Ot Constraints Are Categorical: Scholarworks@Umass Amherst
Ot Constraints Are Categorical: Scholarworks@Umass Amherst
Ot Constraints Are Categorical: Scholarworks@Umass Amherst
ScholarWorks@UMass Amherst
Linguistics Department Faculty Publication Series Linguistics
January 2003
Recommended Citation
McCarthy, John J., "OT constraints are categorical" (2003). Phonology. 52.
10.1017/S0952675703004470
This Article is brought to you for free and open access by the Linguistics at ScholarWorks@UMass Amherst. It has been accepted for inclusion in
Linguistics Department Faculty Publication Series by an authorized administrator of ScholarWorks@UMass Amherst. For more information, please
contact [email protected].
OT Constraints Are Categorical
Author(s): John J. McCarthy
Source: Phonology, Vol. 20, No. 1 (2003), pp. 75-138
Published by: Cambridge University Press
Stable URL: http://www.jstor.org/stable/4420242
Accessed: 25/06/2009 18:33
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at
http://www.jstor.org/action/showPublisher?publisherCode=cup.
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the
scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that
promotes the discovery and use of these resources. For more information about JSTOR, please contact [email protected].
Cambridge University Press is collaborating with JSTOR to digitize, preserve and extend access to Phonology.
http://www.jstor.org
Phonology 20 (2003) 75-138. X 2003 Cambridge University Press
DOI: 10.1017/S0952675703004470 Printed in the United Kingdom
1 Introduction
In Optimality Theory (Prince & Smolensky 1993), a constraint can assign
multiple violation-marks to a candidate. This happens in two situations.
First, there can be several places where the constraint is violated in a single
candidate, as when ONSET assigns two marks to the form [a.pa.i]. Second,
some constraints measure the extent of a candidate's deviance from some
norm. For instance, the constraint ALIGN(Ft, R; Wd, R) assigns three
violation-marks to [(pi.ta)Ft.ka.ti.ma]wd, one mark for each syllable that
separates the right foot edge from the right word edge.
Constraints of the first type are called CATEGORICAL.The majority of
OT constraints that have been proposed are categorical. Categorical con-
straints never assign more than one violation-mark, unless the candidate
'Gradient' has sometimes been used in senses other than the one employed here.
For example: Harrikari (1999) uses the phrase 'gradient OCP' to refer to a set of
OCP constraints distinguished by locality, and not to a single gradient constraint.
Constraints that assess forms continuously (i.e. numerical optimisation) have also
been called 'gradient'.
0 T constraints are categorical 77
assignment of main stress. These are shown to have a similar basis: cat-
egorical constraints on the location of the head foot that are descendants of
Prince's (1983) End Rule. Here again the argument for categoricality and
against gradience comes from Occam's Razor: gradient alignment is doing
no necessary work, its functions having been usurped by categorical con-
straints, which are needed anyway.
The last substantive section (? 7) switches to autosegmental phenomena.
Docking of morphemic features and tones (?7.2) arguably falls under the
same rubric as infixation. Flop or reassociation processes (?7.3) exemplify
the effects of categorical COINCIDE constraints (Zoll 1996). These con-
straints, I will argue, avoid an unwanted typological prediction of gradient
alignment under ranking permutation. Finally, ?7.4 looks at autosegmen-
tal spreading processes. A novel constraint is proposed that combines the
properties of two current non-alignment-based approaches to spreading.
To sum up, the thesis of this article is that the known applications of
gradient constraints in OT can and in many cases should be reanalysed
with categorical constraints. Many of the categorical constraints that step
into this role have been proposed previously; those that are novel here are
independently motivated. Overall, this argument is the natural sequel to a
remark by Prince & Smolensky (1993: 88): 'the division of constraints
into those which are binary and those which are not ... is not in fact as
theoretically fundamental as it may at this point appear'. Gradient con-
straints are not an essential element of OT; they are an imposition on it, as
is apparent once we set out to define what it is that OT constraints do.
2
See Eisner (1999) and Potts & Pullum (2002) for other developments along these
general lines. See McCarthy (2002a, 2003a) for another application of the notion
'locus of violation'.
78 John J. McCarthy
The letter x is mnemonic for the LOCUS of violation. It is the phonological
constituent that the markedness constraint militates against (compare the
'focus' of a constraint in Crowhurst & Hewitt 1997). As noted in ? 1,
categorical constraints may assign multiple violation-marks when there
are multiple loci of violation in the form under evaluation. It is, then, a
general fact about categorical markedness constraints that when two can-
didates candl and cand2 contain equal numbers of loci of violation of
some constraint C, C assigns an equal number of violation-marks to candl
and cand2. The principal thesis of this article is that the theory of CON
limits all markedness constraints to schema (1).
Schema (1) requires a couple of remarks before we go on. First, (1)
requires all markedness constraints to be formulated negatively; they are
prohibitions rather than admonitions. This is consistent with generally
accepted practice, and has even been argued to be necessary (de Lacy
2002). Second, certain constraints have a symmetric character that makes
the choice of a locus of violation arbitrary, as Maria Gouskova and Alan
Prince have pointed out. This is true, for instance, of the *LAPSE con-
straints in (31) below, which are violated by sequences of unstressed syl-
lables. This arbitrariness, though, is only a problem for the analyst seeking
to translate various previously proposed markedness constraints into a
consistent format like (1). A theory of CON is not obliged to make that
translation easy or even fully determinate.
Gradient markedness constraints, of which alignment is the principal
example, cannot in general be stated within the strictures of (1). Here is a
definition of alignment, expanding on McCarthy & Prince (1993a: 80),
that makes the assignment of violation-marks fully explicit (cf. Ellison
1994, Zoll 1996):
(2) Expanded schema for alignment constraint3
ALIGN(Catl, Edgel; Cat2, Edge2; Cat3)--
VCatl if 3Cat2, assign one violation-mark VCat3 that intervenes
between Edgel of Catl and the nearest Edge2 of some Cat2,
where
Catl, Cat2 are prosodic or morphological categories, Cat3 is a
prosodic category and Edgel, Edge2 E (Right, Left}.
This formulation makes explicit what is usually implicit in analyses that
use alignment: some specific unit of distance (Cat3) is used to determine
the extent of violation.4 Reference to the nearest Cat2 also makes explicit
In (3a), for instance, the three feet are misaligned by five, three and one
syllables, respectively. So this candidate receives nine violation-marks
from ALLFTR.
Perhaps the most remarkable thing about gradient ALLFTR and
ALLFTL is that they are able to distinguish between (3b) and (3c). This
property is important when other constraints rule out (3a) and (3d), so that
(3b) and (3c) compete directly. As we will see in ?6.2, though, empirical
findings about stress typology do not support this aspect of the gradient
theory (Kager 2001); the (3b)/(3c) distinction does not seem to be an
authentic, independent difference between languages.
Significantly, there is no reasonable way of reproducing this apparently
unnecessary distinction using the categorical schema in (1). Consider
hypothetical categorical constraints against non-peripheral feet: *FT / aO
and *FT / _ a. These constraints do not differentiate (3b) and (3c); they
assign equal marks to both. Or consider hypothetical categorical con-
straints against a syllable that is preceded or followed at any distance by a
foot: *r / FT ... and *a / _ ... FT. Again, these constraints cannot dis-
tinguish (3b) from (3c), because both have five syllables that are preceded
by some foot and five that are followed by some foot.
80 John J. McCarthy
Several anonymous reviewers have raised an objection that goes some-
thing like this. Suppose CON supplies categorical constraints against feet
that are preceded or followed by at least three syllables (cf. Karttunen
1998, who adopts a similar artifice to deal with multiple loci of violation):
*FT / Orr- and *FT / c_ ocr. These constraints can distinguish (3b)
from (3c), since (3b) has two feet, each of which is followed by at least
three syllables, but (3c) has only one foot meeting that condition. It would
appear, so the objection goes, that the power of gradient ALLFTR/L has
been reproduced using only a categorical schema.
There are two answers to this objection. The first is that the full power
of gradient ALLFTR/L has not really been recaptured. ALLFTR/L can
make similar distinctions in even longer words, but the counting con-
straints *FT / Uocrn_ and *FT / -oor cannot. More counting constraints
could be added, but since CON is finite, it will never be possible to repro-
duce the full effect of ALLFTR/L. What we seek are analyses of stress
systems, not analyses of individual words that we happen to encounter.
Clearly, counting constraints like *FT / arxr_ do not analyse the same
systems that ALLFTR/L does.5
An even more telling response to the reviewers' objection is that con-
straints like *FT / orCx- contravene a widely assumed (though often tacit)
principle of linguistic metatheory: rules and constraints are local, a re-
quirement often expressed by saying that rules or constraints do not count
beyond two in their definitions (Chomsky 1965: 55, Hayes 1995,
McCarthy & Prince 1986, Nelson & Toivonen 2001). For example, no lan-
guage requires the presence of at least three round vowels to initiate round-
ing harmony, nor do we ever find that complementisers may be doubly but
not trebly filled. The impossibility of constraints like *FT / CraUx is
therefore quite independent of alignment, OT and even phonology.
In fact, the impossibility of such constraints is already implicit in the
markedness constraint schema (1). The locus of violation k is a single
phonological constituent. Paul Smolensky suggests that the contextual
condition C also be limited to mentioning a single phonological constitu-
ent that is separate from K. This is a strong claim; if it proves correct, then
constraints are inherently local because they can never mention more than
two distinct constituents and a relation between them, such as adjacency
or shared membership in a superordinate constituent.
Now consider faithfulness constraints. In correspondence theory
(McCarthy & Prince 1995, 1999), the standard faithfulness constraints are
inherently categorical in their assessments. For example, MAX assigns a
violation-mark for each input segment without an output correspondent.
DEP does the same, but with input and output transposed. IDENT(F)
5 It might be objected that OT itself involves counting violations. As has been em-
phasised repeatedly (Prince & Smolensky 1993), the key notion in OT is compari-
son, not counting. Furthermore, this objection blurs an important distinction
within OT between the constraints and EVAL. Constraints like *FT / aa_ build
counting right into their definitions; this has nothing to do with how EvAL compares
candidates.
OT constraints are categorical 81
assigns a mark for each input/output segmental pair differing in the value
of feature F. The less familiar UNIFORMITY and INTEGRITY, which pro-
hibit segmental coalescence and diphthongisation respectively, are also
inherently categorical: they assign one violation-mark for each segment
that has multiple correspondents. I-CONTIGUITY, which prohibits internal
deletion, O-CONTIGUITY, which prohibits internal epenthesis, and the
ANCHORconstraints that prohibit peripheral deletion and epenthesis are
contextually restricted versions of MAX and DEP, so they are categorical
just as MAX and DEP are. (For more on the ANCHOR constraints, see ?4.)
Overall, then, these faithfulness constraints are in accordance with the
categorical markedness schema (1), except that their loci of violation are
mappings of input and/or output constituents, rather than output con-
stituents themselves.
This leaves LINEARITYas the only constraint whose status vis-a-vis
gradience is as yet unclear. LINEARITYforbids metathesis, and the in-
tuition we seek to capture is that the non-local metathetic mapping /a,By/
-[ya)a] is less faithful than its local counterpart /a,y/37/[ay/3]. Thus,
non-local metathesis, which is notably rare (non-existent according to
Poser 1982), can occur only when local metathesis is unsatisfactory.
The need to distinguish local and non-local metathesis leads Hume
(1998) to describe LINEARITY as a gradient constraint (cf. Carpenter 2002),
but this conclusion does not necessarily follow. The input /a/3yl can be
regarded as asserting three linear-precedence relations: a > /B, /3> y and
a> y. Categorical LINEARITY assigns a violation-mark for each input pre-
cedence relation that the output contradicts: two marks for [yacp/]and one
mark for [ay,B]. Though this sort of constraint goes beyond the highly
limiting schema (1), it is nonetheless categorical in its assessments.
Local constraint conjunction (Smolensky 1995) creates new constraints,
markedness or faithfulness.6 The local conjunction of constraints Cl and
C2 in domain D, written [C1&C2]D, is violated if and only if both Cl
and C2 are violated by the same instance of D. Given how local conjunc-
tion is defined, [C1&C2]D is necessarily categorical, even if Cl, C2 or both
are gradient. Therefore, local conjunction is not a potential source of new
gradient constraints and may be safely set aside.
To sum up, the claim that all OT constraints are categorical has here
been reduced to the claim that all markedness constraints conform with
the constraint schema (1) and all faithfulness constraints are of the stan-
dard categorical types in correspondence theory. Gradient constraints,
most prominently alignment, are not compatible with (1), nor can all the ef-
fects of gradience be obtained using (1). If it proves true, as I argue below,
that gradient constraints are not required in OT, then it is possible to
maintain a restrictive claim about CON: all markedness constraints are
based on (1), and faithfulness constraints assign at most one mark for each
unfaithful mapping.
6 For related work on local conjunction, see the references cited in McCarthy
(2002b: 43).
82 John J. McCarthy
Before venturing into the realm of the empirical, it may be necessary to
clarify a limit on the goals of this article. The proposal made here is that
(1) sets down a standard that all markedness constraints must meet; the
proposal does not say that everything meeting this standard is an actual
constraint in CON. In other words, (1) presents necessary but not sufficient
conditions for valid markedness constraints. Like other constraint sche-
mata in the literature (e.g. McCarthy & Prince 1993a, Smolensky 1995,
Eisner 1999, Bakovic & Wilson 2000, Wilson 2000, 2001, Potts & Pullum
2002, Smith 2002), this one does not obviate the need for other formal or
substantive limits on what constraints are possible.
3 Bounded gradience
Gradient constraints in the OT literature are not limited to alignment.
A taxonomy of attested types of gradience is useful to organise the dis-
cussion. In (4), the known types of gradient constraints are classified ac-
cording to the dimension along which violations are assessed.
(4) Types of gradience in the OT literature
a. Horizontal gradience
Assign violation-marks in proportion to distance in the segmental
string. Example: ALIGN(Ft, R; Wd, R; or), ALIGN(PfX, L; Wd, L;
Seg) (used in infixation - see ? 5).
b. Vertical gradience
Assign violation-marks in proportion to levels in a hierarchy. Ex-
ample (Prince & Smolensky 1993: ch. 4, (66)): NON-FINALITY no
head of Wd is final in Wd. The Wd is headed by its main-stressed
foot and recursively by the head syllable of that foot. One violation-
mark is assigned for each of these that is final in Wd. E.g. Latin
*[a(mo:)] gets two marks and [(aimo)] gets only one.
c. Collective gradience
Assign violation-marks in proportion to the cardinality of a set.
Example (Padgett 1995a, 2002): CoNsTRAINT(Class) assign one
violation-mark for each member of the feature-class Class that does
not satisfy CONSTRAINT. E.g. Assim[Place] assigns two marks to
[angba], one to [augba] and none to [aijmgba].
d. Scalar gradience
Assign violation-marks in proportion to the length of a linguistic
scale. Example (Prince & Smolensky 1993: 16): HNuc 'a higher
sonority nucleus is more harmonic than one of lower sonority', i.e.
assign a nucleus one violation-mark for each degree of sonority less
than the sonority of a.
el b. [[(nata)]t5nchi] * ***
c. [[(nata)]5nchi] #l **
b.rantas
c. rantasa *! *
But with FINALC ranked above DEP, underlying vowel-final words should
also get epenthetic [?]. This is incorrect, as shown by forms like /lompo/
[lompo], * [lompo?] 'big'. We have here a ranking paradox: [rantasa?]
requires FINALC> DEP, but [lompo] requires DEP> FINALC.
This paradox leads McCarthy & Prince (1994) to propose that alignment
is what blocks [?]-epenthesis in [lompo]. Ranked above FINALC, ALIGN
(Stem, R; a, R) blocks epenthesis in [lompo]; ranked below CODACOND,
88 John j. McCarthy
it does not block epenthesis in [rantasa?]. Tableau (9) shows how this
analysis works:
> ALIGN(Stem, R; a, R) >
(9) CODACOND FINALC > DEP in Makassarese
a. /rantas/ CODACOND
ALIGN(St,r) FINALCDEP
.rantasa? * I *#
ii. rantas j ! _
iii. rantasa j * * *
b. /lompo/
W i.lompo *
ii.lompo? j I , I
*
C. [WdCVCVC-VCVCV] _ . _
d. [WdCVCVC-[wdVCVCV] *! *
Peperkamp (1997: 81) assumes that a structure analogous to (lOa, b) is correct for
Italian, but she doesn't defend this assumption or consider candidates analogous to
(l Oc).
90 John 7. McCarthy
The subcategorisational or different-edge alignment constraints, like
SFX-TO-PRWD in Axininca, are replaced by another type of ANCHORcon-
straint:
(12) D-ANCHOR(C1,CO, E)
If x = Edge(Cl, E) and y = Edge(C0, E), then x9x' and x' is im-
mediately adjacent to y.
'Any element at the designated edge of CI has a correspondent that is
adjacent to an element at the opposite edge of CO.'
So, for example, the alignment constraint SFX-TO-PRWD is replaced by
ANCHOR(SfX, Wd, L).
Constraints based on these schemata are inherently categorical, in the
same way that faithfulness constraints in general are categorical. ANCHOR
is satisfied or not, depending on whether the required correspondence
relation exists. Likewise, D-ANCHOR is satisfied or not, depending on
whether the required correspondence and adjacency relations exist. That
these constraints should be preferred to alignment follows from the overall
argument of this section: requirements that the edges of morphological
and prosodic constituents coincide or be adjacent are, in known cases,
always enforced categorically, and gradient evaluation leads to implausible
predictions about language typology.
Applications and extensions of this idea appear in McCarthy & Prince (1993a, b),
Noyer (1993), Akinlabi (1996), Urbanczyk (1996), Buckley (1997), Fulmer (1997),
Spaelti (1997), Boersma (1998), Carlson (1998), de Lacy (1999), Stemberger &
Bernhardt (1999) and McCarthy (2000a). For a different approach, which replaces
alignment with faithfulness, see Horwood (to appear).
The existence of constraints like ALIGN(-um-, Wd, L) is sometimes offered as
proof that OT has language-particular constraints. This point is somewhat jesuiti-
cal. ALIGN(Pfx, Wd, L) and ALI(;N(Sfx, Wd, R) offer a universal framework for
stating constraints on affix placement. That individual affixes must somehow be
identified as prefixes or suffixes on a language-particular basis comes as no surprise.
A real 'language-particular constraint', if any exist, would presumably have the
character of the language-particular rules in other theories: a one-time ad hoc
statement with no typological commitments whatsoever.
0 T constraints are categorical 91
For example, in Prince & Smolensky's analysis of Tagalog, infixation of
the actor-focus morpheme -um- is attributed to a constraint hierarchy
where NOCODA crucially dominates gradient ALIGN(-Um-,L; Wd, L).
This ranking leads to less-than-perfect alignment with consonant-initial
words like [sumulat] 'to write' or [prumeno] 'to brake'. The tableau in
(13) shows how infixation is achieved:
/um-preno/ |NOCODA*COMPLEXONS
oE i. prumeno *
ii.pumreno *!
b. *COMPLEXONS> NOCODA
NoCODA
/'um-preno/II*COMPLEXONs
i.prumenoj
ii. pumreno *
Since only a fraction of all verbs are lexically marked to take -um- as their actor-
focus marker, and since there are other actor-focus markers like ma-, mag- and
may-, it is no loss for a verb to be blocked from having an -um- form for phono-
logical reasons. See Schachter & Otanes (1972: 284ff).
12 Complex onsets may be permitted only initially; there is some reason to think that
the same clusters are heterosyllabic word-medially. Schachter & Otanes (1972: 29)
cite the word [libro] 'book' as evidence that 'the preference for short vowels in
closed syllables is reflected in the pronunciation of certain loan words ... in which a
vowel that is stressed in the language of origin is short in the Tagalog borrowing'.
In short, this word is syllabified [lib.r6].
13 There is a large body of work applying the idea of partially ordered or tied con-
straints to problems of phonological variation. For references, see McCarthy
(2002b: 233).
0 T constraints are categorical 93
If these constraints are formally tied in the grammar of Tagalog, and if a
specific ranking is chosen at each application of EVAL, then the observed
variation can be obtained.
The real focus of Orgun & Sprouse's analysis, however, is the role of
labial sonorants in blocking -um- affixation. They propose a constraint,
here called OCP[labial], that forbids sonorant labials in successive onsets.
Most of the starred forms in (14c, d) violate this constraint: *[mumahal],
*[wumalow], *[smumajl], *[swumirj]. They argue that merely ranking
OCP[labial] among the other constraints is insufficient to block -um-
affixation entirely with such words. Instead, OCP[labial] is promoted, on
a language-particular basis, to a new grammatical component called
CONTROL. The control component applies to the output of EVAL, blocking
some candidates that EVAL has judged as optimal. Thus, constraints in the
CONTROL component are inviolable and can cause derivations to crash.
Their analysis is that EVAL proper emits *[mumahal] as the most har-
monic form, but then the derivation crashes when OCP[labial] sees
*[mumahal] in the CONTROL component.
Orgun & Sprouse's argument for enriching OT in this way comes from
the impossibility of deeper infixation to satisfy OCP[labial]. The problem
is that *[mumahal]'s violation of OCP[labial] can be avoided by moving
the infix further away from the initial [m], as in *[mahumal] or *[maha-
lum]. In a conventional OT analysis, without the CONTROL component,
*[mahumal] should be fine because it violates only low-ranking ALIGN
(-um-, Wd, L). The CONTROL component sidesteps this issue: the prob-
lematic candidate *[mahumal] gets no benefit from satisfying OCP[labial]
because it has already lost in the EVAL component by virtue of its poor
alignment.
Orgun & Sprouse hint, however, that the special, post-EvAL application
of OCP[labial] could be avoided 'if ALIGN were supplemented with a con-
straint limiting -um- to the first syllable' (Orgun & Sprouse 1999: 207), a
move they reject on the grounds that 'it clearly is not in the spirit of the
alignment approach to infixation'. This critique seems apt if gradient align-
ment is supplemented with a categorical constraint, but not if it is replaced
by a categorical constraint, as I will argue shortly. But first, I will present
some necessary theoretical background to the reanalysis of Tagalog.
Wd
i.e. -af- is not preceded by a syllable within
*-af- /o r the prosodic word
c. PREFIX/FT(-af-)
Wd
/\"~ i.e. -af- is not preceded by a foot within the
*-af- / Ft prosodic word
These constraint formulations assume the categorical schema (1) and some
additional notational conventions. The category label Wd and the lines
indicating constituent membership should be understood as saying that
Wd dominates both seg and -af- in (17a), and furthermore that there is no
other Wd that dominates either seg or -af- but not both. (This is roughly
equivalent to the 'nearest Edge2 of some Cat2' clause in the alignment
definition (2).) In addition, joint membership in the Wd constituent is
enough; for example, it is not intended that or and -af- are necessarily
adjacent for (17b) to be violated, only that some arprecede -af- within Wd.
For any affix -af-, there will be the full suite of constraints in (17), if -af-
is a prefix, or the SUFFIXcounterparts of (17), if -af- is a suffix. Similar
constraints exist for morphemes that are affixed to prosodic constituents
like the head foot, rather than the word (see ? 5.4).
The PREFIX constraints form a stringency hierarchy in the sense of
Prince (1998): violation of (17c) entails violation of (17b) entails violation
of (17a). (This presupposes, as an anonymous reviewer points out, that the
headedness requirement on prosodic constituents is wired into GEN, SO
that any constituent at level n is guaranteed to contain at least one con-
stituent at level n- 1.) Prince shows that constraints in a stringency relation
never conflict, so they are never directly rankable. They can be ranked
indirectly, through transitivity, as will be shown in (19).
These constraints build the unit of violation into the definition of the
constraint, as has sometimes been assumed for gradient alignment (see (2)
above and ?5.3 below). But they operate categorically: the locus of viola-
tion is the prefix, and so none can assign more marks than there are prefixes
in the form under evaluation. The distance between prefix and word edge
OT constraints are categorical 95
is relevant only to determining whether or not the constraint is violated,
not how much it is violated.
The constraints in (17) will reappear in ?7.2, when we examine the
phonology of floating feature or tone morphemes.
14
In proposing a categorical approach to Tagalog infixation, I have been anticipated
by Boersma (1998: 196-200). Boersma proposes a family of *SHIFT constraints
defined as follows:
(i) *SHIFT(f: t;g: u; d) ... A pair of contours (edges) at times t and u, defined on
two perceptual tiers f and g and simultaneous in their specification, are not fur-
ther apart in the output (if they occur there) than by any positive distance d.
s To be specific, suppose that the correspondence relation maps every segment of the
input to one or more segments of the output, or otherwise to the empty string e.
Mappings of input segments to e violate MAX; all other mappings obey it. The null
candidate has no phonological content and no mappings from the input, because the
correspondence relation is undefined. Therefore, it vacuously satisfies MAX, unlike
a candidate with one or more true segmental deletions. For further development
and applications of the null output as a candidate, see the references cited in
McCarthy (2002b: 230).
96 John J. McCarthy
structure - like NOCODA- or they require structure, when present, to
have certain properties - like ONSET or many alignment constraints. Since
O has no structure whatsoever, it is never in danger of violating either kind
of markedness constraint. Furthermore, because its input-output corre-
spondence relation is undefined, (0 vacuously satisfies all faithfulness
constraints. (Faithfulness constraints are defined on correspondence re-
lations; if the correspondence relation of some candidate is undefined,
then no faithfulness constraint can possibly be violated. See note 15.) By
assumption, 0 violates just one constraint, which Prince & Smolensky call
MPARSE.
To be specific, the constraint MPARSE(-Um-) is violated by the candi-
date 0 whenever the input contains the morpheme -um-. Since verbs with
-um- do sometimes have codas or complex onsets, we can infer that
MPARSE(-um-) dominates NOCODAand COMPLEXONS (see (18a)). Fur-
thermore, since -um- is misaligned by one or more segments, we can
conclude that MPARSE(-um-) also dominates PREFIX(-Um-) (see (18b)).16
pru.me.no *
b. MPARSE(-UM-) >PREFIX(-Um-)
/um-sulat/ MPARSE(-UM-)PREFIX(-Um-)
ii. su.mu.lat *
Ih In tableau (1 8a), the ' = ' symbol and the absence of a vertical line indicate that two
constraints are formally tied.
0 T constraints are categorical 97
/um-mahal/ OCP[lab]:
PREFIX/o(-Urn-)
MPARSE(-Urn-)
PREFnX(-UM-)
BWa. ) *
b. mu.ma.hal * *
c. ma.hu.mal *
This tableau shows a key result. We know from (18) that MPARSE(-Um-)
dominates PREFIX(-Um-), since otherwise -um- would never be infixed. To
this, (19) adds the information that PREFIX/cr(-um-) dominates MPARSE.
Therefore, MPARSE separates the two PREFIX constraints in the hierarchy.
This shows that they must indeed be separate constraints, as proposed in
?5.2.1' (To complete the argument at the level of analytic detail, it is also
necessary to consider dissimilated candidates like *[munahal], which
show that IDENT[Place] dominates MPARSE(-UM-).)
This categorical approach is usually regarded as incompatible with
gradient alignment theory, whence Orgun & Sprouse's argument for a post-
EVAL check by OCP[labial]. If there is a single gradient alignment
constraint ALIGN(-um-, L; Wd, L), then it must either dominate MPARSE
(-um-) or be dominated by it. Either way, the wrong result is obtained.
Gradient alignment theory could be modified to achieve similar results
by building the counting unit into the definition of the constraint. This is,
in fact, the implication of the Cat3 argument in (2). If two otherwise
identical gradient alignment constraints can differ only in the quantum of
violation, as (2) implies, then gradient ALIGN(-Um-,L; Wd, L; a) can be
ranked above MPARSE and gradient ALIGN(-Um-,L; Wd, L; Seg) can
be ranked below it. This move might be seen as the easiest answer to vexed
questions about how to count violations of gradient constraints: for every
gradient constraint, there are several versions distinguished solely by the
counted unit.
If Tagalog is to be analysed within the strictures of standard input/GEN/
EvAL/output OT, then either categorical PREFIX/a(-UM-) or gradient
ALIGN(-Um-, L; Wd, L; a) is needed. The categorical constraints are also
sufficient for Tagalog, as shown in (19) above and (22)-(24) below. The
enriched gradient alignment theory may work in Tagalog, but the gradi-
ence part of it plays no actual role. Categorical constraints are needed any-
way; their existence in OT is not in doubt. Since categorical constraints are
also sufficient, as I have argued here and elsewhere in this article, then stan-
dard Occamite reasoning demands that gradient constraints be eliminated.
It remains only to clear up a few remaining points about Tagalog and
to show the efficacy of the entire analysis before moving on to another
example. If deep infixation a la *[mahumal] is not an option, then why not
17
As Klein (2002: 9-10) points out, PREFIX(-Um-) is never visibly active in Tagalog.
But this scarcely supports his conclusion that it can be dropped from the analysis.
By a central premise of OT, constraints may be low-ranked, but they are never
literally absent from the grammar of any language.
98 Jtohn 7. McCarthy
skip infixation entirely with such words, opting for *[?ummahal] or
*[Pumwalow]? There is a local explanation for the ill-formedness of
*[Pummahal] - mm clusters aren't allowed (see note 10) - but there is no
such explanation for *[?umwalow]. In fact, we know that -um- words
specifically can contain mw clusters because of examples like [sumwirj]. So
*[?umwalow] must be ruled out for another reason: its epenthetic initial
consonant.
b. um.wa.low !
c. ?um.wa.low
rw a. su.mu.lat * *
b. su.lu.mat * *
c. um.su.lat *! **
d.t um.su.lat ** *
e . 0_ _ _ _ _ _ _ _ _
T constraints are categorical 99
0
In (22), candidates without infixation are ruled out by the undominated
constraints ONSET or DEP(C). Excessive infixation is excluded by PREFIX/
cr(-um-). Since there is a form (22a) that violates none of these constraints,
there is an alternative to the null output.
w b. pru.me.no * *
c. um.pre.no *! * *
d. um.pre.no | * * l
eQ
b. wu.ma.low * *
c. wa.lu.mow *! * *
d. um.wa.low *! **
e. ?um.wa.low ** *
In (24), the null output wins because the alternatives are all worse:
a [wumV] sequence (24b) that violates OCP[labial]; deep infixation (24c),
contrary to the dictates of PREFIX/ur(-Um-);or the usual problems with
ONSET and DEP(C) (24d, e)."8
A final remark. This analysis requires that the underlying form of
[?abot] is /?abot/ and not /abot/, treating this word exactly on a par with
Alan Prince raises an important typological question: is deep infixation ever pos-
sible in any language under any ranking (cf. McCarthy & Prince 1993b)? Samek-
Lodovici (1993) finds a possible example involving a geminating (i.e. mora) infix,
though cases similar to [walumow] are not known to me. With further refinement,
the theory developed here may offer an explanation for this typological gap. If
shallow infixation is ruled out, then deep infixation competes with suffixation (cf.
Noyer 1993, Fulmer 1997). Deep infixation and suffixation tie on the constraints
PREFIx and PREFIX/I; if some other constraint, such as morpheme CONTIGUITY
(Kenstowicz 1994, McCarthy & Prince 1999), disfavours infixation, then suffixation
must win. This cannot be the full story, however, because other constraints may
also militate against suffixation (e.g. the OCP rules out both [wumalow] and
[walowum]).
100 John 7. McCarthy
/sulat/. Independently, there is good reason to assume that the underlying
form is indeed /?abot/: there are no [i]/@alternations, and the root-initial
[?] shows up even after consonant-final prefixes. Of course, under richness
of the base (Prince & Smolensky 1993), the grammar of Tagalog is also
responsible for correctly disposing of hypothetical V-initial roots. They
must be treated unfaithfully, because Tagalog words never begin with a
vowel, but no active alternations show how they are treated - the tra-
ditional assumption that hypothetical /apak/ becomes [?apak] is without
empirical support. If we nonetheless assume in the absence of evidence
that /apak/-- [?apak] is the right disposition of V-initial words, then the
ranking [ONSET? DEP(C)J must be added to the grammar in (21). Hypo-
thetical /apak/ then would surface as [?apak]. This verb would have no
-um- form, because DEP(C) dominates MPARSE(-UM-) (see (20)). Do such
verbs exist? They could: some surface [?]-initial verbs don't take -um-, as
predicted, but that could also be because verbs in general are lexically
marked to take -um-. In any case, richness of the base does not challenge
the analysis presented here.
The size of the entire word, and not just the size of the root, is decisive.
For instance, the last example is based on the disyllabic root [gile] 'to sift'.
9 There are additional alternations in the form of the infix. It is i before 1- or r-initial
roots. Its vowel is u before u or o in the next syllable. And its consonant is r in
agreement with an r in the next syllable (cf. Cohn 1992).
OT constraints are categorical 101
This rather puzzling distribution of the -il- and -la alternants can
be made sense of when it is recalled that Nakanai has penult stress. The
-il- alternant is attracted to the left edge of the word - it is a formal pre-
fix - like Tagalog -um-. It is also attracted to the main-stressed syllable. In
disyllables, the main-stressed syllable is also the initial syllable, so both
desiderata for -ii- placement can be more or less satisfied. In longer words,
however, there is no way to attach -ii- to the penult main-stressed syllable
and also keep it close to the beginning of the word.
To get the analysis rolling, we first need to make some assumptions
about the source of the -il-1-la alternation. The form of these two alter-
nants is not phonologically predictable, though their distribution is. There
is a large literature on this kind of allomorphy. The basic idea is that
allomorphs are listed together in the lexicon, so an underlying represen-
tation will contain a set of alternants, such as / {il, la} -sagege/.20 When GEN
constructs candidates, it uses both input alternants. This means that
[silagege], [sagegela], [lasagege], [salagege] and [sagegeil] are among
the candidates that incur no faithfulness violations. (Of course, unfaithful
candidates like [sulagege] or [sagegea] are also in the mix.) The choice of
the winning candidate - and hence the selection of -il- or -la - is as usual
the responsibility of EVAL.
The allomorph -il- is a formal prefix with its distribution under the
control of undominated PREFIX/Cr(-il-). The allomorph -la is a for-
mal suffix, and since it is never infixed, its distribution is governed by
undominated SUFFIx(-la). To understand the -il-l-la alternation, we first
need to get a handle on a couple of descriptive problems: -la functions as
kind of default, occurring only when -il- is blocked, and -il- is attracted to
the stressed syllable.
The first thing to address is -la's default status. Because -la does not
occur with disyllables (*[tagala]), there must be some cost associated with
it. The cost is not input-output faithfulness, however, since neither -il-
nor -la is more faithful. One possibility is that the affixal alternants are
lexically prioritised, as Bonet et al. (2003) have argued for Catalan.
Another possibility is that -la's cost is measured by output-output faith-
fulness to stress (Kenstowicz 1996, 1997, Benua 1997, Alber 1998, Kager
2000, McCarthy 2000b, Pater 2000) or paradigm uniformity (Raffelsiefen
1995, 1999, Kenstowicz 1996: 385, McCarthy 1998). Kager proposes the
following constraint:
20 The idea of lexical entries as sets of allomorphic alternants originated with Hudson
(1974) and is adopted by Hooper (1976). There is a considerable literature applying
OT to problems in allomorphy or lexical selection, much of it cited in McCarthy
(2002b: 183-184).
102 John j. McCarthy
When a word takes the suffix -la or almost any other suffix in Nakanai, the
stress shifts to the new penult: [sagege]/[sagegela]. Stress shift is a violation
of OO-PK-MAX: the stress peak e4 in [sja2g3&4g5e6]does not stand in cor-
respondence with a segment that is a stress peak in [sla2g3e4g5061a].The
infix -il- does not affect stress placement, since it falls to the left of
the stressed nucleus: [ta'ga]/[ti1aga]. On grounds of OO-PK-MAx alone,
the -il- alternant is favoured.
While -la is suffixed, -il- is attracted to the main stress. It is not unusual
to find infixes that are tropic to stress: reduplication in Samoan targets the
main-stressed syllable (Marsack 1962, Broselow & McCarthy 1983), as does
possessive suffixation in Ulwa (Hale & Lacayo Blanco 1989, McCarthy
& Prince 1990). The central analytic idea is that affixes may be prefixed
or suffixed to the head foot rather than the prosodic word (Broselow
& McCarthy 1983, Inkelas 1989, McCarthy & Prince 1990, 1993b).
The responsible constraints follow the same general pattern as (17): AFX-
TO-HD(-il-) is violated by an il-containing candidate where -il- is sep-
arated by one or more segments from the head foot. This constraint is
undominated in Nakanai, since -il- is never found in any other context.
(Forms like [iltaga] are ruled out by another undominated constraint,
NOCODA.)
With these preliminaries taken care of, we are now in a position to ex-
plain the conditions on the -il-l-la alternation. The key idea is that -il-
cannot stray from the first syllable of the word because the categorical
constraint PREFIX/o(-il-) is undominated. When AFX-TO-HD(-il-) and
PREFIX/cr(-il-) cannot both be satisfied, as is the case with trisyllabic and
longer words, then the -la allomorph appears instead, even though it is
dispreferred by OO-PK-MAX. This allomorph sidesteps both of the prob-
lematic constraints, since they pertain only to -ii- and not to -la.
One element of the analysis, then, is crucial domination of la-dis-
favouring OO-PK-MAx by PREFIX/c(-il-):
(27) PREFIX/o(-il-) >OO-PK-MAX
/1{il,la}-sagege/ PREFIX/o(-il-) OO-PK-MAX
(cf. sagege)
w a. sa.ge.ge.la
b. sa.gi.ie.ge
Choosing the suffixed -la alternant in (27) leads to stress shift on the 00
dimension, but the alternative of placing the -il- alternant more than a
syllable away from the left word edge is ruled out by an undominated
constraint.21
21 Other undominated constraints exclude some plausible competitors for the winner
in (27). In *[sagegela], 00-PK-MAX is satisfied by treating -la as a stress-neutral
suffix. But -la-, like nearly all Nakanai suffixes, is stress-determining, not stress-
neutral. This means that the metrical constraints responsible for penult stress must
dominate 00-PK-MAX. Another reasonable-looking competitor is [salagege], with
OT constraints are categorical 103
Another element of the analysis is crucial domination of 00-PK-MAX
by AFX-TO-HD(-il-):
b. si.la.ge.ge
Taken together, the ranking arguments in (27) and (28) establish the con-
ditions for choice between the -il- and -la allomorphs. For -il- to occur, it
cannot be displaced by as much as a syllable from the beginning of the
word or at all from the main stress. If these conditions are not satisfied,
then the -la allomorph occurs instead, even though its presence forces a
stress shift in violation of 00-PK-MAX.
For present purposes, the most important thing about the prefix -il- is
that does not fall exactly at the left word edge in forms like [tilaga] (see
(29)). Deviation by a whole syllable is not possible, as the tableau (27)
shows, but deviation by just a segment is tolerated. This demonstrates that
00-PK-MAX dominates PREFIX(-il-):
infixed -la. But -la is never infixed, so SUFFIX(-la) is undominated, also crucially
ranked above OO-PK-MAX.
104 John J. McCarthy
simultaneously attracted to the left word edge and the main stress. It
cannot be displaced from the word edge by a syllable or more, nor from
the head foot, so when a word is longer than a single foot, the -il- allo-
morph fails completely and -la takes its place. But -la has a cost: because it
is a stress-determining suffix in a language with penultimate stress, it
produces a stress alternation. The allomorph -il- avoids this alternation,
and that option is taken when -il- can get close enough to its preferred
locus so that it violates only the low-ranking PREFIXconstraint.
5.5 Summary
Nakanai and Tagalog show that affix-position constraints must categori-
cally distinguish the extent to which an affix is malpositioned. Classic gradi-
ent alignment constraints cannot do this. The separate ranking required in
Nakanai and Tagalog is not an option with classic alignment. (See Klein
2002 for a further argument in support of quantisation based on evidence
from infixation in Chamorro.)
As I noted in ?5.3, it is certainly possible to construct a theory with
gradient alignment and violation quanta. Indeed, such a theory is con-
templated in the formalisation of gradient alignment in (2). Tagalog and
Nakanai could be analysed in this theory, substituting gradient ALIGN
(-um-/-il-, L; Wd, L; a) in tableaux (24) and (27).
The problem with this gradience-cum-quantisation theory is not de-
scriptive coverage - it is a richer theory, after all - but parsimony. The
categorical constraints in (17) are sufficient for Tagalog and Nakanai.
These constraints take over the actually observed functions of gradience.
For example, Prince & Smolensky (1993) attribute the ill-formedness
of words like *[prenumo] to gradient ALIGN. But categorical PREFIX/a
(-um-) is sufficient to rule out *[prenumo], as shown in (23). So, although
gradience and quantisation are not logically incompatible, they compete
for the same explanatory turf. Constraints like PREFIx/a(-um-) have
the violation quanta without the trappings of gradience. As we have seen
in Tagalog and Nakanai, and as I argue elsewhere in this article, the
need for gradience is very much in doubt. Since OT indisputably has
categorical constraints, and plenty of them. Occamite reasoning de-
mands that we rid the theory of gradient constraints if categorical ones are
sufficient.
The constraint *LAPSE militates against stress lapses generally, but the
constraints LAPSE-AT-END and LAPSE-AT-PEAK can license lapses in those
specific environments by disallowing them everywhere else. Conversely,
*INITIALLAPSE disfavours lapses word-initially. Directionality effects are
obtained from the interaction of these categorical constraints, with a better
fit to observation than the gradient ALIGN(Ft, Wd) constraints.22
Table I shows how the effects of directionality follow in Kager's sys-
tem. This table is an unranked tableau. It provides information about
candidate performance on the various *LAPSE constraints without con-
sidering the constraints' ranking. In addition, the left and right ALIGN
(Wd, Ft) constraints are included. I will discuss these constraints later in
this section.
Table I uses seven-syllable words, since directionality is usually visible
only in odd-parity words, and seven syllables are need to show the full
range of observed patterns. Exhaustive parsing is assumed up to de-
generacy- that is, FTBIN is undominated. The candidates are grouped
according to the position of main stress (indicated by the numeral 1) and
whether feet are iambic or trochaic. Each group, such as (a)-(d), contains
a set of candidates that compete with one another, holding main-stress
location and foot type constant. In effect, the candidates in each group
compete in directionality only. Competition across groups also occurs, of
22
The suggestion that directionality effects are reducible to constraints on the position
of unfooted syllables was made to me in 1993 by Junko It6 and Armin Mester. At
the time, I summarily (and, in retrospect, foolishly) rejected this idea.
O T constraints are categorical 107
*LAPSE LAPSE- LAPSE- *INIT ALIGN ALIGN
AT-END AT-PEAK LAPSE (Wd,L;Ft,L) (Wd,R;Ft,R)
Trochaic:
Main Stress Left
a. (10)(20)(20)0
7777M. 77 7:-M ., ,'-- ....~~~~~~~~~~~~~~~~~~
.. :'(')t
....-6.e.......:...
?'..............
((----
C. j(10)0(20)(20) }
I
* ______ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ __....
_ _
d. j 0(10)(20)(20)
Trochaic:
Main Stress Right
e.
... i'.m...................
g, ,,.?'' R.
0(20)(20)(10)
n
T . .?.:.:R.
...- .
T I _ _ T _
I _ _
g ......... (20)(20)0(10) .
h.
................
o---i::';g-::-.'"g...
........ :''.:::'
.:--j:''
'.''''e-
(20)(20)( 10)0
::
*
* I _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
Iamzbic:
M ain Stress L eft _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
j.
0.)k.
BB._ B ._ ._ ._. . ,.
(01)(02)0(02)0
( 0 ( 0 2 ) ( 0
.
2 ) . ". '. .. .i.
*
*
_ _
_
_
_
_ _
_
_
_ _ _
_
_ _
_
_
I _
_
_
_ _
_
_ _
_
_ _
_
_
_ _ _ _
1. (01)(02)(02) * *
Iambic:
Main Stress Right
O. (02)(02)0(01) I* I I
P (02)(02)(01)0
b. (02)(2)(02)(01)**********! ***********
c. (02)(02)(2)(01)***********! __________
d. (02)(02)(02)(1)
***********1! *********
24
In a theory with various *LAPSE constraints, it might seem that PARSE-5 and its
positional variants in (33) are superfluous. This does not seem to be true. There is a
basic difference in form and function between *LAPSE and PARSE-U; the former is
part of the prominential system and is completely indifferent to foot structure; the
latter is part of the prosodic-hierarchy system and is completely indifferent to
prominence.
T constraints are categorical 111
0
of the general schema (1), the End Rule can be translated into the con-
straints given in (35):
b. ER-R
Wd
i.e. the head foot is not followed by
*Hd(Wd) I Ft another foot within the prosodic word
(The notational conventions assumed here are the same as in (17).) ER-L
is satisfied if the head foot is the first foot in the prosodic word, regardless
of whether it is literally initial in the word. ER-R is the same, with mirror
symmetry. There is nothing new about either of these constraints; except
that they are violable and non-parametrised, they are identical to Prince's
End Rule.
In ?6.3.2 and ?6.3.3, I apply ER-L and ER-R to the one-foot-per-word
phenomenon and main-stress placement.25
6.3.2 One foot per word. There are two situations of interest. One in-
volves languages that are said to lack secondary stress, from which it is
inferred that words contain no feet except for the head. The other involves
the one-foot minimal-word template that is frequently encountered in
reduplication. Each will be discussed in turn.26
As was just noted, gradient ALIGN(Ft, Wd), if it dominates PARSE-or,
will prevent iterative foot-parsing, since every additional foot contributes
more alignment violations. Claims about non-iterative footing should be
approached sceptically, however. Often, they rely on the original analyst's
silence about secondary stress. For example, Latin has been described as
having non-iterative stress, but this is only because the Latin pronunci-
ation tradition does not include any information about secondary stress. In
25
It is in principle possible to mimic the effects of ALIGN(Hd(Wd), Wd) using a
categorical constraint, as has been pointed out to me by Colin Wilson and several
anonymous reviewers: *a / )Hd ... - ... ]Wd. Why is that possible for this alignment
constraint but not others (see ?2)? Because heads are guaranteed to be unique in
Wd, so the first V in the alignment definition (2) can be ignored. This shows, as I
noted in ?2, that categoricality is not a sufficient condition for licit constraints (e.g.
locality conditions are likely to be relevant; cf. Eisner 1999).
26 A direct assault on one-foot-per-word is also possible by invoking a constraint of
the *STRUJC family (Prince & Smolensky 1993), *STRUc(Ft), which assigns one
violation-mark for every foot in the candidate under evaluation. But the need for
and desirability of such constraints have been impeached on typological and other
grounds by Gouskova (2003).
112 John J. McCarthy
fact, though, we know from work by Mester (1994) that Latin did have
iterative footing. The second syllable of words like [pudi:citiam] 'chastity
(ACC SG)' or [vere.bamini] 'you (PL) were afraid', although underlyingly
long, is observed to scan as short: [pudicitiam], [verebamini]. This short-
ening process makes sense if these words are parsed into a succession of
bimoraic feet: [(piudi)(citi)am]. The case of Cairene Arabic is also instruc-
tive. In my observation, it does not have systematic secondary stress
(though cf. Kenstowicz 1980, Welden 1980, Harms 1981), but there can
be little doubt that there is an iterative foot-parse, since otherwise the
position of main stress could not be explained (McCarthy 1979, Hayes
1995: 64-71).
This apparent disconnection between metrical structure and observed
secondary prominence has led to various formal proposals (Halle &
Vergnaud 1987, Blevins 1992, Crowhurst 1996, de Lacy 1998), though it
seems equally reasonable to see it as an aspect of the phonetics-phonology
mapping. As Hayes (1995: 119) writes, 'we might suppose that the pho-
netic and phonological rules of the language just happen not to provide
any means of manifesting foot structure. This solution is viable, given
what we have seen ... concerning the language-specific phonetic realiza-
tion of stress.' The point is that even solid evidence for the absence of
secondary stress, if such is possible, does not permit the inference that
words have only one foot, because the range of ways in which metrical
structure can be realised phonetically is so broad.
With these empirical caveats aside, the End Rule constraints in (35)
offer a way of limiting words to a single foot. If the head foot is obliged to
be both the leftmost and rightmost foot in the word, then it cannot be
preceded or followed by other feet. Therefore, if these constraints are
ranked above PARSE-cTand its positional variants, then the head foot will
also be the only foot. Tableau (36) shows this result:
b. [(10)(20)]
c. [(20)(10)]
/RED+rjandawalka/ ALIGN(Ft,Wd):
PARSE-5MAX-BR
rw a. [(rja'nda)]-[(i3Anda)(wa1ka)] **
b. [(tjAnda)wa]-[(rAnda)(walka)] ** *!
c. [(rjanda)(walka)]-[(rjanda)(walka)] ****!
Austin presents evidence from stress and allophony that the reduplicant
is a separate prosodic word in Diyari, as indicated by the [ ] brackets.
The role of gradient ALIGN(Ft, L; Wd, L; Ca)in this analysis is to rule
out (37c), with total reduplication. (It also ensures left-to-right foot-
parsing in unreduplicated roots.) Total reduplication of a quadrisyllabic
root produces a dipodal reduplicant, which is worse aligned than the
monopodal reduplicant (37a). Ranked between MAX-1O and MAX-BR,
ALIGN(Ft, Wd) controls the size of the reduplicant without affecting the
size of the base.
Suitably ranked, the categorical End Rule constraints can produce
the same result. Since Diyari is a language with main stress on the first
foot and iterative footing, ER-L and PARSE-5 (or *LAPSE) must dominate
ER-R:
b. [(rj'anda)wa]-[(ijAnda)(walka)] *! ***
c. [(randa)(walka)]-[(randa)(walka)] ##t I
d. [(anda)walka]-[(ijanda)(walka)] **! *
114 J7ohn_7.McCarthy
In the failed candidate (39c), the reduplicant contains a foot that is sepa-
rated by another foot from the left word edge. Since this violation-mark is
avoidable by less zealous copying, and since MAX-BR is low-ranked, the
first candidate wins. Gradient foot alignment is not crucial in accounting
for the minimal-word template.
Before leaving the topic of non-iterative footing, it is worth pointing out
a related empirical prediction that follows from eliminating gradient
alignment. Imagine a language with the non-iterative stress pattern
[aco('xar)T] or its moraic equivalent. The standard approach with gradi-
ent alignment posits the ranking [NoN-FIN(Ft)?ALIGN(Ft, R; Wd, R;
) > PARSE-Or]. Under this ranking, the final syllable is unfooted, and, since
every word must contain at least one foot, the sole foot is aligned as far to
the right as possible.
The iterative version of this stress pattern does not require gradient
alignment- see (a), (h) in Table I. But the non-iterative version is po-
tentially a problem for the theory sketched here. The intended output
[cara('aa)or] ties with its competitors *[ora('ac)cra] and *[cr('oa)ooaa], and
fares worse than *[('o)oxraoo], on the positional PARSE-5 constraints in
(33). The *LAPSE constraints in (31) are of no help either. Macedonian is
said to exemplify this stress pattern (Franks 1989, Hammond 1989, Halle
& Kenstowicz 1991), though again the inference that words contain only a
single foot is insecure. If the empirical caveats raised at the beginning of
this subsection can be resolved, so that the case for [aooacr('a)c] metrical
structure is on more solid footing, then the place to look for an analysis is
in extensions to the *LAPSE constraints in (31). For example, the associate
editor points out that a constraint against long lapses (66a) word-finally
would suffice (though see ?2 on locality and counting in phonological
constraints).
6.3.3 The location of main stress. Gradient alignment has also been used
to assign main stress to the rightmost or leftmost foot. In rhythmic stress
systems like those exemplified in Table I, main stress usually falls on the
leftmost or rightmost foot, which need not be in absolute word-initial or
word-final position. In prominence-driven stress systems, main stress falls
on the leftmost or rightmost heavy syllable, with no limit on how far it can
be displaced from the word edge. Standardly, minimal violation of
gradient ALIGN(Hd, Wd) (the erstwhile EDGEMOST)is the source of these
effects (McCarthy & Prince 1993b, Prince & Smolensky 1993).
Following Prince (1983), I have proposed that the End Rule constraints
(35) are responsible for the location of main stress, and that these con-
straints give a categorical advantage to the first or last foot. There are two
potential challenges to this view. First, prominence-driven stress systems
have sometimes been analysed with syllable-counting ALIGN(Hd, Wd)
constraints, and the End Rule constraints cannot directly reproduce this
effect. Second, foot-extrametricality phenomena (Hayes 1995) appear to
show a typical gradient pattern: main stress cannot appear on the final
foot, so it goes on the one next to it.
O T constraints are categorical 115
In general, prominence-driven stress systems can be analysed using the
categorical End Rule constraints, if the foot structure is properly under-
stood. Prince (1985) and Bakovic (1998) propose that prominence-driven
stress systems, which had sometimes been attributed to unbounded feet in
the past (Halle & Vergnaud 1978, Hayes 1980), actually involve binary
feet. Unlike rhythmic stress systems, though, feet are rather sparse in
prominence-driven stress: feet parse all the heavy syllables, and otherwise
they parse a pair of light syllables at the default edge. As in rhythmic stress
systems, the first or last foot is singled out for main stress. Schematically,
the stress patterns are like these:
SWP PARSE-5
B a. LL(H)LL(H)LL
b. (LL)(H)(LL)(H)(LL) ***!
27 SWP has been called Prokosch's Law (Prokosch 1939, Vennemann 1972), Obliga-
tory Branching (Hayes 1980, Hammond 1986) and STRESS-TO-WEIGHT (Fitzgerald
1997). Cf. Prince (1990).
1 6 John 7. McCarthy
depending on which of the positional PARSE-or constraints in (33) is ranked
higher.
The rankings just presented will derive the right foot-parsing. De-
ploying the main stress on the first or last foot is then just a matter of
whether ER-L or ER-R is ranked higher, as the following partial factorial
typology shows:
28 Two right-to-left trochaic versions of the foot-extrametricality pattern are also de-
scribed by Hayes (1995): 'the Grierson/Fairbanks stress rule' for Hindi (162) and
Paamese (178). These cases seem less convincing. There are serious problems es-
tablishing what the Hindi stress facts really are (Ohala 1977, Hayes 1995). The foot-
extrametricality analysis of Paamese is one approach to a rather tricky and not yet
fully understood problem in lexical conditioning of stress (cf. Goldsmith 1990:
215-216). Another example: Buckley (1994) proposes that the invisibility of initial
CVV syllables to stress in Kashaya is an effect of initial foot extrametricality. That
analysis presents many complications, however, that go well beyond the issue of
gradience.
O T constraints are categorical 117
foot from word-final position, must be ranked above gradient ALIGN-
(Hd(Wd), R; Wd, R; cr):
b. [(02)(02)(01)] *
c.
lo [(01)(02)(02)] *
With categorical constraints, if the final foot is skipped, then stress appears
on the initial foot instead. Indeed, it would seem that the categorical End
Rule constraints not only cannot produce the foot-extrametricality pat-
tern, which is attested, but also predict an unattested pattern: penult stress
in odd-parity words (e.g. [(02)(01)0]) contrasting with peninitial stress in
even-parity words (e.g. [(01 )(02)(02)]).
Both of these unwanted predictions turn out to be artefacts of the at-
tempt to reproduce the foot-extrametricality analysis. What's needed is a
way of obtaining the descriptive effects of foot extrametricality, though
not the construct itself. The key idea is that the putative extrametrical foot
is not there at all, so the output forms are actually odd-parity [(02)(01)0]
and even-parity [(02)(01)00]. Importantly, the even-parity word ends in a
lapse rather than a secondary-stressed foot. The responsible constraint is
not NoN-FIN(Hd(Wd)), but rather a similar constraint that replaces it,
NoN-FIN(Ft) (e.g. Kager 1999: 151):
(46) NON-FIN(Ft)
*Ft/ ]Wd
'Word-final feet are prohibited.'
118 John J. McCarthy
If NON-FIN(Ft) is ranked above *LAPSE, then the last two syllables will
remain unfooted in even-parity words:
a. [(02)(01)00] _
b. [(02)(02)(01)] *!
c. [(02)(01)(02)] *!
d. [(02)0(01)0)] _ _ _ _
NoN-FIN(Ft) rules out the full-parsing candidates (47b, c). The surviving
candidates (47a, d) have a stress lapse. With LAPSE-AT-END anywhere in
the hierarchy, this lapse must fall on the last two syllables. Since the last
two syllables are adjacent to both the peak and the end, there is no better
place for that lapse to go.
The winning candidate, (47a), has all of the observed descriptive effects
of foot extrametricality without the extrametrical foot. The extrametrical
foot is an analytic artefact, freely dispensed with when a better analysis
comes along. Indeed, in none of the foot-extrametricality languages do we
find reports of a word-final secondary stress as overt evidence of the ex-
trametrical foot.
This line of analysis provides a categorical treatment of the foot-extra-
metricality pattern. It also avoids the problem noted below (45); with
NoN-FIN(Ft) replacing NoN-FIN(Hd(Wd)), there is no way of getting (45c)
to win in a language with penult stress in odd-parity words.
Another type of foot extrametricality is also robustly exemplified in
Hayes (1995). This is extrametricality in clash. When moraic trochees
parse HLL sequences, they produce a stress clash. According to Hayes,
Bani-Hassan Arabic (366), Maithili (153), Manam (182) and Turkish
(262) declare the final foot to be extrametrical when it is in clash with the
penultimate foot: [(H)((LL))] (vs. clashless [L(LL)]).
Buckley (1998) presents an analysis of Manam that avoids foot extra-
metricality or its OT equivalent. In words of the form [(H)LL], there
is a final lapse because *CLASH disfavours the alternative, *[(H)('LL)].
In addition, WSP rules out the opposite way of resolving the clash,
*[H(LL)].
Even within the Hayes (1995) framework, foot extrametricality in clash
seems like a dubious move. In Hayes' general schema for extrametricality
rules (p. 58), the only context that can be mentioned is the right edge
of some domain, such as the word, and he notes that even this much
contextual information is redundant, because of factors like the Periph-
erality Condition (Harris 1983). A rule assigning extrametricality only
in clash is therefore a big leap in expressive power, and this ought to
encourage scepticism. In fact, there is good reason to be sceptical, since, as
we have just seen, a plausible alternative exists.
0 T constraints are categorical 1 19
Before closing out this section, two other potential cases of foot extra-
metricality should be mentioned. First, Hayes (1995: 262-263) very
briefly describes a right-to-left iambic stress pattern with foot extra-
metricality in clash. The languages, all related (two very closely), are
Javanese, Malay and Sarangani Manobo. Stress falls on the penult unless
it contains schwa, in which case stress falls on the ultima. The proposal is
that all syllables are heavy except those containing schwa and that feet are
iambic, with the final foot extrametrical in clash: [(H)(H)((H))],
[(H)(H)L], [(H)(LH)] and [(H)(LL)].
There are several reasons to doubt this analysis. Indonesian, a very close
relative of Javanese and Malay, is clearly trochaic (Cohn 1989, Cohn &
McCarthy 1994) and a variant of Sarangani Manobo is also trochaic
(Hayes 1995: 178-180). A feature of Indonesian and presumably these
other languages is a well-documented dispreference for stressed schwa
(Urbanczyk 1996, de Lacy 2002). Final stress in words like Javanese
[banar] 'correct' and [gatalan] 'itch' may simply reflect a bias toward
stressing closed syllables when stressed schwa is unavoidable.
Second, the assignment of primary stress in English should be men-
tioned as a potential counterexample to the claims made here, as Joe Pater
has pointed out. According to 'Schane's Rule' (Schane 1972), primary
stress goes on the rightmost non-final stressed syllable. This generalisa-
tion evolved into the Lexical Category Prominence Rule of Liberman &
Prince (1977), which assigns main stress to the final foot unless it is
monosyllabic (non-branching), in which case main stress goes on the
penultimate foot: (Aga)(mem)(non) vs. (Hali)(cair)(nassus).
This principle for assigning main stress falls well outside the typology
documented in Hayes (1995), and it is not surprising that it has received
little attention in Optimality Theory. Further problems, all well known,
are presented by disyllables, which follow other generalisations, by out-
right exceptions like Laidef6ged, and by the many forms that require a
highly opaque derivation to conform to the generalisation: commentary,
obligatory, naircolepsy, salamander, caterpillar, Arist6tle, puimpernickel,etc.
On the whole, then, English primary stress seems much more like a re-
search problem than a prima facie counterexample to categoricality.
6.4 Summary
Gradient alignment constraints are ubiquitous in analyses of metrical
phonology within OT. Nonetheless, the case for gradient alignment of feet
and word-heads is not persuasive. Directionally iterative foot-parsing has
been persuasively argued by Kager (2001) to reflect constraints on lapses
and clashes rather than gradient constraints on foot alignment. *LAPSE and
its congeners provide a better match between prediction and observation
than the gradient foot-parsing constraints ALLFTL/R. Nor does the evi-
dence from one-foot-per-word and main-stress phenomena provide sup-
port for gradient constraints; the principal patterns can be analysed with
simple categorical constraints modelled after the End Rule of Prince (1983).
120 John j. McCarthy
7 Alignment in autosegmental phonology
7.1 Introduction
Perhaps the last bastion of gradient alignment is the autosegmental pho-
nology of features and tones. Gradient alignment constraints have been
applied to three kinds of autosegmental processes, docking, flop and
spreading. In docking, a featural or tonal morpheme that is not associated
with a segment or syllable in underlying representation becomes linked in
the output. The gradience of standard alignment constraints has been
used to produce docking that is near a word edge, even if it does not lie
exactly at the edge. In flop, a feature or tone is reassociated onto a different
segment or syllable. With gradient alignment, the flopped feature or tone
may be attracted toward a word edge without actually reaching it. And in
spreading, an element that is linked to one segment or syllable in under-
lying representation extends its domain in one or both directions, often
over an unbounded distance. Gradient alignment has been identified as
the constraint that compels spreading by forcing a feature or tone to ex-
tend its reach toward a word edge, even when the edge itself is unattain-
able for other reasons.
It is obviously impossible within the scope of this article to locate and
reanalyse every known application of gradient alignment to docking, flop
and spreading phenomena. Since that cannot be done, I will pursue the
more modest goal of proposing alternatives, illustrating them with a few
examples, and highlighting any empirical differences from gradient
alignment.
7.2 Docking
Docking is what happens to a floating feature or tone. Floating elements
are typically affixal; like other affixes, they tend toward one edge of the
word or the other, but may be displaced from it for phonological reasons.
Two examples are given in (48); see Akinlabi (1996), Zoll (1996) and
Piggott (2000) for general discussion.
b.josa *
b.pokYo
7.3 Flop
Flop is the reassociation of a feature or tone from one segment or syllable
to another. In the flop processes of interest here, the feature or tone is
attracted toward the word edge or the head foot, though it may not always
make it there.
In Cuzco Quechua (Parker & Weber 1996, Parker 1997, MacEachern
1999), the three-way lexical contrast plain/aspirated/glottalised is possible
only in the leftmost non-coda obstruent stop of the root.
31 In their analysis of Chaha's close relative Ennemor, Hetzron & Habte (1966: 28)
observe: 'la seule chose concrete qu'on peut noter au sujet des verbes quadriliteres
est que g y semble resister a la labialisation plus que les autres consonnes'.
124 3ohn Y. McCarthy
(51) Cuzco Quechua laryngeal contrasts
qhata ' mountainside'
hap?iy 'to light (a fire)'
lap2ay ' to lick up (said of dog)'
warak?a 'sling made of wood'
The restriction of glottalisation and aspiration to obstruent stop onsets is
unremarkable; the responsible markedness constraints are undominated
in Cuzco Quechua. The important thing for present purposes is that
words like Parker & Weber's hypothetical *[poq?a] are prohibited; it
would have to be realised as [p?oqa] instead.
This looks like an obvious case for gradient alignment (Parker 1997).
ALIGN(Laryngeal, L; Stem, L; Seg) will correctly favour [p oqa] over
*[poq?a]. There is an alternative, however: the same result can be ob-
tained from a categorical COINCIDEconstraint requiring aspirated and
glottalised consonants to be literally initial in the root: COINCIDE
(Laryngeal, Root, L). That constraint, ranked above the faithfulness
constraint NoFLOP (McCarthy 2002a), which bans featural reassociation,
will also correctly favour [p?oqa] over *[poq?a]:
b. poq?a _
b. waraka *!
c. w araka
b. tambalalitsa *! **
c. tambalalitsa *****!
The alternative advanced here says that H is attracted to the final syllable
by COINCIDE(H,Stem, R). This constraint rules out the same candidates
non-gradiently.32
Now, imagine a language identical to Chichewa except that the ranking
of MAx(tone) and ALIGN(H, R; Stem, R; a) is reversed. In this hypo-
thetical language, all high tones should pile up at the right edge of the
stem, so every stem will end in a sequence of high-toned syllables equal in
length to the number of high tones in the input: /i-ma-kii-tambalal-a/
-* [imakut'ambalAal]. Although this pattern is readily predicted by gradi-
ent alignment, there is no way to get it using categorical COINCIDE
(H, Stem, R). No categorical constraint will, for example, induce the first
H to move from the initial syllable to the preantepenult, simply because
the preantepenult is closer to the final or head syllable.
32
In certain tenses, Chichewa flops the high tone onto the penult. In that case, Co-
[NCIDE(H, Hd(Wd)) is the active constraint, under the assumption that Chichewa
has a right-aligned trochaic foot, as suggested by reduplicated forms like [chikulu-
piriro-riro] 'real faith' and parallels elsewhere in Bantu (Myers 1987).
T constraints are categorical 127
0
I know of no language that displays the tonal or featural piling-up effect
that is predicted by gradient alignment. If indeed no such language exists,
then it must count against the gradient theory and in favour of categori-
cality, since the existence of such patterns is an unavoidable prediction of
gradient constraints aligning tones and features. The absence of such pile-
up effects in phonology is all the more striking because they have been well
documented in work on OT syntax (Legendre 1999, 2000, Gouskova
2001, Grimshaw 2001).
There is a broader moral to be drawn here, one that emerges from Zoll's
(1996: 141) discussion of a different but related range of examples:
'licensing of marked structure never involves an injunction to be as close
to a strong position as possible. Rather, licensing always constitutes an all-
or-nothing proposition whereby marked structures are licit in licensed
positions but ill-formed everywhere else.' If something like this remark is
correct, as I have argued here, then categorical constraints truly are the
right way to deal with positional markedness restrictions on features and
tones.
7.4 Spreading
In Kirchner (1993), Smolensky (1993), Archangeli & Pulleyblank (1994a),
Cole & Kisseberth (1995), Pulleyblank (1996) and much other work,
gradient alignment constraints are assigned primary responsibility for
autosegmental spreading of features and tones. Imagine, for example, a
language where all H tones spread rightward to the end of the word, ex-
cept that the OCP prevents any H from spreading onto a syllable that
precedes another H. (This is approximately Shona (Myers 1987, 1997),
with low-toned syllables analysed as toneless.) The gradient constraint
ALIGN(H, R; Wd, R; a) ensures that each H tone maximises its spreading
domain by counting the syllables between the right edge of each high-tone
domain and the end of the word, as shown in (56).
(56) Gradient ALIGN(H, R; Wd, R; a) in tone spreading
H H OCP AilIGN(H,Wd)
IW a. H H
N-, N-,
b. H H
C. H H *!
aaaooaaa
d. H H
I N-a
128 John J. McCarthy
ALIGN'S behaviour in this example is virtually identical to the behaviour of
ALLFTR in (3). Both constraints gradiently evaluate each instance of a
tone or foot for its distance in syllables from the right word edge. Not just
the rightmost tone but all tones must be checked to ensure the victory of
(56a) over (56d). Furthermore, just like (3), the winning candidate (56a)
abundantly violates the responsible alignment constraint; it simply per-
forms better on alignment than the viable alternatives.
The hypothetical example (56) was chosen because it fairly represents
some of the problems that a theory of spreading processes must address.
As I showed in the discussion of (3) in ?2, there is no way to use categorical
constraints to mimic directly the effects of gradient alignment in cases like
this. I therefore explore a more distant alternative that is still capable of
making the crucial distinctions in (56) and similar cases.
The constraints MATCH-R(F) and MATCH-L(F) demand agreement in
F-value between a segment or syllable and any preceding/following seg-
ment (in the case of features) or syllable (in the case of tones). They are de-
fined as follows, where F is a feature or tone and x is a segment or syllable:
X_-F/ XF
b. MATCH-L(F)
Wd
X-F x/\x
/ XF
33MATCH evaluates the candidates in (58) under the assumption that a toneless syl-
lable, which is effectively low-toned, mismatches a high-toned syllable. See Arch-
angeli & Pulleyblank (1994b: 105-106) on why such an assumption is a necessary
concomitant of underspecified representations.
O T constraints are categorical 129
(58) Tableau (56) with categorical MATCH-R(H)
H H OCP MATCH-R(H)
W a. H H *
N-1 N-1
b. H H
C. H H *!
I v-
d. H H
8 Conclusion
In this article, I have argued for a particular view of how OT constraints
work. Constraints militate against structural configurations (markedness
constraints) or non-identical mappings (faithfulness constraints). Con-
straints do so categorically: it is sufficient for any constraint to assign one
violation-mark for each instance of the marked structure or unfaithful
mapping in the candidate under evaluation. The definitional frame of a
constraint, then, is 'Assign one violation-mark for every k meeting con-
dition C', where x is an output structure or a non-identical mapping. No
greater complexity of constraint definition is required or desirable.
This proposal stands at odds with a widely accepted view of OT
constraints, that some are categorical and some are gradient. Gradient
constraints assess goodness of fit over some range. In a review of the
literature, two main types of gradience were identified, those constraints
where the range is bounded and those where it is unbounded. Bounded
gradience is met with sporadically in the OT literature, in certain con-
straints on hierarchies, scales and classes. Bounded gradience is unneces-
sary; any boundedly gradient constraint can, and in some cases must, be
replaced by a set of categorical constraints (?3). Unbounded gradience is
in all likelihood limited to alignment constraints, which have been impor-
tant in analysing infixation, stress and various autosegmental processes.
I have argued (??5-7) that gradient alignment constraints can be dis-
pensed with because their effects are subsumed by other, categorical
constraints, many of which have been previously proposed and indepen-
dently motivated. Moreover, in some cases gradient alignment predicts
patterns that are not observed, and these unwanted predictions can be
avoided by adopting categorical constraints instead.
REFERENCES