressed in
language. Hence, also, not all sentences or constructions of natural language
have a ‘logical form’, i.e. not all can be adequately represented and
distinguished by expressions in mathematical logic.
There are crucial similarities and differences between mathematics and
natural language in respect to the syntactic status of the symbols (or words).
The zero-level arguments of mathematics and logic are the numerals—the
15 The detailed syntax is a bit different if the notation of mathematics is freed from the
linearity of the symbol-sequences. Certain properties of operations, for example, can then be
stated less redundantly and more essentially. Thus instead of defining commutativity as the
equivalence of two orderings (a+b=b+a), we can say that when the operation + on a set maps
a subset of members onto its image the subset is unordered: the unordered pair a, b is mapped
onto its image c. And instead of expressing associativity by grouping—a+{b+c) = (a + b)+c—
we can say that applications of pairwise mappings of members of a subset onto their image are
unordered: i.e. in a, 6, c, any ordered pair (a, b or b, c) is mapped onto its image in the set, and
then the pair consisting of that image with the remaining member of a, b, c is mapped onto its
image in turn.
1(1 One might also say that the statements of different fields within mathematics and logic are
intertranslatable, in their notation, in the case when the fields are abstractly equivalent. Andre
Lentin has called my attention to the possibilities of translation between two mathematical
systems which are not isomorphic. Cf. ch. 6, sect. 11, ‘Polymorphisms; Crypto-isomorphisms'
in G. Birkhoff, Lattice Theory, American Mathematical Society Colloquium Publications 25,
3rd edn. Providence, R.I , 1967, pp. 153 ff. Note here the comment (p. 154) about the
possibility of defining the same abstract algebra in several non-polyisomorphic ways.
names of the numbers—and variables defined for each text over a specified
domain of values. Signs such as plus operate on two zero-level arguments,
the result being again an argument of the same level; such an operator does
not exist in natural language. The negation sign operates on one sentence
and produces one sentence out of it (like the Oa of language). The equals
and greater than signs operate on two arguments, with the result being a
sentence (like Onn in language). The implication sign operates on two
sentences (propositions), with the result being a sentence (like O00 in
language).
To see more specific differences from natural language, note for example
that the symbol e for membership of an element in a class cannot be 0„„, as it
would be in language, since under e the second argument (a class) cannot be
of the same sort as the first argument (N: an individual). In contrast, the first
argument of is in such sentences as An opossum is a marsupial can also be its
second argument, as in This is an opossum: this is is closer to is a case of than
to is a member of. In the grammar of logic, we have two choices. We can say
that names of individuals and names of classes are two different sets of zero-
level arguments (something which can be defined in the external meta¬
language of logic). Alternatively, we can say that the names of classes are On
predicates, with e being an integral part of them (their operator indicator,
G37, 54): is-a-marsupial. Then is-a-member-of-class-X is an operator, as is
is-large or sleeps. The inclusion relation d then has to be an Oot} operator,
since each of its arguments is a class, hence an O. This makes it syntactically
possible for the logic of propositions to be the same as the logic of classes: the
same notation applies for the ability of a class to include another as for a
proposition to be implied by another. In the case of mathematical notation,
the variables function somewhat similarly to different indefinite nouns, that
is, as zero-level N arguments having no limitation of likelihood in respect to
their operators (G70). Operators that map a set A, or its Cartesian product
A xA, into the same A, can be defined in mathematics on various sets; but in
natural language they can be defined only on sentences, i.e. on operators
with their arguments, and not on zero-level arguments (N).
Because of its differences from natural language, the notation of math¬
ematics is a separate linguistic system, a science language, rather than a
science sublanguage of natural language. But the grammar of mathematics,
its external metalanguage, is a science sublanguage, which has to be stated in
natural language. The natural-language discussion in mathematics, e.g. the
talk around proofs, or about discovering proofs, would presumably also be a
sublanguage. Although mathematics is a discovery of relations which exist
on their own, while language would seem rather to be an invention (even if
evolving and unintended), the system of mathematics requires the existence
of natural language in order to be formulatable; but for it to be exhibitable
(in its own notation) does not presuppose natural language. Note however
that the exhibited notation, like science languages in general, does not have
Sublanguages 307
the structure of a mathematical object (such as might aid in its deciphering in

the absence of a natural-language metalanguage).
By the side of the crucial similarities in sentence formation and the
differences in argument and operator classes, there are also encapsulated
differences in sentence structure. One is the status of variables as indefinite
N. The meanings of TV (other than indefinites) in natural language, to the
extent that they can be obtained in logic and mathematics, are provided by
class-membership predicates (compare All men are mortal with For allx, x is
a man implies x is mortal). The fact that the TV-variable position in logic and
mathematics formulas (sentences) is occupied only by the equivalent of an
indefinite noun, which contributes no likelihood (selection) limitations or
meaning to the sentence, makes the information in the sentences of logic and
mathematics arise entirely from the operators—unary predicates and binary
relations. In terms of syntax, it is this that makes mathematics a science of
relations, and makes a mathematical object a system defined by its relations
‘saying more and more about less and less’. In natural language too, there is
a possible grammatical description providing only one TV, an indefinite, so
that, for example, A book fell would be from Something which is a book fell
from Something fell; that something is a book, and similarly John phoned
from Someone who is called John phoned. But this complicates the status of
the is in Something is a book, etc.; and the class-membership predicates
would then be restricted by the same selection properties as the TV they
replace, though perhaps with compensatory advantages.
Another overt sentence-structure difference is in quantification, which in
logic is an operator on sentences (as in the for all x above). While the details
do not have to be discussed here, the language apparatus is close to that of
logic since plural, alf each, any, and the like are analyzed here not as
modifiers of nouns (as they appear overtly in language) but as derived from
operators (such as for each) on sentences (G51, 169, 252, 328, 337).
In reference, too, the difference is not major. Logic and mathematics
have conventionalized cross-reference under the scope of a quantifier, and
in the identity of symbols for variables within a proof or specified sections
thereof. The pronouns of language too serve for cross-reference rather than
for reference, with demonstrative pronouns being derivable from cross-
reference ones. But in the case of language, the great variety of positions for
antecedents of the cross-reference requires a special locating apparatus such
as that of the relative clause, which is absent (and not needed) in the syntax
of logic.
Aside from the lack of selection differences in nouns (variables), the
overriding difference between natural language on the one hand and the set-
theoretic systems—logic and mathematics—on the other is not so much in
constraints on sentence-structure but rather in constraints on sentence-
sequence. Although many things were known in ancient and non-written
mathematics without benefit of proof, the great edifice of mathematics and
logic rests on the methods of proof; the problems that have accumulated
around mathematics in the last hundred years, even in the criticisms made by
the intuitionists, have not involved the basic apparatus of proof (other than
the marginal method of the excluded middle). In the structure of proof—
with slightly different strategies in different mathematical areas—the barest
essentials are that in a sequence, certain initial sentences of a fixed kind be
evaluated by axiomatic status or by truth tables, with the constraint on the
sequence (or partially ordered set of sequences) being such that its last
sentence, the QED, is not less true than the initial ones. These essentials can
be stated a priori and inspectably only in conditions such as those that obtain
in mathematics: where the truth value of certain elementary (axiomatic)
sentences, underlying the initial sentences of a proof, can be determined by
their syntactic composition, or by listing, or in case of difficult infinite sets,
by constructive formulation; and then the truth transmission of the funda¬
mental operators is given explicitly in the truth tables. In natural languages,
these conditions are not met, largely because of the likelihood inequalities
among members of a word class, tied to the shifting knowledge or perception
of the real world.
Even within a science, the likelihood inequalities of operators and argu¬
ments in respect to each other are too numerous and too changeable for us to
establish any useful set of axiomatic sentences whose truth value can serve
non-trivial proof-sequences. However, as noted at the end of 10.5.4, weaker
types of sequence-constraints, suggesting consistency and consequence,
may be statable for science languages, largely based on regularities of word
choice within the co-occurring operators and arguments.
In a different way, the language in which computer programs are written,
and the compilers which mediate between them and the binary code of the
computer, all constitute science languages or linguistic systems not identical
with natural language. Each program is a text in such a language, following
constraints which lead—largely in a partially ordered way—from one step
to a later step. The flowcharts that were used to design programs, and the
line indentations used in writing programs, are examples of partial ordering
of the sentences of the text. Despite the similarities of programming
languages to natural language, the lack of ‘logical form’ in natural language
(cf. at the beginning of 10.6), and the present lack of any adequate
constraints that would relate conjunctions to the particular words under
them (10.5.4), exclude any present simulation, representation, or equivalence
of computer programs to natural language and to the thinking it presents.
10.7. Non-linguistic systems; music
It may be appropriate to consider whether there is any comparability

between natural language and other systems of notation and symbols, such
Sublanguages 309
as music, the language arts, animal communication, and deaf signing. The
major case here is music. As above in respect to mathematics, no attempt is
made here to characterize music except as to its difference from (and
similarity to) what has been seen here about language. In comparing the two
systems, we will see that while both are built out of constraints, the kind of
constraints and their relations—and their content—are entirely different. It
should be noted from the start that whereas all languages have in some
respects a common underlying system, and all fields of mathematics do (in
other respects), there is no evidence that there is something systemic
common to all music.
First we consider the physical elements. The first difference is that in
language the phonemes are essentially discrete objects because they are
defined by distinctions (7.1), whereas the notes of music are discrete partly
as a matter of convenient notation (except in the case of stopped instruments).
In both language and music, the physical elements are sounds, variously
combined in largely linear (or multilinear) order through time, and each has
attained discrete notation. One might think of comparing the notes of
music—written as discrete sounds with fixed lengths—with the phonemes of
language.17 But the phonemes are irreducible distinctions in respect to
repeatability and distinguishability of utterances. In contrast, musical notes
are direct cuts in what is first of all a continuous scale of musical sound,
although the relation to overtones is discrete. Between two successive notes
in a scale other notes (or musical sounds) can be recognized even if lacking a
notation; the like is meaningless between phonemes.
Notes are not arbitrary components of a musical phrase: they and their
relations in the scale are the inherent material of the phrase or chord even
though the whole can be shifted to a different key, that is to different notes in
a comparable scale relation. In contrast, the phonemes by which words are
recognized have no relation to the content of the word, and the same content
can be given synonymously or suppletively (or in different languages) by
completely unrelated phonemes; also the same phonemes can identify
homonymously a completely different word. The similarity of notes to
phonemes is slightly greater in the case of onomatopoetic words, but these
are necessarily irregular rarities in language (11.1). If the human ear could
distinguish as many single sounds as there are words in the language, there
would be no need for a separate stage of phonemic composition of words,
17 Differences among notes are differences in pitch, i.e. in sound-wave frequency. Differences
amono phonemes, however, are, acoustically, differences in timbre, i.e. in amplitude of the
various overtones, as are the differences between musical instruments. Hence the recognizability
of, say, the phoneme a, and its difference from e, is constant no matter at what pitch it is
pronounced (as in men’s and women’s voices). The existence of notes as distinct notations, or
discrete structures in musical instruments, and even the availability of more than one musical
instrument and of different timbres, are all limited to particular musical cultures and traditions.
Both the scope of the present book and the experience of the present writer unfortunately limit
the discussion here to the relatively recent culture of the ’Western world’.
but a comparable situation would not reduce musical phrases to a single

note. Also, the meaning of words and sentences can be given fully in writing
(including in ideograms) which can be understood by a reader who does not
know the sounds of the spoken form. In music, however, a score can indeed
be read without being heard, but this is done necessarily with respect to the
known sounds.
Aside from its sound elements, music differs from language in being able
to have a space of so many standardized note-lengths per bar, in which these
sounds are located; this is actualized primarily by differences in stress,
attack, etc. The distinction between this systemic timing and physical time is
seen in the fact that a ritardando can stretch the final phrases over more time
without altering their relations to the bars or (in many respects) to the
composition. This distinction is also seen more generally in the fact that the
same music can be played slower or faster than the indicated time, usually
without losing essential properties of the composition.
We next consider the functional entities of the system (i.e. the entities in
respect to further combination), which are formed from these elements. In
all languages, we have two segmentations: words as fixed sequences of
phonemes, and sentences as regular sequences of words—regular in respect
to the combinability of those words in all sentences in which they appear. In
music, short musical phrases can be considered comparable to words (with
the relevant chords serving as classifiers of a sort); this especially if the
composition uses them as explicit elements, e.g. by structured repetition or
variation of them. However, if we do not limit the consideration to particular
styles or compositional forms, music does not show two levels of segment¬
ation such as word and sentence, each recognized by systemic regularities in
the level below it. All this entails the lack in music of anything like
predication, since the basic regularity of sentences in respect to words, the
operator-argument dependence which creates the sentences, is unavailable
in music. Many varied relations exist between a longer musical line and its
subsegments, but not the single fixed dependence of words to sentence.
Furthermore, where in language expanded sentences are made out of
elementary sentences in much the same way that elementary sentences are
made out of words, there is no regular such relation of larger to shorter
stretches of musical composition.
As to effect, or meaning, of all these constructions: in language, there is a
fundamental difference, ultimately a single one (predication), between the
meaning of a sentence and that of the words which compose it. In music,
there may be noone essential difference between the effect of a composition,
short or long, and that of the musical phrases or lines of which it is
composed.
Another structural consideration is the relation of the sentences, paragraphs,
and chapters to the discourse, or in the case of music, the relation of the
larger musical segments to the whole composition. Both in linguistic systems
Sublanguages 311
and in much music, there are initial or underlying sentences or themes which
are repeated and modified in making the discourse or composition; and in
both there are subsections with internal regularities different from how the
subsections relate to the whole. In language we saw the double-array
requirement of repeating the word-relations (9.1,3); in science languages, a
few stronger similarities and progressions; in mathematics, the sharp con¬
straints that characterize proof. In music, on the other hand, a great variety
of complex modifications and developments of the themes and chords enter,
together with other elements, into the total musical work. These ‘discourse’
structures are obvious for example in rounds, fugues, the sonata form, and
differently in the demand to return to the tonic. One vague and second-
order similarity between compositions and texts is that, in both of them, the
structural relations among successive short segments (musical phrases;
sentences) are more stringent and ‘convincing’ than those among long
segments (movements; paragraphs or chapters).
In addition to the specific differences such as those noted above, the
overall structural difference is of two kinds. One is that linguistic systems—
including mathematics—are built by a few and permanent constraints,
whereas in music the constraints are more variegated, and differ from one
culture and period to another.1* The other difference is that in linguistic
systems most of the complexity of sentences and discourses arises from non-
informational elements—the transformations in shape. The meaning-effect
that this complexity contributes is quite different from that of the initial
informational sentence-making constraints. These last are simple in them¬
selves. In contrast, in music complexity can arise from the intricacy and
interweaving of the initial constraints themselves. The effect of the com¬
plexity is not necessarily different in kind from that of the smaller com¬
positional elements on which the complexity is carried out. The nearest
thing to grammatical transformations is in the musical form of variations
upon a theme; but the compositional status of this form is not different from
other compositional complexities.
A major issue at this point is the relation of the structure to the content,
to what is being said. In the case of language, it seems that the content
is achieved by the real-world associations of both the word-choice and
the initial constraints (11.4, 6). We cannot say that something comparable
holds for music, but we know that a composition has an effect, which
may be related to the affective-experience association of its sounds and
sound-sequences, its harmonic structure, and the way its compositional
constraints are developed. We may also suspect that just as certain gram¬
matical features are unrelated to the information of the sentence (11.3.3), so
lx The permanence of basic constraints in natural language and in mathematics is what

makes translation possible between languages that have in detail different grammars. In the
case of mathematics, intertranslatability is what is called ‘abstract equivalence' between
mathematical fields. In music, there is no intertranslatability.
certain compositional features may be unrelated to the content or effect of

the music. Nevertheless, having seen the contribution that the sentential
word-dependence and the further constraints make to the informational
capacity of language (11.3.2), we can understand that the structure of a
musical composition—as much or more than the themes—can be a direct
bearer of what is expressed in the composition. Indeed, some musical
themes may have little expressive effect (‘meaning’) until the composition
works on them, as may be the case for the aria da capo in a set of variations.
The relevance of combinatorial constraints (rules of composition) to content
can be seen in the need to ‘solve problems' in composing, and to change
styles in music and art in order to express fresh things: not to compose today
in the style of Mozart or to paint in the style of Rembrandt, no matter how
much the originals may still be appreciated today. Given this we can
understand how a composition or painting can express to us meanings, by
virtue of the structure we see in it, that the maker may not have felt
explicitly. And we can understand that new developments in the formalism—
the ‘language’—of music and the other arts can contribute to what is sayable
in those arts. New devices, local or global, in the language or style of an art
can carry or facilitate new meanings. With all this, the rules of composition
no more suffice to make interesting music than the rules of grammar suffice
to make interesting discourses. But we can see how popular art and ‘high’ art
can be housed in the same system, as ordinary conversations and great
literature can all be housed in natural language.19
We now turn to the content of all these systems. We have seen that the
specific content of natural language is transmissible information, a capacity
provided by its essential discreteness, its word meanings, its dependence
(predication) structure, and its likelihood inequalities. Comparably, the
content of mathematics is preservation of truth values through a sentence-
sequence, a capacity provided by its having only the indefinite N (variables)
with no likelihood variation, its structurally characterized truth tables, and
its constraints on proof-sequence.
Although the universe to which language can refer is roughly the whole
perceived universe—language can talk ‘about everything’ including itself—-
what language expresses structurally is just ‘about’ this universe, i.e. just the
information of it (ch. 11). Furthermore, the discreteness of phonemes and
the conventional pre-set finite stock of words at any one time assures the
largely error-free transmissibility of language, so that what the final hearer
receives is in principle the same substantive information that the original
speaker issued. Thus language cannot indicate quite everything that is
19 The contentual power of a language in art, corresponding to that of a grammar in natural
language, may perhaps be seen in the need to change the language of music after Bach, and in
the inability to paint purely cubist paintings after cubism reached its peak. And the difference
between a language of art and an expression made in that language may perhaps be seen in the
similarities and differences between corresponding cubist paintings of the otherwise very
different Braque and Picasso.
Sublanguages 313
indicated, say, in a picture. It cannot express in a direct and full manner the
whole range of human meaning—what a person ‘means’ to himself, what an
experience means to him, what precisely is the sensation of pain or whatever
that he has. In particular, language cannot give by its structure any direct
expression to emotion, except by bringing in extra-grammatical material,
i.e. items which have no regularities of co-occurrence in respect to other
items of the language: intonations such as of sarcasm or anger, or marginal
constructions and words such as How wonderful, Gee/, Ouch. Feelings are
carried in the language only by informational sentences, possibly false,
which state that the speaker has the feeling (/ am happy), or else by art-
activity (on which more below) of manipulating language, in ways, even
non-grammatical, that carry emotional or aesthetic impact.
In 11.6 it is suggested that the capacity of language to have an effect or to
‘mean’ is by virtue of the association of its elementary parts with something
in the world of experience, and by the constraints on combination of these
parts—or rather, by the association of these constraints with aspects of these
experiences. Further, the effect that language structure has is not the whole
range of meaning but only information, and more specifically the kind of
public information that can be transmitted with little or no error. Music too
has parts, and constraints on their combination: notes as compared with
phonemes, musical phrases as compared with words, methods of composition
compared with syntax. Despite some physical similarities between the parts
and constraints of language and of music, they differ deeply in their
structural properties and in their content—in the aspects of experience with
which they may be associated. However, the very fact that meaning—one
might say subject matter—in language is limited to information leaves room
for important areas of meaning or expression to be covered elsewhere.
Indeed, if language is specifically an information system (to express or
communicate) rather than generally an expression system (for all content),
then all types of expressing other than informational ones must be lodged in
other modes of expression and other types of activity.20
As to music and the arts, the content which they indubitably have must be
related to human experience, since art is made by people and responded to
by people. Many features in the constraints of art and the response to it, for
example the difference in response to art as against the response to decoration,
suggests that here, as in language, the effect (meaning) may be in response
to the kind and complexity of constraint (or stabilities, instabilities, and
organizings) which is met with in the course of human experience—different
from the constraints relevant to language (11.6), but nevertheless constraints
20 How different are informational expression and non-informational expression is seen in

the fact that most language expression is communicated only once, after which the recipient
hearer can use it accretively with other received information. In contrast, the expressions of
music and other arts are often accepted many times by the same recipient, and are not
combinable with other expressions in any regular way to comprise some joint result.
on experience. These meanings in art do not have to be of the same kind as

those talked about in language, for there may be feelings or sensings which
are too private or subterranean or non-informational to have been identified
in the fixed, publicly recognizable, and informational meanings of words.-1
The difference between the meanings of music and of language can be
seen most sharply when language and music collaborate, as in songs, arias,
and recitatives. We may think that much of the music in Bach's St Matthew’s
Passion fits the words marvellously well, or rather fits what Bach undoubtedly
felt about those words; but people of the modern world can respond perhaps
as fully to the music without responding comparably to the words. A given
piece of music can support a variety of texts expressing quite different
information (e.g. in recitatives), and different specific emotions (as when
revolutionary poems replace the words of folk-songs). Similarly, a given text
can be set to music in quite different ways (for example, the Mass). Indeed,
in order to increase the aptness of music to words, the Lieder tradition uses
items of theater—word-phrasings, facial expression, etc.—which no matter
how they enrich the performance are additions to, rather than parts of, the
musical composition.
These rather obvious observations are brought here only in order to point
out that pieces of music cannot be identified with those things that language
can identify—certainly not with information, but even not with those
feelings that are connected with specific information or specific situations
discernible in language. This is all the more obvious in the 'pure' instrumental
compositions, which can have great emotional and aesthetic impact without
our being able to characterize it in terms of language or specific content.
This holds even though many musical compositions are felt to be ‘saying’
(or rather, expressing) something (e.g. the late quartets and sonatas of
Beethoven), and some compositions more deeply than others, as some
language discourses are deeper than others. The inadequacy of language for
music (and the converse) holds even when what music is ‘saying’ may seem
to be indicated by the words of its title or by epigraphs. Music does not carry
or express information in the sense that language does, and neither music
nor language can be translated—even roughly—into the other (as mathe¬
matical notation can be translated into language).
The fact that the informational specialization of language leaves room for
other things to be otherwise expressed does not open the door for the non-
rational to intervene in the semantic domain of language. Music can indeed
say things, but not things that can be put into words. Its essential divorce
from information suggests that music—and art in general—does not provide us
with some ‘immediate knowledge’ of the world such as could be added to our
language-borne knowledge. Language and reasoning are effected by con-
21 The fact that to this day we cannot say—or agree—what it isthat music expresses serves as
a cautionary tale about the otherwise invisible limits of what can be said in language, or at least
about the difference in content between language and music.
Sublanguages 315
straints on the form and on the sequence of predications, none of which are
available to art or to feeling. It is not that these last cannot complement what
we have in language, but they do not supplement it within its own universe
by supplying any immanent material that can then enter into, and correct,
the specific predicational and sequential-argument structure (10.5.4) which
constitutes the informational power of language. Indeed, as language is
important to us in giving us some structuring of the mass of objects and
events which we experience, music and art may be important to us in giving
us some structured expression of the sensings and feelings which we experience
both internally and externally.
In this sense, and given the difference in structure between language and
the arts, Pascal’s The heart has reasons of its own, unknown to reason’ is
questionable: any extra-rational status or validity ascribed to ‘the heart’ (for
Pascal, faith) cannot add to or correct what has been reached by the rational
structure of language. Whatever it is that man finds in art, or for that matter
in art-like response to nature, it cannot contribute to what man finds through
language; it cannot make him wiser in respect to his public problems,
whether material or social. Conversely, art and feelings cannot be said to be
by their nature ‘true’. In the matter of validation, the rationality of language
may be considered to be a way of dealing with the world, with successful
(effective) outcome; but the arts may be considered to be affective responses
to the world, with ‘success’ being either undefined or highly indirect. It is not
the case that what is said may be erroneous or false whereas what is felt is
ipso facto authentic and valid. There may be ‘false’ emotions, and cliche
emotions, which a person may indeed feel but which may be an artefact of
inculcation, or of institutional and public pressures which led him to feel so.
What is more, while the validity of ideas and statements in language can be
tested by reasonably explicit, direct, and public methods of critique (10.5.4-
5), the validity of feelings can be tested only by such complex considerations
as those introduced by Freud and by Marx. Somewhat as one may consider a
discourse superficial, one may consider a composition superficial. And as
one may question an idea stated in language, one might question a feeling,
for example awe, which one might consider no longer befitting man’s
relation to nature and society. More generally, as language cannot directly
express emotions, art cannot directly express ideas. Artists may be influenced
by cultural winds and social movements, but that content appears in their
work only in secondary and often amorphous ways.
The difference between information and feeling may also clarify some of
the difficulties in discussions of mystical experiences. The latter may consist
of real emotional experience associated with some particular informational
(factual) content, but where no publicly relevant connection can be made
between the experience and the informational content (whereas in language
such connection exists, 11.6). The difficulties may not be so great if we
understood that the experiences exist as such, due to whatever conditions.
but that the associated factual content which triggered the emotions in that
experience may not be contentually related to that experience, any more
than the words of a song are contentually related to its music.
A similarity between language and music is that in both of these the
creator of the discourse expresses a meaning and the hearer receives a
meaning, an activity quite different from, say, play or work. Music however
is unlike language in the matching of the author's act and the recipient’s
perception. In language, by virtue of its publicly established phonemic
distinctions and pre-set vocabulary, the hearer is expected to get out of a
sentence precisely what the speaker put into it; any appreciable difference
constitutes a failure in language use—in information transmitting. An
important failure, for example, is the inability of the speaker to indicate, in a
general way within the grammatical apparatus, which word is the antecedent
for a particular occurrence of he or the like. There are also types of
mismatches which are not due to grammar inadequacy. One type is when the
speaker uses circumlocutions or euphemisms because of custom, or of
personal or social discomfort with what he has to say, in which case the
hearer is expected to recognize the shift and to correct for it, understanding
the statement to mean what is intended rather than what is actually said.
Another is when the speaker speaks with indirection, hoping to mislead his
hearer into accepting what he actually says rather than perceiving his intent.
But aside from such special cases, a sentence is correctly understood only
when the hearer receives as closely as possible what the speaker meant or
intended.
In contrast, the arts have no apparatus to ensure that what the maker
meant is what the recipient receives. It is accepted in music and the arts that
the recipient—viewer, hearer, performer—may perceive something different
from (though presumably related to) what the maker meant or intended. It
is acceptable (at least to many) that different historical periods have
different ways of playing Bach, and that music written for Baroque instru¬
ments may be played on modern ones, and certainly that different performers
may interpret a composition differently—that is, play it differently, see
different things in it—even when the composer’s own performance is
known, as for example in Stravinsky’s conducting of his own works. The
importance of performing tradition and individual performance over and
above the score is not just a matter of incomplete notation.
This difference between reception in language and reception in the arts is
understandable when we keep in mind that language transmits more or less
explicit information from person to person, in a system based on full
repeatability of pre-set material, whereas in the arts, the maker expresses an
emotion or a sense of something in ways that call out feelings or sensings, not
necessarily identical, in the recipient; the recipient’s respondent feelings
depend in part on his own emotional nature and experience. Indeed, the
maker can manipulate materials and symbols as a technical or random
Sublanguages 317
activity, without an intended expression of feeling, and yet the recipient may
react with feeling, as he may react to unintended nature. By the same token,
when the work of art contains expression of feeling or intentional structure
from the maker, it is not necessarily the case that the feelings it evokes, or
the structure that is perceived, in the recipient are precisely the same as the
maker’s, or the same for all recipients. Thus expression and communication
are not identical in art, and transmission is only a secondary use of it. In the
absence of systemic control over repetition and sameness, such as phonemic
distinctions provide for language, authentication of what the maker did or
meant, and the communicational problem of maker-recipient differences,
are not systematically resolvable.22
Several activities other than music should be mentioned here in order to
distinguish them from linguistic systems. One is the use of language itself for
purposes of art. This is not a second function of language, but a re-use or
manipulation of language. The language remains in its original and only
definition: words with meanings and operator-argument statuses, sentences
as informational operator-argument combinations. The art use of language
adds to this—and also alters it—by introducing not only additional features
of discourse structure but also non-grammatical features: sound similarities
(onomatopoeia, alliteration, rhyme), time-intervals (syllable count, stress,
rhythm, rhyme), play on word meanings (including allusion, nonce extension
of selection). Modern literary devices also violate some of the grammatical
requirements, from word combinations that are not within selection to ones
that clash with the operator-argument structure or linearization. All this
may be interconnected with the meanings of language material, that is with
its ordinary use; or it may be virtually independent of that. In any case, it
does not constitute an independent system, but rather a set of structural
modifications whose purpose is tangential to the ‘function’ of language or
totally independent of it. These modifications may, of course, follow a
systematic or even conventionalized way of manipulating language or
playing upon it, with emotional or aesthetic effect. The property of acting
within and against rules of a system is more pronounced in music and the
language arts, as it is in puzzles and games, than it is in such arts as painting
and sculpture where the artist creates his independent arrangements of
objects or symbols.
Another activity that has to be mentioned is animal communication,
whether cries or standardized behavior (postures, etc.). There is little
question that some of these are expressive for the individual, and some are in
effect communicative for the group. The only thing that has to be said here is
that these activities differ from human speech not only in degree—the
vocabulary being minute—but also in kind, because there is no operator
argument relation, no predication, no organized likelihood-inequalities, or
22 On authenticity in art, and also on notations in art (10.5.5 above), cf. Nelson Goodman,
Languages of Art, Hackett, Indianapolis, 1976.
anything approaching the structure that makes sentences (12.4.1). It should

be noted that a two or three orders of magnitude increase in vocabulary is in
itself a difference in kind, because the whole vocabulary can then no longer
be efficiently utilized without classification, and because its possible com¬
binations are then so many as to require constraints and regularities.
Finally, it seems that the sign language of the deaf does not have an
explicit operator-argument partial ordering, nor an internal metalanguage,
but rests upon a direct juxtaposition of the relevant referents. This applies to
autonomous sign languages, developed by the deaf without instruction from
people who know spoken language.
V
Interpretation
•i
11
Information
11.0. Introduction
Language is clearly and above all a bearer of meaning. Not, however, of all
meaning. Many human activities and states have meaning for us, and only
some of this can be expressed in language: feelings and vague sensings can be
referred to only indirectly (10.7); non-public information, such as pro¬
prioceptive sensations, can be named only with difficulty; certain kinds of
non-language information can be translated directly into language (e.g.
graphs and charts); but other kinds (e.g. photographs) can be represented in
language only loosely and selectively. Meaning itself is a concept of no clear
definition, and the parts of it that can be expressed in language are not an
otherwise characterized subset of meaning. Indeed, since language-borne
meaning cannot be organized or structured except in respect to words and
sentences of language (2.4), we cannot correlate the structure and meanings
of language with any independently known catalogue or structure of meaning.
In each language, we do not know a priori which specific aspects of mean¬
ing will be referred to by words, and how much will be included in the
meaning of a single word. Even when the meanings are well-defined, it is not
always possible for words of a language to mirror in their structural relations
the relation among the referents. For example, the relation among the
integers is given in Peano’s axioms, but language cannot thereupon name
them as one, successor of one, successor of successor of one, etc. What
languages do in this case is to give arbitrary names to the first ten (or six, or
whatever), and then to use these and other names to indicate successive
multiples of the initial set: the integers are thus named modulo that initial
set, by a language-based decimal or other expansion. Such cases suggest that
we have to study the specific words and structures of a language if we wish to
see what meanings they cover, and how.
As has been seen (5.4), the phonemes are irrelevant to meaning though
they underlie communicability: the phoneme-sequences that constitute
words are not in general constrained in any way that is relevant to the
meanings of the words. The words are thus a fresh set of primitive elements,
which can be identified without phonemes (spoken or written) as in ideo¬
graphic writing, and which are associated with meanings, in a manner to be
discussed in 11.1 and 11.6. The meanings of words are modified by their
322 Interpretation
likelihood relations to their operators and arguments, and in some cases

differ according to their operators and arguments (11.1). And given words
with their meanings, it will be seen in 11.2-3 that the constraints on word
co-occurrence create further, ‘grammatical’, meanings. By separating all
occurrences of language into occurrences of words and constraints on their
co-occurrence (rather than syntax being simply a classification of word com¬
binations), we can show (1) that the meaning of a sentence or discourse is the
meaning of its words plus the meanings of its ordered constraints; and (2)
that each of these meanings is constant for all occurrences of the given word
or constraint in all sentences (even under transformations and zeroings,
3.3); that is, no meaning beyond these is carried by the structure of
grammatical transmissible language. Within a sentence, these constraints
are all included in the partial order of words in respect to sentences (3.1-2).
For discourses and sublanguages, the constraints have been described in
Chapters 9 and 10 and their meanings will be described in 11.3.2.
All of the above yields structural locatings of specific contributions to
the meaning of a sentence (11.4). In addition, it will be seen that language
is limited to particular kinds of meaning (11.6), and that information
has in certain respects a language-like structure (11.7). The relevance
of information to language is seen also in the fact that non-contributing of
information is the grounds for zeroing (3.2.6).
11.1. How words carry meaning
The first question is: what are the least parts (forms) in language to which
meaning adheres? Not phonemes: there are no regular ways in which
the meaning of the words of a language can be obtained as combinations
of meanings assigned to the phonemes of those words. Indeed, phonemes do
not in general have any linguistic meanings; and the pair test which establishes
phonemic distinctions is related not to the meanings of the tested words but
only to the recognizability of their repetitions (7.1).1 If in all the occurrences
of a word the phonemes were replaced by others, we would simply have a
variant form of the same word. But if we replaced some or all of the
occurrences of a word by a word which had different selection, i.e. whose
normal occurrences were different, we would have a different meaning.
1 lence onomatopoeia, in which the sounds of a word suggest its meaning, is a
rarity in language. By the same token, the cases in which different morphemes
that are similar in sound are also similar in meaning must be haphazard and
1 As noted at the end of 7.4, the phonemic composition of words is needed because there are
not enough pronounceable and audible distinctions among fixed single sounds to distinguish all
the vocabulary of a language. The structural and semantic properties of words can be carried
even by having a single sign for each word, as in ideographic writing.
Information 323
not regular; and these sound-meaning correlations are then not usable in
any regular way in the grammar.
However, in each language there are certain listable phoneme-sequences,
called words or morphemes, which carry fixed meanings. It has been seen
that these sequences can be isolated out of the flow of speech even without
any knowledge that they are the words of the language, or what their
meaning is (7.3, 11.3). A sequence of phonemes cannot be established as a
word simply on grounds of being judged to have meaning of its own; it must
also have regularities of occurrence relative to established sets of other
words. There are cases of a phoneme-sequence which seems to have a
characteristic meaning but whose immediate environments do not them¬
selves combine in a sufficiently regular way, as morphemes. The morphemic
status of such phonemic sequences may therefore be border-line. Such are,
in English, the si- of slick, slip, slither, slide, slimy, slink, sling, slog, slosh,
slouch, slow, etc., and the gl- of gleam, glow, gloom, glisten, glitter, glare,
glide, and the//- of flick, flip, flit, float, flash, flare, flap, flop, flutter, fling,
flow, flee, fly (7.4). Such is also the case for -le in scores of words; handle,
dazzle, nozzle, nuzzle, muzzle, etc., even though in some words the -le had
been a more regularly combining English morpheme. On the other hand, a
phoneme-sequence may have to be accepted as a morpheme even when it
has no assignable meaning, just on grounds of its regularity of combination;
e.g. the re- of recommend.
In view of all this it is necessary to consider what is the source and status of
word meanings, and how these relate to the combinational regularities of the
words: how do words carry meaning? First, it is clear that most words have
some association with meaning independently of their occurrence in sentences
(i.e. in combination). Since a large vocabulary was presumably in use before
sentences developed (12.3.1), these words must have had meanings in
isolation. In addition, many words that enter the language at various times
get their initial meanings from the experiences that gave rise to them rather
than from any grammatical combinations. I his applies to newly made words
such as gas and boysenberry, to borrowings whose meanings may be adapted
or specialized from the meanings in the source language (e.g. opera,piano),
and also to word combinations that are not made on a syntactic basis (e.g.
flip-flop, wishy-washy, wheeler-dealer). In addition, many words change or
specialize their meanings in the course of time on the basis of cultural and
historical developments, without apparent regard to their etymology or to
the syntactic combinations into which they otherwise enter: examples are
countless, e.g. Quaker, verb, plane. And of course very many syntactic word
combinations have a specialized meaning beyond their composition: whereas
school-books are any books specifically for school, snow-shoes are not merely
any shoes specifically for snow but a particular type ot object.
How words are associated with their non-syntax-based meaning will be
discussed in 11.6. Before reaching that question, however, it is important to
324 Interpretation
note that this association of a word is in general not to a unique object,

event, state, or experience, but to a set of these—a set which outside of a few
special conditions (such as in science) is not fully delimited and defined
either extensionally or intensionally.
As to uniqueness of referent: there is nothing to prevent proper names
from being pluralized (hence able to refer to more than one object), or from
being used about more than one referent. A pet can be called Zeus or
Napoleon; and while the names thus used contain allusions to their original
bearers, it is hard to say that they retain the original bearer as referent.
Similarly, geographic names can be used for more than one location or
formation. As to the extension of referents, it may be impossible to
determine the complete set of individuals which would constitute the
denotation of a given word. It thus seems that words cannot be limited to a
single referent or a closed and unextendable set of referents. They cannot be
defined by their extension.
As to the intension of a word, for most words it is difficult to state a
definition, or a set of alternative definitions, which would precisely characterize
all cases in which the word can be used; this even though one might be able to
state definitions that would suffice to distinguish the meaning of the given
word from that of all other words (a much more delimited task). As will be
seen, words in general refer to properties and classifications (of objects or
events) which would not necessarily be otherwise distinguished—i.e. except
for their being named by the same word. Words thus select properties, and
classify their bearers (referents).
There are also various problems in fitting together these two aspects of
word meaning: the set of referents for which a word is used—its denotation
or extension—and the sense which it conveys—its intension or dictionary
definitions. Not only can a word used in a single meaning refer to different
individuals who may not otherwise be identical or equivalent, but also two
words or phrases with different senses can unbeknown to speaker and hearer
or known to one or both refer to the same individual: Sir Walter Scott and
the author of the Waverley novels, New Amsterdam and New York, Franklin
D. Roosevelt and That Man (as he was called by opponents), the evening star
and the morning star ?
For some words there is just one specific meaning as in boysenberry, or
closely related meanings as in time. For many words the meaning is a set of
referents with a fuzzy boundary: e.g. as to what is included in chair (a chaise
longue?, a one-legged spectator seat?, a rock hollowed out for sitting?).
A word may have a single more or less continuous range of meanings, as for
chair; or more than one range of meanings, as for station (rail and bus, radio
and television, but not telephone; meteorology, and perhaps not astronomy,
etc.), or for field (agricultural, sports, region, range of interest, and as
: This difference is involved also in words, such as unicorn (11.6.2), whose set of referents in
the real world is null, hut whose sense or meaning is not empty.
Information 325
defined in physics and in mathematics). The distinctions and boundaries of

meaning-range are complicated and vague. It will be seen that meaning is
more easily stated as a property of word combinations (or of words in
combination) than of words by themselves.
Although meanings exist for most words in isolation, it can readily be seen
that this is not all there is to say about word meanings. In the first place, a
phonemic sequence may have different meanings, and be considered different
words, when in the environment of different words, e.g. two words pen in
Fill the pen with pigs and Fill the pen with ink, in contrast to one word spill in
The pigs spilled out of the pen and The ink spilled out of the pen. This occurs
also for many words which may be considered the same in their different
occurrences despite their different meanings: e.g. for spur in The spur on the
riding boots was too short and The spur inside the flower was too short and a
railroad spur, or policy in Fire losses (or: Foreign adventures) are not
included in our policy and Foreign adventures (or: Industry bail-outs) are not
included in our policy. Also, two words which are very different in meaning
may be almost synonymous in a particular environment, although each
keeps its own meaning throughout: e.g. the cells divide and the cells multiply.
The environment is thus a factor in the meaning of words.3
The dependence of meaning on environment clarifies somewhat the
problem of meaning-range. Once we see that the meaning of a word in a
particular occurrence depends on its environment, the categorization of
meaning-ranges can be replaced by a categorization of the environing
words. Note that when we judge the meaning of a word-occurrence by what
can replace it in the text, we are really judging the meaning by textual
environment. We need not ask the virtually uninvestigable question: is the
meaning-range of all occurrences of a given word a single continuous one, or
a composite of several ranges? Instead we ask: do the environing words, of
the various occurrences of our given word, fall into distinct sets? The
environments of spring of metal and spring of gushing water are readily
distinguishable, even if we can find a few common to both (e.g. The force of
the spring is surprising)', this even though these words share an etymology
(‘are the same word etymologically’). Such is the case also for table as a piece
of furniture and table as presentation of data, although the awareness of a
metaphor-like relation is stronger here. In contrast, a clear separation may
not be possible among the environments of all occurrences of field (above).
Environment distinctions, which are more measurable than pure meaning-
distinctions, can make it possible to distinguish homonyms from separate
meanings of one word, and the latter from an indivisible, coherent meaning-
range.
3 Even for the many words whose meaning is constant the semantic effect may differ in
different environments. When said of an unknown object, small, large may be used relative to
the size of speakers and hearers; hut said of a known object, these are relative to its class of
objects: small elephant, large flea.
326 Interpretation
Observation of the data makes it clear that the relevant environment for
meaning-specification of a word-occurrence is its immediate operator or
argument, more rarely its farther operators or arguments and its co¬
argument; modifiers of a word are counted as operators on it in a secondary
sentence (3.4.4). Because the environment of a word delimits its meaning,
the meaning of a syntactic construction is sharper than the meaning-ranges
of the component words taken separately. Thus the correlation of meaning
with syntactic constructions of words is more precise than it is with words
alone; and indeed, a dictionary is frequently forced to refer to grammatical
environments if it is to present word-meanings with any precision. The
meaning is in the words, but the precision for each occurrence often comes
with the combination. Cases in which environment differences do not match
meaning-differences are only apparently such: for example, there is much
similarity in the environments of certain words whose meanings are opposites
(left!right, small!large, etc.); it is relevant here that many opposites (good/
bad, young/old, mother!daughter) are known to have important semantic
features in common. Indeed, certain aphasias and slips of tongue can replace
a word by its opposite.
New extendings of environment and meaning are constantly being made.
They are not determined merely by the speaker's intent. In many cases there
are alternative possibilities for the environment, and the extending is
affected by various factors. Thus when a new operator was needed for the
motion of a plane in the air, fly was not the only possible choice (the
Zeppelins sailed): similarly for fly as the moving of a flag. When words are
needed for incapacitating a tank, or for eliminating a bill in Congress, kill is
not the only word with a somewhat appropriate meaning. And steamships
and motor-ships sail, but motor-boats do not. In the many such cases of new
word combinations, the meaning-range of the chosen word (fly, kill, sail,
etc.) changes to accommodate the new use.
A case of this is seen when metaphors, originating as a simile, yield a new
meaning of a word, as in the case of see in the sense of ‘understand’ (/ see
what you're saying), or in the meaning of understand itself as it developed
out of its etymological source. Metaphor is an extreme case of an apparent,
or de facto, change of meaning in a word. It arises in a simile when indefinite
nouns and verbs, and also the as of the simile itself, have been zeroed,
leaving the remaining word to carry the meaning of the zeroed material in
addition to its own. If the metaphoric see above is derived from some source
such as / treat what you're saying as one's seeing things (through an unsaid /
treat-as-one's-seeing-things what you're saying, reduced to I treat-as-seeing
what you're saying, G405), then the metaphoric see has the meaning of the
original physical see plus the meaning of as (while the verb-classifier treat and
the indefinites one, things, carry little meaning), ‘I as-though-see what
you’re saying’. Other cases of words which are left to carry the meanings of
zeroed operators on them are seen for example in the ‘zero causative' of He
Information 327
walked the dog from loughly He took the dog walking (through an unsaid but
grammatically constructible He took walking the dog). Such situations can
even lead to words having apparently opposite meanings, as in He dusted the
furniture from He treated the furniture for (its) dust as against He dusted the
crops from He treated the crops with (insecticidal) dust. A different case is
that of abstract nouns, as nominalized operators, which arise when the
indefinite arguments of an operator are zeroed in a nominalized sentence
(i.e. a sentence under a further operator): e.g. hope from people’s hoping
things (in Hope springs eternal), vehemence from one’s being vehement (as in
Vehemence is uncalled for here, 11.2). ‘Lexical shift’ (i.e. meaning-shift in
words) is thus simply the effect of environment extension plus zeroing.
The most general factor in the varied and changing meanings of words is
simply the constant though small change in likelihoods—what words are
chosen as operators and arguments of other words, and how frequently they
are thus used. Beyond the argument-requirement dependence, the grammar
does not limit each word to particular other words as its operator or
arguments: any On operator can make a sentence—even if a nonsensical
one—with any N word, and so on. Given this, the particular choice (or
grading of likelihoods) of words depends largely on what makes sense (or
intended nonsense), given the current meanings of the words; but it depends
also on marginal distinctions that the speaker wants to make, on attention-
getting surprise, on euphemism, on the analogy of how related words
combine, and on other factors.
The changing frequency and acceptance of certain combinations affect
what comes to be seen as the main meaning of a word, as do increasing
specializations (i.e. frequencies of particular meanings) of word use. Over
time one sees many changes in combination which produce recognizable
differences in meaning, e.g. as between / was flying in a 747 (an extension in
the meaning of fly) and A moth was flying in the 747. There are words in
which successively extending the frequency of the environment-choices has
resulted in large differences in meaning: for example, within the history of
English, fond from ‘tasteless’ to ‘affectionate’, like from ‘to please’ to ‘to be
pleased by, to favor’; in wear, by the side of the continuing meaning ‘to be
clothed in’ there was also a development to ‘to deteriorate’ and ‘to pass’ (The
night wore on).4
Although the meaning of any word at a particular time is a factor in
contemporary choices of operators or arguments for it, the meaning is
clearly not sufficient to determine those choices. But those choices in turn
affect what is understood as the meaning of that word. After new choices of
operator or argument have been made for a word, we can often find an
4 Changes may he so natural, given the changing culture, as not to be noticeable. Thus, the
old verities (literally 'the old truths’) once referred to what users of the language must have
understood as, indeed, ‘the old truths’. By now our suspicions as to the once-accepted truths are
such that when we say the old verities we are close to meaning the old falsehoods .
328 Interpretation
extended core-meaning for that word, such as would fit both its old and its
new environments (e.g. discharge an employee, an obligation, an electric
potential difference). But we would not necessarily have come to this
meaning-core before the new environments had come to be used: consider
to charge the enemy, or the accused, or the purchase, and to charge a battery.
That is, knowing the meaning of a word does not suffice for us to predict
what new environments would or would not be used with that word.
Different resolutions in this respect are found in different languages:
thus although English and French have rather similar systems of kinship
nomenclature, seemingly opposite choices appear in grand-daughter as
against petite-file, each being a reasonable way of adding ‘one stage farther’
to daughter, but grandmother is paralleled by grand'mere (and aieule). (Note
that great-granddaughter is arriere-petite-fille, and for great-grandmother
one can say not only arriere-grand'mere but also bisaieule.) Extending to
new environments is all the more complex when we consider nonce forms,6
metaphor, analogy (especially that), and literary turns of phrase. All these
many-faceted similarities and differences in meanings of words can create
regularities and patterns of meaning-relation among words, which are a
source of interest in lexicography and literary criticism.
Attempts to categorize and to find regularities in extending and changing
meanings of words have not fared well. But what is important is that such
attempts have succeeded far better when directed to the choices of environ¬
ment rather to any inherent dynamics (such as narrowing or widening) in
meanings as such. That is, a word is extended from one environment to
another on environmental grounds, e.g. on the basis of the environments in
which that word already occurs in comparison with the environments in
which variously related words occur.6
The meaning that a word has with a particular operator or argument is
constant for all its occurrences with that operator or argument. Thus not
only is this local meaning more sharply defined than the general meaning-
range of the word, but also it is this meaning that is preserved under further
operators. This meaning is also generally preserved under transformations.
However, a word may seem to change meaning when it incorporates the
meaning of attached words which have been zeroed, as in expect John in the
sense of expect his arrival, and in She’s expecting for being pregnant. In
contrast, when words come to be used in greatly broadened selections, their
meaning becomes less specific (and in some cases their form is reduced)
when they are in the broadened selection. For example, o/is derived from
off, and have is weakened from ‘possess’ to ‘be in (a situation)’ when it
In nonce forms and certain jokes a word is used in a new environment with the under¬
standing that this is not a precedent, that the meaning of the word should not be modified to fit
the new environment.
6 H. M. Hoenigswald, Language Change and Linguistic Reconstruction, Univ. of Chicago
Press, Chicago, 1960.
Information 329
occurs with all verbs to make the perfect tense. (Note that in the French
perfect tense, e.g. J'ai fini hier, the ai is shown by the hier to be no longer
present tense, but in English 1 have finished by now, where there is no I have
finished yesterday, the have is largely still present tense.) Furthermore, in
some cases a word combination changes its meaning so that it no longer
follows in a simple way from the meanings of the component words: these
are the idioms (e.g. kick the bucket, fix his feet). This may take place even if
the non-idiomatic meaning of the combination continues to exist, as in take
care of him both in the sense of caring for him and in the sense of neutralizing
him.
It would seem to follow from all this that meaning has to be stated as a
property of word-occurrences (i.e. of a word ‘token’) rather than of word
(i.e. of its ‘type’). But this is wasteful, since very many word-occurrences
have the same meaning. The regularity of meaning is in terms not of word as
such nor of word-occurrence, but of word combination: the word has fixed
different meanings when with different subsets of its selection (immediate,
or somewhat farther).
In sum: first, the meaning of a word in isolation is in most cases the
starting-point, and its meaning affects its current selection of operators and
arguments, though only in a loose way. Second, the operator-argument
selection of a word, while affected by its current meaning, can be changed by
the pattern of the selections of variously related words (especially in
analogic extension), which in turn affects its ensuing meaning. But it must be
kept in mind here that the selection of a word is not simply the set of
operators and arguments with which it occurs, but the set it occurs with in
above-average frequency (hence e.g. not as jokes); we include here cases
when the frequent co-occurrent is somewhat distant in the dependence
chain. It is thus that selection can be considered an indicator, and indeed a
measure, of meaning. Its approximate conformity to meaning is seen in that
we can expect that for any three words, if two of them are closer in meaning
to each other than they are to the third, they will also be closer in their
selection of operators and arguments.
By speaking of a word’s selection instead of a word’s meaning in isolation
we leave room, first, for different meanings (taken from the range in
isolation, or extended out from it) in the presence of different operators or
arguments, and second, for meanings to change and to change differently in
particular environments. The latter has been shown to occur not only on the
basis of processes of meaning change but also on the basis of environmental
similarities and differences. For example, a word can change meaning in
suffix position while it does not change in its free-word occurrences (compare
-ly with like).
Selection is objectively investigable and explicitly statable and subdivid-
able in a way that is not possible for meanings—whether as extension and
referents or as sense and definition. Indeed, one can take a reasonably large
330 Interpretation
sample of sentences or short discourses and list the complete selections for
various words, although in many cases there may be uncertainty as to
whether a particular operator or argument has selectional frequency for the
given word or is a rarer co-occurrent which does not affect its meaning.
Characterizing words by their selection allows for considering the kind and
degree of overlap, inclusion, and difference between words in respect to
their selection sets—something which might lead to syntax-based semantic
graphs (e.g. in kinship terms), and even to possibilities of decomposition
(factoring) for particular sets of words. Such structurings in the set of words
are possible because in most cases the selection of a word includes one or
more coherent ranges of selection (3.2). The effect of the coherent ranges is
that there is a clustering of particular operators around clusterings of
particular arguments, somewhat as in the sociometric clusterings of acquaint¬
ance and status (e.g. in charting who visits whom within a community).
Selection sidesteps certain boundary difficulties that are met with in
considering meaning. For example, words whose referents have at all times
null extension (unicorn, centaur, spiritualists’ ectoplasm, or for that matter
laissez-faire capitalism) present no problem here, because they each have
characteristic and stable selections, and ones which adequately distinguish,
say, unicorn from centaur. Such words can also be characterized by their
definitions. Indeed, the selection of a word can be used to generate the set of
its partial definitions: a chair is something on which one can sit, which in
most cases has four legs, etc. As to different words which are not really
synonyms but have the same referent (e.g. Sir Walter Scott and the author of
Waverley), these have or can acceptably have the same selection—but not in
sentences where their difference is involved (e.g. not in The king did not
know that Sir Walter Scott was the author of the Waverley novels). Words
whose definitions are not readily statable, or which may have no definition
or meaning, such as the spiritualists’ ectoplasm or there in such sentences as
There's a man here, can nevertheless be characterized by their syntactic
position and their selection.
Selection also treats directly the cases where meaning and phonemic
composition do not correlate: in suppletion, different phoneme-sequences
have the same meaning; here we note that they have the same environments
except for being complementary in respect to a particular part of the
environment (e.g. for is, am, are)—a situation which has a distant similarity
to the issue of Sir Walter Scott above. In free variation, partially different
phoneme-sequences have the same meaning; here we note that they have
the same environments (e.g. the two pronunciations of economics, or the not
and -n’t, and will and wo-, of will not, won't). In science languages (10.5) a
given word symbol, with fixed sentence-type positions, has a sequence of
English phonemes in English articles and a different sequence of French
phonemes in French articles. In homonymy, the same phoneme-sequence
can have different meanings; here the phoneme-sequence occurs in two or
Information 331
more unrelated selections—environments of kinds which do not otherwise

have the same phoneme-sequence selecting them coherently (3.2.1, e.g. see,
sea, Holy See). In synonymy, to the extent that it exists, different phoneme-
sequences have the same meaning and the same selection.
More generally, words which have certain kinds of selection have (or get)
certain semantic properties. For example, words which occur under virtually
all operators with little likelihood difference have the meaning of indefinites
(e.g. the set someone, something), and, interestingly enough, do not have
different meanings under different operators. Also, a small set of words or
affixes which have this indefinite-like property but are approximately com¬
plementary in respect to some set of further operators usually constitutes a
grammatical category or paradigm: e.g. the tenses in respect to adjoined
time words (Someone will act tomorrow, Someone acted yesterday), or the
plural and singular affixes in respect to adjoined number words (Three things
seemed important, One thing seemed important). Words that come to be
used with a large variety of arguments lose some of their individual meaning:
e.g. take, give, do, have in take a walk or a look, give a look or thought, do
work or a jig, have looked or have a look, etc. More closely, bi-sentential
operators (words or phrases) which become frequent on a wide variety of
sentence-pairs lose some of their specific meanings: e.g. / will go provided
(or: providing) he doesn't, as against the specific meaning in His setting up a
trust fund provided for her going to college-, similarly for is due to and other
conjunctional phrases (My writing you is due to her, or ... to her insistence,
as against There is still some money which is due to her). A sharper case of
this loss of specific meaning is in the prepositions, where the largely locative
original meanings (in the box, on the table) are almost lost in their much-
widened range of environments: interest in a problem, people in charge,
thoughts on a subject, bets on a horse.
We see then that meaning is specified not by a word but by a word with its
choice of operator or argument (usually its immediate ones) (cf. n. 9 below).
Indeed the vocabulary of a language is not large enough for the variety of
experience, and this is partially remedied by the meaning-range of words in
different environments (which include the secondary sentences that become
modifiers of a word and so can specify submeanings of it, as in steamship
from something like ship using steam, medical examination which is quite
different from doctoral examination, etc.). In general, the continuous nature
of much of experience does not always lend itself to discrete representation
by words. Words do divide up the world of meanings, and much of the
division is much the same in many languages, greatly facilitating translation.
But there are given words in one language which cannot be translated, or not
precisely, into another—if there are no words in the other language having
similar coherent selection to the given words in the first one; there are cases
of events and actions for which the other language has no word. We cannot
make words mean what we want, because we cannot restrict them to the
332 Interpretation
particular selection which would support that meaning. Nor can we reach
words in any organized way by starting from their meanings, quite apart
from the fact that there is no known way in which we can objectively and
consensually organize the world of meanings. But for the most part, the
argument-requirement of a word is preserved under translation: that is, On
words (e.g. walk) usually translate into On words, O0 words (e.g. occur) into
00, and so on. And for each word, what the learner or analyzer of a language
does is not to think deeply about what the word ‘really means’ but to see how
it is used, keeping in mind the fact that the various life-situations in which it
is used are largely reflected in the various word combinations in which it
occurs. The grammar cannot be built out of meanings; but it can be built out
of word combinations.
11.2. Information in sentences: Predication
One can readily see that the meaning of a sentence is more than a selection of
the meanings of its words: John saw Bill does not mean the same as Bill saw
John, and the meanings which we can garner from collecting the words John,
Mary, Tom, Smith, call, tell, will, to, and are not the explicit, and varied,
information that we get from Smith will tell John to call Mary and Tom and
Mary will call John Smith and tell Tom to. The sentence information over
and above the meaning of its words is given by the meaning of the operator-
argument relation. To see what this meaning is we note that the structural
property is not merely co-occurrence, or even frequent co-occurrence, but
rather dependence of a word on a set: an operator word does not appear in a
sentence unless a word—one or another—of its argument set is there (or has
been zeroed there). When that relation is satisfied in a word-sequence, the
words constitute a sentence. This statement does not exclude other words or
combinations, especially zero-level words, from constituting exceptional
types of ‘sentence’, as perhaps in Fire\
The operator is thus not something one says by itself, but rather something
one says with—about—the argument, a predication on it.7 A word that
requires the presence of one or another of certain other words means—in
addition to its own lexical meaning—that it waits on those other words, i.e.
that it is said about them. Hence each occurrence of the dependence relation
contributes a predication meaning (between its participants) to the sentence.
Thus the meaning or information of a sentence is not just a sum or
juxtaposition of the meanings of its words. And since the predication
relations in a sentence are a partial ordering of its words, we see that the
meaning of each partial ordering of words (i.e. each sentence) is that same
7 Predication is a relation between two meanings, and is the meaning of a relation between
two word-occurrences. It is not the same as assertion (5.5). In John s accepting is uncertain, the
accept is predicate on John, but not asserted, as also in John will accept if he is invited.
Information 333
partial ordering (i.e. predication relation) of the meanings of the words.

Thus not only the words but also their partial order in a sentence have a
semantic interpretation.
More generally, all grammatical relations among words in a sentence
contribute their particular meanings to the total meaning of the sentence,
over and above the meanings of the words (which are identified by their
phoneme-sequences). This holds for the various ‘modifier’ meanings of
relative clauses, adjectives, adverbs, secondary clauses, all as modifiers of
their hosts; and the meanings of pronouns as second occurrences of nearby-
occurring arguments.
Similarly, -ness, which makes a noun out of an adjective, is not just a suffix
meaning ‘abstract’; it has that meaning because in changing a predicated
property (e.g. serious) into an argument of further predication, the suffix is
attached to the adjective alone, independent of its subject, so that it seems to
name (nominalize) the property by itself: John is serious, John's seriousness
impresses people, or (1) Seriousness impresses people (from (2) Anyone's
seriousness impresses people). In the present theory, all relations other than
predication can be derived from the predication relation (e.g. for English in
G, passim). Their semantic contribution, however, is observable (or felt)
independently of this derivation: (1) is seen as speaking of seriousness in the
abstract, while in (2) it is speaking of the property of a specified subject,
which is the indefinite and hence zeroable, yielding (1).
Significantly, every grammatical relation except predication can be para¬
phrased by word combinations that do not u§e the relation in question: e.g.,
for the modifier relation: A new book arrived paraphrased by A book
arrived; it was new, or more precisely by A book—the book was new-
arrived. And what is most translatable between languages is the partially
ordered predication relations in a sentence: which words are predications on
which other words: the ordered predications are a constant of the two
paraphrastic (intertranslatable) sentences. In contrast, many meanings which
are carried by other grammatical relations or constructions in one language
are carried (and translated) in another language by some combination of
words without that special construction. It is thus impossible to express the
meaning of predication, in a given language or by translation into another,
except by word combinations that are constructed with predication. This is
because the dependence that creates the operator-argument relation does
not just carry a meaning like all word meanings. Its meaning produces
something new, which we may call language-information (11.7)-—the dif¬
ference between the meaning of words and the meaning of sentences. As will
be seen in 11.6, we can say that a given word, say tree, has its meaning by
virtue of its association with something in the world of experience and by its
range of combination with other meaning-bearing words in sentences.
Comparably in the case of predication, we can say that its characteristic
dependence has a meaning over and above the meaning of its words by
334 Interpretation
virtue of its association with specific conditions in the world of experience, so

that Trees fell does, not mean just any juxtaposition of these two meanings—
e.g. that the speaker saw or thought of trees and that something then fell—
but specifically that falling is a condition of the trees. This aboutness is what
creates an item of information out of a collection of meanings. Information
differs from meaning not only in that it reflects a situation, a fact, rather than
an object or a property or event, but also in that sentences may be felt
to ‘give’ information whereas words ‘have’ meaning: information charac¬
teristically records and communicates, where meaning indicates.
11.3. How language carries information
We have seen that when the constraints on form, i.e. on word combination,
are isolated we can distinguish between constraints that carry information
and a (mostly reductional) residue that does not. The constraints on
combination constitute a linguistic system by themselves, the base sub¬
language (4.1), but the residue does not. Thus the information in language is
the information in a structurally distinguished sublanguage, which consists
of base forms of each sentence in the whole language.
11.3.1. Where the information is located
In the base the relation of form to information is much sharper than in the
whole language, and here we can ask precisely what kinds of form carry what
kinds of information. We can take a base sentence, ultimately an elementary
sentence no proper part of which is a sentence, and adjoin to it various word-
sequences which make it into a longer sentence, and then ask what information
has been added by each additional part. (We disregard here the removal
of parts of a sentence, by zeroing, which does not remove any of its
information—up to ambiguity.) A more organized account of such accretion of
information is given by the semilattice structure of the sentence. Here every
word that enters into the making of the sentence has a fixed point in the
partial order, at which its form and its meaning are contributed. Here, then,
every leg of the partial order contributes its own grammatical (ultimately
predicational) meaning at that location; and here it is found that each type of
meaning has its characteristic location. Typically, these contributions include:
zero-level nouns; first-level operators, as their properties and relations; the
linear order of the first (‘subject’) and second (‘object’) arguments of
a relation; the difference between things (nouns) and states or events
Information 335
(sentences); second-level operators, as properties and relations of sentences.

The further contributions to information are noted in 11.3.2.
Translation, which seeks in a second language the same (paraphrastic)
information as in given sentences of a first one, shows that some features of
form are not universally specific to information. Chief among these are: the
phonemes of words in one language or another; the meaning-ranges of
particular words; even, up to some limit, the size of the vocabulary (the
number of nouns in one language can be smaller than in the other, since
existing nouns can be given modifiers (secondary predicates) to make up for
a semantically lacking noun); morphology (the affixes of one language can
be translated into free-word operators of another); linearization of the
partial order.
In contrast, some features are preserved under translation and hence are
features of information. Thus the operators in the original sentence are
generally translated into operators in the other language. Specifically, for
the most part, the various argument-requiring classes are preserved under
translation, so that N translates into N, On into (9„, Ono into Ono, etc. And
the semilattice—that is, the partial ordering of the words—in the translation
is generally found to be very similar to that of the original sentence.8
11.3.2. Successive stages of information-bearing constraints
It should be clear first of all that every vocabulary or grammar choice made
in constructing a message (sentence or discourse) is an application of a
constraint defined as being available at the given point in the construction.
Each chosen constraint makes a fixed contribution, at the point of its
application, to the substantive information carried in the message. For a
given type of choice, either the contribution is always null (e.g. trans¬
formations, 11.3.3), or it is accretively meaningful. Furthermore, the amount
of choice at a given point may be graded: the more expected a given choice-
decision is at that point, the less information is contributed by the expected
choice. And if what has been chosen can be determined (reconstructed)
from the environing choices, no free choice has been exercised, and no
meaning contributed, at that point.
The meanings of all words and grammatical entities can be indicated via
the constraints listed below, which may be grouped as follows:
(1) words in isolation (with their meanings), viewed as a non¬
syntactic constraint on phonemes;
(2-5) the constraints involved in environment-based specifications of
word meaning;
s Cf. the English and Korean examples in Z. Harris, Mathematical Structures of Language,
Wiley-Interscience, New York, 1968, pp. 110, 112.
336 Interpretation
(6-8) the constraints that underlie the meanings of sentences over and
above the meanings of their words;
(8-10) the constraints that yield the further meanings of sequences of
sentences;
(11) the constraints that yield the meanings of word subclasses and
sentence types in sublanguages.
Loosely, the element of meaning common to all the constraints is that in

each sentence the entities between which an ordered constraint holds are
relevant to each other at that stage in the construction of that sentence. It
will be seen that each additional constraint adds informational capacity to
the language, and that this capacity is constant for all occurrences of the
constraint.
(1) As an initial approximation, we note that words have a range of
meaning when said in isolation or in irregular combinations (such as might
have existed before syntax became established, 12.3.1). These meanings in
isolation cannot be due to regular constraints on word combination. They
can however be described as the meaning associated with individual con¬
straints on phoneme combination, although there is little regularity in this
constraint. Such a constraint would be defined on the resultants of the
phoneme-establishing constraints (7.1, 2).
(2) In limited cases, when a given word occurs in a particular subset of
environments, it has a different meaning than it has elsewhere. That the
difference is due at least in part to the environmental constraint can be seen
from the relation between the meaning-difference of the given word-
occurrences and the meaning of the environing words in these occurrences
(The field was ploughed for the new planting'. Afield is defined as having two
operations). The ability of words to have different meanings under different
operators is a special case of word meanings being characterized by their
selections, (4) below. Such ability would be inexplicable without (4), i.e. if
the meaning of words is assigned only as an inherent property of words in
isolation, as it is to a first approximation in (1). The extreme case is the
idiom, in which a word-sequence (construction) has a meaning not currently
related to the meanings of its component words.
(3) The ability of words to refer to words has been seen (ch. 5) to
create several interrelated informational capacities: the metalanguage, cross-
reference (for word-repetition), and the performative I say for sentences as
said. Each of these capacities involves the constraint of unusual environ¬
ments: for the metalanguage, the zero-level positions of quoted (‘mentioned’)
words; for the performatives, the ability of l say to occur on every sentence;
and for cross-reference the fact that fixed-antecedent cross-reference is
limited to positions where repetition has high likelihood (a constraint).
(4) The fact that different words have different selections, i.e. have
different operators or arguments with which they occur more frequently
Information 337
than average, is a constraint: that constraint is the existence of relatively

stable frequency differences for word co-occurrences.
(5) Various regularities in the frequency yield the meanings of special
kinds of likelihood, such as: the likelihood of a word’s being repeated, which
makes referential pronouns; of having roughly equal likelihood under all
operators, which makes indefinite nouns (something); of a word’s having the
same operator likelihood as its modifying noun has, which makes fragment-
nouns and set-nouns (piece, crowd G204).
(6) The dependence that creates the operator-argument structure has
been seen in 11.2 to bear the meaning of predication.
(7) There is also a common meaning to each argument-requirement set
of operators: On as ‘intransitive’ states or properties of something, Onn
as relations and grammatically ‘transitive’ acts, O0 as aspects of events
(sentences), even Ono as largely purposive (‘human’) relations to situations
or events, and so on. Furthermore, there is a common meaning-difference
between being the subject and object of an operator, which is given by the
linear ordering of the arguments under an operator. This last is a dependence
not between word classes but between word-occurrences. These meanings
are clearly related to the constraint that creates the class or structure.
(8) Certain operators on sentence-pairs impose a constraint favoring that
some word or other in the second sentence under them be the same as some
word in the first. This gives these operators a meaning of ‘connective’ (e.g.
but, because, if), and makes the two sentences under the connective more
relevant to each other (9.1).
(9) Discourse has been seen (9.3) to be characterized by a constraint,
whereby operator-argument combinations (or the equivalent) of word
subclasses special to the discourse repeat in its successive sentences. This
repetition of sentence types expresses a relevance among the sentences, thus
advancing from the recording of solitary observations to the focusing of
various items of information around a topic in such a way as to constitute a
discussion.
(10) There is a further constraint within parts of discourses, governing the
relation of certain successive sentences, so as to create argumentation and,
in the extreme case, proof. This constraint is formulated explicitly over the
main discourses of mathematics and formal logic. In natural language as a
whole it exists only as an inexplicit and often tenuous constraint, appearing
episodically in various sections of various discourses, more particularly in
science. There it may be a constraint in the making. It creates a relation
between the information of the initial sentences of the sequence and its last
sentence: the last is a ‘conclusion’ from the initial ones, i.e. not less true, or
not less plausible, than they are. The fixing of syntactic constraints which
ensure that the last sentence is no less true or plausible than the initial ones is
an addition to the meaning-bearing constraints in language.
338 Interpretation
(11) In natural language as a whole, the constraint that creates syntax, i.e.
sentences, is the operator-argument partition of words or word-occurrences
into argument-requirement classes, while the likelihood constraint on word
combination in (4) above is simply a not-well-defined variation within this
partition. However, in science sublanguages word combination is so sharply
limited that the place of likelihoods is taken by definite constraints which
further partition word-occurrences into subclasses of operators and argu¬
ments. This seems a novelty, changing the non-grammatically-differentiating
likelihoods of (4) into grammatically-differentiating subclassifications. How¬
ever, it is actually no more than an extension of the conventionalization of
use which had produced the original operator-argument partition; cf. (5)
above. In the whole language the use of the dependent words was distinct
and conventionally fixed enough to create distinct word classes, but likeli¬
hoods were not; in science sublanguages the likelihoods, i.e. the use of
specific words, are distinct and conventionally fixed enough to create
combinatorially and semantically distinct subclasses of words within the
whole-language classes. This is the major information-bearing constraint of
sublanguages, especially of the languages of sciences.
The result is a characteristic structure for the semilattices of sentences in
each science language, with a strong connection to the information in the
science, and with the various word subclasses occupying particular positions
within the semilattices—the meta-science material, if present, being at the
top of the lattice of each sentence as secondary sentence. In addition, a
science language may have new argument-requirement classes which do not
appear as such in the whole language, especially in respect to quantity
relations (e.g. ratio). In all, the word classes, sentence structures, and
discourse structures of science languages provide not only the specific
information of the science but also its characteristic overall information
structure, much as the structure of the whole language provides both the
information reported in the language and also—in the operator-argument
system and its secondary grammatical relations—the structure of language
information in general.
11.3.3. Non-informational constraints
Of the various constraints, on sounds and on morphemes, some have no

regular effect on which words combine with which, and hence no effect on
the further structure of language or on its information. Such are the specific
phonemes of a language, and the phonemic composition of each word,
morphophonemics (the change of phonemes when a word is in particular
operator-argument and reductional environments), assimilation and dis¬
similation (the change of sounds when a word is in particular linear phonetic
environments). Such also are the automatic morphemes, such as English -5,
Information 339
-mg, which have been analyzed for the present theory as being merely
indicators of operator status or of argument status, rather than being
reduced words such as enter a sentence by virtue of their argument-
requirement status (G37, 40).
Also non-informational are the reductions, as noted above (3.3). Since
these are changes only in the phonemic shape of words, or in their position,
we can say that the word remains in the sentence even if in altered or zero
shape, and that its grammatical relations are unaffected, so that reduction
when so analyzed makes no change in word combination or in information.
Even if we consider the overt appearance of sentences, where reduction of a
word to zero would be taken as absence of it, the loss of the word would not
entail a loss of information because words that are reduced in this way are
reconstructible from the environing words, and thus had contributed no new
information to their sentence. The non-loss of information is also seen in the
fact that the bulk of transformations, and especially reductions, are optional, so
that both the original and the transformed shapes exist, often in the same
discourse-environments, and the transformed are clearly paraphrastic to the
original. In morphophonemic reductions, too, the paraphrastic character is
clear: I will and I'll, going to and gonna, etc.
Nevertheless, reductions can indicate meanings, without contributing
new word meanings. This happens because reductions indicate high likeli¬
hood of a combination and even, as in the case of the compound form or the
adjective order, the relative stability of a combination (G231, 233). When
operators which have high likelihood of undergoing tense variation get the
tense affixed to them, and so become ‘verbs’ (He aged), while for the other
operators the tense is given in a separate word, leaving them as ‘adjectives'
(He was old), the different localization of the tense creates a fixed distinction
between less and more stative operators (G188).
Some reductions affect or indicate the speaker’s or hearer’s relation to the
information in a sentence, without adding to the information or subtracting
from it. The effect is primarily in the way the information is presented. One
such effect is in the amount of information which comes through at a given
point in the sentence. This results even from some morphophonemics: given
the am, is, are forms of the operator be, the person who hears Am /
responsible? knows already from the first word that the argument will be I.
(But for the continued presence of /, one could have analyzed am as is
plus I.) Another effect on presentation of information is the degeneracy
of certain reductions, whereby different reductions in different sentences
produce identical word-sequences, which are thereupon ambiguous (hom¬
onymous): He drove the senator from Ohio to Washington is produced by
one reduction from He drove to Washington the senator who was from Ohio,
and also by another from His driving of the senator was from Ohio to
Washington. The ambiguity did not exist in the sources, before the reductions.
The alternative linearizations of the partially ordered words of a sentence
340 Interpretation
can bring words of particular grammatical statuses in the sentence to its

beginning, as its ‘topic’, without altering the grammatical relations: e.g. This
l deny as against I deny this. There is also the interrupting of a sentence so as
to insert a secondary sentence as a modifier (relative clause) within the first
(3.4.4).
The capacity for paraphrase in a language is relevant here, because
paraphrase, like translation, is intended to preserve information. We can
therefore ask what features of form are preserved under paraphrase, thus
correlating with information. Paraphrase is of several kinds. One is reduction,
as in I expect him from / expect him to come, where the zeroed word
contributed little or no new information. Another is the transformational
recasting of a sentence, as in John is the author as against The author is John,
or Its wording was mine as against I worded it, where the relation of subject
to verb is seemingly reversed. A third paraphrase capacity is synonymy or
addition of modifiers, as in The plane almost crashed and The aircraft almost
crashed, or in The igloo was hardly visible and The snow house was hardly
visible, where words or phrases with similar definitions and selections are
interchanged.
Reductions can also contribute to certain episodic classification of infor¬
mation, which may not enter into the perception of the users of the
language. One example is when a reduction applies to a particular domain,
e.g. the be here, come, etc., which are zeroable under expect, or in English
the fuzzy domain of the male humans and higher organisms which are
pronouned by he, and the same (female) plus traditional other (pleasurable,
etc.) objects such as ships which are pronouned by she. all in contrast to it.
Such categorization is seen also in the case of information which is conveyed
in specially reduced or patterned forms, such as that covered by affixes in
morphological languages, or in the time-order and plural covered by such
paradigms as conjugations and declensions, or in the fuzzy set of English
auxiliary verbs (can, may), or the set of ‘count’ nouns in English which
require a in the singular (a tree, a man, etc., as against water, or man in the
sense of ‘mankind’, etc.). Such categorization is stronger when there is a
paradigmatic requirement, i.e. when words of a given set (e.g. nouns) must
occur with one or another morpheme of the category (e.g. gender, or the
distinctions made in the Bantu noun classes), or when absence of a form
means not just absence of its meaning but presence of an opposite meaning
(e.g. when absence of the plural suffix means not unstated number but
specifically singular, as in English). Although such structures are very
prominent in use, not all languages have them.
There remain scattered additional similarities and structures which have
some meaning-properties but no explicit information. Some of these are in
grammar (which includes all regularities) but not in syntax (which specifies
word-class combination); such are occasional onomatopoeia (e.g. teeny), or
cases of phonemic and semantic similarities among words (e.g. mother!
Information 341
father!brother!sister, or the si- and fl- words). More important, but still not
explicitly informational, are such literary devices as poetic meter, rhyme,
and alliteration, all of which relate one word to a particular other in the text
(whatever else they do); or allusion to absent words by phonemic similarity
to them in words that are present in the sentence (in literature, jokes, swear¬
words), and various kinds of literary and popular word-play.
11.3.4. The meanings of word relations
In sum: if we now survey the meaning-effects listed in 11.3.2-3 we see

that combinatorial constraints (11.3.2) affect the meaning of the combina¬
tions, chiefly by suggesting relevance between their parts, while the other
relations, largely reductions, rearrangements, and phonetic similarities,
which do not change the words' co-occurrences, do not affect the substantive
information, but only indicate the speaker’s attitude to it (e.g. by the topic
linearization, 3.4.2) or affect the hearer’s receipt of the information. In
particular: in (2) of 11.3.2, the occurrence of a word under a particular
operator selects a relevant meaning for that word.9 In (3), the ability of
words to refer to words comes from extending the selection property of (2);
and it creates new metalinguistic meanings appropriate to the new kinds of
co-occurrence (as in Mary is a word. The word which is Mary has four
letters). In (4), the above-average frequency of each operator is constrained
to particular argument words; and, except for nonce forms, whenever a
word is used with some other operator or argument its meaning is adjusted
to be relevant to that new environment. In (5), similarities and patternings
among the co-occurrence likelihoods of various words correlate with similar¬
ities and patternings in their types of meaning. In (6), the partition in respect
to co-occurrence that creates the operator-argument relation among the
words of a sentence yields the meaning of predication which connects the
meaning of the operator to that of the argument. In (7), the similarities and
differences in respect to argument types make similarities and differences in
the predicational scope. In (8), the likelihood of word-repetition in the two
arguments of an Ooa operator brings out the expectation that the two
argument-sentences be relevant to each other in meaning. In (9), the
likelihood of word-class operator-argument repetition in connected dis¬
course creates the relevance of the sentences to each other. In (10), the
conventionalized requirements of sequence and repetition in the sentences
of an argumentation creates the meaning of ‘leading to a consequence’.
y One might say that the environing words select this relevant meaning from among the
meanings that the word has. However, the only evidence that the word has the given meaning is
that it has it in the given environment or in reference to such an environment. Hence we cannot
say that the word in and of itself has a stock of meanings from which a given environment
chooses, but only that the word is accorded a given meaning in a given word-environment.
342 Interpretation
In all these, the meaning of syntactic ('institutionalized', 12.3.5) word co¬

occurrences is that the occurrence of each word has relevance to the
occurrence of the others.
In contrast, the word relations of 11.3.3 do not add or subtract word co¬
occurrences, as is shown by the reconstructability of the base (unreduced)
form; and they do not yield new information. Some phonemic changes and
reductions in words change the point in the sentence at which the hearer gets
a particular bit of information, or create degeneracies for him when two
different sentences receive the same form. Some permutations call attention
to particular words in the sentence (thus creating a secondary kind of
relevance). Many transformations give a sentence alternative shapes which
may yield stylistic benefits at a cost of complicating the decoding of the
information and the correlation of form with content. In some cases trans¬
formations that are based on broad likelihoods of word-occurrences create
‘grammatical’ categories such as quantity, time, gender, or the verb-adjective
distinction (again creating kinds of relevance). But these are just collectings
of words with phonetic or other similarities into sets having slightly semantic
properties. Finally, some relations hold between a word in the sentence and
words which may not be in the same sentence or may not be present in the
same discourse. This includes partial phonetic similarity among semantically
related words (e.g. father, mother, etc.), or various kinds of allusion.
Relations within the text include alliteration, and meter and rhyme. In all of
these the sounds or the position of one word call attention to some other
word or some other meaning, and so create new semantic relations and
relevances among words.1" All these meaning-effects, and the forms that
engender them, are distinct from the informational ones seen in 11.3.2.
11.4. Form and content
We can now attempt a summary of the relation between form and content in
language. As in so many issues which are clouded by being conducted in
catch-phrases, the resolution here is not one or the other of form and
content or even both form and content together, but a particular relation
between certain forms (produced by certain constraints) and certain content.
Specifically, certain types of content-bearing elements enter into certain
types of meaning-affecting constraints on their combinings, and vice versa.
At this level of generality, the situation is not too different from the relation
between form artd expression in music and other arts.
The relation between form and content cannot be a correlation in the
ordinary sense of the word—a covarying of members in two independently
Differently, contrastive or emphatic stress does not relate to any particular other word. It
functions as an additional word in the sentence, a synonym of emphatically or of rather than
anything else, or the like.
Information 343
known sets—because content as we know it from language, i.e. as a system

of information, is not known without reference to the form of language.
Some meaning and information—even some that are expressed in speech—
exist without grammatical language. But there are certain kinds of meaning
and information that exist only in words and sentences. It has been seen
above that much of this meaning is allied to particular features of form. This
alliance is seen even in the case of the different kinds of word meanings:
words having the same argument-requirements generally have some aspects
of meaning in commmon. Zero-level words (N) largely refer to concrete
objects; operators with one-place arguments (O,,) are intransitive; and so
on. Nevertheless, there are major cases of different forms for the same
meaning: synonyms (and definitional phrases replacing a word), trans¬
formational paraphrase, morphophonemics, and for that matter the different
phonemic forms of words which have approximately the same meaning
in different languages. And there are major cases of the same form for
different meanings: homonyms, transformationally ambiguous sentences,
different meanings of a word under different operators, and the idiomatic
versus literal meanings of idioms.
More generally, the universes of language and of meaning are not quite
identical. On the one hand it is not the case that, in any particular language,
everything can be said, or said within reasonable limits or directness
(10.3.4). Not only are there extra-grammatical meanings in speaking, such
as those of gesture, glances, and intonations of emotion, but also there are
various language-type meanings which are expressed in no language, or not
in one language or another. For example, if something would involve
complex reference to various parts of a sentence, requiring the addressability of
arbitrary phonemes and morphemes in it, it cannot be said; for even though
such complete addressability is definable within the reference capabilities
(sameness-statements and relative location) of language, it is not provided
by the pronoun systems of actual languages.
Individual languages may also have particular types of meaning that can
be said only by circumlocution, either for lack of vocabulary or because of
grammatical restrictions. There are also form-content mismatches, as in He
caught the glass before it hit the floor (or He died before he got the chance to
see his grandchild) as against the less common He caught the glass before its
hitting the floor. The former seems to include the false assertion that the glass
hit the floor or that the man got the chance to see his grandchild. We may
explain away such a form by appropriate semantic interpretation of tense or
of assertion; but such adjustment may be ad hoc, due precisely to such
‘wrong’ forms.
An indirect type of form-content correspondence is the following: by the
present theory, grammar has to describe not merely the not-well-defined set
of actually said sentences, i.e. the popularly acceptable sentences, but the
set of all possible sentences within a stated structure. This condition on
344 Interpretation
grammar is not just an artifice for convenience of formulation. It has

linguistic support, and it has a specific semantic effect: it accounts for a
certain part—the imaginative and nonsensical part—of the content of
language. For recognizing the ability of any zero-level word (N) to occur as
argument of any first-level operator, and the like, makes it possible to create
word combinations that are beyond any experienced meaning. This is found
not only in some schizophrenic talk, but also in imagination, in fairy-tales
and science fiction, in literary effects, nonsense verse, and word-play—and
in various irresponsible discussions.11 Indeed, one can make a sentence
without quite knowing what its meaning is.
Despite all these reservations, the great bulk of information that is given
in language is associated with specifically relevant features of vocabulary
and syntactic form. Even something which is considered a semantic datum
about sentences, namely how a sentence is characterized by its consequences,
can be largely stated in syntactic and discourse-structure terms. For this last,
we recall that the capacity of a discourse to lead to a logically acceptable
conclusion from given initial sentences is obtained by particular proof-like
constraints on sentence-sequences in the discourse (10.5.4). We recall also
that the meaning of a word can be syntactically approximated by the set
of its more frequent than average environments. We can then say that the
meaning of a sentence is syntactically approximated not only internally by its
construction, but also externally by the set of its successor sentences under
the proof-like constraint, i.e. by the set of its consequences.12
The association of form with meaning can be studied in detail if we start
with sentences of any size (initially, zero) and ask what partially ordered
accretions of words would enable them to carry specified accretions of
information.
In sum: when the methods of analysis in Chapters 7 and 8 are applied to
arbitrary samples of natural language we obtain, for reasons discussed in
11.5, a structure that conforms to our general experience of the world (11.6).
When the same methods are applied to samples of a relatively closed subject
matter, we obtain a structure that reflects the objects, relations, and events
of that subject matter (ch. 10). When they are applied to a single discourse
by itself, we obtain a structure which in many cases shows what the discourse
had set out to do (9.3). In each case, the structure reveals something about
the burden of its universe of data.
11 In general, situations which call not for information but for some other kind of expression
may occasion sentences whose overt information is not what is really being communicated. For
example, if people have encountered each other several times in one day, they may say ‘It's a
small world or simply, with a smile, ‘We meet again’, where what is being communicated is not
the facts that are said, but the recognition of the unlikelihood of so many encounters—the
content of the smile.
12 H. Hiz, ‘On the Rules of Consequence for a Natural Language’, The Monist, 57 (1973),
312—27, and Information Semantics and Antinomies', in G. Dorn and P. Weingartner, eds.,
Foundations of Logic and Linguistics, Plenum, New York, 1985.
Information 345
More particularly: the numbered form-meaning features of sentences in

11.3.2, 4 are each a constraint which acts on the resultants of prior constraints,
and which brings a further informational capacity not available in its
absence. Since such features and those noted in 11.3.3 cover the structural
devices of the language, and since each one adds its own kind of informa¬
tional capacity to the language, we see that indeed the structure of the
language is made out of an accretion of specific constraints which produce a
stepwise accretion of informational capacity and properties.
This accretional analysis of structure with its informational capacity
underlies the result noted above that the meaning of the sentence as an
ordered set of constraints (including the word-choice as a constraint) is that
same ordering of the meaning of the constraints (including the chosen
words) in the sentence (9.1).1' Such a statement is the more understandable
if we see that information is not just a correlation or interpretation of the
constraints, but is an intrinsic effect of them. This view of the step by step
correspondence between form and content in language (where both form
and content are more explicitly analyzable than, say, in many arts) has
various applications. For one, it is possible to make critiques of the language
of a science to see if its vocabulary is more meaningful or teleological than
the specific entities and relations or events of the science require. For
another, it is possible to see why in the case of an art, the introduction
of a fruitful new ‘language’ or apparatus may be not less important than
the production of new works. In a time-slice, the step by step correspondence
shows form effecting (‘informing’) content, but developmentally we have
form arising not only from its own dynamics but also from the conventional¬
ization of its largely content-oriented use (12.3.5).
The situation of form effecting content and of content as a factor in
favoring the development of form has some similarities—for all the essential
differences—with the relation of morphology to physiological function and
to behavior in biological systems and their evolution.14
11.5. A theory of language or a theory of information?
Given all this, we can consider the question: is the theory presented here a
theory of language, or of information? It is in any case a theory of language
structure, since it attempts a reasonably complete survey of the constraints
in language, some of which are essential to it (and must be defined as
independently of each other as possible) because of the lack of any external
metalanguage in which the structure of language could be stated. And it is a
13 Even idioms are included here, their constraint being a special case of a word having
different meanings under different operators.
14 It also has an echo in the use of the term ‘semantics’ as ‘what a system can do' in analyzing
computer programming languages.
346 Interpretation
theory in which meaning was not used in making the analysis. It also
produces a structure of language-borne information because a distinguished
subset of these constraints—ultimately the operator-argument dependence
and the word likelihood within it—are characteristically associated with an
informational interpretation; they each have the effect of information.
These constraints thus create the essential nucleus of language, and also
create its information. The relevant question is whether in choosing the
methods that yielded this theory, other methods might have been available
that would not have yielded this close informational interpretation. But the
search for a least grammar was justified in 2.2 without any consideration of
meaning. That the constraints on equiprobability which yield the least
grammar carry informational interpretation is due to the nature of information:
both word combination and information are characterizable by departures
from randomness. Features which do not give new information, such as the
linearity of language, or the time and quantity categories that are created by
paradigms, are distinguishable structurally from the informational ones.
There follows from this, what is in any case borne out by experience, that
we cannot in general impose our own categories of information upon
language. In a limited way we can arrive at certain categories of information
independently of language, but we cannot locate these types of information
in any regular way in a language unless the language has a structural feature
whose interpretation this information is. We cannot determine in an a priori
way the ‘logical form’ of all sentences. We cannot determine their truth
except to the degree that we can locate them in a proof-like constraint
on sentence-sequences, in respect to known prior sentences or later con¬
sequences. We certainly cannot map them in any regular and non-subjective
way into any informational framework independently and arbitrarily chosen
by us.
The other side of the coin is that we cannot disregard any information that
is carried by a structural feature of a particular language. Various languages
express time-order and quantity with various distinctions and relations. We
need not go only by the most overt expression in the vocabulary but can also
seek second-order evidence, in more subtle restrictions on the environments
of this vocabulary, to see what time relations are expressed in the language.
But we are bound to the regularities of word combination in the given
language. For example, some languages have overt forms (suffixes and the
like) for the ‘aspect’ of verbs (mostly features of duration), while others
mark this only by regular differences in word combination, especially in
respect to time words, as against yet other languages which have no regular
indication of aspect (G274). An informational category of ‘aspect’ either
must be or must not be recognized for the language, accordingly. And
the non-informational constraints (11.3.3) too remain important for each
language.
If in spite of the informational dependence upon the demonstrable
Information 347
structure of the particular language, excellent approximate translation is so

freely possible among all languages, it is because the basic partial-ordering
relation is identical in all languages, and the meaning (equivalently, likeli¬
hood) ranges of words, or words with suitable modifiers, are sufficiently
close to permit the translation.
To sum up: both language and information are necessarily structured by
departures from equiprobability acting upon departures from equiprobability
(which is why languages are informationally intertranslatable), although
many kinds of information are not expressible in language, and some items
of language structure do not express information. In addition, language
universally develops informational facilities that do not arise directly from
dealing with the real world. One such is the metalanguage; another is the
capacity for saying nonsense. The similarities and differences between form
and content surveyed in 11.4 show in what ways the theory of language
structure can be or cannot be a theory of information (or of the language-
borne subset of information).
11.6. Whence the power to ‘mean’?
What gives language the power to mean something? We have seen that the
meaning of a sentence is not a primitive unanalyzable entity, but rather
the product of accretions from the fixed meanings of its words and the
constraints on their combination. We therefore have to ask what gives the
words and the constraints the property of having these accretive meanings.
The usual view is that words are associated (as symbols) with some selectively
perceived aspects of experience (as referents); a less selective representation,
such as a photograph, would not be considered a symbol of what it represents.
As to the meanings of grammatical relations, they have been little studied
and are presumed to be a new primitive element created by language.
In what follows we will see: first, that the dependence relation of words
reflects what one might consider dependences within man’s perceivable
world (11.6.1); and second, that word meanings and co-occurrence selections
express a categorizing of perceptions of the world (11.6.2).
11.6.1. Co-occurrence and the perceived world
Rather than begin the discussion from word meanings in language as a

whole, we start from the analysis of word combinations in science languages.
Here, especially in sub-sciences, we see something new: words classified
together by their having similar word-choice environments, especially in
348 Interpretation
their immediate operators or arguments, sometimes despite differences in

their meanings in the science as a whole (10.3-4). When we consider these
word classes, we see that they reflect the actual objects and relations of the
science. If antigen names, antibody names, and cell names constitute
different classes in immunology language, so are antigens, antibodies, and
cells different classes of objects in the immunology laboratory. If there are
sentence types in which antibody names are co-arguments of cell names
under a specific class of operators (contain, produce, secrete), so are there in
the laboratory such specific relations and events which hold between anti¬
bodies and cells (Cells contain antibodies). To a large extent, a given
sublanguage class of zero-level arguments (N) has as referent a class of
objects which can indeed be put into one class in that they have much the
same status in respect to the activities and observations in the immunology
laboratory; and the sublanguage operators on, say, two N classes indeed
refer to the relations between the objects named in one class and the objects
named in the other class, in the real world of the laboratory. Thus each class
has a class meaning (e.g. ‘cell’, of whatever kind, or ‘contain’) which is
characterized by its total set of sublanguage operators or arguments, and
also by its laboratory referents. As to each word within a class, it has a
meaning identified by all the laboratory experience with the relevant object
or event—all of which can be expressed by the various operators and
arguments co-occurring with that word in the science language.
What we see here is not only a relation between the meaning of a word and
its co-occurrences with other words but also something of a correspondence
between the co-occurrences of words and the perceived co-occurrences of
objects and events about which those words speak. This correspondence
is sufficiently consistent, so that the types of word co-occurrence can
be considered to correspond to types of co-occurrence among objects,
properties, relations and events in the world described. In the real world,
properties or events ‘co-occur’ with the objects which participate in them.15
This is reflected in the basic co-occurrence types of language: the dependence
of an operator on its arguments.16 The co-occurrence or connection (in an
appropriate ‘action space’, not necessarily in time or locale) of the objects
15 When we recognize (above) the meaning common to a class of words as being related to
the meanings of the word-environment which is common to those words, we can also recognize
that the different meanings that a word has in different environments are related to the different
meanings of those environments. Thus (a) word meanings that depend on word-environment
are not different in kind from (b) word meanings that depend on association with the real world
(11.1 and n. 9 above); in case (a) the co-occurrence of a word and its environment reflects the
real-world co-occurrence of the here-selected referent of the word with the referent of its
environing words.
16 Given that cat is a zero-level word and smile a first-level word (operator), one can see a
Cheshire cat with no smile or other distinguishable expression, but whether one can see a smile
without a Cheshire cat or some other bearer is questionable, Lewis Carroll’s claim to the
contrary notwithstanding. Similarly, when an abstract noun is taken as a nominalization of
sentences having indefinite arguments (11.2, 11.6.2), it presupposes the existence of such
sentences, even if their arguments are unknown or fictitious.
Information 349
participating in a relation is reflected in the linear ordering of the arguments

(co-arguments) under an operator. The connections between properties,
relations, and events between which causality or any other relation holds in
the real world are reflected in the second-level operators whose arguments
are operators, hence sentences.
The correspondence can even be seen in more global relations. For
example, just as a prior science must exist before one can base oneself on it
in another science, so the sentences of a prior science appear as already-
present zero-level arguments under the operators of the other science, i.e. as
descriptively earlier in the partial order of the sentences in the other science
(10.3). And the grammatical fact that the meta-science material, which
describes the scientists’ actions and thoughts, appears as the top operator
over the separately existing science sentences conforms to the real-world
fact that the scientist is an external factor in respect to the separately
existing—again generally prior—events of the science.
The correspondence, specifically in respect to co-occurrence, between
language and the world it describes is not surprising, since each of these can
be described within the same universe—that of departures from random¬
ness.17 Each co-occurrence, or joint involvement of entities in a relation,
can be viewed as a departure from randomness. On the one hand, each
contribution to the meaning-bearing structure of language has been seen to
consist in a constraint on combination of phonemes and of words. On the
other hand, each object, property, relation, and event in the world can be
considered an unstated constraint or departure from randomness in the
physical world. It is not surprising that some of the constraints which
constitute the perceived world should be reflected in constraints on the
word-occurrences with which we speak about it. It is this correspondence
that makes information about the world carryable by a sentence, and in a
manner that is objective rather than subjective and private. The whole
correspondence is reminiscent of the relation between art and experience,
that colors our response to works of art.
The explanation above of the language-world correspondence in science
languages shows that the correspondence is based on properties of language
in general rather than on any special properties of science sublanguages.
And indeed, more loosely and less demonstrably, the correspondence holds
for the basic types of co-occurrence in language in general. The fact that
these basic types of constraint are rooted in properties of the physical world,
which are experienced in common by all humans, explains why the partial
ordering (and to a lesser extent some of its secondary constraints) is
apparently universal for language, and while different from one language
culture to another is nevertheless far more similar among languages than
the choice of what phoneme-sequences make words and carry particular
meanings.
17 The necessary caveat is given in 11.6.2.
350 Interpretation
Most generally: the mathematical object created by the dependence-on-

dependence relation can be considered to reflect—to ‘mean’—the co¬
occurrence relations of objects, situations, properties, in the perceived
world. Beyond this, we find various types of co-occurrences of word-
dependence: repetition of a word in sentences under certain conjunctions,
repetition of grammatical relations between particular word subclasses in
discourse and sublanguages, constraints on sentence-sequence in proof.
These can be looked upon as reflecting subtle dependences among the
co-occurrences in the real world, or else as building something new—
‘information’—about the real world. Information is thus not an entity, but a
relation of symbol to the world: the recording and communicating of the
occurrence of perceived objects and events. And the structure of language,
and of information, reflects not the structure of the world but only this much
about it. It should be noted that if we want to build an alternative or a further
representation of the real world, which will ‘mean’ something further about
the world, we cannot do it by simply defining new word classes, but must
define new grammatical relations such as third-level operators (that depend
on operators that depend on operators) to house more complex concepts
such as ratio or co-vary.
What we have here is not merely that meanings can be expressed by
grammar, i.e. by arrangements of words as well as by words themselves, but
rather that when we get down to the elementary constraints we see that each
constraint has a fixed meaning (8.2.2) as each lexical element has a fixed
meaning, although a different kind of meaning: the words reflect (or image)
objects and situations in the world, while the constraints reflect the co¬
occurrence of these in the world.
Finally, language and information, because of their limited relation to the
world, make possible certain structures about the world that could not be
made directly with the vastly more complex and never fully known actual
events of the world. Such informational structures are: generalizations,
imaginative extrapolation, questioning, discourse, and controlled argument¬
ation. Somewhat similarly, mathematics and logic, because of their yet more
limited relation to the real world, make possible the controlled structure of
proof, which is true about the world but could hardly be obtained directly
from observable relations of events.
11.6.2. Language and the perceived world
Given what has been seen in 11.6.1 of the correspondence of language

constraints with constraints in the perceived world, we now ask what is the
case for the whole vocabulary and structure of language. As has been noted,
words are associated first of all with objects, properties, and events of the
perceived world. We may call these the referents of the words, but this does
Information 351
not have to be restricted to their extension in the logical sense. Unicorn is

associated with certain more or less specific perceptions of a horned horse,
and co-occurs with a particular selection of operators, even if its extension in
the real world is null. Most language-expressed meanings which do not
consist in directly perceived aspects of the world are found to be grammatical
derivations of words with more concrete meanings. Simple cases of this are
the pronouns, and metaphoric meanings. More complex cases are: the
meaning common to a class of words that have similar word-environments
(11.6.1 and n. 15 above); the fact that most ‘abstract’ nouns are nominal-
izations of operators whose arguments have been zeroed as being indefinites
(i.e. a disjunction over all members of the argument class, G225)—thus
coercion is derived from anyone's coercing someone’, and even the meaning
'not-less-true-than' which is produced by the sentence-sequence constraints
of proof structure.
However, even for the many words which seem directly associated with
objects and events in the world, the association is a complicated relation.
For a word to mean something in the world involves isolating and choosing a
particular (not always delimitable) part of the world or a (not always
specifiable) property in the world. It further involves the de facto classifying
of various such parts or properties as are referred to by the same word, since
words do not have unique referents (9.1). Such distinctions and classifications
may not have counterparts in the real world, and may be fuzzy within
one language and may differ as among languages.Is They arise because
languages—their speakers—have no other way of usefully associating symbols
(the words and their combinings) with the perceived world. To this we must
add the compositional properties of words: that they and their components,
the phonemes, are discrete, distinguishably repeatable, finite in number at
any one time and finite in length, and pre-set within a language community.
All these create a stable, public, and transmissible structure for the words.
In this way Procrustean symbols are formed onto which aspects of the world
are projected, although the word-symbols can then be extended into less
symbolizable (e.g. abstract, relational, or private) uses.
The meaning of a word is known not only as an association with aspects of
an experience, but simultaneously as a particular selection of operators on it
or arguments under it with which it occurs in higher than average frequency.
If a word has several distinct meanings, it has them in corresponding distinct
(‘coherent’) subsets of selection. To use a word in a different meaning is to
use it in a different selection, so that to establish a new meaning for a word is
not merely to announce the new meaning but to get a new selectional subset
for the word. In accord with an observation of Henry Hoenigswald, a new
definition for a word is in itself a new operator-argument for it, and it entails
1X There are nevertheless many similarities in the meaning ranges of words, as among
unrelated languages. Many unrelated cultures have quite similar species distinctions in their
words for flora and fauna.
352 Interpretation
the occurrence of the word in other new environments; which are coherent
with those in the new definition. It follows that not only word meanings, but
also word likelihoods in various combinations, are associated with aspects of
the perceived world, except for those word combinations that are fostered
by social and institutional factors (fashion, fiat, etc.) or by intra-linguistic
factors such as analogy; or communicational efficiency. The relation to
people's experience of the world holds also for a word's extra-high likelihood in
particular combinations, that permits reduction of the word.
That meanings of a word are for the most part determinable both by
association with the real world and also by the ‘selection' of the word
becomes explicable if we say that it is the sentence, or word combination,
that has meaning by association with experience; and the parts of the
sentence and also their arrangement have meaning by association with
the perceived parts of that experience and with the way these parts co-occur
in it.
On the basic and universal language structure thus formed, with its
component meanings thus rooted in the perceived world, there appear
various further structures, not all of them universal and not all related in a
direct way to properties of the world's phenomena.
Some of these are local developments based on non-informational regular¬
ities such as phonemic similarities among words: e.g. morphophonemics, or
the differing declensions and conjugations in some languages. Here we must
include the use of a single word for various meanings, differently in different
languages. Others are conveniences in presenting information. This includes
reductions and other grammatical transformations, which clearly do not
reflect anything in the referents of the words and sentences. It also includes
grammatical developments that bring distinguished kinds of information
into common form: e.g. the cross-reference pronouns (5.3.5), or the para¬
digmatic arrangements of tense, person, and number. These last categories,
for example, are certainly involved in the physical world, but not in the same
relation to objects and events in the world as they have to nouns and verbs in
the paradigms of many languages. A slightly different case is that of
transformations that are based on word likelihoods but that create new
grammatical distinctions and partitions out of graded differences in experience-
reflecting likelihoods: e.g. the partitioning of operators into verbs and
adjectives (as in English) by establishing two distances for the tense on an
operator, possibly on the basis of how frequent is tense-change within a
discourse for the different operators (5.5.3, G265).
Still other further structures consist of new constraints on co-occurrences,
expressing new information, but not based on any immediately recognizable
constraints in the physical world. These—all of them constraints on sentence-
sequence—are the word-repetition around conjunctions (9.1), the operator-
argument class-repetition discourses (9.3), and the proof-structure sequences
in mathematics and some science writing (10.5.4). One might say that these
Information 353
constraints do not reflect or record perceptions of the world, but rather

make possible our useful manipulation of that information. But one might
also wish to argue that this usefulness in language constraints reflects
corresponding regularities among events in the real world.
Finally, there are further structures and capacities which arise not from
representing the world or from manipulating that representation, but from
the developmental possibilities of the representational structure. Such is the
constraining of the dependence relation into a dependence-on-dependence
and into a mathematical object (3.1.2). This dependence-on-dependence
has the further effect of divorcing the predicational constraint (the partial
order of 3.1) from the likelihood constraint (the selection of 3.2), with the
result that the real-world predicational relation can act on non-real-world
(non-selectional) pairings of words. This admits as-though-experiential
meanings: imaginative, jocular, nonsensical, and other novel statements,
‘sensibly" grammatical yet freed from the need to reflect experience. Another
systemic factor going beyond real-world experience is the development of
metalinguistic capacities, including the metalanguage and the performative
/ say. And the development of metalinguistic and systemic concepts such as
the ‘relevant to each other’ above.
Given all this, it is clear that the correspondence of language with the
world—in the co-occurrence of 11.6.1 differently from in the association of
11.6.2—is a frame for thinking about meaning rather than an explicit
relation between language and the world. For one thing, the departures
from randomness in the case of language are stated constraints acting in a
stated hierarchy; nothing so specific can be said for the world referred to.
For another, many words refer to particular selected aspects or classifications of
objects and their relations, rather than to referents which would necessarily
be distinguished or selected in the absence of language—this even though
many languages make much the same referent-choices in the world, as can
be seen in the structural simplicity of much of their intertranslatability.
Finally, an untold amount of the world is not referfed to in any language—
not only because it is not public, or not discernible at the given time, or not
of interest to record and communicate, but also because it would be more
than human beings—finite and of a particular size—can note or respond to.
As a result, we can not relate the structure of the referred-to world to the
structure of the referring language; this was already noted in 2.4, where the
comparison was with meaning rather than with the world referred to.
Nevertheless, we know that whatever structure the world has is a matter of
departures from randomness. It is only in this vague sense that we can
suggest that when an occurrence of language refers to a situation in the
world, the departures from equiprobability which constitute that occurrence
might be said to have as their referent something of whatever departures
from randomness constituted the situation in the world. The correspondence
mentioned here differs therefore from any correspondence theory within
354 Interpretation
philosophic realism. Indeed, rather than a general correspondence with

the world, what we have in language is, first, a correspondence between
the proper property-object co-occurrence in perceived nature and the
operator-argument non-zero probability in language; and, second, a cor¬
respondence between the relative likelihoods of property-object associations
and operator-argument co-occurrences. Thus rather than reflecting the
structure of the world, the co-occurrence structure that creates language
creates the concept of 'information' about the world.
We have seen that in the absence of any independent knowledge of what
information is, the information of language can be considered to be the
meanings of words and of the constraints on their co-occurrence. We have
also seen that precisely these parts of language are related to aspects of the
perceived world, as symbols (in part) in the case of words, and as cor¬
respondence (in part) in the case of the constraints. In addition to this, there
is information which cannot be expressed directly in language or in its
grammar; and there are structures in language which add no information.
More importantly, the known structure of the world is far beyond what is
directly expressed in the structure of language. Also, many structures in
language are only indirect expressions of aspects of the world: for example,
the modern French perfect has taken over meanings and selection of the past
tense while retaining the structure of a present tense, while the English
perfect has done so only in special environments.
Hence if one says that form follows upon content in language and also
content follows upon form (11.4), this is by virtue of two meanings of
‘content’. In ‘form follows upon content’, the word-choices and co-occurrence
constraints are based, in two different ways, upon the objects, properties,
and events of the perceived world and upon their co-occurrences. In
‘content follows upon form’, the information or meaning of a sentence is
given by the grammatical relations of its word-choices. The statable informa¬
tion about the world is thus mediated through the isolating and classifying
properties of words as symbols, and the devices of co-occurrence constraints
in syntax.
The commonplace that while mathematics is a discovery (of something
about the world), language is an invention or convention of man is thus not
quite so. Language is indeed a convention, but one that associates sound-
combinations with perceived aspects—objects and their relations—of the
world, and is therefore constrained by what constitutes successful per¬
ceptions in dealing with the world—thus a discovery.
11.7. The structure of information
We begin with information as an undefined term, with no known structure.

In language, we distinguish certain accretional steps in constructing a
Information 355
sentence and discourse, ultimately the partial ordering of words and con¬
straints thereon, with words themselves characterizable by their status and
likelihoods in the partial ordering. These steps are found to yield in a fixed
way a distinguished part—the lion’s share—of the information carried in
the given sentence and discourse. Thus the information itself is structured
by such accretional steps. Indeed it has been seen that given a system
or message with certain constraints and certain information, the defining
of additional constraints in it makes possible corresponding additional
information (11.3.2, 12.3.3).
First, the choice of a particular word at a given point in constructing a
sentence is a departure from equiprobability of phonemes and words there,
and yields a meaning.
Then, the relation of departures from equiprobability to informational
contribution is seen throughout the grammar. For example, words lessen
their information as they broaden their selection (i.e. as choosing them
constitutes less of a constraint on their argument or operator, 11.1), e.g. in
the process of provided, providing becoming conjunctions (10.3, and n. 2
there). Words with broadest selection add least information, as in the case of
the indefinites (3.2). And words which are reconstructible from the environ¬
ment, i.e. add no information, are zeroable (3.3.2). In addition to word-
choice, each constraint that creates the partial order of words in the sentence
is also a departure from randomness in this language universe, and yields a
meaning. The information in a sentence or a discourse is thus formed by
departures from equiprobability acting upon departures from equiprobability.
The information is not simply an unordered collection of departures from
equiprobability, but a specific structure consisting of one step acting on
another.19
In the mathematical theory of communication (Information Theory), the
total departures from equiprobability (the redundancy) of a message or a
channel is discussed as its informational capacity, indicating the amount of
information it can carry.2" The analysis of amount does not identify or
19 The analysis of language information as successive steps of constraint makes possible

comparison with the production of meaning-controlling sequences of steps for computability, in
Post Productions (E. L. Post, ‘Finite Combinatory Processes: Formulation F, Journal of
Symbolic Logic 1 (1936) 103-5), and in the Turing Machine (A. M. Turing, ‘On Computable
Numbers, with an Application to the Entscheidungsproblem', Proceedings of the London
Mathematical Society2 42 (1936-7) 230-65, 43 (1937) 544-6; R. Herken, ed., The Universal
Turing Machine: A Half-Century Survey, Oxford Univ. Press, Oxford, 1988). However,
computability cannot be fruitfully compared with externally imposed arbitrary semantic repre¬
sentations of language, but only with the structural steps that produce the meanings of a
sentence. As was seen in the comparison of natural language and mathematics (10.6), the
essential difference between the two is in the variety of likelihoods within a set of symbols in
respect to the next step in the constructions.
c. E. Shannon and W. Weaver, The Mathematical Theory of Communication, Univ. of
Illinois Press, Urbana, 1949; A. I. Khinchin, Mathematical Foundations of Information Theory,
Dover, New York, 1957; A. M. Fano, Transmission of Information, MIT/Wiley, New York,
356 Interpretation
distinguish the contributions to the total information, or the structure of

their combination if any. It is clear that a random set can carry or express no
information except via an external code-book. A set organized by one
repeated system of departures from randomness, e.g. a pure crystal, can
carry only very limited information. This can also be seen in precisely
repeated simple decoration, e.g. a geometrical frieze, as distinguished from
art (where the issue is paucity of expression rather than paucity of information).
When we decompose the form, and with it the meaning, of sentences and
discourses into the accretional departures from equiprobability, we are
exhibiting that same structure for the information therein. In this way we are
able to speak of the structure of information, at least of the kind of
information expressed by language. We see that information is not produced,
except in a trivial way, by a single departure from equiprobability or
an unstructured aggregate of such departures. Rather, in a system which
embodies or produces, and not merely carries, information, the total
amount of information is obtained out of a specifiable structure of certain
departures from equiprobability acting upon other departures from equiprob¬
ability.
1969; A. Kolmogorov, ‘Logical Basis for Information Theory and Probability Theory’, IEEE
Transactions on Information Theory, IT-14 (1968) 662-4; C. J. Chaitin, Algorithmic Informa¬
tion Theory, Cambridge Univ. Press, Cambridge, 1987.
12
The Nature and Development

of Language
12.0. Introduction
The syntactic theory presented here enables us to arrive at some under¬

standing of the nature of language. To be able to draw conclusions from the
structural properties of language, it is necessary to have an organized and
reasonably complete view of the whole structure of a language; otherwise
the discussion can be supported only by episodic examples. It will be seen
that certain properties are apparently common to the known languages,
others belong to speech as distinguished from language structure, and still
others are limited to the grammar of particular languages or language
families. There are also certain kinds of phenomena which do not combine
in any regular way with the main body of material in a language, and which
therefore have to be left out of any precise language structure: for example,
the voice of each speaker, coughs of hesitation, and intonations of anger or
surprise.
12.1. Overview of language structure
As for the phenomena which show regularities, they are describable, for
each language, by the following relations and processes:
(1) Words. Sequences of phonemes (the latter defined by phonemic
distinctions, 7.1), bearing (in the great bulk of cases) meanings which are
associated with aspects of the perceived world, and are characterized by
their likelihood of combining with particular other words in the dependence
relation. The elements of communicational distinction are phonemes, but
the elements of meaning-carrying are words and their relations. The words,
however, are not simply symbols assigned by their users. Their ability to
have a particular meaning (i.e. to be such a symbol) is identical with
their having particular clustered (coherent) selections of their co-occurring
operators or arguments, (3) below.
358 Interpretation
(2) Dependence. A relation of ‘dependence on dependence' (3.1.2),

which partitions the set of word occurrences in utterances into dependence
(argument-requirement) classes of operators and their arguments. A sen¬
tence, or minimal discourse, can he defined as the locus of the satisfactions
of the dependence relation: when every word in a sequence has its argument
satisfied therein, the sequence is a sentence. At each stage of sentence¬
making, the words between which a direct dependence holds are next to
each other, except for the effect of prior sentence-making steps, and of later
transformations. All other grammatical relations are either paraphrastic
rearrangements of this dependence, or else further constraints on the
resultants of this dependence.
As was seen in Chapter 3, the dependence relation can best be recognized
as a constraint, which states what is excluded from the language. In a less
essential way the other basic constructions, (1) and (3)-(4) here, can also be
recognized as constraints.
(3) Likelihood. The dependence-on-dependence says nothing about how
likely the occurring of a particular word in an operator class is in respect to
particular words of its argument class, and how constant such likelihood is.
An additional constraint in language provides that each word indeed has a
particular, though fuzzy, likelihood (even down to nearly zero) of appearing
as operator on each individual word of its argument class, and that this
likelihood is roughly constant within particular communities and periods.
In the syntactic theory, words have no intrinsic properties other than this
dependence status, and, within that, their likelihood of particular com¬
binations. To these two constraints in sentence-making, the meaning of the
sentence is closely related. A word’s phonemic composition identifies it, but
not intrinsically. This is because other words (with other combinatorial
likelihoods) may homonymously have the same phonemic composition (as
in hart vs. heart)-, and because different phoneme-sequences may have the
same meanings and the same combinatorial likelihoods (as in the two
pronunciations of economics), or the same likelihoods with a particular
complementary difference (as in / am, he is).
(4) Reduction. In stable operator-argument environments, certain words
which have highest likelihood of occurring there may be reduced in their
phonemic composition, even down to zero; also, the distance between
words that are grammatically relevant to each other may be reduced (as in
modifier position and in affixation). The situations in question include both
being the most likely in respect to the particular neighbors in the given
sentence, and being normally likely in respect to exceptionally many possible
neighbors in any sentence.
(5) Miscellaneous. A few independent processes are found in particular
situations: pre-syntactic morphology, other than that which can be considered
to be produced by reduction (12.3.1); asyntactic sentence-making as in Fire!
The Nature and Development of Language 359
or Hey you/; constructions of one sentence on the analogy of another, as in

We had a quick cup of coffee, built like We had a quick hath (from We had a
bit of quick bathing from We bathed quickly, G377); sentence structures
which are paraphrastic (‘transforms’) to another structure without being
reductions of it (8.4, 11.3.3); and occasional foreign borrowings, euphemistic
word-replacements, or the preservation of now non-grammatical idioms or
proverbs (e.g. The more the merrier).
These last constructions differ from the preceding processes above in that
these last may have no recognizable relation to an operator-argument
source: we may derive have a quick bath from bathe quickly, but have no
such source for have a quick cup of coffee. This lack of a now-recognized
relation may hold even for forms which have historically been derived from
operator-argument sources, as in the case of goodbye from God be with
you.
Conversely, sentences which arose historically by analogy may be inter¬
pretable as though directly reduced from operator-argument sources. This
is evidenced by the possibility of finding for most of these ‘miscellaneous’
exceptional sentences an as-though source formed in accordance with
operator-argument relations, from which the exceptional form could have
been derived by suitable reductions of known types (e.g. have a quick coffee
from an artificial have a quick coffee-drinking from drink coffee quickly). In
such cases analogy or pre-syntactic morphology, or the like, is the process
that created the sentence in question (mostly on the basis of some existing
operator-argument sentences), but the operator-argument structure is the
framework in respect to which the word-sequence in question is understood
as a sentence. Only a reasonably small number of word-sequences are used
without being referable to the operator-argument framework. Other than
these last, the ‘miscellaneous’ cases which are indeed referable to operator-
argument examples are systemically secondary to the operator-argument
process which has created the framework for them.
There is another respect in which the operator-argument process is
central to sentence-making. This is that the sentences of the grammar, i.e.
the metalanguage, which constitute a subset of the*sentences of the language
itself, cannot themselves be formed without the operator- argument process
of (1 )-(3) above; but they can be formed without any of the other sentence¬
making processes (5.2). Thus all the structure needed to form every sentence of
the language can be defined in terms of dependence-on-dependence and
combinatorial likelihoods alone, or through metalinguistic sentences that
are so formed. The lack of a non-language-like metalanguage (one which
itself would not have (l)-(3) above) leaves these basic (l)-(3) processes to
be characterizable only by exhibition of their constraints.
In addition, the processes of (l)-(3) above each embody a particular kind
of information; and in constituting a sentence, each application of one of
these processes contributes a fixed meaning to the sentence. Furthermore,
360 Interpretation
these processes suffice for all the information in language. Every sentence in
a language is paraphrastically derivable from a sentence found only by (1)—
(3), or is paraphrasable by such a sentence.
(6) Conditions of speech. There are certain additional properties of
language, which are due not to the processes of making sentences out
of words, but to the conditions of speaking. One is the arbitrariness of
phonemes, i.e. their lack of regular relation to the meaning of words or the
structure of sentences.1 Another is the contiguity of constructions (2.5.4). A
third is the linearity of constructions (2.5.3). Yet another property is the
ability of words to refer not just to extra-linguistic objects and events but
also to words in sentences: this makes a metalanguage possible as a subset of
language. And nothing prevents a word in a sentence from referring to the
occurrence of a word in its neighborhood: this makes cross-reference
(pronouns with antecedent) possible.
(7) Finally, language changes (12.2). Different parts of the structure—
sounds, vocabulary, likelihoods, reductions—change in different ways at
different rates. Languages change somewhat differently in different—
especially in separated—communities, thus creating dialects and ‘daughter’
languages and related languages.
Given the list above, we may note that the processes of sentence-making,
in keeping with the conditions of speech above, not only produce the
sentences of a language but also yield the particular constructions and
grammatical relations of each language, as regularities of those sentences.
These include: major word classes such as conjunctions, adjectives, adverbs;
special small word classes such as quantifiers or the English auxiliaries
(where characterizing them as special combinations of general constraints
circumvents the problem of their peculiar character and fuzzy membership);
modifiers (from secondary sentences); morphology (that this can be derived
via reductions explains how it can be a major yet not universal apparatus);
paradigms and their special categories (person, time, number); special
rules, and exceptions to rules, and in general the special domains of certain
reductions.
In spite of the specificity of the processes and conditions, the set of
sentences in a language is not well defined: there are many marginal
sentences, largely because the normal likelihoods of certain grammatically
important words (e.g. take a —, give a —, 12.2), and the domains of certain
reductions, are only roughly statable. There may also be cases of constructions
and regularities that do not apply to particular words in their expected
domain. Overtly, this situation resembles the ‘exceptions’ so common in the
traditional rules of grammar. However, not only do far fewer ‘exceptions’
remain after reduction analysis than in traditional grammar, but those that
1 The fact that speech is in general continuous plays no role here, because language is built
out of distinctions in speech, which constitute discrete entities.
remain have more assignable causes or systemic properties. Partly, they are
due to lack of regularization, by uneven applications of analogy or by
retention of old, frequently used forms (as in the morphophonemics of is-
am-are). And partly they are due to language being constantly in process of
change at one point or another, so that when the description is sufficiently
detailed one is bound to come upon reductions and analogies that are in the
process of extending their domain or of becoming reinterpreted into the
operator framework—and perhaps differently among different speakers.
They may also be due to arrested syntactic change, as in the vagaries of do in
the course of the history of English." And they may be due in complex ways
to the overlapping domains of different reductions.
Although non-regular constructions, and exceptions to the regularities,
thus remain, it is necessary to analyze each language in respect to its
operator structure as far as this analysis will go. Only then can we see for
what forms, if any, the operator-argument reconstructions are impossible,
or too costly in complicated use of the basic apparatus, so that we may prefer
to accept the form as part of the finite material that is outside the operator-
argument system. We can also see how numerous these residual forms are—
far fewer than the exceptions and inexplicable forms found in traditional
grammar—and what characteristics and sources they have. And above all,
as is seen below, for the rest of the language we obtain a simple and unified
and procedural—even computable—structure, with everything referrable
to first principles, something which could not have been obtained by eclectic
and episodic analyses or by purely semantic considerations.
In contrast to the constraints of (l)-(4) above, which delimit what can be
and what cannot be in language, the processes described thereafter, including
the special constructions in particular languages, indicate specific structures
in the language without in general saying what is excluded.
In contrast to the derivational theory developed in the preceding chapters, it
is possible to think of a maximally organized accounting of the similarities
among sentences, somewhat along the lines of cladistic classification in
evolutionary theory. Such an approach would not assume the primacy
of initial sentence-formation over analogic back-formation. Rather, all
similarities would be treated on a par. Nevertheless, the fact that most
analogic formations are on the model of existing syntactic word combinations
makes the pre-analogic, derivational, constructions the key syntax. This
derivational syntax has been found to be best described as generated by a set
of constraints, if for no other reason than that the set of word combinations
of a language is too large and fuzzy to be listable, whereas the constraints
that generate them are interrelatable and listable even if sometimes with
fuzzy domains. However, the constraints with their domains do not quite
match the set of word combinations in the sentences of a language, a large
2 Otto Jespersen, Growth and Structure of the English Language, Macmillan, London, 1948,
pp. 218 ff.
362 Interpretation
part of the mismatch being due to analogic processes over parts of the
domains. We are thus back to recognizing the constraints as primary
processes and analogy as secondary. As to the derivations which describe
the relations among the sentences created by the constraints, they are
largely accretive and recursive: one sentence is derived from another by the
accretion of a constraint on that other, even if the constraint itself consisted
of removing something from the sentence rather than adding to it.
12.2. Structural considerations from language change
In modern linguistics it has been a tenet that synchronic descriptions are

separate from diachronic ones, because many systemic relations are relevant
only among forms that are used together, in various sentences that are
available to the speaker and hearer. However, since at any one time various
forms and uses in language may be in process of change (2.5.7), a sufficiently
detailed description of a language may be expected to have various points at
which something is not quite in accord with the rest of the system. In purely
synchronic description, such irregularities-in-process would be fitted, as one
kind of deviation or another, into the regular material of the language,
just as we would try to do with permanent irregularities. In many of these
cases history shows a more regular source, or suggests a more regular future
form which is being approached. Such past or future regularities may reveal
the present irregularity to be a transition between states that are regular
in the given respect, and thus may locate it among the constraints that create
the regular structure of the language. In some of these cases, historical
considerations show how forms that appear to be syntactically peculiar can
be restated as peculiar morphophonemics (or peculiar transformational
domains) on syntactically regular forms. English examples are: the auxiliary
can, etc. (G298), the ‘pro-verb' do (as in / do not want it, G297), the
periphrastic tenses (He has gone. He is going, G290,294), suffixes like -hood
or -ful which are close to being compound nouns (G173). At any one time in
a language there may be peculiar forms which may be cases of words moving
between one regular status and another: for example, the partially tense-like
and partially auxiliary-like used to, as in / used to go, where there is no / use
to go, I will use to go, l have used to go.
In order to bring earlier or later historical stages to bear upon the analysis
of a given form, we must know how to recognize which forms constitute the
earlier or later stages. This requires knowledge of the mechanisms of change
in phonemes, meaning ranges, and syntax, and also of the possible sources
and directions of the largely unpredictable changes. Such knowledge is
much too vast to be summarized here, but there are two properties of change
in word use and in syntax which are particularly relevant. These are.
following Manu Leumann and more recently Henry Hoenigswald:3 first,

that word-change is not directly word-replacement but a matter of change in
use of sentences, and second, that new' sentences do not suddenly replace
old sentences, but rather the ‘new’ sentences come to be used for the same
meaning as other sentences and so come to compete with those others for a
particular informational niche (i.e. for the same word likelihoods around
them), sometimes with one winning out and the other dropping from use
(cf. in 12.4.4).
Some changes yield no appreciable change in the structure of the language:
such are the effects of most sound-changes on the syntax of the language,
and the effects of most changes of word likelihoods, or the introduction of
new vocabulary (by borrowing or word-coinage), or certain replacements of
one word by another (as of in wit by conscience). Other changes have greater
effects: large extendings of a word's selection (i.e. of the list of those words
with which it is likely to occur) can significantly alter its meaning (e.g.
provided, 11.1), and can even create a new (secondary) grammatical relation
and meaning, as in the case of the adverbial suffix -ly which is a reduced form
(with greatly expanded selection and modified meaning) of earlier dice ‘with
the body or form of (G191,307). Many other suffixes, too, show phonemic,
likelihood, and semantic changes in contrast to their known or suspected
sources. As has been noted, some words show large shifts in meaning and in
grammatical environment, as in fond, which earlier meant ‘foolish’; or like,
which changed from 'pleases' to ‘is pleased by’ (the earlier form of / like it
was It likes me). Even smaller but regular changes in selection can create
new semantic or grammatical effects, as in the well-known replacement of
English cow, calf in matters of eating by Norman-English beef, veal, which
created a limited distinction between animal words and food words; or the
use of meat for flesh in matters of eating, so that meat ceased to be used for
‘food’ in general except in meat and drink and in sweetmeats (in this case
the selections of flesh, meat, food were redistributed, leaving irregular
boundaries).
In many of these changes the earlier structure is still relevant, though less
visible. When give and take have a nominalized verb as object (give a good
look at it, take a good look at it), both their selection and their meaning has
been extended, but their original character can be traced not only in their
precise meaning here but also in the existence of give it a good look (lik ogive
him a new idea) as against the unavailability of take it a good look (cor¬
responding to the unacceptability or marginality of take him a good idea).
More importantly, as noted, the ‘perfect’ tense in English (/ have written)
can still be analyzed to a large extent—with historical justification—as a
-1 Manu Leumann, 'Zum Mechanismus des Bedeutungswandels', htdogertnanische For-

schungen 45 (1927) 105-18; H. M. Hoenigswald, ‘Ancestry and Descent: An overview of
Historical Linguistics’, also ‘Obsolescence, Neologism, and Replacement', both in The Oxford
International Encyclopedia of Linguistics, 1990.
364 Interpretation
present tense have with state, condition (as presumed ‘source’ of -en) for
object (G290): one can hardly say (1) / have written three letters yesterday
evening (as compared with I wrote three letters yesterday evening; but in
French one says II a ecrit trois lettres hier soir)\ however, one can say I have
written three letters by now, like I have considerable experience by now, and I
have written three letters since yesterday and I have this condition since
yesterday. Differently from English, the French perfect has already beome a
past tense. Note that in English, the inability of nominalized sentences to
carry tense is circumvented by using have, which is structurally tenseless
(‘present’), to carry adverbs of past time: My having written three letters
yesterday evening left me with fewer chores today, where have must have
been added to the nominalized My writing three letters yesterday, since one
does not say (1).
The English auxiliary verbs also show various traces of their being the
ordinary verbs that they once were (G298).
Nevertheless, some constructions and distinctions may get lost from
language: in languages which have lost their case-endings on the arguments
of a verb, not all the distinctions have been taken over by prepositions or by
relative positions of the arguments. In addition, some new grammatical
situations may be formed. One example is the English auxiliary verbs which
could once apparently accept on the verb under them a different subject
than their own (as though one said / can for John to win); this is no longer
possible, and we have only I can win (from a required non-extant or no-
longer-extant source I can for me to win). Another example is the definite
article, which is apparently a relatively recent development, and which has
different grammatical requirements in different languages. Also, the English
-ing is a conflation of two very different earlier morphemes (a verbal
adjective and a verbal noun), but the single modern phonemic form has been
broadening its selection so as to become increasingly a single construction
(G41, 294).
We see that any theory of language, even just of language structure, has to
provide for change in language (2.5.7). It is also clear that change is partly
related to the existing structure, though not determined by it, and that it in
turn affects the structure in various ways and degrees. In the present theory,
there are intrinsic relations between structure and development (12.3.0).
Formulating the structure in terms of constraints is different from formulating it
in terms of objects or constructions that are defined by lists or by external
properties. The constraints leave room for particular elements to be unclear
as to their satisfying (i.e. being in the domain of) one constraint or another.
They even leave room for particular elements to satisfy a constraint in more
than one way (depending on how it is analyzed)—either because it is in
process of changing, or simply because it is irregular or not fully domesticated
(reinterpreted) into the system.
Furthermore, the history of sentences, words, and constructions is the
history of their use, that is, of their combinations in speech. But it will be
seen below (12.3.5) that it is the very use of language that becomes
conventionalized and is finally standardized into the constraints of language
structure, so that the structure of language is an end-product of the history of
language, which indeed could hardly be otherwise. The relation between
structure and history makes it possible to evaluate and compare the change
in words and in grammatical combinations of words (i.e. constructions): for
example, to say wherein a word or construction is or is not the same as a past
word, or the same as a cognate word (i.e. a historically related word in
another language).
There are certain things which apparently do not change. These are: the
dependence-on-dependence relation as the formative relation for sentences,
the existence of stable though changeable likelihood differences among
words in respect to co-occurrence, the ability of words to refer to words, and
apparently the existence of low-information reductions. One indication of
their permanence is that these are apparently universal, and so must
have gone unchanged since their coming to be. A consideration for this
permanence is that if these relations are what created sentences as we see
them now, then they cannot be replaced by any other process except through
an innovation as basic as the ongoing structure of language—something
which is not likely to happen as long as language proves roughly adequate.
The essential adequacy and stability of these basic relations depends in
part on their being hospitable to changes and differences in words and
constructions. If the dependence relation that makes sentences held just
among fixed classes of words, defined by lists or external properties, and
if likelihoods of co-occurrence were fixed, then new words—entering by
borrowing or cultural change, or altered in their interrelations by phonemic
change—might be unworkable within the existing syntax. A pressure might
then grow to develop other ways of forming sentences, or informational
expressions, with these new words. But the fact that the relation is a
dependence-on-dependence of words, that the fuzzy likelihoods are change¬
able and not mutually exclusive, that reductions are made by low-information
properties rather than by a fixed listing, enables new words to enter and
function in the system. It is this that makes the system stable, and unchanging in
its fundamentals.
12.3. Formation and development of language
12.3.0. From descriptive order to time order
The origin of language, more specifically of sentences, is not normally

associated with a theory of language structure, and is a subject diligently
avoided by modern scientists. However, the present structural theory is
366 Interpretation
intrinsically relevant to issues concerning primitive—or in any case, early—

formation of sentences, and this in two respects. First, language structure is
obtained here not in terms of composition but in terms of processes—the
processes that constrain. These processes rest on the standardization of use
in speech (12.3.5), and are thus of the same kind as the processes of language
change (12.2). The latter we can follow back historically for a few thousand
years. But these processes have inherently the potential not only of regular¬
izing an already existing (and possibly differently structured) speech, but
also of creating sentences out of an unstructured stock of separate or
arbitrarily combined words. Hence it is natural to push the investigation of
these processes backward, and ask how the relations which they create
might have begun.
Second, the structure ascribes a descriptive order to its main constituents:
first-level operators are defined as operating on zero-level arguments to
create sentences, and perhaps conversely; second-level operators are defined
asoperatingon first- or second-level ones to expand the sentence; reductions
are carried out on existing sentences. When we are dealing with a least
description (2.2), covering almost all the structures of a language in a single
system of constraints, and especially when the core of the system carries the
responsibility of being a mathematical object, then the descriptive order of
entities bids fair to be an essential definitional order, not an artefact of a
particular approach. True, it is possible that the description in terms of
operators and arguments may well be some kind of large restructuring
of previously existing otherwise-formed sentence-like word combinations,
with the operator-argument form being attained by disuse and modification
of any word-sequences that did not conform to it. The language we know
may have been preceded by very different kinds of speech, possibly different
kinds in different human communities. But in any case at some period
speech took on a form which can be characterized by the above descriptive
order of processes, which is their definitional order: B>A in descriptive
order if B is defined in terms of A, and A is not defined in terms of B.
In respect to time order, the descriptive order suggests two possibilities.
1 he descriptively later may have existed first, with the descriptively
earlier being a regularization of its structure. Such a time order is reasonable
in the case of phonemes: although words are defined as sequences of
phonemes, we can hardly think that the phonemes existed without words;
rather, we would suppose that a stock of words existed without benefit of
phonemic explicitness, and that the distinctions developed as a way of
increasing their transmissibility.
Somewhat differently, defining likelihood inequalities on the resultants of
the dependence relation (i.e. on operator-argument pairs) is an artefact of
the theory: the datum is the likelihood of combination, and the dependence
is a name for all likelihoods which the theory considers to be positive, no
matter how small.
In the case of sentences, however, it is more reasonable to think that the

descriptive order is also the time order, with elementary sentences being
formed first, by a preferred dependence in the co-occurrence of words. Even
if there also existed pre-vocabulary utterances in which similar segments
were gradually distinguished as separate words, the definitional order
suggests that, rather than a systematization creating language, there may
have been temporally different developments of successive vocabulary
classifications: first zero-level and first-level (forming elementary sentences),
then second-level words. At the very least it suggests how language could
have grown, by a recursive repetition of a simple dependence relation that
created the elementary sentences. And the considerations of 12.3.5, about
how such a relation could come to be used, suggests that the successive
processes proposed in the present theory would not have been momentaneous
inventions, but the gradual conventionalization, through time, of com-
municationally successful habits.
12.3.1. Pre-syntactic speech
The descriptive order of operators and arguments refers to their status in

sentences. Both types of word may have existed pre-syntactically, i.e. before
the development of sentences, but not with these classificatory statuses.
There are certain things that most likely existed in speech before sentences
came to exist. There must have been words, for sentences to be formed of.
This means in the first place conventional sound-sequences associated with
particular referent meanings, i.e. with particular types of objects and
experiences. It may well be that the sounds were already then subject to
phonemic distinction, because otherwise distinction and transmission of a
large vocabulary could be inefficient (7.1); phonemic distinction can be
characterized simply as those sound-features which are preserved under
transmission. From the explanation of the early development of sentences
(12.3.2), it would seem to follow that word combinations must have already
been in use before syntax, with different likelihoods for a given word to
combine with different other words. Or very many distinct utterances must
have been in use, which were or became sufficiently similar in some of their
parts to become decomposable into segments which we now call words. Such
explanations suggest that a large vocabulary was already in existence, so that
the standardization implicit in sentence origin would have served to bring
some order into an existing stock of word combinations.
The lack of fully structured sentences in this presumed developmental
stage does not preclude that many speech communities may already have
had morphology, that is, particular phoneme-sequences with particular
meanings which could be affixed to certain words, e.g. for plural, time,
masculine-feminine. If all speech communities did, then the morphology
368 Interpretation
must have been lost in various languages, since not all languages now have
morphology; forgetting is not unknown in language history, as in culture
history and in politics. It is also reasonable to suppose that pre-syntactic
speech had intonations or special words or affixes for command and question
and the like.
12.3.2. Early development of sentences
If words are said in combination, certain combinations will readily add up to

some fuller interrelated meaning, over and above that of the words alone
(boy sleeps or sleeps boy), and certain others will not (walks sleeps). This
arises from the actual relations of the words’ referents in the speakers’
experience. There are different likelihoods for different conbinations in
what people would normally say. When a great many word combinations are
said in life situations, each word will be found to occur mostly with one or
another of a particular set of other words: sleeps and walks perhaps with boy,
dog, etc.; flows perhaps with water, milk, sand, etc.; see with the pair / and
any of the above. Such combinations may occur far more frequently than
sleeps with just walks or sees. It is reasonable to suppose that the less useful
word combinations were less said, as not constituting effective commun¬
ication. The meaningfulness of word combination is especially important in
the case of transmission: in face-to-face communication the hearer can judge
the meaning of a word combination with the aid of some of the extra-
linguistic situational information that the speaker has. But in transmission
beyond the immediate situation, the words are on their own. If speakers
tended to have certain words occur with any of a set of other words but not
alone, then that preference could be conventionalized into those words not
being said except with one of the other set. This standardization would then
constitute a constraint on how those words combine.
We thus reach a particular constraint, a dependence among words. The
dependence may impose relative positions in the linear order of speech:
the dependent word may appear usually in particular positions in respect to
the words on which it depends. Within this dependence, the likelihood of a
particular word occurring with particular words of the other set remains as it
was before the standardization into a constraint. This dependence, which is
found in the grammar of individual languages, is thus not an invention or a
plan; it could be neither, since the concept of such a relation could not have
preceded the existence of the relation itself. Rather it is a conventionalization
of natural limitations on the use of word combinations, which are imposed
by the need for effectiveness in communication.
Since the limitation on use was due to the lack of usefulness of particular
word combinations, the dependence which was created thereby carries, and
expresses, a meaningfulness in the very co-occurrence of the participating
words, over and above the meanings of the individual words. The meaning
of combining is that the parts are relevant to each other. And since the
form of the dependence is that the dependent words are said only in the
presence of the others, the interpretation is that the dependent word is said
about those others, that it predicates something about them. The dependent
word is then the predicator or operator on the words on whose presence it
depends; the latter are its arguments. This is not to say that the idea of
predication underlies the dependence relation. Remembering Piaget, one
does not have to understand in order to do. People do not need an awareness
of predication in order to speak or understand language. They need only to
know that certain words are not said except in the presence of certain
others—aside from later reductions—and that this means that the dependent
word (recognized by its position as well as by its likelihoods) is said ‘about’
those others. The awareness of the relation could only arise from experience
with its use in language, and becomes explicit much later, in grammar and in
logic.
The dependence relation creates sentencehood. Every combining of a
word with words on which it depends constitutes a sentence, and saying it
has the meaning of asserting—either asserting the words themselves, or
asserting that the speaker wishes, wonders about, demands, or asks the
contents of those words. It has thus created, out of a particular dependence
relation of co-occurrence among words, a new object (sentence), with a new
kind of meaning (predication), and with new relations (of requirement)
within its set. Non-predicational utterances exist in languages, chiefly one-
word ones such as Fire!, Ouch, Hello, John (the latter may be a case of
calling, or surprise, or pointing and the like). Whether these one-word
utterances should be called sentences or non-sentential utterances of the
grammar is a matter of terminology. It should also be pointed out that non¬
dependent (i.e. non-sentential) word combinations may or may not have
some rough meaning, according to the semantic possibilities of the words:
more in John Bill Mary saw with than in Up vacuum light eats. But a
sentence, i.e. a satisfaction of the dependence relation, always has meaning
and has it explicitly even if nonsensical: John saw Mary with Bill, Light eats
up vacuum.
The evolving of this dependence relation is the major step in the formation of
language structure as we see it now. Not only does it make sentences, all on
the basis of a single relation among words, but it also brings into existence
structural criteria: criteria for recognizing sentences and relations among
sentences, and criteria for distinguishing nlh-level word classes. It establishes a
structural constraint whose meaning is relevance or aboutness. In addition,
it establishes the concept of information as a particular (predicational)
relation among word meanings.
The separation of pre-syntactic vocabulary (12.3.1) from syntax (12.3.2)
makes it possible to think that different parts and functions of the brain were
370 Interpretation
involved in the development of these two contributions to language, and

that the development of the two in the brain took place at different times. It
also specifies the difference between these two contributions to the existence
of language: a large set of symbols, as simply sounds associated with
experience, and a relation of dependence in their use (their co-occurrence in
speech), in which the symbols become modified in meaning and in effect
redefined as cogs in respect to a system.
Beyond the formation of the elementary sentences come the complex
sentences, formed by words which depend on words that themselves depend
on something: e.g. continue on sleep which in turn depends on boy or dog,
etc. Such words as continue therefore in effect depend on sentences; and
with them they construct an expanded sentence. In the same way, know
depends on the pair of a zero-level argument and a sentence (/ know the boy
likes the dog); and if depends on a pair of sentences (The dog barks if it sees a
cat); and so on. In principle, the possibility of defining further dependences
for words is unlimited.
Another further event or constraint, which can be looked upon either as
making other (derived) sentences or as making new forms of existing
sentences, is the reduction in phonemic content (or the change in position)
of certain words in certain dependence and likelihood relations to other
words in the same sentence. As we know, the conditions for reduction are
mostly that the word in question adds little or no information to what is
carried by the other words in their given dependence relation; these other
words can be looked upon as being present in the sentence (as it is being built
up) prior to being joined by the new word. Each reduction has a particular
domain, i.e. a particular set or type of words on which it can be carried out;
and each language has only certain reductions and conditions for reduction
which are special to it. Most reductions are optional, so that the unreduced
sentences continue to exist in the language, or at least to satisfy any
reasonable grammar that can be made of the language, even if some
unreduced sentences are wildly complicated and never naturally said. For
the reductions, their basis in use is the convenience of word and sentence
abbreviation and of identifying the antecedent for cross-reference words
(5.3.2), weighed against the ambiguities that many reductions create for the
hearer. The effects of reductions create various special structures that come
to characterize certain portions of sentences in one language or another, and
that may package the presentation of information to the hearer in various
useful ways. An example is the creation of modifiers (relative clauses,
ad jectives, adverbs) out of a secondary sentence which has become attached
to something in the primary that has been repeated in the secondary (3.4.4).
12.3.3. Further development of syntax
The arising of structural effects from specializations of language use does not
cease with the initial formation of syntax. Structural development continues, in
long-range trends of change which can be characterized only vaguely. There
is some overall evidence of a trend to syntactic regularization, yielding
simpler grammars. Such simplification may be withstood by some of the
most frequently used entities (e.g. the variants of he: is, am, are, was, were)-,
this is to be attributed to frequency, not to any special semantic importance
of be. There is also evidence that structural irregularities, usually encapsulated,
are created on various occasions as by-products of historical changes.
Types of irregularities include the disappearance of cumbersome or
stylistically disfavored sentences that are the unreduced source of existing
reduced sentences; thereupon the base set of sentences lacks the material of
the reduced sentences (and so no longer contains all information of the
language) unless artificial source sentences are reintroduced, with the
necessary reductions, in formulating a grammar of the language. Another
type of new irregularity is the restricting of the arguments of particular
operator words. For example, it has been seen that the auxiliary verbs
(which are Olla) restrict the first argument (subject) of the operator under
them to be the same as their own subject: in lean solve it, the subject of solve
is / and the reconstructed source has to be l can for me to solve it. This is not a
restriction on the argument domain itself, which would violate the math¬
ematically important condition that the dependence of an operator is only in
respect to the dependence properties of the argument words and not in
respect to any external property or listing. And indeed, solve can have as
subject any zero-level argument (i.e. any word which itself depends on
zero), as can can. The only new requirement is that under can both subjects
must be the same. This more specific restriction is still an irregularity, but a
lesser one: rather than restricting the arguments of solve or can, it restricts
the arguments that solve or any other operator has when under can—and
not to a particular list but to being the same as for that occurrence of can.
There are also variousy other irregularities, as when what is obviously the
same word satisfies the conditions for two different classes (without current
grammatical derivability of one from the other) and has to be multiply
classified: e.g. that as a zero-level argument (in / know that), and that as
argument-indicator (in I hope that he's here) which shows that a sentence
(He's here) is appearing as argument (of hope). An irregularity of a different
type, in the complex conditions for quantity words, is seen in the alternative
forms A small number of people is coming tonight and A small number of
people are coming tonight (in the latter, small number of is a hard-to-derive
adjective on people).
There is another development which could take place only after orderly
372 Interpretation
sentences came to exist, but which modified in detail the nature of sentence
structure. This is the spread of analogic formation and analogic word
combinations, and other extensions of word use, which made word com¬
bination and local structures within sentences (e.g. phrases, paradigms)
more rule-based and less a matter of the semantic appropriateness which
was assumed in 12.3.2 to have underlain the development of the dependence
relation.
In contrast with such slow or small changes, there are great specializations
and further structures in language when the things being talked of (the
subject matter) have specialized relations among their perceived objects.
This is seen in concentrated form in sciences and above all in mathematics,
as noted in 10.2-4. The greater precision and intellectual complexity of
scientific writing is creating minor grammatical innovations, as in the use of
respectively to make cross-connections, e.g. in A and B produce Y and Z
respectively (from A produces Y and B produces Z); such a criss-cross of the
source sentences would be exceptional in natural-language grammar, and
the use of respectively in colloquial English is not restricted to this environ¬
ment.
Other innovations are larger but still can be housed within a generalized
language structure. One innovation arises out of sentence-sequences, i.e.
out of dependence among words of neighboring sentences; these also have
inherently a limiting feature of having a common subject matter. The
simplest case of this is seen under certain (conjunctional) Ono, which
connect two sentences by operating on the pair of them: here the high
likelihood is that there is some word that occurs in both sentences (9.1). A
larger development is that of a double array of word-class dependences in
the sentence-sequences that constitute discourse (9.3). Another develop¬
ment is the singling out of certain constraints on sentence-sequences,
whereby the last sentence has the semantic status of a conclusion from the
others: proof in mathematics (sayable in English if we wish), and—in a far
less specific way—well-constructed argumentation in science. This creates a
new semantic power for language: a distinguishability of non-loss of truth
value or of plausibility (in science sublanguage).
More fundamental innovations in language structure arising from subject-
matter limitation involve the overall properties of the operator-argument
structure, differently in mathematics and in science languages. In mathematics
and mathematical logic, likelihood differences within an argument or operator
domain have been eliminated. In argument status, the differences are
eliminated by having only an indefinite argument, a ‘variable’, meaning in
effect only ‘something’ or ‘a given thing’ (within a domain), it being under¬
stood that within an indicated sequence of sentences each occurrence of the
same variable-symbol is a cross-reference to the first (while each occurrence
of another variable-symbol is not specified as to sameness with the first,
unless stated to be the same or different). And the meaning of operators
there is related not to likelihood differences but to truth tables that specify
the truth value of the resultant sentence from the truth values of its
arguments; likelihood then has no semantic or structural relevance.
In science languages, as was seen in 10.3, the consistent restriction of
subject matter, and the attention in each set of texts to particular properties
or relations within it, is reflected in particular operators being used only on
particular words out of their possible natural-language argument domain.
What is new here to language structure is the ‘only’. As noted, in English as a
whole, the 0,m eat may never have occurred on the N, N pair vacuum, light;
nevertheless. Vacuum eats light is a sentence and we even know what it
means, though the meaning may be nonsensical if eat is taken as ‘nourishes
itself on’, or simply untrue in a general sense if eat is taken as ‘effects
diminution of. In contrast, in immunology the On„ contain cannot occur on
the ordered N, N pair antibody, cell (though it occurs on the pair cell,
antibody): the possible though nonsensical English sentence Antibody contains
cells is not considered a possible sentence of immunological language.
This creates subclasses, largely superseding likelihood gradation: a novel
language-structure situation in which there are several sentence types
(operator-argument structures) at the same level (e.g. the Onn level, or the
Ooa level). And it creates a novel semantic power for language: a dis¬
tinguishing and eliminating of irrelevance and nonsense.
The syntactic systems of science and of mathematics have lost the math¬
ematical-object status (because their dependence is defined on particular lists
of argument words), and they have to be described in an external meta¬
language, that is, in some natural language. Because of the semantic
differences expressed by these various structural innovations, the formal
notations of mathematics and science languages (as in 10.5) can be translated
into natural language, but natural language as a whole cannot be translated
into them.
Both in the regularizations, irregularizations, and changes in the language
as a whole, and also in the larger innovations of science languages and
mathematics, we see that change and furtherance in structure can go on, that
the development of language is not at an end. We also see that at least the
larger changes and innovations arise from the different external conditions
(e.g. subject matter) and internal linguistic pressures (e.g. analogy) under
which language is used, and from the conventionalization of such usage. The
conventionalization is seen, for example, not only in the development of
explicit rules of mathematical well-formedness and proof, but also in the
peer-pressure criticism by colleagues that keeps scientists’ arguments within
inspectable grounds.4 Both the conjectured initial formation of syntax and
its further development are thus institutionalizations of ways of effective
recording and communicating of information about the world.
4 Cf. P. Latour and S. Woolgan, Laboratory Life, Russell Sage, New York, 1979; K. D.
Knorr-Cetina, The Manufacture of Knowledge, Oxford Univ. Press, Oxford, 1981.
374 Interpretation
12.3.4. The limits of language
The further, and presumably continuing, development of grammar does not

mean that everything needed for saying anything can be reached, or that
anything that the syntax makes possible (and that would be helpful for
speaking) will some day be reached. There are many inadequacies in one
language or another, and in all languages.6 In many languages some complex
applications of grammatical relations cannot be expressed by direct use of
the relevant grammatical devices. In the English comparative, one cannot
repeat the comparative form in order to compare a comparative with a fixed
amount: one cannot say He is taller than me than me (or: He is more taller
than me than me), but only He is taller than me by more than my own height ox
the like. Also, direct nominalization of a sentence in English, by argument
indicators placed on the operator of that sentence, leaves no room for the
tense that the sentence was carrying: e.g. in John came yesterday (in past
tense), which is nominalized in John's coming yesterday surprised us (with
no tense on come). The past can nevertheless be expressed here by circum¬
vention, as in John’s having come yesterday (where have is not in the past
tense, and have come is not really in the perfect tense since the perfect would
not accept yesterday here); much more rarely, futurity can be indirectly
expressed in nominalization, as in the rare John's being to come soon, from
John is to come soon which is present tense but indicates a plan for the
future.6
More generally, some features which we would consider of grammatical
relevance may not be distinguished in language. For example, the use-
mention distinction (5.4) is not indicated by any general grammatical
device, although quotation marks are often used for it in writing. Another
example is seen in the case of assertion. In English, assertion is not
5 Of least interest here is the fact that the specific forms of reductions in a particular language
may make it impossible to express certain things in it. For example, a language in which absence
of a plural suffix means 'singular' makes it impossible to leave the difference between singular
and plural unstated. Furthermore, in such a case it becomes unclear whether one can say, for
example. Either John or his co-workers is responsible or Either John or his co-workers are
responsible. Another type of situation is seen when we consider That he failed is clear, formed
by is clear operating on He failed. If now we want another second-level operator (e.g. is not
universally accepted to act on That he failed is clear we find that the grammatically ‘correct’
form, as in That that he failed is clear is not universally accepted, is generally avoided in favor of
That he failed is clear is not universally accepted (which can only be described by an ad hoc
reduction of the two occurrences of that to one).
(1 Another inadequacy involves the antecedents for pronouns. Consider for example the
following sentence: That night he wrote to Tanya telling her that the gipsy woman was in despair
and that he could no longer marry her, Tanya (A. N. Wilson, Tolstoy, Norton, New York,
1988). Here marry her is ambiguous, and its ability to refer to Tanya would have been even
weaker if the author had written saying that instead of telling her that. The alternative, of not
pronouning here, by writing marry Tanya, is virtually unavailable: He wrote to Tanya that he
could no longer marry Tanya would seem to refer to another Tanya. The easiest solution is the
stylistically forced marry her, Tanya.
structurally defined. Whether a sentence is asserted can make a grammatical,

not only a semantic, difference, as is seen in certain structural phenomena:
for example, as noted, we can say That he left is false (or improbable or
uncertain), but not He falsely (or improbably or uncertainly) left, because the
latter can only be a reduction from He left; that he left is false (or improbable
or uncertain), where the negation is too unlikely a sequel to the assertion to
support the reduction (3.2.2, end). Nevertheless, it is hard to find what
structural features indicate assertion. Many higher operators assert or non-
assert their arguments: I assert that he left in contrast to I deny (or think) that
he left. But in the absence of these, the situation is not always clear. It hit the
floor is not asserted in / caught the glass before it hit the floor (but it is asserted
in I timed the glass before it hit the floor), and it is only fuzzily asserted in It hit
the floor, so they say. In I'm here, looking at a house which exactly fits my
needs, the existence of such a house is presumably asserted, but in I'm here
looking for a house that exactly fits my needs, it is not asserted: there may be
no such house.
More pervasive are certain lacks which are structurally characteristic for
natural language, but may be overcome in specialized linguistic systems such
as science languages.
One of these is the lack of a complete addressing system (3.4.4, end). An
effect of this lack is that when a speaker uses free pronouns (he, she, it, etc.),
he knows which neighboring word is the antecedent for it, but cannot
communicate this to the hearer within that sentence (see n. 6 above).
Furthermore, the partial addressing apparatus which is indeed used consists
of words which refer to the positions of other words in the sentence (e.g.
latter, or wh- generally referring to the preceding host noun). The gram¬
matical analysis shows that referential pronouns are not directly cross-
referencing devices, but simply reduced repetitions of a (usually) preceding
word. Since the word has then been reduced by virtue of its repeating the
antecedent, the fact of reduction constitutes a reference to the antecedent.
Since no word or word-sequence can be a repetition of its own occurrence,
no word can refer to its own occurrence: in This sentence is false, the this
refers grammatically to some preceding sentence (as in He says 7 didn’t
know.’ But I assure you, this sentence is false). Thus this form of the
impredicative paradox is unavailable in natural language (when the grammar is
considered in detail); other forms are sayable, but their impredicativeness is
indirect.
Another lack which may be unavoidable is the lack of total regularity in
the vocabulary. Some words may be factorable into a sequence of two or
more words in some grammatical relation to each other, and this on
grammar-simplifying grounds.7 Flowever, in natural language, the openness
7 However, factoring on purely semantic grounds (e.g. kill into cause to die) yields little or
no advantage, and will virtually always run foul of differing combinatorial restrictions or
likelihoods: from The Queen's command caused his words to die unspoken, we cannot obtain
376 Interpretation
of the vocabulary, and the fluidity and multiple determinants of co-occurrence

likelihoods, exclude the possibility of a complete and regular factoring of the
vocabulary into a least set of words which are maximally independent in
respect to co-occurrence. In a science language, where at any one time there
are fixed subclasses (even for new incoming words), the combinatorial
potentialities of the words may be sufficiently regular to permit factoring the
existing words into an organized substrate of vocabulary components.
Also, natural languages have apparently not developepd a set of third-
level operators—ones whose arguments are operators whose arguments in
turn are operators. Science languages, or the scientifically precise use of
certain words in natural language, may as noted above develop such a
dependence: e.g. rate, as in the rate of change of some property or relation
of things.
In spite of the various open features of language, there are certain limits to
the direction of its development. Some obvious limits are due to its discrete¬
ness and linearity. The basic partial ordering, the repetition positions under
Ono and Ooa (G139, 146, 148), and the status of modifiers (3.4.4) are all
simply not indicated in the linearity, though one could imagine, as offshoots
of language, systems of writing or graphing that would be linguistic in
retaining the dependence and likelihood properties but would be freed of
linearity (e.g. the semilattice representation of sentences).
The possible content of language is limited by its associating pre-set words
with objects and properties of the publicly perceived world, and the immutable
partial order of dependence which creates sentences and which is essential to
science languages also. The meaning in languages is thus restricted to what
can be meant by these words and their combinations.8 As a result, language
is limited in what it can express, especially in what it can express structurally,
and directly (rather than just by talking about it). In effect, by its grammar,
language can do directly little more than predicate (which includes, for
example, taking questions and commands as statements that one is asking or
requesting something). Language is thus primarily a system for publicly
communicable information, as seen in Chapter 11. There are many inexplicit
. . . killed his words unspoken; and from I'm just killing time here we cannot obtain I'm just
causing time to die here.
x The particular meanings of extant words are in part accidental, although there are many
meanings that are found with little difference in a great many languages. Some types of
meanings are harder for words to come by. For example, many abstract concepts such as (a)
understanding, important, or (b) truth, magnitude, are available in single words either: (a) only
as a result of metaphoric specialization of words of concrete meaning (G405), especially of
morphologically complex words; or (b) only as a result of operator nominalization (from
nominalizing a sentence with zeroable indefinite arguments: truth from things' being true). In
addition, the meaning of a word can be greatly modified by adding modifiers to it, which
changes the selection since the further operators on the given word are then determined by the
word as modified (as in The snow house melted). (It should be pointed out that, as in all
derivations in the present book, ‘from’ above means specifically ‘derivable by established
reductions or other transformations in the grammar’.)
and private meanings that language cannot express directly. As noted in

10.7, language cannot communicate that which is communicated by music
and other arts. Feelings are indicated either by the possibly false statements
that one has these feelings, which is hardly an expression of them, or
by phonetic activities (intonations, ways of pronouncing) which have no
regularities of combination with anything in grammatical structure and so
are not part of a grammar. Literary, poetic, and even jocular effects in
language are obtained not primarily by special grammatical features (such as
allusion) but by special deviation from standard grammar, or by exceptional
utilization of grammatical features.
Finally, natural languages all have the property of having different
relatively stable yet changeable likelihoods for individual operator words in
respect to their individual argument words and vice versa. This makes it
difficult for words to be classifiers of other words in any definite manner. It
also makes it impossible for language to exclude nonsense—that is, for
nonsensical sentences to be ungrammatical. And it makes it impossible to
identify single sentences by their structure as true or false: even and on a
sentence together with its negation is not necessarily false, as in A politician?
Well, he is one and he isn’t. However, in science languages, where likelihood
properties are replaced by special word subclasses organized into special
sentence types, both of these limitations can be overcome.
12.3.5. From use to constraints: Institutionalization
The conventionalization and institutionalization of language use to the point

of its becoming a constraint on how one speaks one’s language (noted at the
end of 12.3.3) may arise from several sources. For one, the limitations on the
experience that people express or communicate makes it more appropriate
and expected to use certain words together rather than others. Second,
speakers and hearers need to understand each other efficiently, and so come
into accord on what sounds and word occurrences are considered repetitions
of each other, where to expect particular kinds of words (if they are being
said), and in what way to understand the relation of words in a combination,
and finally, when to understand something as being a reduction of another
word. In later specialized constraints, there may also be conscious imitation
and agreement on how words are used, and peer pressure for particular uses;
this may hold for the development of mathematical terminology and notation,
and of styles of argumentation in science reports.
In all this conventionalization, and even in the institutionalization that
makes some forms grammatically ‘wrong’, there is little or none of the
indirection and misleading that make so much of culture an institution
of control. In this respect language differs from much of culture. The
378 Interpretation
conventionalization here will be seen to have something of the character of

biological evolution rather than of power-preserving social structure. It is
also important that individual idiosyncrasies in language, as well as cultural
and literary styles or social and regional divergences, can all be described in
terms of special constraints, even though in a rather different way than the
constraints of language structure, of language-like systems, and of language
change. For this means that in various respects and degrees all of these can
be investigated in similar terms, and therefore can be compared.
The development of constraints from word-specific dependence (i.e. of
each word on a selection of certain others, as in eat on the pair boy, berry) to
the general relation of dependence-on-dependence was crucial to reaching a
stable system. As has been noted, such a system could be generalized so as to
be adequate for any new situation, whether this situation consisted of new
words in existing classes (e.g. other pairs than boy, berry under eat), or
whole new classes (e.g. words, such as continue, that might depend on eat in
the way eat depends on boy, berry). Indeed, defining word classes by their
dependence on the dependence properties of other words creates a self-
contained and self-organizing framework for language and enables it to
expand without limit, but only in terms of this same relation. For example,
an 0o class that operates on O to make a sentence does it in the same way
that O,, operated on N to make elementary sentences (as in The boy’s eating
of berries continued)-, and so does an Ooa class on a pair 0,0 to make one
sentence out of two (e.g. conjunctions).
It is clear that the constraints developed where they did not exist before,
ultimately out of a situation without constraints. However, they did not—
and could not—develop out of nothing, but rather out of tendencies to use
certain kinds of word combination in preference to others. The initial
constraint created the set of orderly elementary sentences where no such
thing had existed before as a fixed structure. But this does not mean a sudden
innovation or even a recognizable innovation—only that types of com¬
bination which were widely in use, in contrast to the non-use of other
combinations, came to be used exclusively. This is in effect the same as
what happens in language change (12.2). A somewhat different situation of
a new structure arising without apparent invention or discovery was seen in
the case of the alphabet (7.2.2).
The initial conventionalizations developed partitions of the set of words
out of fuzzy differences. To give an artificial example: if, say, sleep was used
with child, dog, etc., and perhaps rarely with tree, and virtually never with
berry or walk, a comparison with, say, fall which occurs with all these words
except walk, suggests that berry be included as possible argument of sleep
(with only nonsensical meaning), while walk is excluded as argument of
sleep or fall and is left to be itself an operator. The partition of the set of
words into N, On, etc. is thus a reassignment of virtual non-occurrence into
rare or nonsensical occurrence in the case of berry under On, and exclusion
in the case of walk under O,, (any such occurrence being then called ‘not in
language’ or ‘wrong’).
Changes of usage do not only conventionalize tighter structuring. They
can also create word classes that at least in some respects cut across the O,,,
O0, etc. partition, as in the case of prepositions, which occur in many
operator statuses, and which we can attribute to just one operator class only
by calling upon transformations for deriving all other uses. Indeed, changes
in use can reduce some of the tight structuring of grammar. For example, in
various languages the gradual phonetic subsidence of case-endings, which
had been a structured and required set of operators, left prepositions, which
constitute a less structured set, to fill their syntactic place.
A prominent situation of usage-change-creating tight structure is seen in
the special constraints of special word classes; such processes have some
similarity to what is called phonemicization, wherein changes in the pro¬
nunciation of certain sounds can make them phonemically distinct from
other sounds in ways that they were not before. One syntactic case is the loss
of the second subject under English auxiliary verbs such as can. As has been
seen, those verbs may have had a full sentence as their object (second
argument), as though one could say I can (for) him to do it meaning
something like ‘I know how for him to do it.’ In time these verbs ceased to be
used in cases where the subject of the verb under them was different from
their own subject. As the sameness of subject for can and the verb under it
became a requirement, a new constraint arose in the language. These verbs
also have an unremovable tense, and they zero the to under them. Since
several verbs share these requirements, there rises a small subclass of
‘auxiliary’ verbs different from other verbs. This subclass is thus the product
of a set of constraints acting on the same words. The tight structure here is
lessened because the domain of this set of constraints is a bit fuzzy: some
verbs have only some, not all, of the constraints of the auxiliaries (e.g. I
ought not do it by the side of / ought to do it.)
Another important constraint which presumably grew out of use preference
is the verb-adjective distinction. Both of those word subclasses are operators,
but, as has been seen, those operators that had higher likelihood of tense-
change within a discourse had the tense attached directly to them (e.g.
walked, aged) and are called verbs, while those that did not had the tense
attached to a separate ‘carrier’ be (e.g. was brown, was old) and are
called adjectives (G188). While operators are graded in respect to tense-
changeability, they are partitioned in respect to tense-attachment: sleep is a
verb even for a hibernating bear; ill an adjective even for an intermittent
illness. Note that the same distinction can be made without creating a
grammatical requirement: Chinese has only optional suffixes for ‘past time’,
but these occur more commonly on momentaneous and active predicates
than on stative ones.
In all these cases we see much the same kind of structure-creating process.
380 Interpretation
The particular accumulation of pronunciational change which makes new

phonemic distinctions (‘phonemicization'), such historical changes in grammar
as the formation of compound (periphrastic) tenses, Manu Leumann's
formulation of historical change in terms of use-competition among sentences
(n. 3 above), the formulation of word-use change in terms of combinatorial
patterns as in the work of Henry Hoenigswald (n. 3 above)—all of these in
different ways exhibit grammatical innovation as a result of changing use
becoming institutionalized. The same process can also yield the initial
formation of grammar out of freely occurring, grammar-less word com¬
binations.
It has been noted that some sentences do not conform to the operator-
argument structure. The general and self-contained framework of all these
constraints makes it possible for almost all non-conforming constructions to
be reinterpreted as though they did conform. They are thus domesticated
into the system. This applies both to pre-syntactic forms that may have
existed before dependence and sentences evolved, and to many forms that
arose later by non-dependence processes such as analogy and change.
Reinterpretation of forms as though they were produced by the dependence
relation on known words, with reductions of known types, is possible only
because the predicational meaning of the dependence suffices for almost all
meanings of word combinations, and because known types of reductions
suffice in almost all cases to yield existing forms out of suitable dependence-
structured sources. For some items, such as the perhaps pre-syntactic
tenses and plurals, it is indeed difficult to find reasonable sources (though
dependence-structured paraphrases are always available); but it is easy for
others, including the recent periphrastic tenses (such as the perfect tense,
has caught) as well as the perhaps pre-syntactic interrogative and imperative.
As an example of how the arising of a fixed sentential structure can lead
to reconsidering the grammatical status of linguistic forms, consider the
domestication of word meanings into syntax. Before syntax came to be, the
many words that must have already been in use must have had some range of
meanings as associations with referents, i.e. with objects, events, states in
the perceived world: rock, say, would have meant a stone of big or medium
size, and stone medium or small, and pebble small. Once words are used in
operator-argument combinations, within which the likelihoods are relatively
established, we find a new situation, local synonymy: rock can have stone as
synonym in / dug up the rock, but in They climbed the rock and built houses
on top of it, one would be much less likely to admit stone in place of rock.
Thus the meaning of a word-occurrence in a fixed structure may be only a
segment of the meaning-range of that word, and may be the same as the
meaning, in that position, of some other word which is not otherwise
synonymous with it. This situation makes the meaning of a word depend on
its grammatical co-occurrence, and in a way that is more explicit in detail,
hence more useful, than presenting the full meaning-range as a single entity.
In a dictionary, a word is given a meaning or a range of meaning, although

the set of referents is in many cases not an otherwise existing or useful class.
In a more careful dictionary, a word is given a list of meanings, without the
implication that these form a whole, a continuous range, or even that they
are always interrelated in content. In the occurrence of words in sentences,
however, the situation we find is not that words appear in a whole meaning
with just part of that meaning used in each occurrence, but that each word-
occurrence in a sentential (operator-argument) environment appears in just
a specific meaning which may not be the same as that of another occurrence
of that word in another environment. Starting from meaning by referent, we
have thus reached meaning by operator-argument co-occurrence. The
difference between these two is thus a difference between two stages of
vocabulary: pre-syntactic (or in isolation, independently of sentences), and
in respect to sentences (i.e. when occurring in sentences).
Another way in which syntax readjusts the meaning of words is seen in the
pronouns. It is reasonable to assume that pre-syntactic vocabulary contained
demonstrative words such as he or this, whose individual referent varied
with the situation of speaking. Once words were used in sentence structures,
most occurrences of pronouns became cross-referential, referring to the
same referent as some other word, ultimately some non-demonstrative
words, in the flow of speech. The remaining demonstrative occurrences
(even perhaps in e.g. It's a nice day) could then be artificially reconstructed
as cross-referential to a zeroed metalinguistic We are speaking of something;
that something is a nice day. Starting from the demonstratives as having a
unique kind of referent, approximately an unidentified specific referent in
the extra-linguistic situation, we have thus reached cross-reference within
the sentence-sequence (5.3.4).
In this way, the few essentials of syntactic structure not only constitute a
stable system, but also are consolidated by being able to house almost
everything else in language, including word meanings. Considering that
language also includes its own metalanguage, we see that it is a self-
contained system. It is exhibited and recognized by constraints on repetition
and combination of parts, and is defined within itself. While this analysis
shows something about how language could have come to be, it leaves open
the question of single or multiple origin. The relation of word meanings and
constraint meanings to the phenomena of the world, and the communication-
success criterion for institutionalization, suggest that language could have
had various primary-parallel origins, no less than a single origin, or possibly
a mix of the two.
382 Interpretation
12.4. The nature of language
The theory of language structure and development presented here makes it

possible to draw some conclusions about the nature of language. The
conclusions are the more relevant since the theory finds language to be a self-
sufficient system which includes its metalinguistic apparatus and its informa¬
tional interpretations (all of which have to be accounted for, because any
account of them would all have to be stated in language itself). We have seen
in 12.3 how language could develop out of a situation in which there was as
yet no language. We now consider what the whole theory shows about the
properties and functions of language.
12.4.1. The language system; language and thought
It should first of all be clear that language is not directly a more developed
form of animal communication. Not to enter into the discussions of animal
communication and ethology, we may nevertheless judge that the difference
between human and animal ‘language’ is one of kind, not of degree. Animal
communication contains sounds or behaviors which one might consider the
equivalent of words; but, in present knowledge, it has nothing of the
dependence relation which creates sentences and predication. A less obvious
but fundamental difference is that there is little present evidence of the
respondent animal exercising any choice in type of response (to stimulus or
‘symbol’), other than whether (and perhaps how strongly) to respond.
Existing evidence suggests that animals have developed innate propensities
to particular behavior in response to stimuli; the stimuli include behavior
(sounds, postures, etc.) of others in their species. This can be called
rudimentary communication; but we cannot say that the behaviors are
subjectively signals or symbols, and there is no systemic signalling struc¬
ture with potential for variation and expansion. The apparent behavioral
sequences, as in some aggression and mating behavior, are not innate
sequences within an animal, but are a co-evolved interaction produced by
each given participant responding to the stimulus offered by the other in
response to the animal’s own behavior or presence. The whole is more like
the co-evolution of flowering plants and pollinators such as bees than like the
functioning of human language.'*
y K. von Frisch, The Dance Language and Orientation of Bees, Harvard Univ. Press,
Cambridge, Mass., 1967; G. Baerends, C. Beer, A. Manning, eds.. Function and Evolution in
Behaviour, Oxford Univ. Press, Oxford, 1975; D. J. McFarland, Feedback Mechanisms in
Animal Behaviour, Academic, London, 1971; M. Lindauer, Communication among Social
Bees, Harvard Univ.’Press, Cambridge, Mass., 1971.
As to whether animals can learn language, such work as that with

chimpanzees has not differed in essentials from other animal training,
except for the greater complexities of response made possible by the
chimpanzees’ general intelligence. The impression that there is language
comes from choosing appropriate words as names for the buttons the
chimpanzees are trained to push.
A very different distinction is to be made between languages and thought.
Again without entering into this vast problem, several considerations are
obvious. There is certainly some type of thinking that takes place without
language and is unrelated to its structure.10 And once language is in
existence, there are occurrences of structured language without comparably
organized thought, as perhaps in illusions, in alliteration, in poems of
mood, in nonsense and nonsense-verse, not to mention small-talk. The
decisive factor is that language has additional conditions over and above any
representation or facilitating of thought: being public (and pre-setting a
public stock of words and co-occurrence constraints); efficacy for trans¬
mission; and realizability in speech with its linearity, with the limitations
which are imposed on speech by articulation and auditory distinguishability,
and those which speech imposes on structural representation of partially
ordered and complex relations among words. Some structural features of
language are of little or no relevance to any language-independent thought:
for example, phonemic distinctions and phonemic composition of words,
reduction and other transformations, and the loss of specific meaning for
words with broad selection.
Evidence that thought and information can be independent of certain
grammatical features is seen in the occasional circumventing of grammar
to express meaning (as in expressing time without tense in nominalized
sentences), and more dramatically in the fact that writing in a given science
can be described in the same science-language grammar no matter in which
natural language it is written (10.5). Nevertheless, most orderly thought and
most language go together: the structure developed by language makes
precise, and to an unknown extent may shape, and even direct, thought and
the perception of the world, even while it limits the character and direction
of these.11 This situation of something functioning both for service and for
10 L. Weiskrantz, ed.. Thought without Language, Clarendon Press, Oxford, 1988. In the
large and uneven literature on this subject, special reference should be made to two of the early
modern contributions: the writings of Jean Piaget, especially Le Langage et la pensee chez
I'enfanl, Delachaux et Niestle, Paris 1924, and Jean Piaget and Barbel lnhelder, Memoire et
intelligence. Presses universitairesde France, Paris, 1968; and also L. S. Vygotsky, Thought and
Language, MIT Press, Cambridge, Mass., 1962. For the status of language in respect to
thought, it may be relevant that almost all writing systems, though they start with pictures as
direct representations, develop into signs representing words and their sounds.
11 This vague relation of thought to the basic structure of language is quite different from the
unfounded suggestion that the detailed grammar of a particular language affects or determines
the thinking of its speakers, as in the claim that languages which have no grammatical tenses
might block the development of physicists who need t as an essential variable.
384 Interpretation
control is well known: aside from those pervasive social institutions which
serve solely for control over the people, the bulk of social, cultural, and
personal conventions both facilitate as well as limit—often simultaneously—
the behavior of the people who function within them.
What has been seen in 11.1, 11.6, about the meaning of the structural
entities of language suggests that the system developed in language is
consonant with important features of advanced thinking. In overview, the
relation of language structure to thought may be summarized as folows.
(1) For humans to use words (singly), i.e. sounds that are associated
with (‘mean’) the user’s states or received stimuli, is consonant with
animal behavior (even if the animals are not thereby subjectively ‘sending
messages’); hence, this in itself may not constitute thought as a substantively
new phenomenon peculiar to humans. (2) Using word combinations in the
presence of combinations of the stimuli referred to need not be considered
as essentially different from using single words, even if in the objective world
certain of the words in the combination ‘mean’, roughly, objects, while the
other words mean states or properties of those objects. (3) However, the
fixing of an ‘operator-argument’ relation between the object words and
the property words in each combination, even though it presumably was
only a standardization of a naturally-occurring constraint on the combina¬
tions, created predication as a semantic relation between the words in a
combination, whatever they be. This had two semantic effects. (4) One was
that any new combination had meaning by virtue of the predication relation
between known words, not only by virtue of the association of the combina¬
tion with particular experience; thus, the meaning of words shifted from
being extensional (i.e. the specific associated referents) to being intensional
(i.e. anything that could fit the predicational relations in which the word was
used). (5) The further effect of (3) was that predication could be said
independently of actual experience, since a predication would have the
meaning of one word predicating about another even if there was no
experience with which that word combination was associated. It is reason¬
able to see (4) and (5), both of them due to (3), as important features—
perhaps facilitators—of thought going clearly beyond any pre-human level.
Before stating this explicitly, we survey the structural meanings. As to
words: first, words (or morphemes) are discrete entities, since they are
sequences of discrete phonemes (or of phonemic distinctions): this makes
the meaning-bearing symbols explicit. Second, when discrete entities are
used to represent in a regular way the objects or events (and relations) with
which they are associated (end of 1.4), they necessarily come to mean
selected perceived aspects of the objects or events, rather than the less-
selective representation contained in, say, a picture. Third, since by 11.1 all
words can be used not just for a unique object or event but also for various
ones (any ones that fit the meaning the word already has), each word is in
effect defined not by a specified set of referents (objects or events), but by
the properties that are common either to all its referents or to various subsets
of them. As to sentences: the operator-argument relation over a finite
vocabulary creates sentences that suffice to indicate or represent not the real
world so much as perceptions related to the phonemena of the world.
In addition, the informational structure of language provides structural
possibilities that go beyond information, expressing and perhaps even
fuelling the imaginative capacities of thought. This potentiality in language
arises from the fact that words whose meaning comes from association with
perceptions of the world are naturally restricted in their co-occurrences to
the way perceptions in nature co-occur (11.6.1). As a result, the restriction
of combination (selection) among words reflects the meaning of those
words. But this restriction leaves open the possibility of using the unused
combinations. The unused combinations present new meanings, which may
be nonsensical, interestingly imaginative, etc., in respect to the perceptions
that had restricted those combinations. And old words, or newly made
phoneme-sequences, can have new meanings which are determined by their
new selection of word combinations rather than by direct association with
observed phonemena (e.g. unicorn).
A crucial contribution of language to thought is, then, the separation
of predication (grammaticality) from word-choice (meaning), due here to
distinguishing operator-argument dependence (non-zero co-occurrence like¬
lihood) from all the specific likelihoods. Due to this, when we want to
make a new combination of meanings we find at hand a structure that puts it
into a publicly-recognizable predication, without our having to develop a
new bit of grammar for the new word-combination—a new consensus for
what the new juxtaposition of meanings should mean. This permits one
important type of thought—predication, with the interrogative, optative,
and other modes derivable therefrom—to go beyond the direct meaning-
combinations of our experience. (Compare the lack of useful new meanings
in totally ungrammatical sequences of words.)
A second crucial contribution is the recursive extension of predication to
create predication on predication (more generally, all the second-level
operators, such as Oa). Selectional restrictions on these in respect to their
arguments (especially in the case of connectives) create the machinery for
structured extended thought, discourse, and proof.
We see that the projection of phenomena onto perceptions, the ability to
define by properties rather than by denotational extension, the ability to
imagine beyond what has been observed, all arise in the structuring of
language. But these are also part of the apparatus of advanced thinking. It is
not a question of whether these capacities came from language to thought,
or the converse. One can reasonably suppose that these developments in
language capacity facilitated developments in thinking, and advances in
thinking facilitated the utilization of language-structure potentialities.
It should perhaps be noted here that because of the structural differences
386 Interpretation
between language and logic (primarily due to the different selections that
individual words have, 3.2.1), the attempts to represent the meaningful
structures of language in the structures of mathematical logic have not been
successful. Logic goes beyond language in the capacity to structure proof,
and language goes beyond logic in the capacity to differentiate meaning.
It may also be mentioned that the way complex thinking is expressed in
language structure casts doubt on the claimed potentiality of computers to
carry out an equivalent to thinking. Computers can indeed be programmed
to analyze sentences in respect to the known structure of their language
(ch. 4 n. 1). But, first, we know that this structure is based on the recognition
of phonemic distinctions, the phoneme-sequences that constitute words,
and the operator-argument relation among the words of a sentence. We
have seen (12.3.5) how these could develop in the course of the intricate and
public behavior of conscious living organisms interacting with their environ¬
ment, but we do not know how computers could develop such structures
without being given them in a program. Second, nothing known today would
suggest that computers can carry out, sufficiently for any serious purpose,
the selective perception and the subjective response to the outside world, let
alone the awareness, that determine the ongoing use of human language—
the word-choice of a sentence and the sentence-sequence of a discourse. The
fact that this response and awareness must have material grounds within the
processes of the human organism does not mean that the particular man¬
made mechanisms we know today should be expected to simulate those
processes.12
It follows that language cannot be understood as a further development of
some existing system, but has to be explained de novo. To do so we have to
go by the properties which seem to be necessary and universal for language.
Before doing so, in 12.4.2-4, we note here which universal properties are
not systematically necessary. Some, such as linearity and the limits on
distinguishability of sounds, are, though universal, not in principle necessary to
language but are due to the fact that language developed in speech rather
than, for example, in gesture or markings. Indeed, language has had to
circumvent some properties of speech: to circumvent the continuous sounds
by means of segmentation, the continuity in differences among sounds by
means of distinctions, the paucity of sound distinctions by using phoneme-
sequences for words, and the linearity of word-sequence by taking it as a
projection of a partial order on words (which is retained in writing even
though there a lattice representation would be possible).
Other universal properties are by-products of the linguistic system rather
than being themselves historically formative factors: such are the capacity of
a language to be its own metalanguage, the capacity for cross-reference, the
i: Against the claims that computers can be ‘taught to think’, publicized under the name of
Artificial Intelligence, see Roger Penrose, The Emperor's New Mind, Oxford Univ. Press,
Oxford, 1989.
possibility of classifying words structurally, and the emergence of a math¬

ematical object (6.3).
Another relevant universal property of languages is their intertranslat-
ability. The meaning-ranges of words (their operator-argument selection),
and the allusion based on phonemic similarity, and some reductions, and
also any paradigmatic requirements, are all not directly translatable unless
the two languages in question happen to be similar in the given respect.
However, the inter-language similarities of the selections and meanings of
many words, the modifiability of a word’s meaning by attaching modifiers,
and the virtual identity of the meaning of the partial order (predication, and
the meanings of constraints defined from it), together make approximate
translation possible between any two languages, within reasonable limits of
length for any needed strings of modifiers.
As to content, it has been seen that language developed as a vehicle of
public, transmissible information. But not all information is or can be
carried by the structure of language. More precisely, language is seen to
contain a deposit of publicly recognizable aspects of experience. Further¬
more, the speaker’s intent in a sentence may be to refer to a unique object or
situation, a token, but the sentence as said refers to a class of referents, a
type. To go beyond this public reference one has to go beyond single
sentences into an additional structural dimension—the double array of
discourse or the word subclasses of sublanguages, all of which capture the
meaning of subject matter as going beyond separate observations. Or else
one has to go to the sentence-sequence structure of argumentation, of
reasoning and drawing conclusions. Sentence structure and word-choice
alone cannot give these additional informational meanings, even as they
cannot distinguish truth, subjective ‘truth’, plausibility, or nonsense. But it
is uniquely language structure that can admit of these major further develop¬
ments.
12.4.2. Systematizing factors
To understand the nature of language we may ask what created the order—
the structure—that constitutes that system, and how this structure came to
be. Syntactic order does not come from phonemes or meanings. The
vocabulary by itself has no intrinsic structure that can be stated: the
phonemic structure of words is not relevant to their syntactic behavior, and
their meanings do not fall into a sufficiently sharp and orderly system.
Thereafter, the formation of sentences consists in imposing a meaning¬
bearing ordering (dependence) of word-occurrences. This dependence is
generalized by aggregating words into classes in such a way that all sentences
have the same dependence formation. The formation of elementary sentences
388 Interpretation
is then repeated recursively on the already-formed sentences, to make the

remaining sentences of the language.
Meaning-bearing orderings beyond sentence-formation are created by
the similarities among sentences in sequences or in sublanguages. There are
also miscellaneous structurings in particular languages, which are meaning¬
bearing only in a secondary way if at all.
These are in broadest terms the contributions to making the order, or
structure, which we see in language. We now ask how these contributions
came about.
In 12.4.2-4 we ask what are the formative factors in language structure
and how they determine its nature. In the first place, the fact that language is
a public activity entails its being much the same both for speaker and for
hearer. In this it differs somewhat from, for example, music, which, while it
is often a public activity, can also occur privately, and even when public does
not have to mean the same to producer and recipient. Within this public
status, language developed not so much for face-to-face communication as
for error-free transmission. This can be seen, as noted, in the development
of discrete elements out of the continuous phenomena of speech, in the
establishment of phonemic distinctions among the complexly differing
sounds, in institutionalizing repetition as replacement for imitation, and in
the public pre-setting of a vocabulary of phoneme-sequences—ones which
are arbitrary (not onomatopoetic) in respect to the word meaning.
Transmissibility without error-compounding requires not only that the
elements of the material to be transmitted be discrete, but also that they be
pre-set in sender and receiver. When the speaker and the hearer are
referring to a set of elements known to both, the hearer need receive only
enough of a signal to distinguish a particular element—phoneme or word—
in contrast to all other elements that could occur there. When the hearer
then transmits (repeats) the utterance, he pronounces his own rendition of
the pre-set (i.e. known) elements which he has distinguished. This means
that both must learn to recognize a set of grammatical elements, primarily
particular phonemes (or phonetic distinctions) and secondarily vocabulary
(morphemes, words), in respect to which they speak and perceive utterances.
It is this public institutionalization that makes the transmission of an
utterance a repetition, whereas an attempt to redo or transmit something
whose elements are continuous or not pre-set is an imitation. It is well
known that all languages have to be learned; and that when a speaker says
something which he has heard he is repeating it, not imitating it (whereas he
would be imitating in the case of a yell or other non-linguistic sound). Those
linguistic phenomena which are not represented by discrete pre-set elements
may indeed carry information (e.g. intonations of hesitation or of exaggerated
matter-of-factness), but they do not enter into grammatical relations to
other elements.
The restriction to discrete elements is uniquely appropriate for error-free
transmission of utterances. There has been much discussion as to whether

language is an instrument of expression or of communication. Rather than
either of these, its structure is suited for an instrument of transmission.
Indeed, expression and face-to-face communication are served quite well by
such continuous phenomena as intonations, phonetic characteristics of
individual speakers, gestures, etc.; and the hearer may garner a great deal of
information from such continuous data, over and above what is grammatically
structured in the speaker’s utterance. But when the spoken material is
transmitted, i.e. spoken by the hearer to a new hearer, and so on down the
line, there is a possibility of compounding errors when the continuous
elements are transmitted. In contrast, when the utterance is characterized
by discrete elements only, it can tolerate a considerable amount of ‘noise’ in
transmission from one hearer to the next without producing error.
The non-grammatical status of the non-discrete, non-repetitional elements
underscores the fact that the grammatical elements and relations that arose
were specially suited to the transmissible information. In these circumstances,
what was primarily communicated was not private feelings and desire, or
interpersonal comments and requests, but statements of information. The
private and interpersonal can in various degrees be said, but only by making
the usual kind of transmissible statement about them (/ am happy or How
are you? or Please).
In addition to these factors in the formation of phonemes and words,
much of the structure of syntax is explicable as developments which were
appropriate and effective for such public communication. This is especially
the case for the main creators of syntax, namely, the massive constraints on
word combination. As noted in 12.3.5, these constraints, which state non-
equiprobabilities of combination, are public standardizations of the public
use of words in combination, whereby first of all the combinations that never
had occasion to be used became proscribed, as an institutionalization of
their non-use. Going further, the fact that certain words (here termed
operators) normally appeared in combination with others was then insti¬
tutionalized into a requirement for those operator words, as depending on
their argument words and so creating a sentence with them.
Standardization of behavior is not surprising in a public context. Already
in the pair test that established phonemic distinctions (7.1), what created the
distinction was not the meanings of the words, nor differences in meaning,
and not even the behavior of the speaker and hearer in recognizing the same
phonemic distinction, but rather the institutionalization of these recognitions,
so that they were standardized throughout the population. It is the absence
of such standardized distinctions in areas other than language that makes the
equivalent of phonemes unavailable for culture analysis and for other fields
where export of linguistics’ phonemic success has been sought. In the far
more complex situation of syntax, the driving-force in standardization is the
adequacy and effectiveness of common use and of its institutionalization as a
390 Interpretation
stable system of communication. This systematizing process may be not only

a codification of custom but also a honing of the behavior as a tool, an
increasing of the efficiency of language as a vehicle for information.
Such efficiency-increasing systematizing seems to have taken place not
only universally for the basic constraints of language, but also locally for less
basic ones. Thus in many languages, very common operators and modifiers
(which are operators in attached secondary sentences) are attached as
affixes to the words (arguments) to which they apply: e.g. causatives and
verbalizers, as the prefix in ensure (with its near synonym make sure);
possessive suffixes on nouns and tense affixes on verbs; also affixed arguments,
as in subject and object affixes on verbs in some languages. In some of the
most common situations, these are conventionalized into required paradigms
for the given language. Another reduction in distance is having a modificatory
secondary sentence interrupt a primary sentence at the word being modified
(3.4.4). In addition, there are widespread standardized reductions in the
phonemic shape of high-likelihood word-occurrences, including the dropping
of words which are reconstructible from the environing words, given the
hearer’s expectations about word combinations.
Beyond the establishment of the dependence relation, which sees a word
as requiring one or another member of the set with which it in fact occurs,
there was a further systematization which increased the generality of the
system and its adequacy for new and changing word combinations. This was
to admit as possible arguments of a given operator A all arguments of
operators whose argument sets intersected the argument set of A. This
procedure creates a partition of the vocabulary into a few operator and
argument classes. We thus move from fuzzy argument sets of each operator
to a fixed set of arguments from which each operator of a given level selects
its more likely arguments. This generalization makes possible the move
from dependence to the dependence-on-dependence of 3.1.2, creating the
mathematical object of Chapter 6. It gives the linguistic system stability and
flexibility in the face of change and innovation.
One other stabilizing flexibility is given to language by the close relation
between the meaning of a word and its higher-than-average co-occurrence
(selection) with particular operators and arguments: the meanings of a word
can be extended or modified simply by extending or modifying their selection
(i.e. what they combine with) without having to exhibit to the hearers the
new aspects of the perceived world with which the speaker is now associating
the word.
At this point it may be useful to attempt an orderly survey of the motive
sources for the shape and change of language.
1. There are initial representational factors: in associating short sound-

sequences with perceived objects and events; in associating a dependence
(of operator on argument) with the co-occurrence of events and relations on
the one hand and objects on the other; in the different but relatively stable
likelihoods of particular operators in respect to particular arguments.
2. There are ongoing communicational factors. These are primarily in the
service of effectiveness; in instituting distinctions (first of all, yielding
phonemes); in serving public consensus, so that speaker and hearer under¬
stand language approximately in the same way; to facilitate transmission
beyond face-to-face; to reduce the length (and possibly, in some cases,
the complexity) of messages without loss in their information. Some com¬
municational changes may also be in the service of social convenience: e.g.
the introduction of more attention-catching forms, as in the periphrastic
tenses; or the avoidance of socially undesired forms, as in euphemisms.
These motive sources may come into conflict with each other, as when
certain reductions create ambiguities (primarily degeneracies) that complicate
the hearer’s decoding of what the speaker meant, i.e. complicate their
consensus.
3. There are ongoing influences of one set of forms on another, within the
forms of language and their ways of combining. These include the develop¬
ment of patterns within grammar, and also the many kinds of analogical
formation, including the cases where word-frequency and word combinations
are extended on the basis of how other words are used.
The third type above may be considered ‘internal’, the first two ‘external’
and in various senses ‘adaptive’.
12.4.3. A self-organizing system
That language is self-contained we know from its containing its own meta¬
language, and from the properties seen in the preceding chapters and in
12.4.1-2. Here we will survey briefly how this condition, though roughly
stable, nevertheless made possible the growth of the complicated systems
which are to be found in language. It will be seen in 12.4.4 that language
structure is accretional, and it has been seen in 12.3.4-5 that both language
structure and language information stay within certain limits. In this sense
we can say that language makes itself, as indeed everything that is not made
or planned by an outside agent must do. The overall self-organizing condition
in language is that on the one hand the existing forms limit what is available
for use, and also to some extent affect (direct) this use, while on the other
hand use favors which forms are preserved or dropped, and in part also how
they change; this includes the ability to reinterpret borrowed and other
external material as though structured by the existing system of the language.
Throughout this study, we have seen that the non-random combinations
of words in sentences are best described by constraints, and that each
constraint is defined on previous constraints, down to the unstructured
392 Interpretation
sounds. Unless—surprisingly—all constraints arose simultaneously, we

have to judge that they came to be as successive stages of structure. When
later structures develop out of earlier ones without being due to external
intervention, we may consider that self-organizing features are displayed.
As most notably in Darwinian, Marxist, and Freudian theories, we find, in
describing what coexists in an organism or a structured aggregate, that
further and explanatory regularities are reached if we ask what processes
form the observed interrelating phenomena.
(1) The self-organizing process begins after the existence of some (arbitrary)
stock of words. In phonemics we see a structure that developed around its
existing elements. Phonemes could not have existed before words; hence
words must have originally existed as sound-sequences without phonemic
distinctions. The development of phonemic distinctions among these sound-
sequences has taken place in all languages, apparently, and has made
specific distinctions among words. In so doing, it has created fundamental
features of structure: discreteness of the ultimate elements of language
(since a difference is a discrete entity), fixedness of word-form, and repetition
as against imitation in transmission of speech (and writing, in contrast to
pictorial messages).
(2) The chief process (in 12.3.5) begins after the existence of word
combinations, which were presumably at first casual and unstructured. The
institutionalization of the actual non-occurrence of combinations into a
restriction on combining creates a structural requirement (dependence) out
of a meaning-related customary behavior.
(3) The reduction and zeroing of words whose occurrence is predictable,

originally also a meaning-related customary behavior, became an item of
structure as particular domains of words came to be reducible in particular
environments.
(4) In various languages, particular restrictions on word combination

created new word classes and structure, again largely by developing meaning-
related customary combinations into requirements. This is seen, for example,
in particular domains of transformations, as in the verb-adjective distinction,
and the English auxiliaries. It is also seen in the requirement of certain small
subclasses of words (often in affix form), such as the time, person, and
number affixes of conjugations and declensions, and in requiring the subject-
repetition under the auxiliaries.
(5) At some point language came to predicating about sentences much as

in the existing (elementary) sentences the operators predicate about their
word (N) arguments. For this, the same dependence that characterized the
structure of the elementary sentences was extended to the further operators
in respect to the elementary sentences. This created second-level operators
and sentence expansion.
(6) This extended dependence invited generalization from dependence on

specified word lists to dependence on the dependence properties of words,
creating a mathematical object at the core of language structure, as a stable
criterion for that structure and for further self-organization within it.
(7) Although the dependence structure would appear to be just an
abstract construct, which constitutes a description of regularities in the
speech events of the real world, it had sufficient reality in people’s word-
choice behavior so that new combinations, which were not born out of this
dependence, were formed on its analogy. Analogic construction, which is
not only apparently universal but also very prominent in language, thus
extends the domain of the dependence relation, even where it was not the
sentence-forming process on the underlying elementary words. In addition,
there is evidence that sentences and sentence-parts which were not formed
by the dependence relation (borrowed, reduced beyond recognition as in
goodbye, proverbs carried over from past grammar, etc.) can be reinterpreted
as resultants (or exhibitors) of dependence.
(8) The later regularities are not within the sentence, which has already
reached a flexible stability, but within subsets of sentences, such as sentence-
sequences (ch. 9) and referent-restrictions (or, equivalently, restrictions in
communicationa! situations, ch. 10). The well-formedness of sentences and
the structural conditions for sentence-sequences, as in proofs in logic
and mathematics, are special examples of requirements that institutionalize,
and in this instance codify, the common behavior found to be effective for
discourse in those fields.
This brief survey has shown a few self-organizing features: the develop¬
ment from differences to explicit distinctions (1); the institutionalization of
customary behavior into requirement and structure (2, 3, 4, 8); and the
extending and generalizing of relations (5, 6, 7).
In all, we have here various examples of how structure (and, in some
cases, the impression of purpose) arises out of non-structured (or non-
purposive) events—as was also seen for the alphabet (7.1.2), and as is a
known issue in the theory of evolution. In some cases, the events create not
directly the structured forms, but the conditions for structured forms: thus
the establishment of phonemic distinctions created the discreteness of
phonemes and words which made it easy for the relations among these to
form classes of them.13 As to the phonemes, the roles and physiologic forms
of the vocal organs (vocal chords, mouth, etc.) allowed particular variations
and combinations of sounds which yield particular cross-classifications
(‘phonetic patterning’) of the phonemes. In syntax, parallel events on
similar domains can create subclassification, as in the bundle of reductions
13 Events can also disturb or modify structure, as in phonemicization, when a phonetic

change in an allophone (a sound which is, in a particular environment, the phonetic value of a
phoneme) can make it become a different phoneme.
394 Interpretation
that creates the English auxiliary verbs, where some of the reductions acted
on can, may, need, etc., while others acted on can, may but not on need, and
soon. And the utility of repeating a grammatical relation between particular
words in successive sentences of a discourse can lead, over the whole
discourse, to a double array of its sentences.
Furthermore, the accretional property, whereby events act on the resultants
of prior events, has in many situations the effect of preserving the structure
created by the prior events, and even of strengthening its importance by
providing substructures within it—in some cases creating the effect of a
direction of development. For example, it has been seen that second-level
operators act on first-level operators (hence on the elementary sentences
created by the latter) in the same way as the first-level operators had acted
on their zero-level arguments in forming those elementary sentences (and
comparably when the second-level operators act on second-level operators).
This recursivity establishes a single structural process for all sentences.
Thereafter, transformations are able to act in the same way on all sentences,
in terms of this common structure, thus strengthening the importance of the
structure. A different kind of example is the extension of prepositions from
their original argument-requirement status to many other syntactic statuses.
Each such extension was probably fostered by all the previous extensions of
the prepositions.
We thus have a positive feedback, in which the formation of structure
favors accretions—including subsidiaries—of various kinds to that structure,
all of which further establishes the structure.
The further development of language constraints—in discourse structure
(9.3), argumentation (10.5.4,10.6), and sublanguage (ch. 10)—has a property
relevant to this survey. For while all these further structures are clearly
fostered by what is being talked about—by the nature of discourse, argument,
and subject matter—it can be seen that the apparatus which is here developed
is not something new unrelated to the language structure so far, but is rather
a set of further constraints in the spirit of language structure and within
openings for further meaningful constraints that had been available in
language structure.
12.4.4. An evolving system
We now consider the motive sources which, given the existence of words and
structures as in 12.4.3, directed the many steps of structural development,
and affected their survival. It will be seen that these steps have features of
evolutionary processes.
(1) Toward efficacy of communication. First, many of the developmental
steps increased the effectiveness of communication and transmission of
words and then of sentences. When words developed from being sound-
sequences to being phoneme-sequences, they became more distinguished

one from the other, and their transmission (as repetition rather than
imitation) became more error-free. The use of words in various combinations
permitted more meanings to be expressed with a given stock of words.
When word-combination likelihood is institutionalized into word-class
dependence, with the unused being now excluded, we have a facilitating
(and limiting) contribution toward expressive and communicative effective¬
ness.14 This is seen also in various later developments: in the fixed (word-
subset) domains for many reductions; in the replacing of custom by require¬
ment (including the case when the absence of an affix means its opposite or
complement, as when absence of plural means singular); in the creation of
word subclasses for science sublanguages by excluding other words from the
subclass's environment; in the requiring of specified sentence-structures and
sentence sequences for proof in the language of mathematics.
Communicative and transmissional efficacy is also gained from abbreviation.
The grammatical reductions abbreviate certain words when no information
loss results (3.2.6, 3.3). They thus reduce the redundancy in a sentence, but
at the cost of complexity in decoding and at the cost of adding a constraint—
if one first constructs the source sentence and then reduces it.
There is also a question of efficacy in the encoding of a sentence by the
speaker when he makes the sentence, and in its decoding by the hearer in
recognizing it as a sentence. Encoding advantages may perhaps be seen in
the survival of exceptionally frequent forms (e.g. the complementary am,
are, is, was, were, of be), and also of common idiomatic phrases (e.g. For
God’s sake!) which are gestural in their semantic character and in their
inertness in respect to the grammar. (Such possibilities of survival do not
apply to how words are affected by sound-change; the latter affects the
pronunciation of phonemes rather than the use of words.)
Decoding advantages are seen in the reduction of distance between an
operator and its argument, as when the operator becomes an affix on its
argument (child-like from like a child, remorseful from full of remorse). So
also when secondary sentences which modify a word are inserted after the
host word in the primary sentence, as an interruption, rather than being said
after the primary sentence (where they would need an address-laden pronoun
to identify their host, 3.4.4).
There are also encoding and decoding advantages, in the long run, when
the grammatical apparatus becomes as simple as it can be. This is seen in the
possibility of a least grammar (one or more, 2.2-3), and in the non¬
development of a complete addressing system (5.3.2). It is also seen in the
fact that the dependence relation (3.1) is characterized not by complicated
differences among probabilities but by the simple contrast of zero versus all
positive probabilities. And it is seen in the fact that in the base form of many
14 Presumably the immediate advantage of facilitating the communication outweighs, in the
ongoing evolution of language use, the long-range disadvantage of limiting the communication.
396 Interpretation
languages, words have fixed argument-requirements (syntactic classification,

G65) rather than having one classification in one occurrence and another in
another. (The many English words, such as walk, which occur both as nouns
and as verbs are not a counter-example: for each word, one status is
derivable in a regular way from the other, G66.)
There is also an encoding and decoding simplification when grammatical
processes take place as soon as the conditions for them are satisfied, and not
after later additions or processes take place. Reductions can be shown to
take place, in the course of making a sentence, as soon as the two words
between which the high-likelihood condition is satisfied are both present
(G207). Also, the selectional limitation of a word, whereby it has higher-
than-average frequency only with particular co-occurrents, holds mostly
between words and their immediate operators or arguments (G5).
Major systemic simplifications were gained at relatively small structural
cost when the (first-level) operator-argument relation in elementary sentences
was extended to further (second-level) operators acting on the sentence (or
on its operator) as argument, yielding expanded sentences; and when the
domain of referents of words was extended to include words and word-
occurrences, yielding cross-reference (pronouns) and metalanguage.
The generalization of dependence, once both first-level and second-level
operators existed, to being a dependence on dependence, meant that
operators were not defined by specific lists of argument words. This left
room for words to be dropped or added, or to change in meaning (by
changing their selection, or by adding modifiers), and thus made the basic
dependence system flexible and stable in the face of language change and
development.
Finally, there is the use of analogic construction and the reinterpretation
of non-dependence utterances as pseudo operator-argument sentences.
These extend the dependence structure over virtually all utterances, and
thus simplify the carrying out of grammatical variations on them.
(2) The stopping of developments. A telling feature of language is that the

universal structure-making processes can stop when further development
would add little to communicational effectiveness or to information capacity.
For example, the number of positionally successive arguments which operators
have stops generally at two, although a few three-argument operators may
be found (e.g. in English put, attribute). Here, given enough two-argument
operators acting under further operators, one can get the semantic effect of
//-argument operators, to any //: e.g. He placed the book so that the book was
on the table for He put the book on the table. Similarly, there are only two
levels of operators: those that act (depend) on words that depend on
nothing, and those that act on words that depend on something. There
seems to be no clear case in natural language of dependence on dependence
on dependence (e.g. a verb that can act only on such verbs as act only on
verbs). Without the second level, language would have only elementary
sentences. But by having second-level operators act on second-level operators,
we can get the effect of higher-level operators. The reasons for the length
and depth of arguments stopping at two may thus be not very different from
the evolutionary advantages in biology that let sex differentiation stop at
two.
In respect to the stopping of developments, it may be relevant to note a
similarity and difference between the evolving of language and biological
evolution. In language, communicational ease and distinguishability can be
considered to correspond to the success-in-the-environment in the ‘survival
of the fittest' model; and developments toward such efficacy may be expected
to go no further than the environmentally imposed needs. In contrast, the
old question of why some evolutionary developments seem to go on beyond
what is needed for survival in the environment can be met in biological
evolution by taking into account a second ‘cutting-edge’ which differs
somewhat from the demands imposed by the external environment: namely,
the effect of intra-species competition, e.g. for territory or in mating
selection, and the effect of co-evolution between species or between individual
behaviors within a species (e.g. in animal communication, 12.4.1). In this
latter cutting-edge, the favoring of particular ones among individuals already
changed by favoring up to that point can continue in one direction without
obvious brakes. This situation does not arise in language, except possibly in
such favorings as for constructions or word uses that are more attention-
getting than the current usages. Language therefore does not seem to have
runaway directions of development.
(3) Lack of plan in detail. Although the development in (1) above may
give a global direction to language, many detailed developments are less
consistently directional. This limits the regularity and simplicity of the
system. For example, many formations are not fully carried out. The
differences among the sounds (‘allophones’) which are complementary
alternants of a phoneme are in part regular over various similar phonemes
and in part unique; similarly for the morphophonemic alternants of a word.
Reductions are also in some cases similar to others, and in some cases unique
(e.g. in zeroing to come under expect).
At any one time, various structural developments are incomplete, and
some remain so. Special constructions such as the auxiliary verbs can be the
product of cumulative centuries-long changes in form and in restrictions of
use; but not all of the words have all the characteristic properties, so that the
subclass remains a fuzzy one. Regularization is a tendency, not a plan or a
law, and it can stop short at certain points (e.g. at some of the most
frequently used suppletions such as is-am-are) or at half-way stations (as
happened in the development of do toward being a pro-verb). Most generally,
there is the conjecture of Henry Hiz that every word may have a unique
398 Interpretation
choice of transformations in whose domain it belongs,15 supported now by

the discoveries of Maurice Gross on the unique constellations of grammatical
and transformational properties for individual French verbs.16
There is also, in various languages, a scatter of required reductions, going
against the general property of optionality of reductions. And various
languages and language families have grammatical complexities at various
points, in ways that are obviously not essential for language as a whole and
that have no measurable communicative import for the particular language.
In the light of all these irregularities in the grammar of any given language,
it is not surprising that in describing the grammar of a language, or in
analyzing particular constructions, we may find things that have non-unique
analyses—that can be fitted into the grammar, or into the general theory of
language structure, in different alternative ways. This does not mean that we
should not analyze language in terms of specific processes, or that there is
unclarity about the effects of each type of constraint.
Another kind of non-plannedness is the under-utilization of structural
potentialities due to cost or inconvenience of use. Thus, as has been noted, a
complete addressing system for sentences is available both in the partial
order of dependence and in the linear order of speech, but only rudimentary
bits of addressing are used (as in former-latter referring to parts of an
utterance). Instead, the needs of cross-reference are met, for example, by
the roundabout but simpler addressing of interrupting one sentence (after a
given word in it) with a secondary sentence that begins with the same word
(the basis of modifiers, 3.4.4), and by leaving certain pronouns, such as he,
to be ambiguous for the hearer as to the unaddressed antecedent. Also,
language can provide grammatical grounds for factoring certain parts of its
vocabulary into selectionally more independent word-components (e.g. He
made the recruits walk ten miles for the 'causative' He walked the recruits ten
miles), but again such possibilities do not carry through regularly. Many
languages also place a two-argument operator between its two arguments,
although various ambiguities would be avoided if the operator were either
before or after all its arguments (as in Polish notation).
In the same spirit, languages are not fully adequate for all informational
purposes. Typically, each language has some features of information, e.g.
certain time-relations, which are difficult to express in it, especially in a way
that would distinguish them from other possible interpretations. In Endish,
aside from various difficult distinctions in tense, there is no tense in nominal-
ized sentences (i.e. when a sentence is the operand for a further operator):
one says His being there yesterday sufficed, with no past tense, though
H. Hiz, ‘Congrammaticality, Batteries of Transformations, and Grammatical Categories’,

Proceedings, Symposium in Applied Mathematics 12, American Mathematical Society 1961
pp. 43-50.
16 M. Gross, Methodes en Syntaxe, Hermann, Paris, 1975, and Grammaire transformationelle
du franqais: Syntaxe du verbe, Larousse, Paris, 1968.
periphrastic tenses are used to circumvent this in English (12.2, 12.3.4).

Also, English (differently from some other languages) lacks a third-person
imperative: one circumvents this by saying Let him go there, which puts it
into a second-person imperative with an informationally irrelevant let. (Also
a first-person imperative in Let's go.) Many languages lack the form and
meaning of a three-stage comparative: no John is older than Mary is older
than Tom, though one can say (not using the comparative form) John's age
exceeds the excess of Mary's age over Tom's. In many languages, too,
the speaker’s knowledge of the antecedent of a pronoun cannot be com¬
municated to the hearer: in John phoned Tom yesterday; I know because I
bumped into him, the speaker knows to whom him refers, but he cannot tell
the hearer without cumbersome additions.
(4) Accretions. The structure of language is descriptively accretional:
phonemic distinctions, words, word combinations, the dependence relation
in word combinations (which makes elementary sentences), the extension
from dependence on words to dependence on dependence-properties (which
makes a system for complex sentences). The whole system is a resultant of
successive contributory departures from equiprobability. The system is
constructive in that virtually everything is reached by constraints acting on
previous constraints, i.e. built on the resultants of previous constructions.
But it is also accretional in that the previous constructions are virtually never
lost; they retain some presence in the later resultant. Thus the second-level
operators re-enact the process of the first-level operators, and act on the
first-level operators, leaving the constructions of the latter (on zero-level
arguments) intact.
The reductions are shape-changes within the underlying sentence: the
changed—even zeroed—words are grammatically still present, and their
dependence structure is retained (or paraphrastically transformed) in the
reduced form. (In addition, the optionality of most reductions leaves the
original sentences still in the language.) Hence, when we say that a sentence
is derived from another, we could better say that it contains that other. Thus
the expanded sentence, / hope he will go, contains not just the words of He
will go but also the grammatical relations and meanings of the independent
He will go. And so in reduced sentences, l expect John momentarily, from /
expect John to come momentarily, still contains the to come of John is to
come momentarily, since momentarily is selectionally inexplicable unless the
to come is still present even if in zero form. The accretional character is also
seen in the fact that one can so define the later operators that they never have
to alter or correct the information in their operand sentence. In particular,
as noted above, John will not go can be derived not from not on an asserted
John will go (which it then denies) but from / deny, parallel to l say, on a
composed but not yet asserted John will go; I deny acts here in place of Isay.
In this way the source is not nullified or corrected but is added to, gram¬
matically and informationally, by every added word.
400 Interpretation
More generally, the dependence-on-dependence relation is generalized

from the direct dependence relation, but without eliminating the latter:
within each language in each period, the argument-requirement of each
operator can in principle be given as a list of words. Somewhat similarly,
transformations act on expanded sentences in the same way they act on
elementary ones—and this without affecting how they act on the elementary
ones themselves. Also, the later constraints on sentence-sequences, in
conjoined sentences (9.1), in the double array of discourse (9.3), and in
proof structure (10.5.4), leave the internal structure of the component
sentences unaffected.
It should be noted that languages have ongoing changes now too: consider,
in English, such local ongoing developments as the tense-like used to (/used
to go there) and going to (I’m going to go there), or the more systemic trend to
make the ‘perfect' (e.g. has gone) into a past tense, to free the passive from
its completive aspect, and to put tense-indication into nominalized sentences.
Changes of this kind can in the long run have an effect on the general
structure of the language.
(5) Form and function. We have seen that when the whole structure of
language is analyzed into its least constraints, we obtain a step-wise con¬
nection between each contributory constraint and each contribution to the
information in the utterance (11.4). In consequence, the development of
form and the development of informational capacity go hand in hand. There
are some similarities here to the general, phenomenological, relation of
structure and function, especially in living organisms but even elsewhere
in the physical world. In biological evolution, structural changes arise
independently of any function, by mutations, but their ‘survival' depends
largely on the contribution of their engendered physiological activity or
organismic behavior to the survival and procreation of the organism in w'hich
they are located or of individuals with similar genetic equipment. In this
way, structure and ‘function' are evolutionarily related.
More generally, one might raise the well-known question as to how a
direction of development or evolution gets built up in a theory that seems to
speak of arbitrary changes. In the case of language the attempted changes
themselves may not be arbitrary, but may be due to any of several, perhaps
conflicting, tendencies (simplification, colorfulness, precision, etc.), although
no such general sources of change had been definitely established. The key
supplier of direction, however, may be that each change operates on
something in the language as previously changed, so that for example once
dependence has been institutionalized any change in word combination
would have the effect of creating further detail (likelihood change, trans¬
formational shift of word use, etc.) rather than cutting across the previously
established structure.
In language, the types and sources of change are studied in historical
linguistics and in some sociolinguistic investigations; some changes may
arise from the regularizing pressure of the existing system (12.4.2-3). In any
case, the survival of the changes depends in part on their contribution to the
informational and communicational efficacy of the linguistic system. In this
way, the relation of structure and function, which originates in the form-
content correspondence of 11.4 and 11.6, is maintained through the history
of language change.
In (l)-(5) above, we see processes characteristic of evolving systems. The
whole is of course very different from biological evolution, the more so since
language lacks essential properties of organisms, such as necessary death,
descendants, and speciation. The source of variation is not mutational or
‘random' in the case of language; in language we have to explain not only the
survival-value of a form, but also its source—how it came to be. (The
sources of variation would be considered ‘random’ if they are unrelated to
what affects survival.) Also the driving-force or ‘cutting-edge’ of evolutionary
change is different, though with suggestive similarities. In both cases, a
variant could take hold primarily when it satisfied certain external conditions.
In biological evolution, as noted above, one can distinguish on the one hand
natural selection’s species-survival which includes increased opportunity of
having descendants for the individual or for others in his group (some of
whom may be expected to have much the same as his genetic equipment),
and on the other hand mating-favoring properties which also increase the
chances of having descendants, but which are less limited to what is favored
by the environment, and indeed may be in general more unlimited.
In language, a slightly comparable distinction exists between informational
efficacy and competition within an informational niche. The efficacy criterion
holds that survival of a linguistic form (phonemic distinction, word, word
class, construction, reduction, etc.) is favored if it is more efficacious in
communicating information—including here the cost to speaker (especially
in encoding) and to hearer (in decoding). This efficacy can be understood as
success in eliciting a response appropriate to the speaker’s message; this
formulation skirts the subjectivity issue (in animal communication, is the
‘message’-issuing ‘speaker’s’ cry or posture subjectively communicational,
or is it just an expressive behavior which constitutes a stimulus eliciting
a particular response?). In the competition for niche, sentences containing a
particular form may be favored in a particular language community due to a
variety of properties: they may be clearer, more forceful or contrastive,
easier to say or to understand, more comparable systematically to related
sentences, etc.—all in a never-ending series or escalation of innovation.
The changes due to efficacy are for the most part long-range, universal
and basic for language, more regular with very few exceptions, and may
present the long-range picture of a progression. The changes due to niche-
competition are relatively short-range, and particular to individual languages,
language families, or periods. They are less closely bound to—and limited
by—informational efficacy, and if some have a directional character, it is
402 Interpretation
rather like cultural trends, cumulative to the point of constituting the

‘genius’ of a language, or its ‘drift’.
To all this must be added that the definability of a metalanguage and of
cross-reference within a language and the double array of discourse structure,
the specializing of science sublanguages, and the defining of logic and
mathematics as linguistic-type systems, all show that major informational
capacities can still be evolved out of the sentence-making structure.
Throughout, we can see that the structure of language relates in part to its
history, not only in the specific relation of language change to language
structure but also in the fundamental relation of the structure to the
evolution of language.
More generally, while language is a subset of general cultural behavior, it
develops under the shaping force of having it measure up against what would
be effective in communicating information about the perceived world. Both
of these conditions give the impetus and direction for the origin and
development of language. Different factors thus shape the development of
language than that of other behaviors, and occurrences of language are so
readily distinguishable that they can be studied separately from other
events, to see their systemic character.
12.4.5. What is needed for language to exist?
The properties of the self-organizing and evolving systems discussed above

show language to be a self-contained stable system, resting on no prior
framework, not even on a prior space in which its events are located, and
creating in itself the metalanguage that can describe it. However, whereas a
metalanguage that describes a system can be created after the system exists,
a metalanguage that forms or generates a system would have to exist before
the system does. We have seen that the latter is impossible in the case of
language because the metalanguage of a language is itself a linguistic system
(2.1). In accord with this, we have seen that the non-equiprobability of word
combination exhibits the elements and their combinabilities even in the
absence of a metalinguistic generator. Then in the absence of a metalinguistic
generator we understand the various departures from equiprobability,
which together comprise the total non-equiprobability, as being each formed
by a particular constraint. We thus see language not as a single object or
whole system, but as the cumulative product of constraints on word-
occurrence, which could arise out of a language-less situation, and which
developed in the further constraints of Chapters 9 and 10 above, and in the
course of historical changes (such as are studied in historical and comparative
linguistics, and in linguistic change).17
17 In considering the developing use of language both in the species and in the individual, it
may be noted that the descriptive stages of sentence construction, which provide the apparatus
This picture is of some relevance if one considers the early development of

language, about which virtually no evidence exists. There could of course
have been a single source, developing with the development of early man
from some point onward, with the secondary (and often less regular)
constraints of particular language families arising later in separated com¬
munities. However, the self-organizing and evolutionary character of almost
all the structural developments suggests that the universal constraints in
language could possibly have arisen independently, or quasi-independently,
in separated communities.18 In either case, it is plausible to think that the
development of sentencehood—the dependence relation—accompanied a
major social or productional development in early man.
In any case, it was the arising of constraints that constituted the develop¬
ment from words and unstructured co-occurrences of words to sentences
and to the existing language structure, in a situation where there was as
yet no language having that structure. These constraints are more regular
than the word combinations they produce, because complexities and non¬
regularities can be relegated to their particular domains.
Furthermore, while language structure is certainly unique, the constraints
that create it are not unique except in detail. They are not so different from
some constraints in other aspects of human life. For example, grammatical
relations, such as being the subject or object of a verb, exist nowhere but in
language. But to have the occurrence of certain things depend on classes of
other things is a kind of dependence that may not be unique to language—
yet such is the constraint that makes particular word-occurrences into
subject or object. Also, language is a structure of demands: some com¬
binations are in it and therefore correct, while others are not in it and
therefore wrong. But these demands, as was seen in 12.3.5, can be seen
as institutionalizations of less demanding and more naturally-occurring
customary uses in word combining. There are demands that are unique to
language; but they can be reached by institutionalization of custom—of
customary combining of particular bits of behavior. That bits of behavior
should be combined, and in customary ways, is not unique. And for custom
of On (operator on sentence) and O00 (operator on two sentences), can be considered to

parallel the grammatical apparatus needed to describe neural processing and recall of language
material. In neural processing and recall, the need to recognize, classify, and locate each
sentence parallels the O,, operator class, and any processing of associations between two
sentences parallels the 0„0. In the spirit of Harry Stack Sullivan's view that what is unconscious
is what has not had to be raised to consciousness, the apparatus of bringing memories into
awareness and recall involves the ability to process the memories with an apparatus paralleling
O,, and Oao\ and it could be facilitated by initial sentential deposit of the memories. This is in
contrast to involuntary recall after brain injury, where no specific association (0„„) seems to be
involved.
18 This picture of how language could have developed does not mean that a different system
could not have arisen instead to fill the same functional niche. Language is not like the integers,
which have properties independent of people knowing them, and for which we cannot imagine
alternative structures to have been able to develop in any human community.
404 Interpretation
to be institutionalized, whether for convenience of living or for social

control, is no novelty in culture and in social organization. The effect of this
view of language is to avoid both the non-explanation of saying that
language is just something sui generis, and also the reductionism of saying
that language is no more than some simpler communicational system.
The lack of uniqueness is seen not only in the constraints that create
language but also in the knowledge or capacities needed to speak and
understand language, in the absence of any metalinguistic direction. The
constraints of Chapters 3-5 indicate all the kinds of departures from
equiprobability which carry the grammatical structure and the information
of a language. They are ordered in respect to each other, so that we know
which has a more pervasive effect, and so that we can avoid counting any
constraint or its effect twice.
We can therefore count all the kinds of constraint in a language; and in any
sentence we can count all the constraints that go into making it. We have to
count, for a particular language: (a) its phonemes (its phonemic distinctions
and their allophones); (b) a finite stock (up to some limit) of those phoneme-
sequences that constitute words or morphemes at a given time, with the
main meanings associated with each; (c) the argument-requirement for each
word (in most cases fixed for all base-occurrences of the word); (d) roughly
the selection for each (i.e. with what operator or argument words the given
word has higher-than average co-occurrence, yielding its environment-
specific meanings; this is primarily a set of word-pairs); (e) the main
reductions in phonemic shape, with their conditions for application and their
domains (which may be individual words, or all words in a given position);
(f) also the linearization of the operator-argument pairs or triples, and the
alternative linearizations.19 While no one will count all this, it is clear that
the amount of knowledge, and—what is more important—the mental
capacity required for using and sequencing the particular kinds of knowledge,
can be within the ordinary mental capacities of human beings.2" Indeed, all
of these contributions to the sentences of a language are specifically learned
by each speaker and hearer (over a given vocabulary), first simply by long
19 This catalog is not entirely the same as the least-grammar form of characterization
required in the present theory. And neither of these is entirely the same as the way an infant or
an adult learns a language, or the way language came to be historically, or the way language
information is processed in the brain.
2(1 How the brain processes both the speaking and the recognizing of speech, and what
specific capacities of the brain are involved, is still little known. Clinical research in aphasia and
neurological research have both shed some light on the problems. For relevant evidence from
aphasia, see in particular: A. R. Luria, Traumatic Aphasia, Mouton, The Hague, 1970; M.
Critchley, Aphasiology and Other Aspects of Language, Arnold, London, 1970; K. Goldstein.
Language and Language Disturbances, Grune and Stratton, New York, 1948; H. Head,
Aphasia and Kindred Disorders of Speech, Macmillan, New York, 1926. In any case, what has
been seen in the present book about the relevant structure of language suggests that it is not a
direct reflection of the structure of neural action in the brain. Rather, language may be
understood as a particular relation of its own—constraints, above all dependence—among
neural events and mechanisms.
exposure to the constrained word combinations in the language, and perhaps

later by metalinguistic (grammatical) description.21
This Gedankenexperiment shows that there is no single ‘language capacity’
involved in the ability to use a language, but rather that language use is a
learned activity drawing on various otherwise-existing mental capacities. It
also shows that perhaps none of the mental capacities involved are solely
linguistic. If one of them is, it is the association of particular words (or
sound-sequences) with certain perceptions of the world; but this is the one
language-type capacity that animals have too (though to a far more limited
extent, and normally as a direct response to a stimulus).22
In spite of the mass of detail and of complex resultant forms, what
characterizes language is a small set of relatively simple and regular con¬
straints which created phonemes and sentences and further constructions on
the basis of words, and which constitute the set of behavior patterns which
a person has to acquire in order to use a language freely. No special
structuralism or capacity is required; in any case, it would be inexplicable
how such could come into existence before language itself existed.
:i This survey indicates what is needed in order to use a whole language. To understand an
individual sentence, all that is essential is a recognition of the phonemes of the language and the
argument-requirements of the words present, hence (a) and (c) above. This is seen, for
example, in the fact that any finite syllable-satisfying phoneme-sequence of a language is
acceptable as a sentence if we can take an initial sub-sequence in it to be some name or newly
created technical object and the remainder to be some action or state of that person or object;
preferably either the initial segment should end in the phonemes of the plural suffix or the
second segment should end in the phonemes of a singular tense suffix.
22 The distinction made in 12.4.1 between human and animal language does not mean that
human language requires a totally different suigeneris principle, nor does it deny the possibility
of evolutionary continuity from animal communication (cf. D. R. Griffin, The Question of
Animal Awareness, Rockefeller Univ. Press, New York, 1981). The fundamental new relation
(dependence) among ‘symbols’ (words), and the far greater stock of symbols (with their varying
phonemic shapes), can presumably all be learned within the general structure of the so much
more massive and complex human brain. It should be noted in this connection that there is no
present evidence as to how old is human language as a syntactic system. Since a large stock of
words may have existed for a long time before the syntax-creating dependence relation
developed, such evidence as there is for the existence of Broca’s area in the brain of early man
(R. E. Leakey and R. Lewin, Origins, Dutton, New York, 1977, p. 190) may mean only that a
large stock of words was then growing. Syntactic language, with its facilitation of complex
sustained thought, might have developed only much later, perhaps at the time of the rapid
growth of material culture and co-operative production.
.
‘
Index
Abbreviation 371,395 Adverb 100, 183

Aberrant situations 108 on performatives 142
About 60-1,102,312,332,334,368 Adverbial modifier 241
Aboutness 271,334 Affix 80,82,85-6,98,105, 107, 178,181,
Absence: 232,242,244,249
of expected 209 Affixal reduction 107
of a form 340 Affixless 55,108
of required 209 After 143
Abstract 333,351 Agreement 151,178,182
noun 36,229,327,349,372 Algorithmic presentation 151,251
Abstract equivalence 311 Alignment 259-69
Abusde langage 136,258-9 of repeated subclasses 271
Acceptability 200,235 Alliteration 342
available for grammar 202 Allophone 46,397
different 204 Allusion 342
grading 201-2 Alphabet 48,378
independent of truth 201 extensions 170
qualified 202 original 170
Acceptable 70,88,242 phonemic 168,176
less than components 256-7 Alternant 33
not due to meaning 201 free 179
Accidentality 90 see also variant
Accretion 22,26,344-7,355-6,361,391, Ambiguity 22, 88,99, 110, 116,219,282,
394,399 340,371
of information 334 grammatical 229
Acquisition of language 404 lexica! 229
Acrophony 169-70 Ambiguous 46,106, 127, 157,215,244,250
Active 12,197 n-way 229
Adaptive 391 Amount of choice 335
Addition 209 Amount of information 356
Address 248 Analogic 372,396
Addressability 343 processes 79, 111, 114
Addressing: Analogized 200
apparatus 103,120,130 Analogy 27,84, 112,121,206,236,241,328,
capacity 39,104 352,359,361,374,391,393
system 117,120, 132,248,375,396,398 on model 115
Adequacy 365 Analysis 9,35,48
Adjective 8,76,85,93, 101, 112, 121, 178, correction 38
181,213,233,241,352 Analyze 119,151,250
order 89-90 And 54,69-70,83,88,93,129,141,212,256
Adjectivized 197 not only binary 144
Adjoin 183 4nr/-residues 101
Adjunct 99,101-2, 183, 198 Animal communication 309,317,382,405
similarity to centre 199 Answer 140
single word 212 Antecedent 39,69, 101-3, 184,252,374
Adjunct string 154,83-4 fixed 129,131
Adjunct-host relation 102 not fixed 129,132
408 Index
Antecedent position 130,307 180,197,235,237,241,273,282,334,371

unaddressed 398 difficulties 107
Anticipatory contamination 97 has all sources 240-1
see also slips; Spoonerisms least constraints 116
Any 142,244 more regular 115
Aphasia 326,404 not all in language 240-1
Applied mathematics 146 not all information 111
Arbitrary change 400 notin 107
Arbitrary signals 148 with information 95,108
Argot 79 Basic structure 250
Argument 55,57-9,118,233,369 Beat 118
first 55 Before 143
free 60,234 Behavior 166
free for operator 111 of word 37,39,45,1^2,230
multiple kinds 113 Behavioral 60
order 100,349 Best fit 35,291,300
primitive 17 Beyond constraints 122
Argument dependence 292 Bi-sentential 70,302
Argument-indicator 99. 106, 112,372 predicates 212
Argument-repeating 108 Binary 144,215
Argument-requirement 17,24-5,59, 180, compositions 158
“ 182,280 Borrowed 239
classes 153 Borrowing 80,112,359
partition 119 Boundaries 118
set 337 of a science 298
Argumentation 257,303,337,351,378,387 sentence 154
constraint 303 word 154
Art 27,309 Brain 369
high 312 see also neural
language of 312,317 Broad selection 110
Artificial intelligence 386 But 9,10,244
At 114,326 demands two differences 256
As though 236,241,246,250,353 By 11-14
Ascending chain 268
Aside-intonation 102 Can 74-5,93,107,113,239,379
Ask 85,91 Cancellation 190
Asked: Cannot occur 105
what can be 243 Canonical form 259,268,291
which positions can be 141 Capacities needed to speak 404
Aspect 66,75,347 Carrier 108,113
Asserting 369 Case 121,245
Assertion 153,197,332,375 Case-endings 60,240-1,364,379
Assimilation 236 Categorial grammar 21,28,60
Associative 144,157 Categories 86
Asyntactic 239,359 Categorization 340,347
Attention-getting 397 Causality 303,349
Authentic 315 Causative 80,83,87,114,177
Automaton: Centerstring 154,183,185
cycling 190 Chain of equivalence 268
Autonomy of syntax 146 Change 24,27-8,36,41,44,47,53, 121, 150,
Auxiliaries 74-5,93,107,113, 116,239,241, 327,360,365,378,380,396
243,364,371,379 in form 232
Awareness 369 marginal 155
Axiomatic sentences 308 ongoing 361
in process 364
Back-formation 361 rate of 151
Bar 44,310 in science 297,304
Base 54-5,78-80,105-8,111,115,157-60, sources 391
Index 409
and structure 364 Combinatorial 282
in the structure 363,393 likelihoods 360
in word use 107,379 simplicity 37
Choice 335 Combine 180
determinate 84 Combining-relations 16, 145
likely 84 Comma 89,102, 164,251
Chord 310 Common knowledge 71,87,89,257
Circumlocution 343 Communication 358,367-8,374,389,391,
Circumventing grammar 74,374,386-7,399 395
Cite 81,125,128 effective 389
location 120 efficacy 401
Cladistic 64,361 efficiency 352
Class 45,57,181,391 Communicative 317
Classes 56,181 Community 26
of more-constrained 146 Commutative 144
new 122 sub-sequence 215
by similarity 117 Comparative 93,244,374
Classifiability 152 Compare structures 245
Classificand 78 Competing 107
Classification 18.34,37,45-6, 73, 117,163, Competition for niche 402
317 Complementary 46,330,331,397
criteria 16 variants 86,167, 170,175
multiple 34-5, 106, 119, 181,228,241, Completive 76
244,396 Complexity 27,251
of vocabulary 106 in composition 311
Classificatory 62,67 of the environment 268
Classifier 66,75,78,81-2, 87, 108, 114,274, measurable 245
281,289,298,326,377 measure 112
Classifying 351 non-informational 311
Classifying sentence 78 Component sentence 9,70,89,160,183,
Clause 237 196,213,247
Clusteiing 45,293,330 differentiating 224
multiple 34-5 unchangeable 106
Co-argument 59,67,77,80,156,396 Component transformations 12
Co-evolution 383 Composite member 45
Co-occurrence 18,60,370 Composition 365
Co-occurrence relation 17 of text 270
Co-occurrent 41,58 Compound 96
change of 239 noun 87,92,207
overlapping 64 word 82,85,101
Codify 393 Compounding-permutation 110
Cognate 365 Computability 119, 151, 185,195,355,361
Cognition 21,23 Computer 273,288,300
Coherent 330-1 analysis of sentences 386
environments 351 processing 20,23,112
range 330 programs 45,127,250,308
selection 332 simulation of thought 386
Colloquial 81,91-2,94,275 Computing the derivation 111
Color 181 Concatenation 145,157
Combination-constraint 303 Concepts:
Combinations 17,37 about language 60
catalog 32 abstract 151
listing 53 Conclusion 296,301,337
not all occur 117,174,385 from preceding 270
notequiprobable 32 Concord 151
not excludable 116 Configurational pressure 250
of reductions 120 Conflation 364
unusual 67 Conjoined sentences 213
410 Index
Conjugalional subclasses 121 Constructive formulation 308

Conjugations 3,120,154,340 Constructivist methods 145
Conjunctional sentences 88 Contained sentence 115
Conjunctional use 85 Containers 66
Conjunctions 54,66,100, 102,122, 181,233, Contaminations:
241 intentional 97
new 120,280 Context 30,148
Connective 81,337 narrow range 77
Consequence 28,302,303,308,344 variety 77
Consensus 332,391 Contiguity 44,45,53,55,100.150
Consistency 270 correction on 152
Consonant 169, 177 Contiguous 55, 118-19, 178, 182-3, 190
Constituents: to operand 45
marginal 183 to string 183
Constrained 292,302 Continuous 44, 167. 182-3,389
sequence 302 Continuously differing grade 147
Constraint 53-4,63,74,104-5,117,173, Contradictory 67
230,236,278-9,289,298,301,307,309, Contrast:
311,317,334-9,352,358,364,368,370, in pronouncings 166
378,403 Contributions:
as constructs 54 to non-equiprobability 33
on combination 4,168 to simplicity 36,38
common 109 Control 384
on constraints 345,399 Controversy 284,287-8
effect of 4 Convention 165
first 7 Conventionalization 25, 113,250,275.338,
how satisfied 364 345,264,367,374,379
in the making 337 Correct 403
non-repetition 63 Correspondence:
preserved under reduction 96 language-world 353
primitive 230 theory 354
unique 54 with experience 348
on word choice 255 Corresponding position 70,268
Constraint system 117,356 Count nouns 121,241,340
Constraints 21,25-7,33,37,40,48,53, Cover 59,247
245-6,255,361 Create:
act on each other 53 conditions 394
beyond sentence 255 dependence 392
combinations of 15 distinction 339
and communication 23 further detail 401
counted 404 mathematical object 393
fewer 245 new 334,338,347,351,369,372,
limit combinations 153 379
meaning of 23,350 order 388
non-informational 338 sentence 366,390
not unique 403 structure 392
regular 403 Criteria 294
successive 392 Cross-reference 103,128,133-5,243 276
survey 5 291,307,336,371,387
Construct 54,62 apparatus 131
Construction 105, 185 see also reference
absence of 249 Cross-referent 382
local 237 Cultural change 297
preserved 244 Culture 377,402
utilized fully 244 Culture-analysis:
Construction-difference 230 not phonemic 390
Constructive 16,36,43,45, 118-19, 145, Cumulative 397
182 Cut 48.147
Index 411
Data 6,15,30,43,48,53,62,68,237,272, cost 112

303 non-optional 107
corpus 19 preserves meaning 111
and theory in sublanguage 301-2 retains dependence 243
word choice 200 vs. similarities 351
Datum 366 Derivational source 231
Deaf 317 complex 112
sign-language 28 Derive 237
signing 309 Derived 55,81,105-6,111,251
Deciphering 307 from the base 109
Decision procedure 49,268 grammatically 85
Declensional subclasses 121 historically 85
Declensions 3,154,340 means contained 399
Decoding 342,395-6,401 as though 121
Decomposable 367 Description.
Decomposition 183,215-17 incomplete 248
differences 222 Descriptive 49
of words 330 gain 180
Defined: order 366
by relations 164 Descriptively prior 97,349
within itself 382 Development 25,132,371
Definite article 364 of form 345,400
Definition 127,206,257,324,330 further 394
ordered 59 of in formation 400
Definitional: of language 369
order 366 ongoing 374
relation 78 runaway 397
Degeneracy 22,46, 88,97, 106, 110, 116, stopping 396
219,339 structural 353
local 250 Diachronic 47
from metasentences 259 Dialect 27
see also ambiguous Dictionary 326,381
Degenerate 244,251-2 definitions 42,71
Degrees of freedom 168,245 Direct discourse 124,137
Deictic 133, 144 Direction 400,402
Demonstrative 133-4,242,381 Disagreements 289
Denotation 324,330,385 Disambiguate 224-5,250,252
Denumerability 157 Discontinuous segment 152, 182
see also enumerable Discourse 25, 189,259,272,283, 387
Depend 55 analysis 278
Dependence 27,54-5,62,79,105,114,234, grammar 280
235,247,276,292,332,349,360,368, structure 18,298
382,400 types 201
on dependence 17-18, 28, 54, 58, 154, Discovery 72,171,268
272, 276, 291-2, 350, 353, 358, 378, Discrete 41, 147, 167,360
391 cuts 44
of later sentences 269 distinctions 44
of operator on arguments 59,65 elements 388
order 60 meaning-bearing 384
properties 58 Discussion 289,298,302,337
recursive 367 Disjunction 140
satisfied 93 Dissimilation 98
Dependings 247 Distance 44,89, 182-3,185,248,358
Derivable 60,360 fixed 103
Derivation 7,11,13, 15,22,34-5,38,49,55, grammatical 130
68,79,85,90,213,231,233,237,239, linear 130
280,376,399 not defined 118
admits further reduction 110 Distant 100
412 Index
Distinction 147,387,391 analyses 361

of sounds 164,168 examples 357
Do 93,397 Equiprobability 93,104
Domain 13,38,45-6,48,86, 111, 121, 152, departures from 3,22-3,30-3,49,53,61,
205,214,229,244,340,361,364,371 105, 145,245,345-6,353-6,402
extended 393 non- 33
fuzzy 54 Equivalence relation 158,268
monotonically decreasing 160 Equivalent in discourse 269
of reduction 64 Erroneous 315
in respect to domains 205 Error compounding 148,388
restricted 206 Etymologically related 112
special 199,242 Etymology 11,35,85,228,324,325
syntactically recognizable 205 Euphemisms 79
Domestication 49,359,364,380-1 Every 9, 15
Double array 269,283,298,302 Evidential operators 302
Drift 250 Evolution 171,345,378,393-4,397,400,401
Driving force 401 of language 402
Duration 90-1,206 Evolutionary process 395
Durative 63,66,75,90,113,121 Exaggerations 79
Dynamics 345 Except 9, 10
Exception 121,151,159,360
Each other 68,89 Exceptionless 42
Early development 403 Excising 185
Early man 403 Excluded 56,57
Eclectic 361 Exhibitable 306
Economics 46 Expand in same way 378
Efficacy of communication 395 Expanded sentences 310
Elementary sentence 114-15, 159, 190,214, Expansion 190,196,199
228-9,273,334,367 Expect 68,89,328
expandable 154 Expectancy:
from words 231 high 245
Elements 38 Experience:
higher order 146 aspects of 79
ultimate 165 of the world 344,347
Emotion 313 Explanation 42
see also feelings Expression 389
Empirical 49,54,56 non-informational 313,344
Encapsulated 108, 152,294,297,307 system 313
Encoding 395-6,401 Expressive 317
Entities: features 22
additional 235 Extend 69,324
stock 118 environment 326
Entry of operator 157 use 206,372
Enumerable 226-7 Extension (of referents) 48, 320,351,384,
see also denumerability 396
Envelope 272 Extra-grammatical 313
Environing 174 Extra-linguistic 134
sentences 201
Environment 37,40-1,46,86, 175,278,325 Face-to-face 134
sentential 32 Factoring of vocabulary 330,375
Environment distinctions 198,325 False emotions 315
Environmental constraint 336 Falsely 67
Environments: Falsification 301
all 11 Favored 211
set of 179 Feedback:
similar 181 positive 394
unique 9 Feelings 314,321,377
Episodic: see also emotion
Index 413
Feminine 178 low 91
Fictitious arguments 349 relative 245
Final sentence 303 of word pairs 235
Finitary grammar 150 Frequency within dependence 155
Finite 43,244 Frequent 87, 368
information 227 Front position 102,106
sentences 150 Fronted secondary sentence 103
speakers 150 Fronting 99,100, 102,211
structure 118 blocked 108
users 118 Frozen expressions 54,67,239
Finite-state 171,214 Function 345
First principles 243 Functional 26
First-level 59, 114, 118 Functor 28,60
Fit into the discourse 270 Further operators
Flexibility: future arguments 108
stabilizing 391 preserve selection 111
Flowchart 308 Fuzzy 48,324,358,397
For 89 domains 53
For each one 140,142 membership in language 53
For... to 98 set 75
Forgetting 367 subclasses 243
Form 31 subset 75,108-9,113
behavior of 37
and content 295,342-3,354 Gender 121,154,178,240
and function 400 agreement 80
and information 242,300 Generality 390
Formalism 312 Generalization 305,350,396
Formation 365 Genius of a language 402
of grammar 380 Gestural 396
of language 369 Gesture 343,387
Forms: Give 84,363
behavior 39 Globally degenerate 250
co-exist 47 Going to 94
future and past 37 Government 151
Formula 287,291,303 Graded 113,380
structure 295 sentence 215-16
types 300 Grading 200,208,225
Formulaic language 294,296-7,305 different 204
Foundations of mathematics 145 more than one 203
Fragment 66,244 not of preservation 204
Fragment-nouns 337 preservation 206
Free: same sentence 203
for the operator 234 syntactic properties 206
variation 330 Grammar 8,34,118, 121, 128,241,273,278,
word 80,105,107,239,249 279,332
Frequency 65,77,329 common 160
absolute high 84 criterion for 21
of combination 327 fuller 244
higher than average 76 of language pairs 227-8
low 78 not included 44,147,377
in reconstructed sentences 78 redundant 93
relative 77 sublanguage 285
relative to operator-argument 78 traditional 8,163, 180,182, 189, 199,228,
zero 56,62 237
Frequency differences 337 word-class 198
exceptionally high 84 Grammatical:
high 87,113 analysis 189
high total 84 categories 63,249,331
414 Index
character 249-50 Imitation 148, 165,378, 388, 389

construction 118,121,293 repeating 176
environment 326 Immanent material 315
innovations 372 Immediate constituent 182,189
juncture 81 Immediate knowledge 315
meaning 25,321,333,389 Immunology 284
processing 403 Imperative 85,122,139,153,241
relation 99,268,358,363 first person 139
rule 242 third person 139,399
status 122 Imperfective 75-6,121
system 6-7 Imprecision 297
Grammatical relations 44,80,97, 100, 104, Impredicative paradox 375-6
173,190,333,339,403 Inadequacies 374
preserved 99 Inborn 165
recast 83 Inclusion of propositions 225
reconstructed 97 Incomplete 397
with same words 196 Increment 231
Grammatically ambiguous 159 Indefinite 66, 108,110,114,189,307,326,
Grammatically conditioned 273 327,331
Graph 225 noun 75,87
of set of propositions 226-7 Independence of the elements 152
Grid 230 Independent 36,49
Grouping 99 Indicator 13,106, 108,236,240-1,251,339
Indirect objects 241
Have 74,363 Individual:
He 38-9,66,88, 104, 133,340,398 languages 26,121
Hearer 22,26-7,30,46,49,84,88,97-8, same 252
130, 132-4, 148,237,251-2,339,342, Inequalities 203
371,388,390,395 oflikelihood 155
supplied by 84 preserved 203-4
Here 134 Inference from environment 84
Hierarchical 189 Infinite regress 114,277
Hieroglyphic 169 of my 138
High-expectancy zeroing 9,92 Infinitive 241
Higher entities 163 Inflectional 249
Higher than average 62 Information 21-2,25,27-8,49, 54,56,68,
Historical 107,112,121, 144, 180,229,239, 78.80,85-6,91, 106, 108, 144,273,294,
241 299-300,307,342,351,354,369,389
change 380 all 105-6,131,371
linguistics 401 all in base 108,122,240-1,334
replacement 242 amount 335,356
source 81,110 choice 335
History 81,101,107,131,364 contribution 79,84
Homonym 35.64, 125, 135, 165, 168, 175, covert 344
177, 179,252,309,325,331 and form 335
Homonymy 97,106,294 expression of 313
Host 80-2,86,92, 100, 102-3, 107 of a field 294
Host string 183 finite 227
Human 337 and dependence 24
N 63 framework 300,346
from constraint 345
/ 133,138 grammatically created 334
Identifiable 84 interrelated 283
Identity 158 location 288
Idiom 43,67,88, 115, 183,329,336,345,359 low 63,84,86,91,370
Imagination 344,350,385 no 79,83,86-7, 138, 140
Imbedding 45, 115, 119 not all in language 387,398
see also nesting not in base 110
Index 415
not in language 354 Intonation 85, 101,102,106,140,168,202,

preserved 80,112,340 217,233,367
processing 283 Intransitive 132,187,244,337
representation 284,296 Invention 171,355,367,368
retrieval 23 Irregular 9,109,241,246-7
of science 286,338 Irregularities 92,362,371,398
of structural features 131,346 due to phonemes 180
structure 112,288,356 Isolation 174
substantive 22,27,83.105 Iteration 157,227
use of 353
vs. meaning 334
Jocular 97,353
Information processing 23,283,288,297
Join-semilattice 247
Information Theory 16,23,79, 84,154,
Joke 109,133,202,244,328-9
356
Judgment differences 168
Information-bearing 26, 174 linguistic 239
Informational capacity 18,23,336,345 Juncture 82,117
change 296
contribution to 244
efficacy 401 Kernel 159,216,240
facilities 347 Kernel-sentence 217,240
Kinship 328
interconnection 298
niche 401 Knowledge change 289,308
structure 294
Informationally maximal sentence 227 Language:
Initial 169,172,174,178 of art 345
segments 174 capacity 405
sentences 302 community 47
wh- words 102 de novo 387
Innovation 372,379,402 differences 24,47,227,249
Insert 184 invention or convention 355
Insertion 185 makes itself 392
Institutional factors 352 manipulation 317
Institutionalization 27, 342,378,389, 395, must be learned 389
403 pairs 227
Institutions of control 378 similarities 351
Integers 321,403 and thought 21
Intension of a word 324,384 without thought 383
Inter-sentence relation 282 Language-information 334
Interaction 383 Language-independent structures 228,
Intercalation 184-5,239 259-60,291
Interchange 100 Language-less situation 403
Intermediate 112,140 Language-like systems 24,378
non-extant 110 Language-wide property 49
Internal conjunction 286,293 Language-world correspondence 350
Interpretation 60 Languages 154,219,351
Interrogative 85,122, 140,153,197,232,241 Lattice 217
disjunction 140 differences 224
explanation of 141 representation 387
from non-interrogative 141 Lattices, set of 217
Interrupt 131,398 Laws of a science 274
Interrupting sentence 117 Learned by each speaker 26,405
Interruption 45,89,101,103, 106, 119,211, Least:
396 constraint 168,291
relevance 103 description 366
not alternative linearization 248 grammar 38,49,70,93,153,228,237,346,
not merging 120 396,404
of sentences 247-8 structure 119,119-20
Intertranslatability 387 Least-restriction grammar *95
416 Index
Length 147, 167 List 55,58,146,200,205,243,245,249,281,

double 147 308,323,364,365,372,374,396,400
multiple 168 Listing 53
Length permutation 101, 105, 110,211 Literary devices 317,341
Less restricted 168 Locate 103,132
Lexicalshift 36,116,327 antecedents 104
Lexicography 328 Location 39,81, 124, 125,130, 182,276,297
Lieder 314 words for 128
Likelihood 78-9,180,235,291,305,307, Logic 153,306,350
347,358,367-8,377 of classes 306
change 61,327 and language 386
of combination 366 of propositions 306
constraint 61,338,353 Logical:
difference 281,365,373 form 305,308,346,384,386
to distant argument 63 relations 141,144
grades 92,105,275 Longer sentence 334
grading 62-3,68-9,71, 108, 153 Lost:
in grammar 61,63 from language 364
high 63,68-9,89-90,129,139,339 from the vocabulary 107
high total 79
higher than average 79 Marginal 31,32,39,109,152,168,179,200,
highest 76,78-9,91,131,358 201,229
inequalities 61,76,158,308 Marginal sentence 105
low 73,218 Marker 98
lowest 76 Markov process 171
normal 77 Mass nouns 121
of pairs 61 Material implication 301
positive 56,61 Materials and Methods 289,298
preservation 63,68,158 Mathematical:
presumed not zero 115 description 151,153
reconstructed 68 formulas 227
relative to operator-argument 77 formulation 53-4,239
similarities 68 logic 305,373
stable 61 object 18,146,154,238,281,306,366,
unrestricted 65 374,387
zero 57,73 properties 15-17
zero grade 74 structures 18
Likelihoods in the world 153,351,352 Mathematics 32,153,155,291,301,303,
Likely 70 305,312,372,374
not 69 translatability 311
Limits 376,392 Mean what we want 332
Linear 252 Meaning 8,11,28,31,37,40-1,58,66,
order 59,128,149,368 69-71,132,73,77,80,84,87,96,104,
projection 45,98,387 106,112-13,122,125,137,148,155,165,
representation 99 173,179,180,198,201,212,228,235,
speech 97 244,292,305,307,310,314,321,326,
Linearity 44,149,376 343,358,376,380
Linearization 98,99, 105, 111,116,235,247 ability to have 236
alternative 82,99,104 affects selection 327
different 157 beyond experience 344
freedom of 101 change 65,122,324,327
normal 97,99 coherent 325
same 119 of combination 324,329,368
Linearly ordered 156,168,215,221 and combinations 64,65
Linguistic system 280,290,305, 306, 308, in combinatorial constraints 341
387 common 109
Linguistics: constant 328
structural 163 of constraints 322
Index 417
contextual 325 Meta-science 338,349

in co-occurrence 65,325,329,341,348, Metalanguage 32,40,125, 153,272,274,278,
381 283, 306,336,347,352,359-60,387
correlation with 42 external 31,40,279-81,374
as criterion 42 Metalinguistic 39,40,60,67,109,123-4,
dictionary 239 138,201,382
difference 24,83,86,165,281 capacities 123,135,353
different 31,204 instruction 126, 130
different from source 110 interruption 130-1
discrete 384 meaning 341
distinctions 325 operator 139
in environment 325,329 on phonemes 141,144
extended 391 sentence 130,227,257,276
fixed combination 360 statement 125,127
general 86 Metaphor 36,42,61,64,67,80,83-5,
grammatical 25,239,333 113-14,179,200,243,325-6
implicit 80 Metaphoric 62
independent of language 41 Metascience 286,290,293,299
in isolation 323,329 Metasentence 258
language-based 352-3,385 adjoined 257
language-borne 321 zeroable, as known 257-8
less specific 84 Metasentential 65,67,76,81,100,103,106,
list 41 159
main 327 Metatheory 278
measure 329 Meter 44,118,342
new 385 Method 30,48-9,54,164,180
not in language 343 Middle 122
not in phonemes 321 Minimal sentence:
not predictive 41 within sentences 196
not used 346 Mistakes 92
opposite 340 Model 6,121
organization 332 Modes of expression 313
outside discourse 134 Modificand 82
preserved 79,229 Modifier 78,82,85,89,99, 101-2,104,116,
ranges 228,325,351 189-90,243,285,293,371,387
by referent 381 close 89
same 116,198 merges in primary 119
from selection 324,328,331,336,384 Momentaneous 63,75,121
similarity 64 Morpheme 36,86,149,163
of source 110 bound 171,175,178
specialization 115 boundaries 173,180
starting from 73 end 172
still present 116 free 175,178
substantive 82-3 marginal 180
under different operators 115 variants 245-6
unpredictable 109 Morphological systems 154
unstatable 330 Morphology 16,27,55,105,116,182,229,
unusual 67 236,239,359,367
usefulness 42 same 120
vs. spelling 136 Morphophoneme 27,33,86, 121,126, 176,
of word class 348 235
Meaning-bearing arrangements 229 replacing phoneme 176
Meaningless 74 Morphophonemic 39,46, 167,240,339,397
Measurement 205-6 instruction 126
Memories 403 statements 125
Mention: systems 154
for word 136 Morphophonemics 37,339
Mentioned 137 Moved to before 82
418 Index
Moved to the antecedent 103 Non-ambiguous 225

Moving: Non-contiguity 177, 182, 185
before further operators 106 Non-extant 107,110
of tense 100 form 93
Multilinear 44 intermediate 114
Multiply classified 280,372 Non-hereditary 171
Music 22,44,309,388 Non-occurring 67,69,107
Musical phrase 310 Non-planned 398
Mystical 316 Non-predicational utterances 369
Non-randomness 33
N 57 exhibits structure 402
human 63 Non-real-world 353
Name 289,324,333,366 Non-unique analyses 398
of phonemes 124 Nonce 200,244,328,341
of word-sound 123 Nonsense 73,201,301,344,347,353,373,
Natural language 274,291,297,305,306, 377,383
308,312 intended 327
Natural mapping 216 Nonsensical 88,353,369
Necessary 135,387 Not 9-11,77,109,140-1,242
to language 124 Not in language 379
Negation 64 Notation 39,303,306,309
Neighbor 44,103,117,179 derived 304
Neighborhood 182 systems 20
Neighboring sciences 299 Note-lengths 310
Nesting 149,243 Notes 309
of adverbs 119 Noun 36,57,99,178,233
see also imbedding Noun phrase 181-2
Neural: Now 134
action in language 404 Nuance 211,229
processing 403-4 Null extension 330
see also brain Number 178, 181
Never 56,181
almost 56,69 Q 57
New constraints 338,352 O...0 59
New grammatical 364 Observable 62,68
in sublanguage 293 Observing the regularities 275
New grammatical distinctions 352 Obsolete forms 251
New grammatical relations 351 Occurrence:
New irregularity 371 never 379
New meanings 352 nonsensical 379
structure for 384-5 Occurrence-dependences 18
New members in class 280 Occurs alone 58
New semantic power 373 Omission 71-2,132,208
New situation 378,381 evidence 84
New word classes 351,392 position 73
New words 365 On 57
Next 103,358 O,, „ 59
neighbor 172 One-dimensional structure 270
to each other 103,131 Ongoing changes 400
Niche 121-2,363 Only 9-10, 181,244
of language 401
Noise in transmission 148
Nominalization 14,34,198,212,241,374 Ono 57,71
Nominalized 12,13,34,178, 197,212,233, O,loo 59
327,333 On 57
sentence 34,36,121,364 Oon 59
Nominalizing 86 Ooa 57,70,102
morphemes 86 prefers a repetition 256
Index 419
Open features 376 Parentheses 99,217
Operator 17,55,57,59,80,85,107,233, Parsimony of analysis 38
240,366,369 Partial order 8,28,32,59,82,83,97,100,
binary second-level 158 102, 105-6,115, 128, 146, 156,234
carries arguments 106 constraint 65
further 198 items of 99-100
higher 71,132 position 104
immediate 156 preserved 111
level limit 396 represented linearly 97,99
lower 71,132 same 98
non-ordinary 141-2 Partial ordering 215-17,308,347,350
on an operator 238 Partial transformation 158,208
on sentence 85-6,153 Partition 48,59,75, 156, 164, 181,208,216,
unary second-level 157 338,379,380,390
on verb 153 Passes 171
zeroed 238-9 Passive 11,13,80,83,101,110, 120, 197,
Operator occurrences 59 199,203,205-6,211-14,232,242,244
Operator-argument 28,102,292,310,332 Passive adjective 12,12-13
construction 244-5 Past 364
form 111 Patterning 27,391
framework 359 Peer-pressure 374
relation 105,107,111 Perception:
source 234 selective 355,384-5,391
tangential to 246 successful 355
theory 232,237 Perfecttense 12-14,110,329,363
word pairs 235 Perfective 75-6, 121
Operator-indicator 106, 108 Performative 91,137,336
Opposite meaning 326 Period 185, 197
Opposites 326 Period punctuation 54
Optional 104-5,160,211,240,371 Periods within a science 299
product of required steps 110 Periphrastic tenses 79,235,363,381
Or 140-1 Permanence 89,365
demands one difference 256 Permitted 184-5
not only binary 144 Permutation 12-13,44,82,94,96,99,101,
Order: 110,185,209
co-arguments 97 apparent 82,211
descriptive 24 length 87, 100
operator-argument 97 limited 243
time 25 process 82
Ordered: Permuted to before 83
description 3-4 Person 119-20
traces 216 Phoneme 33,37,46,154, 167,176,309,366
word-entries 159 in syntax 245-6
Ordering of adjectives 89 as type 48, 123
Organism 26 unrelated to meaning 123-4,135,148,173
Oriented 217 in word 146,245-6, 322
Origin 382 Phoneme distinction 123
of language 365 Phoneme replacement 177
Outside syntax 245-6 Phoneme sequence 172,387
Overproduce 244 ambiguous 258
intercalated 177
Pair test 165 into sentence 258
Paradigm 122,237,249,331,340,346,352, new 171
390,393 Phonemic:
Paragraph 269 analysis 164
Paraphrase 27,235,333,340,381 composition 173,180,309,358
Paraphrastic 60,105,107,117,197,339 differences 166
power 231 distance 81
420 Index
distinction 165,367,388 property 38

information 83-4 structure 394
neighborhood 167 under transmission 367
segments 166 Primary sentence 102
sequence 127,274 Primitives 163
shapes 176 Prior not lost 399
structure not syntactic 388 Prior science 40,282,299,349
successors 172 Private 389
system 166 Probability 155
Phonemicization 379,380,393 positive 53-4,354,384
Phonetic: relative 354,384
assimilation 98 zero 73
neighborhood 175 Procedure 163-4
segmentation 166 effective 171
similarities 167 for analyzing 251
Phrase 280 Process 365,392
Pictographs 169 additional 235,246
Place" 84-5 immediate 396
Plan lacking 397 hereditary 171
Plural 85, 120-1, 175,237,239,242-3,307, Processing 273
331 grammatical relations 252
agreement 239 Progression in discourse 271
Pluri-morphemic 116 Pronoun 39,42,69,81,88-9,101,104,129,
Pluri-sentence structure 255 184,213,241,244,307,337
Poetry 44 antecedent 374
Pointing 40-1 free 132,252,375
Polish notation 99,217,398 Pronouncings 165
Popular art 312 Pronouning 92
Position 88, 100, 102 Pronunciations 136
corresponding 88 Proof 301,308,337,350,393
Position changes 251 Properties 38
Postulate 56-7 not essential 241
Postulated 63 Proposition 203,215-17
Power to mean 270-1 Propositional form 208
Pre-set 147,148,151,165,174,376,388 Propositional sentences 208
Pre-syntactic 235,239,367, 369,380 Pro-verb 129
Pre-verb 122,235 Provided 280
Precede immediately 102 Public 35,43,349,388
Preceding noun 89 Public information 313
Precision 168 Public institutionalization 389
Predicate 377 Publicly communicable 377
logic 28 Punctuation 250-1
Predicating about sentences 393
Predication 60,155,310,332,333,353,369, Quantification 307
387 Quantifier 15,66,153,242,290,293
meaning 380 Quantity 299
on predication 385 words 372
unparaphrasable 333 Question 77,213,244,367
Predict 31 disjunction 140
Predictability 152 multiple 141
Predisposition 165 wh- 140
Prefix 80-2 yes-no 140
Preposition 36,59,66-7,81-2,84-5,180-1, Quotes intonation 137
233,241,331,379
Prepositional 99 Randomness 53
Present 364 departure from 23
Presentation of information 371 non- 33,238
Preserved 38 Range 214
Index 421
of meanings 324 informational grounds 90

Rare 115 instructions 126
Rationality 315 not arbitrary 237
Realism 354 optional 93, 105, 109,235
Reasoning 315 ordered by conditions 96
Rebus 168,169 partial order preserves 96
Recall 403 patternings 27
Recasting of sentence 109,340 positional conditions 92
Reception: pre- 79
in the arts 317 required 93,108-9,240
in language 317 residue upon 79
Reciprocal 89,109,243 in semilattice 94,115,157,248
verbs 68 simultaneous 95, 156
Recognition 151,241 small list 120
process 130 successive 83,113
Recognize 112,165,250 system 80,112
Reconstruct 250 unusual 67
source 246 variants of instructions 126
word repetition 257 when applicable 95
Reconstructed 68,69,70,72,78,84,87 when conditions are met 93
sentence 242 in word pairs 93
Reconstructible 87,89,339,355,390 word presence preserved 96
in situ 88 to zero 80
Reconstruction 6-7,108,244,361 Reductional:
Recurring sequences 259-69 derivation as though 131
Recursion 43,189-90,385 path 250
Recursive 145.150,394 subset 113
dependence 388 Reductionism 40,404
equations 154 inapplicable to predication 333
expansion of sentence 154 Reductive 49
Recursively 189 Redundancy 356
applied 118 Refer 81, 102-3,119,123,276
denumerable 43,150 to outside the discourse 133
Reduced 55,101,106 to utterance 125
contains unreduced 115 Reference 252,307,360,375
historically 81 see also cross-reference
in situ 82 Referent 15,347,348,351,367
sentence 109,157 choice 353
to zero 88 class 387
Reducible 69,77, 152 not unique object 324
not 73 null 276,330,604
word occurrences 94 same 213
to zero 88 Referential 88,128,234. 242
Reduction 7,49,55,57,79,98, 105, 107, reduction 66
109-11, 155, 158,197,232,234,239,339, relation 81
358,370,378,392,397 situation derivable 104
admits further selection 108 Rcferentially related 103
to affix 86 Regular 45,55, 146, 147, 163,207
attributing to 238 domains 153
before further operators 94 more 176,183
block further operators 108 Regularities 8, 11, 16,30,34, 125,168, 172,
blocked 92 179,180,238,284,313,317,323,346,
conditions 248 352,353,357,360,377
of distance 396 of change 301
domain 76,116 of classes 196
grid 111 not outside operator structure 249
heavily favored 115 not over discourse 196
in situ 81 of wo'rd-choice 308
422 Index
Regularity 54,270 Reshaping 237

of combination 174 Respectively 68
limited 397 Response:
Regularization 68,151, 152,246,291,300 choice in 382
can stop 397 Restricted:
Regularize 107,108,151,366 beyond base 116
Regularizing-pressure 401 least 233
Reinterpret 97,393,396 less 107,232
Rejection 67,79 words 239
Relation among words 369 Restriction 184-5,233,236
Relations 38 on combination 217
among sentences 225 local 249
defined by 164 only 234
Relative clause 41,81,90,101,104, 184, 190, on phonemes 166
199,213,233,243,307 special 233
develop into 103 Resrictive modifier 38,77,90,292
lattice 248 Restructuring 366
restrictive 38,77,90,292 Results 289,292,298
similarity to sentence 199 Rhetorical 304
structure 102 Rhyme 44,118.342
Relative position 104 Ritardando 310
Relative pronoun 81,102,104 Root 117,239,249
Relevant to each other 336,352,353,368 Rule-based 372
Reliability 293
Repeat 165 Same 103,106-7,147,179,180,365
Repeatable 190 environment 330
Repeating the antecedent 375 individual referent 133
Repetition 81,88-9,92-3, 102,110,128, relation 378
132,147,378,388-9,395 word 133
at beginning of interruption 131 Sameness statement 81, 130-1, 133,135-6
fixed position 88 words for 127-8
of preceding 103 Sample 48,272
reconstruction adjoined 257 Say 91,100,106,124,137
reduction of 128 modifiers on 142
in sentence sequences 19 plus adverb 140,141
of a sentence type 270 Sayable 312
in sublanguages 19 Science 18,272,278,372,374
under and 69,70 articles 283,289,302
under Onno 71,72 boundaries 290
Under O,,,, 13-15,71-2 information 20
under O„0 70 language 246,280,291,297,306,330,348,
under operators 128 376-7
Repetition-supplying sentence 126 of relations 307
Repetitionalzeroing 92,104,110 Science sublanguage 76,284,290,297,306,
Replacement 107,153, 190,246,363 338
Represent: SCSCS 99
information 123 Second-level 59,114,118
the representations 123-4 operator on 118
Request 85 Secondary 106
Required 55,60,115,184,249 intonation 89
arguments 111 operator 78
change 240 sentence 40-1,81,90,101-2,142,233
combinations 121 Segmentation 172-3,387
intermediates 110 intermediate 173
modifier 243 into sentences 186
reduction 109,116,398 Segments:
steps 249 expressing distinctions 166
Requirement 369,380,395 Selection 13-14,26,62,64,68-70,73,77,
Index 423
107, 121, 153,176-7, 179, 197,307, Sense 324

329-30,348,351,363,387,391 Sentence 6,17,24,34,53-5,310,358,369
affects meaning 73,327-9,341 bound 197,213
after reduction 238 boundaries 186
ofany 142-3 characterization 228,230
broad 66,79,83-8,280,331 complex 370
broadened 329,364 conjoined 256
broadest 355 created 106
change 79 derived 109
clustered 64,358 expanded 370
coherent 64,135-6,179 family resemblance 117
complementary 74 free 197,213
extension 65,84-5,113 from non-sentence 60
of grammatical constructions 229 from sentence 60
included 75 long 36,98, 100
increase 84 longest 43
intersecting 75 marginal 361
loss 238 possible 146,237,344
narrow 67 recast 117
negative 67,92 short 36,98
non-coherent 113 shortest 57
normal 78 unreduced 371
of phonemes 124,135 Sentence construction 235,247
relations of 66,67 contained 106
similar 74 decomposition 183
stable 173-4 defined by transformation 216-17
strong 67 definition 182
to high operators 86 differences 111
types 65 elementary 216,228
under interrogative 142 existence 186
undernor 142 extended 240-1
unique 64 finite length 150
unrestricted 65 Sentence information 332
of word 135 informationally maximal 226-7
Selectional: lower 106
likelihood 62 not encompassed in 271
niche 121-2 not said alone 241
property 75 partial 224-6
relations 78 Sentence plus metasentences:
subset 75-6 reconstructed 242
Self-contained 378,382 residual 216
Self-organizing 378,391 same 235
Self-sufficient 127-8 self-contained 258
Semantic 60,63-5,70, 79, 89, 101-2, 110, self-sufficient 127-8
121,136,143,158-9,184-5,240,249, Sentence sequence 19,196-7,272,301,307,
270,273,280,283,304,326,331,333, 372,393
342-4,361,369,375,396 Sentence sets 102,112
change 363 Sentence structure 180-1,199,387
criterion 246 few 119-20
factoring 375 successor 344
property 206-7 tranformational 229
Semantics 9,102,110,233,247-8 Sentence synthesizing 250
Semigroup 17, 157, 159 Sentence type 20, 120-2,270,272,278,279,
Semilattice 59, 109,115,156-60,215-17, 292,337,373
221,246-7,290,299-300,334-5,338, unbounded length 144-5
376 unreduced 17
includes reduction 94 Sentencehood 369
not single (for wh-) 247-8 Sequence 18,44,301
424 Index
Sequence (cont.): 132-3,139,168,237,251-2,326-7,339,

of behaviors 383 368,375,387-8,395
of more-constrained 146 Speaker and hearer 89,378,388
of sentences 19 Special constraint 378,379
Sequencing 163 Special structures 371
Set 86-7 Specific restriction 372
closed under operations 145 Specific situations 314
of propositions 105,208,214-15,228,229 Speech 48,149,164,180,357,360,367,387
of sentences 102, 105,208,214,228-9,242 different 366
of transformations 214 linear 235
with one operation 157 Spelling 39-40,166,172-3, 176
Set-nouns 337 Spoonerisms 97,252
Set-theoretic systems 307 see also Slips
Set-vs-memberconnectives 9 Stability 365
Shape: Stable 63,203-4,378,382,396
changes 55,235 Stages:
syntax correlation 8 successive 180
zero 55 Standardization 367
Shortening 81,84-5 ofuse 365
Sign language 317 Statements 125
Signal 383 Stem 249
Similarities 108 Step:
among words 120 acting on step 355
Simile 326 of analysis 6-7
Simple 173-4,396 derivational 35
Simpler 164,174,232,236-7,251-2 large 35
Simplest 179 small 35
Simplicity 36-9,49,134,167,246,297 Stochastic processes 16-17,150, 154,171,
Simplification 37,396 186
Simultaneous 44, 168 Stress 81-2,85-6
Sir Walter Scott 324,330 change 178
Site 296 emphatic 342
Slips: main 178
unintentional 97 secondary 82-3
see also Spoonerisms types 185
Social structure 378 zero 82-3
Sociometric clustering 604 String 181-5
Something 66 relations 190
Song 314 theory 189
Sound 310 Strong verbs 121
unrelated to meaning 149 Structural:
Sound change 396 criteria 369
Sounds outside language 125 details 120
Source 87,92,96,103,106-7,114,140-1, differences between languages 227-8
158,199,213-14,231-2,234,236-7, irregularities 371
239-40,364,401 linguistics 113,196
artificial 242 relations 311
contained 109 Structuralism 405
different 199 Structurally recognizable 105
for words 246 Structure 36,364,383
missing 105 common 233
not nullified 400 and function 401
reconstructed 110 further 372
sentence 15,109 and history 365,402
as though 359 of individual languages 3
Space 44,45,117,150,184,310,402 of language 3
Space-measuring 90-1 for new meanings 385
Speaker 22,26-7,30,41,46,49,97,112, 115, not fully utilized 120
Index 425
from preference 25 composition 308

in sentence set 214-15 operations 105-6
unique 403 order 388
of utterances 184 relations 25
whole 15 theory 144,358
Structure-creating 380 Syntax 24,26,109,303,355,365,369
Sub-sequence 149,172-3,229 independent of phonemes 180
Subclass 200,279,283,285.380 independent of vocabulary 242
in sublanguage 281 not in lists 242
Subclassification 121 readjusts the meaning 381
Subgrammar 279 relation among words 242
Subject and object 337 starts from vocabulary 246
Subject matter 272,278,372,374 structure 16
Subject-less verb 106-7 Synthesis 9
Subjunctive 108,139 Synthesizing 151,250,276
Sublanguage 18,62, 159-60,246-7,272, degenerate 250
279,283,334,387 System 370,382
analysis 288,294,304 two 80
condition 271 whole 37-8
grammar 273,279 Systemic regularities 310
interrelations 282
of science 20 Tabular:
Subset 63 representation 300
of sentences 393 structure 291
of word sequences 145,149 Tabulation of the structure 269
of words 122 Take 84-5,122,235,363
Subsidiary clauses 197 Taxonomy 37-8
Substantive information 335 Teleological 40,345
Substitutability 189-90 terminology 304
Substitution 301 Tendency 397
Successful perception 355 Tense 85,93,100,112-13, 119-21,134-5,
Successive: 141-2,237,243,329,374,380,400
rows 270 absolOte 143
senteces 103 from non-tensed 143
Successor-restriction 172-3 non-temporal use 144
Suffix 80-1,329,333,363 relative 143
Suffix-alternatiori 86 relative to say 143
Suggestions 271 Terminology 378
Suigeneris 24,404 That 98,108
Summary of the argument 269 The 242
Suppletion 330,397 Themes 311-12
Suppletive 136,179,246 Theory 9,17,22,24,27,30,30-1,33,38-9,
source 110 48,53-4,62,108,151,163,228,231,
variant 107 235,239-41,248,277,366,382
Survival 395,397,400,401 constructs 22-3
Syllabic 249 falsifiable 237-8
Syllable 37,169 of language 364,365
Symbols 347,350,355,357,370,383 of non-elementary sentences 228
for the world 351 of syntax 3,245-6
Synchronic 47 Thinking 308
Synonym 86-7,89,107,137,179,259,281, Thought 21
285,309,325,330,331 and information 385
local 68,381 without language 383
in sublanguage 282 Timbre 309
Synonymity: Time 66,84,89,90,120-1,299
lexical 229 modifiers 120
Synonymy 78,291,294,340 order 25,366
Syntactic 342 slice 47,345
426 Index
Token 123-4,241,329,387 Two-verb expression 74

Tone 168 Type 123-4,150,166,241,329,387
Topic 99, 102, 104,211-12,243,283 verb underand 93
Total rarity 77
Trace 7,85,87-8,98, 111,136, 138,209-10, Un- 233
215-16,221,250-1 Unambiguous 160
identical 222-3 language 221
of zeroing 209-10 Unbounded 43,244-5
Transform 73,196-7,208,216, 359 Uncertainty 39,289
Transformable 69 Under-utilization 398
Transformation 11,63,79-80, 87, 105,206, Understandable 127
208,235,273,291,328,335,340,352, unboundedly many sentences 144-5
379,394,400 Ungrammatical 377
decomposition of propositions 209-10 word-sequence 250-1
optional 339 Unicorn 351
sequences 216 Unimorphemic 55,57,101,108
what is preserved 22 Unique 54,64,177, 180, 186,211,216-17,
Transformational: 225-6,397
decomposition 196,228 environmental restrictions 176
differences 208-9 environments 9
lattice 247 informational capacity 155
method 199 to language 18,24
relation 216-17 ordered product 215-16
system inadequacy 230 referent 387
theory 228 structural appearance 249-50
trace 209 Uniquely restricted 179
Transformationally ambiguous 204 Universal 31,39,47,117,118,273,350,
Transformations 55,82-3, 153,196-7 352,365,387
block further transformations 243 processes 250
elementary 210,222-3,230 Universal 26,31
paraphrastic 229 Unlikelihood 67,69
partial 198 Unordered 36,215-17
products of 214 Unreduced 92-3
simultaneous 214 overshadowed 242
successive 214 Unrestricted 160
Transformed sentence 96 Unstable 75
Transitional probabilities 17,32,149,171 Unusual 67
Transitive 87,244,337 Use 363,364-6,368,379
Transitive verbs 14,132 affects forms 392
Translatability 333,353 changing 11
Translate 160,273,315,374 competition 380
Translation 219,228,305,332,335,347,387 go out of 112
structurally equivalent 227 of phonemes 136
Transmissible information 312 preference 380
Transmission 148,368,388 of words 396
Transparent structure 79 Use-mention 124,135,243-4,375
Transposing 81-2,85,93 Used to 363,400
Transpositions 83 User’s perception 251
Tree 156 Utterance 54
Trends of change 371 distinction 176
True 315,377
Truth 38,201,270,346,388 Variable 307,373
of sentences 377 Variant 107,246
Truth tables 301 forms 45
Truth value 303 less restricted 107
preservation 312 synonymous 108
Truth-transmission 308 see also alternant
Two-dimensional structure 269 Variation 312
Index 427
in grammatical relations 197 Word co-occurrence 342
Verb 8,34,36,99,112-13,121,180-1 233 probabilities 146
352
Word combination 332,367
Verb phrase 181 composite 239-40
Verb-adjective distinction 380 new 326
Verbalizing 85-6 useful 368
Vocabulary 151.239,242,246,274,278,289, Word entry 60
292,317,345.351,367,369 factoring 246
closed 281
indivisible 105
differences 249
internal construction 105
factoring 281,376 same 46
maximally independent 109 Word m-tuples 61
regularity 376 Word meaning 23
utilized 244
Word occurrence 55
Vocabulary-syntax system 246-7 Wordplay 344
Vowel 169.177,239-40 Word repetition:
Vowellessness 170 in co-arguments 257
inacrophony 170 reconstructed 257
status 17
Well defined 152,242 Word replacement 65
Well-formed sentence 189 Word sequence 228
Well-formedness 44,190,393 formation of 235
Wh- 81,89,130-1,213,247-8 Word subclass 19,249,395
Wh- word 102 Word subsets 74-5. 108
Which is 89,101 in grammatical relation 255
Who 81 related in discourse 197
Word 80,178,310,357,358,367 Word-components 246
addressing 243 Word-pair relation 78
derived 116 Word-pair repetition 259
end 124 Word-repetition 179
fit into system 144-5 Word-root 177
fitting into sets 16 Word-semilattice 83
how determined 163 Word-size 171-2, 179
individual properties 109 Words about words 123-4
in isolation 336 Words and the world 351
lists 15,18 Writing 30,48, 149, 168, 180,251
new 15,18, 144-5 Wrong 378,379,403
as phoneme sequence 124
same 125,136
Yes 181,233
of the same subclass 78
You 133, 139
sequences 171
whole sentence 115
without phonemes 321,392 Zero 80,83-4,87, 129
from word sequences 108 as morpheme 177
Word boundaries 171 vs. non-occurrence 243
Word choice 231,355,387 Zero argument 209
constraint 255 Zero causative 326
preserved 198,228 Zero grade 74
repeated 255 Zero-level 59,114,118
same 197 arguments 114
same in transforms 196-7 Zeroable 9,68,71,85,89, 127, 138, 140,355
unstable 203-4 not 92
unstated 196 sentence 257
Word class 16,18,20,55,58,64, 118,180, under and 88
272,280,291 Zeroed 69,82,84,86-7,106-7,138
distinctions 27 erroneously 259
inadequacy 199 repetition 70
sequences 198 under 0„„ 88
428 Index
Zeroed word: Zeroing 82,97, 105, 106, 110, 113, 140, 179,
keeps modifiers 96 210-11
meaning retained 326,328 fixed position 89
status taken over 116 operator 238-9
still present 116,399 under and 82-3,93,101
of which is 96

A Theory of Language and Information A Mathematical Approach

Uploaded by

Copyright:

Available Formats

A Theory of Language and Information A Mathematical Approach

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Theory of Language and Information A Mathematical Approach

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What methods are used to analyze language?

What methods are used to analyze language?

A Theory of Language and Information

CLARENDON PRESS • OXFORD

Oxford University Press, Walton Street, Oxford ox2 6dp

Oxford is a trade mark of Oxford University Press

Published in the United States

© Zellig Harris 1991

All rights reserved. No part of this publication may be reproduced,

British Library Cataloguing in Publication Data

Library of Congress Cataloging in Publication Data

Set by Hope Services (Abingdon) Ltd.

This book presents several interrelated matters: a formal theory of syntax;

List of Figures xii

II. THEORY OF SYNTAX

3.0. Constraints on equiprobability 53

4. Language Structure: The System Created by the Constraints 105

4.0. Introduction 105

5. Metalinguistic Apparatus within Language 123

6. On the Mathematics of Language 145

III. GRAMMATICAL ANALYSIS

7. Analysis of Sentences 163

8. Analysis of the Set of Sentences 196

8.0. Introduction 196

8.1.6. Structures in respect to transformations 219

IV. SUBSETS OF SENTENCES

9. Sentence Sequences 255

9.0. Motivation: constraints between sentences 255

10. Sublanguages 272

10.0. Introduction 272

11. Information 321

11.3. How language carries information 334,

12. The Nature and Development of Language 357

12.0. Introduction 357

7.1. Word boundaries in a sentence 172

All figures are from Z. Harris, Mathematical Structures of Language, Wiley-

given science. These methods provide a procedure for organizing the

1.1. A theory in terms of constraints

1.1.0. The major constraints

The third constraint (3.3) on word-sequence makes sentences more

same component word-sequences as certain other sentences which were

and transformations which create a reduced sentence are ordered by the

1.1.1. Example: The set-vs. -member connectives

To give some impression of how analysis in tqrms of the theory explains

2 For the reference G, see n. 4 below.

1.1.2. Example: The passive construction

(1) By on the subject of a ‘nominalized’ sentence (i.e. a sentence appear¬

(2) The -en suffix occurs independently of the subject-object permutation.

which it is attached has become nominalized as an argument of a further

by my neighbor last month; for the expatriation of a generation of writers

1.2. Mathematical properties

1.3. Science sublanguages

Analyzing languages in terms of constraints on word co-occurrence makes it

within a language. If the grammar of a language were stated in terms of

globulin, anti-ferritin, etc., as their first argument (subject) and having

The relation of language structure to meaning was known to many, especially

1.5. The nature of language