BAKER - Thematic Roles and Syntactic Structure
BAKER - Thematic Roles and Syntactic Structure
BAKER - Thematic Roles and Syntactic Structure
Edited by
LILIANE HAEGEMAN
T H E M AT I C R O L E S A N D S Y N TA C T I C S T R U C T U R E *
One central task for any theory of grammar is to solve the so-called "linking
problem": the problem of discovering regularities in how the participants
of an event are expressed in surface grammatical forms and explaining those
regularities.
Suppose that one adopts a broadly Chomskyan perspective, in which there
is a distinction between the language faculty and other cognitive facul
ties, including what Chomsky has recently called the "Conceptual
Intensional system". Then there must in principle be at least three stages
in this association that need to be understood. First, there is the nonlinguistic
stage of conceptualizing a particular event. · For example, while all of the
participants in an event may be affected by the event in some way or another,
human cognizers typically focus on one or the other of those changes as
being particularly salient or relevant to their interests. This participant is
taken to be the "theme" or "patient" of the event, perhaps in some kind
of nonlingui stic conceptual representation, such as the one developed by
lackendoff ( 1 983 , 1 990b) . Second, this conceptual/thematic representa
tion is associated with a linguistic representation in which the entity seen
as the patient of the event is represented as (say) an NP that is the direct
object of the verb that expresses what kind of an event it was. This is the
interface between language and the conceptual system. Finally, there i s
the possibility of adjusting this representation internally to the language
system, by way of movements, chain formations, Case assignment processes,
or whatever other purely syntactic processes there may be. For example, the
NP that represents the theme and starts out as the direct object of the
verb may become the subject if there is no other subject in the linguistic
representation, either because there was no agent in the conceptual repre
sentation (as with an unaccusative verb), or because it was suppressed (as
with a passive verb).
Since there tre at least these three stages between an event and a surface
linguistic description of it, there is room for a good deal of complexity in
theory and analysis. Therefore, most syntactic theorists assume that at
least one step in the association is relatively trivial. In part, this is a tactical
move, an effort to cut down the number of analytic choices that a theory
must make in order to develop an analysis of any particular phenomena.
However, the need for restrictions also seems to be empirically motivated
by the fact that there are in fact important linking regularities both within
73
Let us begin with the vague UTAH in ( 1 ) and see how one could evaluate
its validity. T, do this, it is helpful to break the linking problem down
into two parts: first, there is the matter of deciding which of the partici
pants in a given event is to be expressed as the subject of the sentence;
second, there is the matter of deciding which of the remaining parti ci
pants is to be expressed as the direct object (if any). Any remaining
participants are then typically expressed as oblique NPs that appear with
adpositions or Case markers that express their thematic role rather directly. 3
In this section, then, I begin by considering the matter of subject choice
in two-place verbs. This is the least controversial aspect of linking theory:
76 M A R K C. B A K E R
there is wide agreement that agents rather than themes are chosen a s subjects
in most languages. However, there is still some disagreement as to whether
this rule of thumb is valid for all clauses and all languages. Moreover,
looking at this issue w ith some care will also give us a valuable model of
how to proceed when we come to the much more controversial question
of how internal arguments are linked.
2. 1 . English
b. * [xp John hit] the table and [xp (so)] did the chair.
* John said he would hit the table, [xp John hit] I guess - did it.
Finally, the agent-subject "has prominence" over the patient-object in a
variety of ways involving anaphora, coreference, and quantification . Thus,
pronouns and anaphors contained in the object can be referentially depen
de n t on the subject, but not vice versa.
(4) a. J ohnj washed himselfj
Johnj washed pictures of himselfj
Every manj washed hisj c ar.
* Hej washed John 'sj car. (ou t by Condi tion C)
b. * Heselfj washed Johnj. (out by Cond ition A)
* Friends of himselfj washed Johnj• (out by Condi tion A)
* Hisj friends w ashed every manj. (out by Weak Crossover)
John 'sj friends washed himj.
THEMATIC ROLES A N D SYNTACTIC STRUCTURE 77
Facts like these from English and similar languages motivate the twin claims
of P&P linking theory built into the UTAH: (i) that agents are always (under
lying) subjects, and (ii) "subject" is a structural notion - it is a position
in phrase structure that is outside the VP (see (3» and hence "higher than"
(c-commanding) the position of the patient/object (see (4» .
The one area of controversy regarding subject selection i n English
concerns predicates describing psychological states. In this particular
semantic domain, there do seem to be (nearly-)synonymous verbs that
have opposite linking patterns, as shown in (5).
(5) a. John likes long novels.
John fears dogs
John worries about the ozone layer
b. Long novels please John.
Dogs frighten John.
The ozone layer worries John.
The general Chomskian framework allows for three possible interpreta
tions of this fact: (i) the data in (5) shows that the UTAH is false, and
different predicates require different linking patterns as an idiosyncratic
lexical property; (ii) the (a) and (b) sentences have similar underlying con
figurations, but at least one of them (probably (b» involves a nontrivial
syntactic derivation; (iii) the thematic roles in the (b) sentenCes are actually
different from those in the (a) sentences.
In fact, no one has advocated option (i) in the recent literature, as far
as I know. It seems wrong to infer from the fact that a restricted class of
predicates appear to be idiosyncratic that all predicates are idiosyncratic.
On this view it would be hard to explain the observation that most verbs
work predictably, apart from this semantically coherent domain.
Option (ii) is advocated in Belletti and Rizzi's well-known article ( 1 988),
as well as much work in the Relational Grammar tradition. Indeed, there
are some important things to be said in its favor: Belletti and Rizzi show
that sentences like (5b) have a number of syntactic peculiarities in Italian
(and also English) that can be explained if they are derived by moving
the "theme" argument into the subject position syntactically. Moreover, their
analysis is compatible with a somewhat weakened version of the UTAH,
as they poin i out. I return to a brief discussion of Belletti and Rizzi 's
theory in section 4.2. 1 below.
However, the correct option seems to be option (iii): denying that long
no vels and John have the same a-roles in both (5a) and (5b). This non
syntactic approach to the problem of psych verbs is defended by Pesetsky
( 1 987, 1 995) and Dowty ( 1 99 1 : 579-580, 586-587). Dowty in particular
sketches an explanation for why only this class of verbs seems to be so
variable in its linking properties: neither participant in the event is an
obvious choice for an agent or a patient. John is an animate and sentient
MARK C. BAKER
not a very felicitous term; a more accurate one would b e Levin and
Rappaport-Hovav 's ( 1 995) "internal cause", which includes agency as a
special case.) Indeed, it is a property of John's psychological make-up -
though not necessarily his will - that causes him to respond in a partic
ular way to dogs; indeed, we might be tempted to consider him a coward.
If these remarks are correct, then fear and frighten are both agent-theme
verbs, once these notions are understood in a broad enough way. The
apparent alternation comes from the fact that there are two way s these
notions can sensibly apply to describe a certain eventuality, although what
exactly is expressed is subtly different depending on the choice. Finally,
it is a consequence of this approach that fear should behave like an ordinary
transitive verb in most respects, and this seems to be true (Belletti and Rizzi
1 988, Grimshaw 1 990). 5
2.2. Ergativity
Next, let us consider to what degree the principles of subject choice seen
in English are universal. Suppose, contrary to the UTAH, that the associ
ation between thematic roles and syntactic positions were arbitrary and
variable, either because agents are not necessarily subjects, or because
subjects are not necessarily "extemal'� positions. Then one would expect
that some naturai languages would make use of exactly. the opposite
association, in which the agent is associated with the VP-internal object
position, and the patient/theme is associated with the structural subject
position, thereby having prominence over the agent. Marantz ( 1 984) and
Levin ( 1 983) clai m that this logical possibility is attested in so-called
"deep ergative" languages, including Dyirbal (Dixon 1 972) and a dialect
of Inuktitut. Similarly, Dowty ( 1 99 1 : 5 8 1 -582) suggests that the basic
rules for associating semantic arguments with grammatical functions are
reversed in some ergative languages,6 pointing out that if so, "this provides
an extremely strong reason why we should not try to collapse the notion
of P[roto]-Agent with grammatical subject and P[roto]-Patient with gram
matical object . . . , or adopt a theory which necessarily correlates them
in this unique way." Thus, whether or not there are any truly deep ergative
languages is a crucial issue for the UTAH.
It is helpfu V to begin by imagining what a deep ergative language would
look like. If everything else were held constant, the allowable syntactic
dependencies i n such a language should be exactly the opposite of those
in a language like English. For example, the agent and the verb should form
a unit to the exclusion of the patient for purp oses of word order, VP deletion,
VP fronting, and the like. Furthermore, the agent in such a language should
be able to be referentially dependent on the patient, but not vice versa -
the reverse of the typical English pattern in (4). F�nal1y, in nonfinite clauses
headed by a morphologically simple transitive verb, it should be the patient
gO MARK C. B A K E R
interpreted the other w ay, (7b) would b e a n instance o f basic OSV word
order; this is a rare or unattested type. (Note, however, that the unmarked
word order is the opposite in Greenlandic, as shown in (8b).) Second,
Dyirbal has control(-like) purposive constructions in which there is a special
subordinate verb form and a missing argument that is understood as co
referential with an argument of the first clause. Significantly, if the verb
in the embedded clause is transitive, the missing NP must be the absolu
tive theme, not the ergative agent. Thus, Dyirbal can express the equivalent
of (9b), but not (9a), whereas the facts in English are the opposite (see Dixon
( 1 972) for the actual Dyirbal examples):
(9) a. The manj climbed up in order -j to see the bird.
OK English, * Dyirbal
b. The manj climbed up in order for the bird to see -i.
OK Dyirbal, *English
Similar patterns pervade the grammar of Dyirbal: for example, only abso
lutive arguments can be "shared" in coordination-like clause chaining
constructions, and only absolutive arguments can be understood as the
head of a relative construction. This last constraint is found in other ergative
languages as well , including Greenlandic (Bittner and Hale 1 996). Finally,
B ittner and Hale ( 1 996: 5 65-568) show that in Greenlandic VP operators
such as negation take scope over the ergative Case agent but not the abso
lutive Case patient; again, this is the exact reverse of the pattern found in
English.
However, while some grammatical properties seem to reverse in these
languages, others do not, as pointed out in Dixon ( 1 979, 1 994: sec 5 .3).
Perhaps the most significant is that the patient phrase can contain an anaphor
that is referentially dependent on the agent phrase as in ( 1 0), but the agent
phrase cannot contai n an anaphor that is dependent on the patient (see
B ittner ( 1 994: ch. 4), Bittner and Hale ( 1 996) for Greenlandic).
( 1 0) Juuna-p qimmi-ni nlflslp-p-a-i. GREENLANDIC
Juuna-erg dog-p1l3srefl feed-ind-trans-3s/3p
"Juunai fed hisj own dogs."
Di xon also ,?entions that in control phenomena involving complement
taking verbs with meanings like "can", "try", "want" and "begin", it is
always the agent argument of the embedded verb that is controlled, not
the theme argument. Again, thi s holds in Greenlandic, with the minor
complication that the matrix verb generally shows up attached to the
embedded verb at PF (B ittner 1 994).
Moreover, B aker ( 1 988a: 427-428) extends Di xon 's list of invariant
properties by pointing out that compounding and incorporation phenomena
work the same way in both accusative languages and ergative ones (see also
82 MARK C. BAKER
Mithun ( 1 984» . I t i s well -known that the theme argument o f the verb root
can appear in English deverbal compounds, but the agent argument cannot:
one can have dish-washing , but not *husband-wash ing (of dishes).
Presumably this is another reflex of the fact that themes are objects and
appear structurally close to the verb in English, but agents are not (Roeper
and Siegel 1 978, Sproat 1 98 5 , Di Sciullo and Williams 1 987, Grimshaw
1 990). If so, one would expect thi s pattern to reverse in a deep ergative
language. However, it does not; i ncorporation is common in Inuit, but the
incorporated argument is always understood as the patient, never as the
agent:
(1 1) Juuna alla-mik ilinniartitsisu-siur-p-u-q.
Juuna other-jnstr teacher-seek-ind-intrans-3sS
(Bittner 1 994: 67)
"Juuna is looking for another teacher." NOT: "Another teacher
is looking for Juuna."
See McKay ( 1 975) for similar patterns in Rembarrnga, a morphologically
ergati ve language re lated to Dyirbal .
On balance, then , we see that some syntactic phenomena reverse in
ergati ve l anguages, but others do not. Dixon ( 1 994: 233) criticizes Marantz
( 1 984) and similar approaches to "deep ergativity" for their inability to
capture this fact in a natural way. Instead, he claims that "the universal
category of subject . . . plays a role in the gramm ar of every l angu age. "
However, he distingu ishes a second type of gramm atical function, called
a "pivot", which does differ significantly across languages: Dyirbal treats
patients of transitive verbs together with the sole argument of intran sitive
verbs as pivots, whereas English treats intransitive subjects and the agents
of transitive verbs as pivots. Pivots are distinct from subjects and belong
to another l ayer of grammatical description, one concerned with the
combining of clauses and coreference across clauses (see also Foley and
Van Val in ( 1 984» .
In fact, Dixon 's insights translate rather directly into what has become
the standard approach to ergativ ity in the P&P framework. 7 For reasons
that are quite independent of ergativity, it has become standard to assume
that the agent argument of a transitive verb in Engl ish is a subject i n (at
least) two ways: it is base-generated as the specifier of a VP projection
where it is directly theta-marked; i t then raises to the spec i fier of an
Inflectional head to receive (or check) its nominative Case. Thus, the agent
is both the subject of VP and the subject of IP. However, these two di stinct
senses of subject may diverge, resulting in a "deep ergative" language. Thus,
suppose that the basic projection of arguments is the same i n l anguages
like Dyirbal or Inu it, but the verbs in these languages cannot license
accusative Case on the underlying object. Then , it is the patient argument
T H E M A T I C R O LE S A N D S Y N T A C T I C S T R U C T U R E 83
of the verb, not the agent, that must move to the specifier position of IP
to receive/check nominative Case and trigger agreement on I (Bok-Bennema
1 99 1 , Campana 1 992, Murasugi 1 992, Bittner 1 994, Bittner and Hale
1 996).8 The agent NP, on the other hand, remains in the specifier of VP
and receives ergative Case by some other means (researchers vary on the
exact mechanisms here). Simple versions of the two basic clause struc
tures are compared in ( 1 2), where I leave open the possibility that there
are additional functional categorie.s and/or a more complex VP-intemal
-
�rucrure.
( 1 2) a. ENGLISH
IP
�
NP· I'
A
the man
�
VP
I �
NOM NP V'
I �
� V NP
I �
hit the woman
I I
ACC
b. DYIRBAUlNUIT
IP
�
NPi I'
� �
the woman Infli VP
1
________ 1 �
NOM NP V'
� �
the man V NP
f
ERG I I
hit tj
Several of the peculiar-looking facts of ergative languages now follow
immediately. For example, the movement of the patient past the subject
shows up overtly in Dyirbal, in the form of unmarked Patient-Agent-Verb
order. Similarly, it follows from these representations that agents are outside
the scope of VP operators in English, whereas patients are in Inuit. Finally,
the purposive inflection in Dyirbal ca� be analyzed as a special form of
M A R K C . B A KER
Infl that licenses "null Case", and hence its spec i fier must b e a n u ll element,
perhaps PRO. This null element is then anaphorically dependent on the
matri x clause. Plausibly rel ative clauses in Dyirbal and Inuit, and top ic
chaining constructions in Dyirbal can be analyzed in a more or less similar
way (see Bittner and Hale ( 1 996) for discussion �f participial relatives).9
However, the structural reversal is not total on this account. In particular,
the inner structure of VP is the same in both ergative and accusative lan
guages on th i s approach. Thu s , for l ingu istic phenomena that have the
b a sic VP as their domain, deep ergative languages and accusative languages
are expec ted to work very much the same. For example, the agent c
comm ands the patient before movement; hence the patient can be
anaphorical ly dependent on the agent but not vice versa (see ( 1 0) ) . (The
mo v ement of the patient past the agent does not change this relationshi p
if that is an A-bar movement; see note 9.) Similarly, we explain why noun
incorporation pattern s are the same in ergative languages as in others :
noun incorporation is an alternative to movement to receive Case ( B aker
1 988a) , and it takes place entirely internal to the VP. Within thi s VP, the
patient is in a close enough relationship to incorporate into the verb, but
the agent is not. Finally, it makes sense that control of complement clauses
al so acts jn the same way, since thi s kind of con trol i s known to be sen
sitive to the themati c roles defined over . VP - internal configurations (see
Bi ttner ( 1 994) for some disc u ss i on) .
I n conc lusion, this P&P approach can b e seen a s a formal development
of D i xon 's i n tu i t i o n that "s u bje c ts" are the same in all languages, while
" pivots" vary. The P&P equiv alent of Di xon 's "subject" i s "subject of
VP", whereas the equivalent of "pivot" i s "subject of IP." Crucially, this
theory not only accounts for the fact that "subject properties" seem mi xed
in ergatjve languages, but it gives insight into exactly what mixtures one
fi n d s . Roughly , e rgati ve languages differ from accu sative languages in
thei r A-bar processes (quantifier scope, relative clauses, topics) and Case
rel ated morphol ogy, but are simi lar in terrr13 of A-processes (an a phora ,
inco rp orat i on , complement control). If this approach is correct, it shows that
ergative languages are not counter-examples to the UTAH after al l. Rather,
the bas i c projection of argu ments is identic a l in the two kinds of lan
guages, and it is s ubsequent move ment processes t h a t diffe r.
1 65- 1 72). Thu s, there is not and probably should not b e a fu lly un ified
theory of nonconfigurational ity. l l However, if this general approach is on
the right track, then the UTAH does apply to these languages, just as much
as it does to ergative and accusative-configurational ones, wi th observ
able empirical consequences.
Having confirmed that the choice o f the subject argument is determ ined
by something like the UTAH, let us tum to direct objects. Here we will
be particularly concerned with three-argu ment verbs, and assume that one
argument (the causer of the event) has already been chosen as the subject.
The crucial question , then, is which of the other two arguments is expressed
as the direct object of the verb, and whether the facts in this domain are
compati ble with the UTAH. The remaining argument of the verb is then
typical ly express�d as the object of a semantically appropriate adposition
or semantic Case marker.
In the matter of subject selection, there is l i ttle controversy that the
causer-to-subject rule applies to most verbs in many l anguages; the only
questions are whether this rule should be extended to all predicates and
all languages. However, in the domain of object selection there is much less
agreement as to what the basic rule is. Part of the reason for thi s lack of
consensus is the fact that a substantial percentage of the three-argument
verbs in Engl ish al ternate as to which argu ment shows up as the direct
object. Two particularly important and well-studied alternations are the
dative al ternation in ( 1 3) and the locative alternation in ( 1 4) .
( 1 3) a . I gave the c andy to the children.
b. I gave the children the candy.
Even those three argument verbs that do not alternate give l i ttle guidance
as to which linking pattern should be taken as basic. Thus, donate appears
only in the ( 1 3a) frame, owe appears only in the ( 1 3b) frame, pour only
in the ( 1 4a) frame , and fill only in the ( 1 4b) frame. Moreover, superficial
comparison w i th other languages simply adds to the confusion. For example,
French has only the equivalent of ( 1 3 a) , while Sesotho has o n l y the equ i v
alent o f ( 1 3 b). 1 2 Given this situation, practi cally every i m aginable view
has adherents: some claim that ( 1 3b) i s derived from ( 1 3 a) (Larson ( 1 988,
1 990); also Baker ( 1 988a), and much work in classical Relational Gram mar);
some believe that ( 1 3a) is derived from ( 1 3b) (Dryer 1 987 , Kiparsky 1 987,
Aoun and Li 1 989); many believe that both ( 1 3a) and ( 1 3b) are base gen
erated (Oerhle 1 975, lackendoff 1 990a, Speas 1 990, Dowty 1 99 1 , Collins
T H E M AT I C R O L E S A N D S Y N T A C T I C S T R U C T U R E 81
and Thrainsson 1 996). Moreover, Dryer ( 1 987) argues that different lan
guages make different choices in this respect: in some languages the
equivalent of ( 1 3a) is basic, and in others the equivalent of ( 1 3b) is basic
- a parameter of variation that he explicitly compares to the issue of
ergativity. Indeed, the only thing that this entire range of researchers agrees
on is that whatever account one has, it should be essentially the same for
both the dative alternation and the locative alternation. 1 3
Before considering the data, it is useful to survey once again the logical
possibilities. Two approaches to these alternations are compatible with the
UTAH. The first is to say that one or the other of the linking patterns in
( 1 3 ) and ( 1 4) is basic, and the other is derived from it by some kind of
movement. The second is to claim that the thematic roles in the (b) sen
tences are subtly different from those in the (a) sentences, and this justifies
constructing two different syntactic structures (see Speas ( 1 990), among
others). The solution that would refute the UTAH is one which said that
the (a) and (b) sentences do not differ in thematic roles but are both base
generated as a result of different subcategorization frames associated with
the verbs in a partially idiosyncratic manner; this was the standard view
in the late 1 970s and early 1 980s (see also lackendoff ( 1 990a» . In fact, I
will argue that the dative alternation and the locative alternation are quite
di fferent in these respects, thereby disagreeing with the standard view.
The dative alternation is a result of movement, with sentences like ( 1 3b)
being derived from underlying structures like ( 1 3a) by a combination of
P-incorporation and NP-movement. In contrast, the locative alternation
results from two different conceptions of the event in question: one in which
the hay is seen as primarily affected, and one in which the truck is seen
as primarily affected (Rappaport and Levin 1 985, Pinker 1 989, Dowty 1 99 1 ).
Once the viewpoint is picked, the affected argument (Le., the theme) is con
sistently generated as the direct object. If this is correct, then the minimal
contrast between the two alternations illustrates elegantly the two options
allowed by the UTAH, while giving reason to think that it is correct to
rule out the intermediate option. Moreover, I will survey some evidence that
suggests that the same is true in other, superficially different languages,
thereby supporting the idea that the basic object linking rule is universal,
contra Dryer ( 1 987).
r
3. 1 . Objects in English
Let us begin with English, considering first. the semantic side of the issue.
The crucial question is whether (1 3a) and ( 1 3b) mean the same thing or not,
and similarly for ( 1 4a) and ( 1 4b). The answer to this question depends in
part on how fussy one wants to be. On a crude level , the (a) and (b) sen
tences can often be used to describe . the same events; in this sense they
are synonymous. On a very detailed level, the (a) and (b) sentences differ
88 M A R K C. B A KER
at least with respect to matters of topic and focus, so that they do not always
sound equally good in every discourse environment. However, the crucial
matter w i th respect to the UTAH is whether the two sentences differ specif
ically in terms of their thematic roles, where this is a proper subpart of a
c o mpl e te semantic and pragmatic charac teri zation. In fact, I bel ieve that
at this mediu m-grai n level of description ( 1 4a) and ( 1 4b) differ, but ( 1 3a)
and ( 1 3b) do not.
Take first the locative alternation. There is a clear intuition that the object
argument in both versions of ( 1 4) i s "totally affected": in ( 1 4a), all the
hay is loaded into the truck, but the truck need not be completely loaded,
whereas in ( 1 4b) the truck is completely loaded, but there may be unloaded
hay left over. In both cases, then, the NP expressed as the direct object i s
seen a s undergoing a change o f state (Pinker 1 989) and thus "measures"
the progress of the event named by the verb (Tenny 1 994). Dowty ( 1 99 1 :
59 1 -592) brings out thi s intu i tion by invoking the fact that verbs which
are normally accomplishments aspectually often become activities when
their theme argument is a bare plural or mass noun. Now, in locative
alternation constructions, the argument that i s expressed as the direct object
determines the aspectual quality of the whole clause; whether the oblique
argument is determined or not has no effec t , as shown by the in an hourlfor
an hour test in ( 1 5 ) (see Dowty 1 979).
Here dative shift seems to have n o effect on the judgments: whether stories
is a bare plural or is determined decides whether the event is an activity
or an accomplishment, and the status of children is irrelevant in both syn
tactic frames. Thus, by parity of reasoning, stories is the incremental theme
in both versions of the dative alternation. Therefore we fail to find evidence
that the two differ in their thematic roles.
In fact, the literature that argues for a base-generation account of the
dative alternation does not claim that the two members of the alternation
differ in delimitedness per see Instead, researchers typically try to bring
out the intuition that the "recipient" is affected by the action in ( 1 3b) in a
somewhat different way. Essentially, the idea is that the transfer of pos
session involved in these kinds of verbs must succeed in the double object
frame, whereas it need not in the NP-PP frame. Thus, ( 1 7b) suggests that
the children actually learned some amount of French more strongly than
( 1 7 a). Similarly ( 1 8b) suggests that the ball actually reached Bill, whereas
( 1 8a) does not, and ( 1 9b) is weird because the dead lover cannot perceive
the song. 1 4
Thus, the (b) sentences suggest that the recipients undergo a kind of change
of state in that they come to possess something (knowledge of French,
the ball, an experience of a song). While I agree that there is something
to this judgment, I think it must be stated at the level of "suggests," rather
than "asserts" or "implies". Thus, the sentences in (20) may be stylisti
cally awkward, but they do not feel to me like contradictions and I can
imagine finding them in texts. 15
(20) a. I taught the children French, but they didn 't learn it at all.
b. I threw John the ball, but it didn 't reach him because of the
stroqg wind.
c. Mary sang her lover a song, but he didn 't hear it because he
had just died.
I n addition, there are many syntactic tests that confirm . that the dative
a lternation has a different status than the locative alternation. I will briefly
survey a variety of these. Their common property is that they distinguish
the shifted benefactive/goal argument of a dative verb from the theme
argument of a dative v erb and the direct object of either version of the
locative alternation. My claim then is that these tests show that the theme
of a dative verb, and the objects of both versions of the locative alterna
tion are underlying direct objects, but benefactive/goal arguments are not. 16
However, sp ace l i m i tations perm i t me to give only very brief di scussions
of the theoretical underpinnings of the tests.
The first test involves adjectival secondary predication. Secondary pred
icates obey a c-command condition, such that they must both c-command
and be c-commanded by their subjects (Williams 1 980: n. 1 ) . As a result,
depictive and resultative APs can be predicated of the direct object, but
not the object of a PP, even when this would be pragmatically plausible:
Now locative alternation verbs work just as one would expect on the basis
of the i r su p erficial syntax: an AP c an be predicated of the direct object
but not of the object of the PP, regardless of which expresses the location
and which the materi al (Williams 1 980: 204) . 1 7
,
I I
� �
Now NP" does not c-command AP from its base position, and it is not c
com manded by AP in its derived position. Hence, the goal NP is never in
the configuration it needs to be to be the subject of the secondary predi
cate. In contrast, the theme NP is generated as the specifier of VP and
IS
remains there, so it can be the subject of a secondary predicate. I n this
way, we eliminate William � ' stipulation, and motivate a derivational
approach to the dative shift. Moreover, the fact that the wagon in (22d)
92 MARK C. BAKER
(26) Which perfume do you think I should give t to/buy t for Mary?
More interestingly, it is also perfectly grammatical to extract the direct object
from ei ther version of a locative alternation:
(27) a. Which boxes do you think I should l oad t onto the truck?
b. Which truck do you think I should load t with hay?
The same asymmetry shows up in rightward movements, such as heavy
NP shift. As is well-known, most direct objects can extrapose to the right
over some other constituent when they are long and/or focused. However,
this movement is sharply i mpossible for the goal object of a double object
construction (Stowell 1 98 1 , Kayne 1 984: ch. 9, Larson 1 988).
(28) a. * I gave t candy every child that came to the door.
b. I gave t to Johnny every piece of candy I could find.
c. I loaded t with hay three carts and one wheelbarrow.
d. I loaded t onto the cart a stack of books that h ad been sent to
the l ibrary.
Again, the goal NP of the double object construction has unique syntactic
behavior.
It is not clear what accounts for the dev i ance of (25) and (28a). One
possibi l ity that is compatible with the current framework i s mentioned in
Kayne ( 1 984: 202 , n . 1 9). Suppose that when the goal NP moves to the
derived object position in a structure l i ke (24), it pied pipes the nu l l
preposition along with it, for some reason . Then extracting the N P from this
moved PP violates the Subject Condition (also known as the Left B ranch
Condition, or the Condition on Extraction Domains): one is mov ing a proper
subpart of the structural subject of a functional category. In contrast, simple
object extraction takes the whole subject (which is allowed); perhaps it even
ori g in ates in the V P i nte m a l theta-position where it is lexical ly governed .
-
b Y a head to an NP i f the head a-marks the NP. S ince belief does not
assign a a-role to John in (33b), the genitive Case expressed by of is
impossible in this situ ation. This theory can be extended to explain why
the examples in (29) are bad. As before, we assume that the goal is a-marked
by a (null) preposition even in the dative shifted construction. As such, it
is not a-marked by the derived noun; therefore, it cannot receive genitive
Case from that noun . Neither can it get Case from the null preposition, so
the structure is ungramm atical. In the locative alternation, however, each
bare NP is a theme argument, directly a-marked by the head; therefore either
one can appear as an of phrase in a nominal.
A fourth difference between the dative alternation and the locative alter
nation is found in synthetic compounds. In compounds headed by a locative
alternation verb, e i ther the m aterial or the location can appear as the
nonhead, as long as other factors are controlled for: 22
(34) a. hay-loading, glassware-packing
b. truck-loading, box-packing
However, the goal argument of a dative shift verb can never be in a syn
thetic compound:
(35) a. secret-telling, book-reading
b. * spy-telling, *child-reading
This asymmetry can be explained in terms of the observation that there i s
n o room for a preposition i n a simple synthetic compound. Thus, it follows
that if a given nominal is a-marked by a preposition, it will not be able
to appear in a compound either. This can be seen apart from dative shift
in the u ngrammati cality of examples like *Relative-depending is unwise
(compare D ep ending on relatives is unwise). The (35b) examples are ruled
out for the same reason, given that goals always receive their a-role from
a (possibly null) preposition. In contrast, (34) shows once again that either
argument of a locative verb can be directly a-marked by the verb. 23
Further evidence that locative alternations are syntactically different from
dative alternations comes from quantifier scope interactions. In an NP-PP
frame either the theme or the goal c an take w ide scope if both are quan
tified expressions. However, in the double object frame, the first NP must
have wide scope with respect to the second. Thus, (3 6a) is ambiguous but
(36b) is not; it can only mean that there is a single student who must do
all the work (Aoun and Li 1 9 89, 1 993, Larson 1 990, Hornste in 1 995).
(36) a. The teacher assigned one problem to every student.
(Larson 1 990: 604)
b. The teacher assigned one student every problem.
However, no such scope-freezing effect is found in the locative alterna
tion. B oth versions of the locative alternation are scopally ambiguous; in
T H E M A T I C R O L E S A N D S Y N TA C T I C S T R U C T U R E 95
objects, including location objects. Most of these are restrictions: the goal
object cannot be the subject of a secondary predicate, cannot undergo wh
movement or heavy NP shift, cannot be assigned genitive in a derived
nominal, cannot be the nonhead of a synthetic compound, must take wide
scope with respect to the second NP, and cannot become the subject of an
unaccusative verb. These restrictions can, to a substantial extent, be under
stood if one says that the goal NP originates within a PP even in double
object constructions; it then moves past the direct object to a distinct
"derived object" position. If all this' is correct, then there is an important
correlation between lexical semantics and syntax: where there are few
semantic differences, there are syntactic oddities pointing to a nontrivial
derivation (the dative alternation); where there are significant semantic
differences, there is no sign of such a derivation (the locative alternation).
This pattern of facts supports the UTAH. Furthermore, we make the strong
prediction that all other complement alternations in English should fall
into one of these two classes. This remains to be tested in full, but seems
consistent with what is known so far. 27
This section has also given evidence that helps us to refine the state
ment of the UTAH. In particular, there is good evidence that the double
object construction is derived from an underlying structure that is isomor
phic to that of the NP-to-NP frame, rather than vice versa. If this is correct,
then we can conclude that for English three-place verbs with a theme and
a goal, the theme i s consistently expressed as an NP in the Specifier of
VP, while the goal is inside a PP that is the complement of V.
In such languages, one could still say that the (a) examples are derived from
a source similar to the (b) examples, but only at the cost of saying that
"dative shift" (however expressed) is obligatory. This looks like an unwar
ranted degree of abstractness, making it desirable to consider alternatives
in which the (a) examples are generated directly. In fact, Dryer ( 1 987)
develops a set of arguments along these lines within a modified Relational
Grammar framework. He draws an explicit parallel between the apparent
crosslinguistic variability of object choice in sentences like these and the
variabil ity in subject choice that results in ergative l anguages. Languages
in which the goal of a ditransitive verb and the theme of a monotransitive
verb seem to map into a single grammatical function Dryer calls "primary
object l anguages"; they contrast with "direct object languages" in which
themes are consistently the underlying object whether or not a goal is
present.
The natural translation of Dryer 's ideas into the current P&P frame
work is to assume that in primary object lan guages the goal projects directly
into the specifier of VP and the theme projects as the complement of V if
and only if there is a goal. 28 Indeed, Larson ( 1 988: 35 1 , n. 1 8) seems open
to the possibility that there may be such l anguages, c i ting Johns' ( 1 984)
work on Inuktitut as a possible case in point. However, I will argue that
there are in fact no "deep primary object languages" in this sense. Rather,
a close look shows that the same projection rule that works in English is
at work in both primary object languages and direct object l anguages. My
strategy for bu ilding this argument is simply to show that the "goal objects"
of tri adic verbs in Mohawk and Sesotho show the same peculiarities of
behavior that goal objects do in Engl ish. Such facts then provide evidence
that the basic principles of constructing the underly ing syntactic represen
tation are also the same across the three languages.
there is clear evidence that points the other way. The standard in an hourlfor
an hour test does not work well in this language (cf. Baker ( 1 995b: 290» ,
but Mohawk has an adverbial particle eso "many" that may modify the event
argument of the VP it attaches to. If the modified verb has an "incre
mental theme" that measures out the event, many events will correspond
to many tokens of the kind referred to by that theme argument. Thus, one
can recognize the theme of the verb in Mohawk by attaching eso to the
VP and seeing which NP (if any) .is understood as there being many of.
These dynamics can be seen with the simple transitive verb in (46).
(46) Eso wa-ha-tshA ri- ' ne onhuhsa' .
a.1ot fact-MsS-find-punc NE egg
"He found a lot of eggs."
Literally, this means that there were many events of him finding an egg;
thus the Mohawk user easily infers that there were many eggs found, since
egg-finding events are naturally individuated by the eggs found. However,
even if the subject agreement prefix were changed to plural, this sentence
would not be interpreted as "a lot of people found the egg," since VPs
are delimited by their themes, not their agents (Tenny 1 994). Now consider
the use of eso with a dative verb:
(47) Eso wa' -kbe-tsiket-a-nut-e' ne rati-ksa' -oku ' a.
a.1ot fact- I sS/3pO-candy-feed-punc NE MpS-child-pl
"I gave a lot of candy to the children."
NOT: "I gave candy to a lot of children."
Here eso clearly implies that much candy was given, not that many children
were given candy. Thus, "candy" is the incremental theme, even when a
goal NP is present, and in spite of the fact that the goal is the "primary
object" for purposes of things like agreement. See Aissen ( 1 987) for similar
facts in Tzotzil (Mayan), another language with "obligatory dative shift".
I do not have direct evidence of themehood and delimitedness in Sesotho.
However, Machobane (personal communication) reports that the semantic
restrictions alleged to hold of the English double object construction (see
examples ( 1 7)-( 1 9» are undetectable in their Sesotho translations. Thus,
,-
there is no evi nce of this kind for a difference in a-roles.
Next, let us tum to the morphosyntactic properties of goal-objects in
Mohawk and Sesotho. The first type of evidence comes from compounding
and noun incorporation. In Mohawk, the sole object of many monotransi
tive verbs can incorporate into the verb to form a kind of compound verb
(Postal 1 979, B aker 1 988a, Baker 1 995b). One can then ask which object
of a dative ditransitive verb is incorporable. The answer is clearly that the
theme-object can be incorporated, but the goal object cannot, as illustrated
in (48).
1 00 M A R K C . BA KER
Note that even in (48a) the goal object triggers object agreement, showing
it to be the "primary object" in Dryer's sense. Thu s, noun incorporation
is sensitive to the direct-indirect object distinction, even though Mohawk
seems to be a primary object language in several other respects. Exactl y
the same pattern of facts i s found in Southern Tiwa, Nahu atl , Mayal i ,
Chukchee, and indeed all langu ages that are known t o have syntactic noun
incorporation. B aker ( 1 988a, 1 995b) analyzes these facts as showing that
the goal NP is the complement of a null P, as in Engl ish. This P prevents
the goal from incorporating into the verb by the Head Movement Constraint.
However, these patterns also show a clear similarity to those involving
synthetic compounds in English (see (35» .29
B antu languages do not have true noun incorporation, but some have a
kind of synthetic compounding. I have no information about this in Sesotho,
but Sproat ( 1 985) discusses compounds in the related language Chichewa.
A simple example is (49 ) , where the theme of a monotransitive verb is
the nonhead in the compound.
This can be interpreted as evidence that the goal is not a direct object
argument of the verb in Chichew a any more than i t is i n En glish or
Mohawk . 30
Another peculiarity that goal objects in Chichewa share with their Engl ish
counterparts is the dev iance of extracting goal objects by wh-movement
(B aker 1 988a). In a si mple transitive clause, the sole object of the verb
can be fronted by clefti ng. When a rec ipient-benefactive is added to the
structure (along with an applied affix), it becomes the primary object of
the verb in the sense that it is adjacent to the verb, it may be replaced by
a clitic pronoun, and it becomes the subject if the verb is passiv ized.
However, the goal object cannot naturally be clefted:
T H E M A T I C R O LES A N D S Y N T A C T I C S T R U C T U R E 101
There is, however, another way to interpret the similarities between English,
Mohawk, and the B antu languages: one could say that English too is a
primary object language. In fact, Dryer ( 1 987) himself argues for this
view, claiming that sentences like ( 1 3b) are basic, sentences like ( 1 3a) being
derived from them by an "antidative" rule. If this is correct, then there might
still be important differences in how languages project their arguments,
but those differences will not show up by comparing English with Mohawk
or Sesotho. Instead, we need to compare these languages to true "direct
object languages", which have no dative shift at all.
In fact, it is not so clear that there are any such languages. Many lan
guages do not show any Case-frame alternation with triadic verbs, the
goal argument invariably showing up in dative Case or its equivalent.
Japanese is a well-studied language of this type. However, even though there
is no Case marking alternation, there is a word order alternation in Japanese:
the theme object can appear before or after the goal argument, as shown
in (58).
4. ON T H E F O R M U L A T I O N OF T H E U T A H
S o far we have surveyed reasons for believing that arguments bearing similar
them atic roles are expressed in similar initial structural positions both within
and across languages. Although there are significant differences between
accu sative , ergative, and non-configurational language s , and between
THEMATIC ROLES AND SYNTACTIC STRUCTURE 1 05
Recent work on the UTAH and related topics often states the relevant
principle in terms of a thematic hierarchy. For example, Larson ( 1 98 8 :
3 8 2 ) proposes the following:
(59) Thematic Hierarchy
Agent > Theme > Goal > Obliques (manner, location, time, . . . )
If a verb a determines 9-roles 91, 92, • , 9n, then the lowest
• •
1 990, Foley and Van Valin 1 984). One source of the controversy comes from
the existence of the dative alternation, and the difficulty of determining
wh ich version is basic and which (if either) is derived. This has already been
dealt with at length. In particular, I consider the fact that goals cannot be
subjects of unaccusative verbs to be strong evidence that they are lower than
themes, given the long historical relationship between thematic hierarchies
and subject choice, traceable back to Fillmore ( 1 968).
What then are the arguments that goals should be ranked higher than
themes? Some of them are e mbedded in frameworks of assumptions that
are too different from the current one to permit easy comparison here.
However, Gri mshaw ( 1 990), working in roughly the same framework, dis
cusses two: light verb constructions in Japanese, and compounding in
English.32 The English data has already been touched on. Gri mshaw points
out the contrast in (60) .
(60) a.(?) gift-giving t o children
b. * child-giving of gifts
She says that (60b) is bad because the goal is assigned in a smaller domain
(the compound word) than the theme, which is assigned in N'. This order
o f assignment contradicts he r version of the thematic h ie rarchy . In contrast,
I attri buted the ungrammaticality of examples like (60b) to the fact that
the goal role is necessarily assigned by a preposition, and Ps cannot appear
in compounds. Evidence that this second interpretation is the correct one
comes from compounds formed from verbs that have an optional theme
argument. According to Grimshaw 's account, it should be possible for the
goal to appear as the nonhead in these circumstances, as long as the theme
role is not assi gned. In fact, goal-verb compounds are still completely
impossible, as shown by examples like * child-reading or *spy-telling. Thus,
English compounds do not support a hierarchy w i th goals higher than
themes.
The facts from Japanese Light Verb Constructions are not as well under
stood. The basic form of Gri mshaw 's argument is thi s : Japanese has
constructions that consist of an argument-taking noun and a dummy verb.
The arguments of the noun can in general be expressed in two ways: either
as genitive phrases inside the NP, or as constituents governed by the dummy
verb. However, there is a constraint: the theme can only appear outside
the NP if the goal does as well. Grimshaw again interprets th is as evidence
that the m e is lower than goal on a thematic hierarchy, toge the r w ith a
princ i p l e l i ke ( 5 9 ) . Howev er, other constructions work in e x a c t l y the
opposite way. For example, B aker ( 1 989) uses similar reasoning in his study
of the Serial Verb Constructions found in West African languages. In these
languages, a phrase headed by a triadic verb such as "give" can appear
embedded within a projection of some other verb. When this happens, the
theme argument of the three-pl ace verb may appear as the object of the
T H E M AT I C R O L E S A N D S Y N T A C T I C S T R U C T U R E 1 07
higher verb, but the goal object must always appear in the first projection
of the dative verb. (6 1 ) is a minimal contrast from the Nigerian language
Edo that shows this.
Thus, Grimshaw 's reasoning does not extend smoothly to all comparable
cases. I claim that (6 1 ) shows the true hierarchical relationship between
the goal and the theme, and that a different kind of account is needed for
the light verb facts (see Baker ( 1 995b: 353-356) for a preliminary sketch
of an alternative account).
One attractive dividend of the view that goals are ranked lower than
themes is that they can then be collapsed with other locative expressions,
such as the PP arguments of verbs of putting. The theme clearly c-commands
the locative argument of such verbs, and there is no dative shift to cloud
the matter. Now if goals ranked higher than themes, and themes ranked
higher than location�, as proposed by Kiparsky ( 1 987), Bresnan and Kanerva
( 1 989), and other LFG work, then the two roles clearly cannot be col
lapsed into a single macrorole. However, there is often a rather smooth
continuum between locative phrases and goal phrases. Given that the PPs
in (62b,c ,d) are locational paths on anyone's theory, it seems artificial to
say that the PP in (62a) is not a locational path as well.
Of course, (62a) differs from (62b,c) in that it may (or even must) undergo
dative shift in,some languages; this is part of the motivation for distin
guishing them. My claim, however, is that this is not a thematic distinction;
rather it is due to other factors, such as the semantic redundancy of the
preposition in (62) (Larson 1 988, B aker 1 992a), and perhaps the f�ct that
animate nouns must bear structural Case in many languages. In this way,
one can avoid proliferating thematic roles that are difficult to distinguish
from one another in practice.
This leads naturally into the other controversial question of how many
thematic roles there are. Current theories range from a high of something
) 08 MARK C. B A K ER
The second area of debate related to the UTAH concerns how rigid the
requ irements it puts on a syntactic structure are . B aker ' s ( 1 988a) original
statement impl ied that particular thematic role s were associ ated w i th
particular syntactic positions in an absolu te sense. However, many other
researchers assume that only the relative positions of the arguments are
important. On this view, it does not matter exactly what syntactic position
(say) a theme phrase is generated in, as long as it is higher than any goal
phrase and lower than any agent phrase in the same clause. We may call
a condition of this kind the Relativized UTAH, or RUTAH. Larson 's (59)
i s a form of RUTAH, as he makes explicit in Larson ( 1 990). Similarly,
THEMATIC ROLES A N D SYNTACTIC STRUCTURE 1 09
Many different reasons have been proposed for adopting the RUTAH rather
than the UTAH. For example, Larson ( 1 990) does so for certain theoret
ical reasons that do not necessarily arise in my approach.34 However, Speas
( 1 990: 73) expresses perhaps the clearest reason that needs to be consid
ered - the fact that in some cases the expression of arguments seems to
be context-dependent. The example she cites involves the recipient role,
which may be a subject (as in (63a» , but only if there is no agent present
in the clause (as in (63b» .
(63) a. John received a package from Baraboo.
b. Mary sent a package to John from Baraboo.
This seems to be a counterexample to the UTAH, but is consistent with a
RUTAH that says that agents rank higher than recipients. Similarly, it is
well-known that instruments can appear in the subject position in English,
but again only if there is no agent.
(64) a. John loaded the truck w ith a crane/pitchfork.
b. The crane/*pitchfork loaded the truck. (Levin 1 993: 80)
Grimshaw ( 1 990) has a similar conception in mind when she points out
that there are no syntactic differences between transitive psych verbs like
hate and ordinary agent-patient verbs like kill, even though the particular
thematic roles seem to be different. This follows if syntax is sensitive to
the relative rankings of the thematic roles, but not their exact value.
(65) a. Mary hates John.
b. Mary killed John.
However, 'fl absolute version of the UTAH can still be maintained in
the face of su�h examples if one adopts Dowty 'S ( 1 99 1 ) idea that the basic
thematic roles are prototype concepts rather than categorically defined ones.
On this view, it is possible to say that John in (63a), the crane in (64b),
and Mary in (65a) are all (proto-)agents, and as such belong in the subject
position. This is justified by the fact that each of these NPs shares certain
semantic entailments with the prototypical agent Mary in (65b), although
they do not all share the same ones. For (64b), in particular, there is good
reason to think' that the subject is a slightly peculiar agent/causer rather than
1 10 MARK C. BAKER
( 1 99 1 ). However, the task is not complete, because we still owe Belletti and
Rizzi an account of why frighten-verbs have so many syntactic peculiari
ties (see also Grimshaw ( 1 990: 1 9-25» . Dowty ( 1 99 1 ) ignores these
peculiarities, and Pesetsky ( 1 987) puts them aside as semantic.
Here is a tentative proposal that has the right general character. Part
of the difficulty with psych verbs is that it is not clear what kind of
lexical/thematic analysis to assign them. So far, I have assumed thatfrighten
means basically "x cause [y to be in FEAR (with respect to z)]". Such a
conceptual representation should give rise to an ordinary transitive clause.
Suppose instead that its lexical semantic representation is more like "x cause
[ [FEAR (of z)] to go to y]". Here "fear" is conceived of as an abstract
thing that goes into the experiencer, rather than an abstract state that the
experiencer enters. The stimulus is the causer, as before, but now the
experiencer is a type of goal, rather than a type of theme, that role being
taken by the emotion itself. Each element of this account can be moti
vated by superficial morphology in some languages: frighten-class verbs are
often causatives morphologically (e.g., in Japanese (Pesetsky 1 995: 7 ,
46-47 » ; experiencers often bear dative Case identical to that of goals
(e.g., many South Asian languages); psych predicates are often nominal
rather than verbal (e.g., Palauan (Georgopoulos 1 987), Warlpiri (Simpson
1 99 1 )). Then, by normal rules of projection, one gets (66).
(66) VP
I VP
I �
Past NP V'
� �
the noise V AspP
I I
CAUSE Asp'
�
Asp VP
�
NP V'
� �
N XP V PP
r
I 6 1 �
FEAR (of x) GO P NP
I I
(to) Mary
Several things then happen to this basic structure. First, the abstract elements
FEAR, GO, and CAUSE somehow combine to give frighten. I leave open
whether this is done by incorporation in the sense of Baker ( 1 988a) applying
1 12 M A R K C . B A K ER
in the syntax , followed by late lexical insertion along the lines of Halle
an d M arantz ( 1 993), or whether the combination happens by a kind of
pre-syntactic incorporation as in Hale and Keyser ( 1 993) (bu t see section
5 ). Second, dative shift takes place just as in (24), incorporating the
preposition into the predicate and moving its object into Spec, AspP.
This proposal goes a long way toward explaining the pecu liarities of
the experiencer-objec t that have been noted in the literature; it has very
much the same properties as the goal object of a double object construc
tion. For example, Bel letti and Rizzi ( 1 988) show that the object of a
frighten-type verb is a mild island for movement processes in Italian, an
effect that is also detectable in English:
( 6 7 ) a.? Which company does international unrest fri ghten [the president
of t].
b. Which company does the international community fear [the
president of t] ?
On the other hand, Pesetsky ( 1 995) and Grimshaw ( 1 990) show that the
stimulus subject also has certain properties of a base-generated subject.
Borrowing l i beral ly from Pesetsky ( 1 995 ), we could potential l y resolve
th is contradiction by saying that the stimulus subject is indeed base-gen
erated in the agent-subject position, but it shares some propert ies with
derived subjects by vi rtue of binding an empty category that is the argument
of the abstract noun that characterizes the emotion. Semanti cally, th is i s
rather natural ; the most obvious reading o f The noise frightened Mary i s
that the noise caused Mary to have fear o f the noise itself (but see Pesetsky
( 1 98 7 , 1 995) for other, putati vely different readings). Thu s , the deri ved
stru cture of (66) is (6 9) .
THEMATIC ROLES A N D SYNTACTIC STRUCTURE 113
(69) IP
�
I VP
I �
Past NPn V'
� �
the noise Vi AspP
I �
frighten ' NPk Asp'
I �
Mary Asp VP
�
NP V'
� �
N XP V PP
I I I �
� eCn � P NP
I I
0i tk
Here, the stimulus subject can be anaphorically dependent on the experi
encer object because even though it is not c-commanded by the object, it
is linked to a position that is (see Barss ( 1 986» . There would be many
important theoretical problems to face in fleshing out this analysis36;
however, it seems to have approximately the right cluster of properties to
explain the behavior of frighten-class psych verbs, as well as being con
sistent with an absolute UTAH.
says that these variations are not possible. As such, it automatically induces
th e so-called Unaccusative Hypothesis: the idea that the sole argument of
an agentive intransi tive verb is a subject at all levels, but the sole argument
of a nonagentive intransitive verb is generated as an initial object (Perlmutter
1 978, B urzio 1 986) . Therefore , if the Unaccusative Hypothesis is true
universally, this supports the UTAH over the RUTAH.
In fact, Perlmutter's Unaccusative Hypothesis is now widely accepted
in P&P-style theories, and has proved to be a very productive idea in the
analysis of many different kinds of languages. There has been some debate
about whether it is true universally, and about whether the unaccusative
predicates of a given l anguage can be predicted on universal lexical semantic
grounds (see Rosen ( 1 984» . The UTAH is committed to a positive answer
to both of these questions, and the bulk of current work tends to support
this, once one is adequately sophisticated in one's treatment of lexical
semantics and the syntax of particular languages. For example, Levin and
Rappaport-Hovav ( 1 995) show that, once the details of lexical semantics
are attended to, which predicates are unaccusative and which are unerga
tive is quite consistent across a range of well-studied languages, including
Engl ish, Dutch, Itali an, Hebrew, and Russian . At the same time, more
and more work is accumulating that shows how insightful analyses of
phenomena in less-studied l anguages c an be given if the Unaccusative
Hypothesis is accepted, supporting the idea that it holds universal l y. I
mention only a few instances from l anguages already mentioned in th i s
article: B aker ( 1 995b) finds five differences between unergative and
unaccu sative verbs in Mohawk: (i) the argument of an unaccusative verb
but not an unergative verb can be incorporated; (ii) the quantifier eso "many"
can be floated off of the argument of an unaccusative verb only; (iii) bene
factive appl i cati ves made from unergati ve verbs can have a tran si tive
agreement prefi x , but those made from unaccusatives c annot; (iv) mor
phological causatives can be formed from u naccusatives only; (v)
morpho logical purposive constructions can be formed from unergatives
only. 37 In Chichewa and other B antu languages, unaccusative verbs allow
a locative inversion construction in which a locative expression becomes
the surface subject (B resnan and Kanerva 1 989); however, benefactive
applicative constructions cannot be formed from unaccusative verbs except
under special circumstances (Alsina and Mchombo 1 98 8 , Machobane 1 989).
In West Afri can languages, unaccusative verbs but not unergative verbs
can appear as the second verb in certain kinds of Serial Verb Constructions
(Baker 1 9 89). And so on. Thus, while there are still languages in which
there is l i ttle or no known evidence for the Unaccusative Hypothesis, they
are beco ming fewer and fewer. Moreover, I am not aware of any emp ir
ical arguments against the Unaccusative Hypothesis in a particular language.
The major remaining controversy concerning the Unaccusative Hypoth
esis is whether the phenomena attributed to it should be explained in terms
THEMATIC ROLES A ND SYNTACTIC STRUCTURE 1 15
Possibly relevant data from English comes from the so-called conative
alternation, in which the verb has two arguments, one of which is an agent
subject and the other of which can be realized as either an NP or a PP (Levin
1 993 : 41 -42).
(70) a. I stabbed the loaf of bread (with a knife).
b. I stabbed at the loaf of bread (with a knife).
If the RUTAH were true, one could imagine this alternation being seman
tically vacuous: since there is only one non-agent argument, it can assume
either the direct object position or the indirect argument position freely.
(Note that the preposition in the conative alternation is fixed as at for
most verbs, and therefore can be considered semantically redundant.)
However, the alternation is clearly not semantically vacuous: rather, (70a)
entails that my knife actually made contact with the bread and probably
penetrated its surface, whereas (70b) only entails that the bread was the
target of my stab. Since the direct object in (70a) is affected by the event,
it is reasonable to consider it a theme, while the PP in (70b) is more like
a goal . Therefore, (70) actually shows that themes systematically show
up as direct objects and goals as PP complements, even in the absence of
another internal argument. This again supports the absolute version of the
UTAH. 38
The last issue concerning the formulation of the UTAH I will consider is
whether it is thematic roles that determine the syntactic position of argu
ments, or whether it is aspectual notions that do so (Tenny 1 994, McClure
1 994, Borer 1 995). Roughly speaking, aspectual notions are those that are
involved in expressing the state of completion of the event referred to by
the verb: whether it is completed, in progress, continuing, repeated, and
the like. The idea that aspectual roles might play a key role in linking theory
grows out of the convergence of several lines of research. The first is
investigations of the locative alternation, where the argument that "measures
out" the event is the direct object in both versions, in spite of the fact that
in other respects the same kind of situation is descri bed (see section 3. 1 ).
T he conative al yernation just discussed is similar in this respect: the internal
argument measures the progress of the event when it is a direct object but
not when it is oblique. Facts like these lead Tenny ( 1 987, 1 994) to the
conclusion that there is a c l ose relati o n s hi p b etween bein g a s yntac ti c
direct object and performing this kind of aspectual role. A second thread
comes from the Van Valin-Zaenen observation that some of the putative
differences between unaccusative verbs and unergative verbs actually boil
down to differences in the aspectual notion of telicity (roughly, whether
the event reaches a natural conclusion or not): telic verbs are unaccusative,
1 18 MARK C . BAKER
an d atelic ones are unergativ e (see also Tenny ( 1 994) , Grimshaw ( 1 990),
B orer ( 1 995 » . The third thread comes from the domain of psych verbs ,
wh ere some researchers who have not recognized a thematic di stinction
between the fea r class and the frig h ten class have recognized an aspec
- -
A close look, however, may reveal that there are some. Levin and
Rappaport-Hovav ( 1 995) explicitly consider this issue as it applies to
intransitive verbs. They discover that the connection between aspect and
unaccusativity is not as tight as has been claimed. (72) shows that clauses
with unaccusative verbs can be either telic or atelic, as shown by the
standard test of whether the clause is compatible with the adverb for X
time (atelic) or in X time (telic).
(72) a. The ball bounced/rolled/spun for five minutesl#in five minutes.
(atelic)
b. The soup cooled/the train descended for five minutes/in five
minutes. (telic/atelic)
c. The vase broke/shattered/arrived #for five minutes/in five
minutes. (telic)
Conversely, unergative verbs are often atelic activities, but examples like
(73b) are telic (McClure 1 994).
I n contrast, it is easy t o state generalizations that cover both (74a) and ( 7 4b)
in thematic terms: John is clearly the agent/causer of the event in both cases,
while a change is asserted of the cart in both cases (a change of state in
one case, a change of position in the other) . Many other instances of thi s
k i n d ari se, where the saIne c l a u se i s ambiguous between a n acti v i ty and
an accompl ishment reading, but the grammatical functions remain constant.
A th ird potential difficulty for aspect-oriented approaches to linking is
that i t is not clear w hether they can be extended to account for stative
transitive verbs. Tenny ( 1 994) explicitly puts stative predicates aside in most
of her di scussion, conjecturing that there are in fact no universal linking
principles that apply in such cases (see especially pp. 1 29- 1 3 0, n. 3 5 ) .
Similarly, Grimshaw ( 1 990: 29-30) points out that her aspect-based theory
of psych verbs faces a problem that stems from the fact that frig h ten -class
verbs may (and sometimes must) have stative readings; these clauses cannot
be distingui shed aspectually from fear-class verbs, which are also stative,
even though the l i nki n g pattern seems different.
(75) a. This dog frightens/pleases John.
b. John fearsllikes this dog.
In contrast, it is at least conceivable that a thematic a cc ou n t of the l inking
patterns in (74) could be extended to explain the patterns in (75) in a unified
fashion, presumably by making use of Dowty 's ( 1 99 1 ) idea that agent and
theme are prototype categories. Such an account would capitalize on the
subtle but real intuition that (75b) comments more on John 's character
and tastes, while (75 a ) comments on a particular qual i ty of the dog, in
the way sketched in section 2. 1 . If this l ine is successful, it confirms the
intuition that lin king in stative predicates is not accidental , even though it
cannot be determined in purely aspectual terms.
5. T H E M E A N I N G OF T H E U T A H A N D I T S PLACE IN G R A M M A R
right number of argu ments (such as NPs) , the system "crashes", failing to
produce something with a usable interpretation. However, it is al so true that
if a functor takes more than one argument, it must have some way to tell
which argument is which; this is necessary in order to distinguish restau
rant reviews ("Man eats shark") from suspense movies ("Shark eats man").
The UTAH performs thi s function of distingui shing the different argu
ments of the verb by way of virtually the only method available in
Chomsky 's very spare system: it "merges" the arguments into the repre
sentation at systemati cally different points.40 Therefore, there does seem
to be a pl ace for the UTAH within the limits of "virtual conceptual neces
si ty. "
Keeping the role of the UTAH in mind allows u s to resolve certain
conflicts between the Minimalist Program and the original versions of the
UTAH. The most obvious conflict stems from the fact that Baker ( 1 988a)
stated the UTAH as a condition that held at the level of D-structure, prior
to any movements. Chomsky ( 1 993), however, argues that there should
be no such level: it is not required by "virtual conceptual necessity," there
were few substantive principles that held uniquely at D-structure, and certain
paradoxes disappear once one is allowed to do movements and merges in
any order. If this is correct, then the UTAH is either false or it is enforced
at some other point. In fact, realizing that conceptually the UTAH is part
of the theory of the relationship between l anguage and the C-I system tells
us where it should fit in the grammar: like the Theta Criterion, it should
be u nderstood as an output condition on LF, the interface level related to
C_I.41 There is one slight complication: to apply the UTAH at LF one must
say that it looks not at simple NPs, but rather at the tai ls of chai ns, in
order to factor out the effects of movement. However, thi s is no real
difficulty, since chains are taken to be the fundamental LF objects in this
system in any case (Chomsky and Lasnik 1 993).
The second way in which specific Minimal i st proposal s seem to impact
negatively on the UTAH concerns Chomsky 's ( 1 995) adoption of a "bare
phrase structure" system. In short, Chomsky proposes to elim inate all
reference to different bar levels, thereby remov ing any theoretical signifi
cance from one-bar level categories, and moving to a pure l y re lative
de fi ni t i on of XO an d xmax level categories . Depending on the details, this
proposal cou ld undercut the ability to define the kinds of positions needed
to state the UTAH correctly. In particular, since "specifier" and "comple
ment" become relative notions, the bare phrase structure system seems more
compati ble wi th the RUTAH than with the UTAH (see, for example, s e c t i o n
4.2.2 for Chom sky 's approach to the unergative-unaccusative distinction) .
Furthermore , B aker ( 1 995a) conjectures that the reason linguistic theory
seems to make use of exactly three proto-thematic roles is because X-bar
theory defines exactly three kinds of positions - sister of Xc, si ster of X',
and outside XP - so thematic distinctions are neutralized to the point that
THEMATIC ROLES A N D SYNTACTIC STRUCTURE 1 23
they fit into this template provided by the language faculty. However, if
X-bar theory is in fact derivative (see also Speas (1 990» , then thi s claim
cannot be maintained.
This technical question about position types turns out to be related to a
deeper conceptual one. Above, I claimed that the function of the UTAH was
to show which NP fills which argument slot in the theta-grid of the verb.
The question is whether it does so according to an arbitrary system or a
principled one. To take a special case, is the fact that agents are found in
positions external to themes and not vice versa an arbitrary choice made
by particular languages, an arbitrary choice made by Universal Grammar,
or a natural and principled choice? We have already ruled out the view
that it is an arbitrary choice of particular languages by arguing that there
are no "deep ergative" languages in the sense of Marantz ( 1 984). It would
be desirable to rule out the view that it is an arbitrary choice of Universal
Grammar as well.
Ironically, Marantz ( 1 984) also contains the seeds of how to do this, when
he proposes that there is a semantic basis for the fact that agents are
projected external to themes (at least in English), using data from idioms
and other semi-idiosyncratic constructions. Formally speaking, within the
(neo-)Larsonian view of phrase structure we have adopted, the agent NP
is an argument of a higher verb and the theme NP is an argument of an
embedded verb. This Larsonian structure dovetails nicely with work on
lexical semantics, which generally decomposes standard transitive verbs into
(at least) two predicates along the lines of (77) (Dowty 1 979, lackendoff
1 983, Foley and Van Valin 1 984, Hale and Keyser 1 993).
(77) [ x cause [ y be/become PREDICATE]]
Here x is defined to be the agent argument and y is defined to be the
theme. Given that these predicate decompositions are independently moti
vated on semantic grounds, it is very attractive to identify the causative
part of the lexical semantic representation with the higher verb of the
Larsonian shell ,42 and the be/become+PREDICATE part with the lower verb
position. This is proposed by Hale and Keyser ( 1 993) and has been adopted
by Chomsky ( 1 995: 3 1 5-3 1 6) and many others.43 If this is correct, then
the agent has prominence over the theme not by the extrinsic stipulation
of some kind c:f thematic hierarchy, but by semantic compositionality: the
agent is the argument of one predicate, the theme is the argument of another
predicate, and the second predicate is an argument of the first. If syntactic
structure is a projection of gross lexical semantic structure in this way
(and if the lexical semantics of verbs i s not grossly different across lan
guages), it follows that there are no deep ergative languages or completely
nonconfigurational languages.
In addition to taking away the apparent arbitrariness of the statements
concerning where the theme and the agent appear, this approach makes it
1 24 MARK C. BAKER
(78) a. Transitive:
VI
D VI
I
John VI V2
I �
CAUSE D V2
� ,
the bread CUT
John cut the bread.
[x cause [y be linearly-separated]]
b. Unergative: b. Unaccusative:
VI V2
� �
o VI D V2
I � I I
John VI V2 John FALL
I I
CAUSE LAUGH
J ohn lau ghed. John fel l .
[x cause [LAUGH]] [x become DOWN)
These representations follow Chomsky ( 1 995) , except for (78b). The idea
is that the abstract element LAUGH is a one-place predicate of events; it
means that laughter happened. John laughed then means that John was
the i m mediate cause of an instance of this kind of event. Th u s (78b) and
,
themes have prominence over goals and other path arguments. This que stion
has not often been considered as such, to my knowledge. However, there
is reason to think it should be possible. The trick is to propose a suitable
semantic decomposition for the state or event that is expressed by the V2max•
Suppose for concreteness that V2max expresses a state.4S States can be viewed
as a pair consisting of an individual and a predicate, such that the predi
cate holds of the individual (cf. Kamp and Reyle ( 1 993 : 673» . The
individual in this pair is the theme, by definition. What then is the goal?
It must be an argument that helps define the predicate that holds of the
individual. For example, John put the book on the shelf refers to an event
which consisted of John causing a state, where the state consisted of a certain
predicate holding of the book, and the predicate was being in a particular
position with respect to the shelf. Translating these relationships into a
syntactic structure gives (79).
(79) VI
D VI
I
J ohn VI V2
I �
CAUSE D V2
� �
the book V2 X
I �
BE X .D
I �
? the shelf
Here I assume that BE is a two-place relation that takes a thing and a
predicate and creates the corresponding state. I leave open the exact nature
of the element X, which takes a thing and creates a predicate; the easiest
assumption would be that X is simply the preposition on, but one may
want to leave room for other kinds of cases.46 In any case, the theme in
(79) has prominence over the goal, because it is semantically a direct
argument of BE and the goal NP is only part of an argument of BE.
Crucially, I sup�ose that it would be hard to give a semantic decomposi
tion for a state in such a way that the goal is an immediate constituent
and the theme is more deeply embedded. If this is so, then we have a second
important convergence between lexical-semantic predicate decomposition
and the results of syntactic tests (see section 3).
If this kind of lexical decomposition approach begun by Hale and Keyser
and brought into the syntax by Chomsky and others is correct, then the
UTAH essentially disappears as a separate condition of grammar. The
basic function of the original UTAH was to regulate where the variou s
1 26 MARK C . BAKER
Jackendoff and Pinker have assumed (see, for example, section 3 , especially
(24), and (66)). If at the same time it could be shown that a somewhat
more restrictive conceptual representation than lackendoff assumes cou ld
adequately support other aspects of cogn ition, identifying the two would
become a real possibility. Clearly, linguists alone will not be able to answer
th i s question fu lly. However, it is interesting that Chomsky himself seems
THEMATIC ROLES AND SYNTACTIC STRUCTURE 1 27
McGill University
N OTES
* Research for this article was supported by the Social Sciences and Humanities Research
Council of Canada, grant 41 0-95-0979, and by FCAR of Quebec, grant 94ER0578. I have had
the opportunity to discuss various issues in this article in a syntax seminar at McGill University
and in colloquium talks at the University of Pennsylvania, USC, and the University of
California-Irvine. In addition, I have benefited from discussing these issues with Lisa Travis,
Nigel Duffield, O. T. Stewart, Miwako Uesako, Hironobu Hosoi, Mika Kizu, and Jim
McGilvary. I thank all these people and groups for their valuable input.
1 Here in order to be explicit I am taking a "language production" viewpoint, in which
the conception is given and the problem is to find a suitable linguistic expression of it. The
same stages would be run through in reverse if one took a "language perception" view
point, where the linguistic expression (Le., the PF) is given and the problem is to find the
matching conceptual representation (LF). Thus, no inherent directionality should be assumed
in these associations.
2 For example, Larson correctly points out that Baker ( 1 988a) does not hold to the strictest
imaginable sense of the UTAH in his analysis of the passive. Baker claimed that the agent
role was assigned to the specifier of IP in normal active sentences, but to the head of I
containing the morpheme -EN in corresponding passive sentences. These are not identical
structural relationships, but they were held to be equivalent positions inasmuch as both are
"minimally external" to the VP (outside VP but inside the next highest maximal projec
tion).
;\ Suppressed here is the possibility that the various PPs might also differ in their syntactic
positions. I suspect that this is the case, but in this article I concentrate on goals and other
paths, except for a brief remark about instruments and comitatives in section 4.
4 There may also be an instructive comparison between fear/like verbs and "representa
tion-source" clauses such as John memorized the poem, discussed by Dowty ( 1 99 J : 569-570).
Dowty argues that the subject and object of memorize are relatively ordinary agents
f
and patients. John liked the poem is similar in that it means that John created (or at
least acquired) a P icular kind of mental representation of the poem - a positively valued
one.
S Perhaps other tran�itive verbs with noncanonical subjects such as the goal-subject verbs
receive, inherit, own can be handled in a similar way (see Dowty ( 1 99 1 : 58 1 ) and section
4.2. 1 below).
6 In addition to DyirbaI, Dowty mentions cenain Mayan languages. However, see Aissen
( 1 995) for evidence that Tzotzil at least is not a deep ergative language.
7 One exception to this near-consensus is Johns ( 1 992), who argues for a position that is
close to Marantz's "deep ergative" view of Inuktitut. However, the evidence for her account
is primarily morphological, not syntactic.
S Essentially the same derivation has been proposed for "patient topic" constructi ons in
1 28 M A R K C. B A K E R
Tagalog and other Austronesian languages b y Guilfoyle, Hung, and Travis ( 1 992). This
captures the intuition that these languages are partially ergative in some sense.
9 Another common property of ergative languages is that the absolutive NP can be extracted
by various kinds of wh-movement whereas the ergative NP often cannot. This can be u nder
stood if one assumes that the movement of the patient NP to Spec, IP, although Case-driven,
is actually a kind of A-bar movement (Campana 1 992, B ittner and Hale 1 996). S i nce the
patient NP is the highest position in the clause, it can readily be moved further, but it blocks
A-bar movement of the ergative subject past it, by some kind of Relativ ized M inimality
(Campana 1 992, Nakamura 1 993).
10 However, the reader should be aware that many researchers draw a very different moral
from these facts abou t nonconfigurational languages: they conclude that anaphor-binding,
control, and incorporation must not determined by c-command and other structural proper
ties, but rather by something else - either a functional representation (Bresnan ( 1 982), Simpson
( 1 99 1 ) and other work in Lexical Functional Grammar), or cognitive and pragmatic consid
erations. Much can be said to compare these approaches, but this is not the place to say it.
11
See Chamorro ( 1 992) and Baker ( l 995b) for some comparison o f the Mohawk-type
of noncon figurationality with the H i nd i/Japanese-type, and B aker ( i n progress) for some
comparison of the Mohawk and Warlpiri types.
12
Less is known about the locative alternation crosslinguistically. I t seems that some lan
guages have only the ( 1 4a) pattern (Mohawk is one such language); it is not clear that any
languages have only the ( 1 4b) version.
1 :1 See also Pesetsky ( 1 995). To be strictly accurate, Dowty ( 1 99 1 ) anticipates a u nified
tre a t m e n t for d a t i v e and locative alternations, b u t doesn ' t completely com m i t himself to
one. Sim ilarly. Tenny ( 1 994: 8 1 -83) expresses a preference for a base-generation approach
to the dative alternation, similar to the one she has for the locative alternation, but leaves i t
out of the formulation o f h e r principles. Dryer ( 1 987) and Collins a n d Thrainsson ( 1 996)
say nothing about the locative alternation.
14
These data and most l ike them can be traced back t o Green ( 1974). However. Green herself
questions the existence of dative movement only for teach (and show), saying that as far as
she can tell nonidiomatic examples with give are essentially synonymous. See notes 23, 32,
and 38 for other peculiarities of teach that are probably related.
IS Compare Gropen et al. ( 1 989: 242), who say that a sentence like (20b) "sounds somewhat
self-contradictory" and say of a sentence like (20c) that the nondative-sh ifted alternate "may
sound a bit less anomalous." Thus, while they draw the opposite conclusion from mine,
they adm i t that the key judgments are far from categorical. I tentatively assume that these
di fferences, to the extent that they are real, are due to the topic-focus differences between
the two versions of the dative shift, and not to differences in the semantic/thematic roles.
16 Th e reader shou ld not be confused by the fact that there are also syntactic tests that
treat all four classes of direct object as the same. For example, all four must be adjacent to
the verb, must become the subject of a passive clause, and may cliticize to the verb (e.g. in
Bantu languages). I assume that these properties are sensitive not to the base object position
(the Spec of the inner VP), but rather to the derived object position (the Spec of Aspect
Phrase).
17
W i l l iams' examples are not a perfect m inimal pair in that green is a depictive predicate
of hay, that describes its state throughout the event, whereas full is a resul tative predicate
that charac terizes the wagon ' s state at the end of the event. The d i fference is presumably
not crucial however. Thus, i n ? I loaded the refrigerator with sodas warm . warm is a depic
tive predicate of refrigerator and not of sodas (although the example is not as felicitous as
one would like. and some speakers reject it).
I II
Thus, the fact that the theme NP can be a subject of predication even in a dative shift
construction argues against Larson ' s original claim that the theme becomes an adjunct, as
pointed out by Pau l Kiparsky (personal communication). It also argues against the claim
THEMATIC ROLES A N D SYNTACT I C STRUCTURE 1 29
that the theme NP in a double object construction is the object of a null preposition, as
proposed by Emonds ( 1 985) and Pesetsky ( 1 995).
19
Another possibility compatible with my framework is Larson 's ( 1 988) analysis of the
paradigm in (28) in tenns of "light predicate raising." Larson' s account does not generalize
to wh e x trac t i on, but this may be appropriate since the heavy-NP shift contrast is much sharper.
-
Baker ( 1 988a, 1 988b) argues that the deviance of (25) and (28a) is not simply a parsing
effect created by the juxtaposition of two unmarked NPs: he shows that the same effect is
found in dative constructions in Chichewa even when no theme NP is present, but it is not
found in superficially similar instrumental constructions that have sequences of two bare NPs.
20
However, Kayne cites Jespersen as haviqg examples like the giving o/ words figurative
meanings. The generative tradition of concentrating on Romance-based derived nominals
rather than -ing nominals can be traced back to Chomsky ( 1 970), who is rather tentative
about extending the lexicalist analysis to -ing nominals. However, this biases the matter -
needlessly, I believe - against both the locative alternation and the dative alternation, which
are largely properties of the Gennanic vocabulary of English, as is well-known.
21
In this I disagree with Pesetsky ( 1 995), who finds the (32) pattern unacceptable on the
basis of a handful of noncanonical locative alternation verbs that (atypically) allow Romance
style derived nominals. I believe that the verbs Pesetsky uses actually participate in a somewhat
different alternation.
22
In particular, one must take verbs where both the material and location arguments are
optional; otherwise the examples are ru led out by Selkirk's ( 1 982) First Order Projection
Condition.
23
A sim ilar effect can be seen in adjectival passives: either the material argument or the
location argument may be the subject of the adjectival passive of a verb like load (the recently
loaded truck, the recently loaded hay; Levin and Rappaport ( 1 986: 634»; however, the goal
cannot be the subject of the adjectival passive of a dative verb (the untold story, *the untold
person; Wasow ( 1 977: 344». Levin and Rappaport ( 1 986) suggest that the ungrammati
cal ity of examples like this last one is due to the fact that there is no place for the preposition
needed to assign the goal role. On the exceptional behavior of verbs like teach, see Wasow
( 1 977) and note 38 below.
24
Note that because of the " total affectedness" effect, the object of with in (37b) must
refer to enough stuff to completely load a cart. I believe that failure to attend to this factor
has caused some confusion in the literature; thus, Larson ( 1 990) and Aoun and Li ( 1 993)
(citing unpublished work by Schneider-Zioga) give flawed examples and wrongly conclude
that the scope-freezing effect is found in locative alternations also. Nevertheless, a few
speakers have reported to me that they still find the wide-scope interpretation of every crate
d i fficult in (37b).
25
See Aoun and Li ( 1 989, 1 993) and Hornstein ( 1 995) for interesting proposals. However,
these proposals have conceptual flaws, do not generalize to the full range of cases, and cru
cially assume syntactic structures for the double object construction that are different from
the Larsonian one that I am arguing for.
26
Pesetsky ( 1 995: 1 24- 1 25) claims that the verb get shows this alternation, sin'ce one can
l
say both The book ot to Sue and Sue got the book. See Baker (I 995a: 30 n. 3) for a reply,
claiming that these two sentences are actually derived from two different senses of the
highly polysemous verb get.
27
See section 4.2.2 for a brief discussion of the so-called conative alternation. Verbs of
fu lfilling such as present undergo the alternation in (i), which has superficial similarities to
both dative shift and the locative alternation.
in which inherent Case on the object i s spelled out as with rather than as , a null Case. Thus,
cla us es l ike (ib) cannot be nom inalized (Kayne 1 984: 1 53, 1 56), show the scope freezing
effect (Larson 1 990: 605 ), and do not have unaccusative versions (Levin 1 993: 66) - all
properties that are reminiscent of the double object construction. Other than this, it seems
like a good bet that all other diatheses in English are the result of differing thematic analyses
of the event, like the locative alternation. See Pinker ( 1 989), Dowty ( 1 99 1 ), Levin ( 1 993),
and Tenny ( 1 994) for discussion of some relevant cases.
28 Dryer himself would probably not approve of this implementation of his idea. Toward
the end of h i s article he points out that many languages are not uniform in how they
distribute "object properties" over the two objects of a ditransitive verb: some processes
pick ou t "primary objects" and some "direct objects", even in the same language. In this
he draws a parallel to the phenomenon of spJit ergativity, which is also m ore common than
"pureu or "deep" ergativity. Therefore, he claims that the gram matical functions Primary
Object and Secondary Object do not replace Indirect Object and Direct Object, but rather
coex ist with them and supplement them. (The simpler idea that the Primary Object is the
"real" Direct Object in some languages he attributes to Comrie ( 1 982), among others.}
2
9 Further, we would expect that if these languages have locative alternations, the material
argument could incorporate when the location argument is expressed as a locative oblique,
and the location argument cou ld incorporate when the material argument is expressed as an
instrumental oblique. Mohawk apparently does not have locative alternation verbs, but the
prediction seems to be correct for Chukchee (Nedjalkov 1 976: 206-209).
30 (50) m ay also be ruled out by Selkirk ' s ( 1 982) First Order Projection Condition. This
problem would be solved if an N representing the theme were also included in the compound,
but Sproat shows that this too is impossible, perhaps because of Case theory. The ideal
example to make my point would be a verb like patsa, where the theme argument is optional;
however, it is l ikely that there is no such verb in Chichewa.
31 For example, they suggest that dative Case has at least two syntactic sources: it is assigned
by a nu l l preposition to to its complement (Emonds 1 985), or it is a variant of accusative,
checked in the specifier of Aspect Phrase. They may also point to instances of NP-movement
that are not Case-driven.
32 In a footnote, she also mentions L i ' s ( 1 990) important study of v-v compounding in
Chinese. In fact, I be lieve that L i ' s paper contains only one example that might show the
goal to be higher than the theme: the compound jiao-dong ·'teach-understand'·, which means
"x taught z to y so that y understood z." Since the y argument of "understand" is higber
than the z argument, the same must be true of the y and z arguments of "teach", according
to Li's principles. However, it is not clear that teach is really (always) an agent-theme-goal
verb (see notes 1 4, 38). Li gives no similar examples with more canonical dative verbs like
give, send, ask, or bring.
33 However, Zushi ( 1 992) g ives interesting evidence that some source phrases - roughly
those that imply the consent of the source - rank higher than themes in Japanese. Presu mably
this follows from the fact that such sources are exercising volition, and hence count as a
kind of (secondary) agent.
34 Specifically, Larson needs the RUTAH in order to maintain his view that adjuncts are
innermost arguments, and because he projects theme NPs in adjoined pos itions in dative
shift structures. I do not adopt either of these positions. On my v iew, the land ing site of dative
m o ve m e nt i s Spec, AspP, not S pec. VP. so there is no competition with the theme for the
Spec, VP position (see (24) and note 1 8).
35
My judgments of frighten -class verbs w ith resu ltative AP predicates are mixed. Most
freely formed expressions sound rather bad, as the theory pred icts: * She was depressed.
but the clown amused her happy, * The loud noise frightened John dead. However, certain
sem i-idiomatic expressions are perfect, as are some non-AP predicates: The economic
situation worried Mary sick, Th e loud noise frightened John to death.
T H E M AT I C R O L E S A N D S Y N T A C T I C S T R U C T U R E 131
36 Among them are: What kind of empty category is the ec in (69)? Does the stimulus subject
fonn a chain with this empty category? (If so, then the chain has two thematic roles; if not,
then the dependency should not count for binding purposes, given standard assumption s.)
Why is dative shift obligatory with these verbs - even in Romance, where dative shift does
not otherwise happen? A hint concerning this last question may come from Green's ( 1 974)
observation that dative shift is obligatory even with give when there is a stimulus subject
and/or an abstract theme. Thus, one can say: The loud noise gave Mary a scare, but not ?? The
loud noise gave a scare to Mary.
37 Significantly, the feature of Mohawk that looks at first like it will give the easiest evidence
for the Unaccusative Hypothesis - the split agreement pattern on intransitive verbs - turns
out not to be an unaccusative diagnostic (Baker 1 995b: ch. 5). I believe that the unrelia
bility of this kind of superficial morphological evidence has contributed greatly to the
impression of Rosen ( 1 984) and others that unaccusativity involves a substantial degree of
idiosyncrasy.
311
The one semantically vacuous NP-PP alternation in English is dative shift. However,
this is usually blocked when no theme is expressed: one has I read to the children but not
#1 read the children, I baked for Mary but not #1 baked Mary. These facts are more
consistent with the UTAH than the RUTAH, but missing is an account of why dative shift
is barred here. One possibility is that the . theme is present as some kind of empty category,
and this must reach the specifier of Aspect Phrase position to be properly licensed. If so,
then the goal cannot move to that position.
There are a few exceptions to this generalization, including teach, pay, feed, serve, write;
thus, one can say I taught the children and I fed the children. In fact, when these verbs
have no theme, the "goal" NP loses most of the syntactic properties associated with goal
objects: for example, it can be in a derived nominal, an adjectival passive, or a synthetic
compound (Wasow 1 977). See also notes 1 4, 23, and 32. Note furthermore that some of
these verbs in Romance languages alternate between having a dative argument and an
accusative argument (Borer 1 984, citing Eric Werli), even though the these languages do
not have dative shift. Finally, it is probably significant that these verbs are morphological1y
causatives in many languages: "teach" is literally "cause-to-know", and "feed" is '�cause
to-eat't (cf. Green ( 1 974». Put all together, these facts suggest that the animate argument
of these verbs may be a theme when there is no other internal argument.
39
In fact Levin and Rappaport-Hovav t s ( 1 995) Hnking rules do not use thematic roles,
t
either; their view is that thematic roles should be eliminated (or defined) in terms of Lexical
Semantic Structures. However, their substantive linking ru les can be interpreted themati
cally without too much difficulty. They are stated as follows:
(They also propose a default linking rule, omitted here.) The DCLR is very similar to the
statement that themes are direct objects; it almost repeats the original Gruber-lackendoff
definition of what a theme is (with a slight refinement). Similarly, the ICLR generalizes
the statement that agents are subjects to a somewhat broader range of causes. Finally, there
seems to be no inherent barrier to collapsing the DCLR and the ELR into a single rule,
1 32 M A R K C . B A KER
given that they are unorderable. and the arguments o f some verbs are redundantly classi
fied as internal arguments by both.
40 Other. more obv ious ways of marking which argument is which are not available at LF
in the Minimalist Program. For example, linear order is defined only at PF in Chomsky ' s
system, while Case features are u n i nterpretable and must b e eliminated before LF. The
idea is that where these exist they are PF reflexes of a more fundamental phrase structure
asymmetry. The kind of data reviewed in sections 2 and 3 seems to support this.
41
An alternative, worth considering, is that the UTAH is built into the inner workings of
Merge, the operation that builds phrase markers. Thus. Merge could stipu late that an argument
combines with a phrase X only if it discharges the kind of theta role associated with X by
the UTAH . As a theory of sentence production. this seems natural, but as a theory of basic
linguistic competence it seems clumsy and stipulative compared to the v iew in which UTAH
is a matter of how fully constructed LFs are interpreted by C-I.
42
Larson himself considers the higher V to be a semantically null position, motivated by
purely fonnal requirements of X-bar theory. However, these requirements become unstate
able within the bare phrase structure theory. (See also Chomsky ( 1 995) for other M inimalist
problems with Larson ' s original proposal.)
43 The idea works nicely for accomplishments in particular. It can be extended to activi
ties. including unergatives, as explained below. The hard case is stative transitives. such as
see or like. These are considered to have simple lexical-conceptual structu res by many
au thors. I have been assuming that they can be included in the same theory if one u ses a
prototype approach and thematic rather than aspectual representations. Perhaps a lexical
decomposition such as [x cause [y be represented visually/favorably (in x's mind)]] wou ld
be appropriate.
44 Here I use "bare phrase structure" labels, with no bar levels. Also. functional categories
are omitted here for simplicity, including the Aspect which would come between the two
verbal projections in the case of a transitive or an unergative. A variant of this proposal would
be to say that V2 can only be a state, not an event. Then (78c) would be the representation
for (say) John stinks, and to get John fell one would embed V2 u nder a m onadic V 1 that
meant BECOME, not CAUSE (Travis, in preparation). Related to this i s the question of
whether V 1 and V2 are mem bers of exactly the same syntactic category, which I leave
open.
45 This is adequate for true goals, which define the endpoint of an event. and hence its
resu lting state. Whether other kinds of paths can also be expressed purely i n tenns of resulting
states is u nc lear. For example. is it adequate to express John threw the ball towards the
tree as [John CAUSE [ball BE [closer to tree]]], or does one need [John CAUSE [ball GO
[toward tre e ] ] ] ? If the latter is required, then the text proposal must be generalized to include
V2 operators such as the Jackendovian GO function in addition to simple BE.
46
In work in progress, I explore the idea that the category adjective can be characterized
crosslinguistically as an element that fills the X position in a representation like (79).
47
I recognized the difference between these two slightly di fferent usages because of
discussion o f the matter with Jim McG ilvary, who I thank.
REFERENCES
Aoun, Joseph and Yen-hui Audrey Li ( 1 993) Syntax of Scope. MIT Press, Cambridge, Mass.
Baker, Mark ( 1 985) Incorporation: A Theory of Grammatical Function Changing. Ph.D.
dissertation, MIT.
Baker, Mark (I 988a) Incorporation: A Theory of Grammatical Function Changing. University
of Chicago Press, Chicago.
Baker, Mark ( l 988b) 'Theta theory and the syntax of applicatives in Chichewa', Natural
Language and Linguistic Theory 6, 353-389.
Baker, Mark ( 1 989) 'Object sharing and projection in serial verb constructions ' , Linguistic
Inquiry 20, 5 1 3-553.
Baker, Mark ( 1 99 1 ) 'On some subject/object non-asymmetries in Mohawk' , Natural Language
and Linguistic Theory 9, 537-576.
Baker, Mark ( 1 992a) 'Thematic conditions on syntactic structures: Evidence from locative
applicatives', in: I. M. Rocca (ed.), Thematic Structure: Its Role in Grammar. Foris, Berlin,
pp. 23-46.
Baker, Mark ( l 992b) 'Unmatched chains and the representation of plural pronouns', Natural
Language Semantics 1, 33-72.
Baker, Mark ( l 992c) 'Why unaccusative verbs cannot dative-shift', in NELS. University of
Massachusetts-Amherst, University of Ottawa.
Baker, Mark ( 1 995a) 'On the structural positions of themes and goals' , in: Joban Rooryck
and Laurie Zauring (eds.), Phrase Structure and the Lexicon. Kluwer, Dordrecht, pp.
7-34.
Baker, Mark ( 1 995b) The Polysynthesis Parameter. Oxford University Press, New York.
Baker, Mark, Kyle Johnson and Ian Roberts ( 1 989) 'Passive arguments raised' , Linguistic
Inquiry 20, 2 1 9-25 1 .
Barss, Andrew ( 1 986) Chains and Anaphoric Dependence: On Reconstruction and Its
Implications. Ph.D. dissertation, MIT.
Barss, Andrew and Howard Lasnik (1 986) 'A note on anaphora and double objects' , Linguistic
Inquiry 17. 347-354.
Belletti, Adriana and Luigi Rizzi ( 1 98 1 ) 'The syntax of ne: some theoretical implications' ,
The Linguistic Review 1, 1 1 7- 1 54.
Belletti, Adriana and Luigi Rizzi ( 1 988) 'Psych-verbs and a-theory ', Natural Language and
Linguistic Theory 6, 29 1-352.
Bittner, Maria ( 1994) Case, Scope, and Binding. Kluwer, Dordrecht.
Bittner, Maria and Kenneth Hale (1 996) 6Ergativity: towards a theory of a heterogeneous
class' , Linguistic Inquiry 27, 5 3 1-604.
Bok-Bennema, Reineke ( 1 99 1 ) Case and Agreement in Inuit. Foris, Berlin.
Borer, Hagit ( 1 984) Parametric Syntax: Case Studies in Semitic and Romance Languages.
Foris, Dordrecht.
Borer, Hagit ( 1 995) 'Passive without theta grids', ms., University of Massachusetts-Amherst.
Bresnan, Joan ( 1 982) 'Control and complementation' , in: Joan Bresnan (ed.), The Mental
Representation of Grammatical Relations. MIT Press, Cambridge, Mass., pp. 282-390.
Bresnan, Joan and Joni Kanerva ( 1 989) 'Locative inversion in Chichewa: A case study of
factorization in grammar', Linguistic Inquiry 20, 1-50.
cf
Bresnan, Joan an Lioba Moshi ( 1 990) 'Object asymmetries in comparative Bantu syntax ' ,
Linguistic Inquiry 21 , 1 47- 1 85 .
Burzio, Luigi ( 1 986) Italian syntax: A Government-Binding Approach. Reidel, Dordrecht.
Campana. Mark ( 1 992) A Movement Theory of .Ergativity. Ph.D. d issertation, McGiII
University.
Chamorro, Adriana ( 1 992) On Free Word Order in Mohawk. M.A. thesis, McGill University.
Chomsky, Noam ( 1 970) 'Remarks on nominalization ', in: R. Jacobs and P. Rosenbaum
(eds.), Readings in English Transformational Grammar. Ginn, Waltham , M ass., pp.
1 84-22 1 .
1 34 MARK C . B AKER
Dowty, David ( 1 99 1 ) 'Thematic proto-roles and argument selection' , Language 67, 547-6 1 9.
Dryer, Matthew ( 1 987) 'On primary objects, secondary objects, and antidative ' , Language
62, 808-845 .
Emonds, Joseph ( 1 985) A Unified Theory .of Syntactic Categories. Foris, Dordrecht.
Everaert, Martin ( 1 990) ' NP-movement "across" secondary objects ' , in: Joan Mascaro and
Marina Nespor (eds.), Grammar in' Progress. Foris, Dordrecht, pp. 1 25- 1 36.
Fillmore, Charles ( 1 968) 'The case for Case ' , in: Emond B ach and R. T. Harms (eds.),
Universals in Linguistic Theory. Holt, Rinehart, and Winston, New York, pp. 1 -88.
Foley, William and Robert Van Valin ( 1 984) Functional Syntax and Universal Grammar.
Cambridge University Press, Cambridge.
Fraser, B. ( 1 970) 'Some rem arks on the action nom i nalization in 'English ' , in: R. Jacobs
and P. Rosenbaum (eds.), Readings in English Tran�formational Grammar. Ginn,
Waltham , Mass., pp. 83-98.
Georgopoulos, Carol ( 1 987) 'Psych .nouns' in: Joyce McDonough and Bernadette Plunkett
t
Halle, Morris and Alec Marantz ( 1 993) 'Distributed morphology and the pieces of inflection' ,
in: Kenneth Hale and S. J. Keyser (eds.), The View From Building 20. MIT Press,
Cambridge, Mass., pp. 1 1 1 - 1 76.
Hoji, H. ( 1 985) Logical Form Constraints and Configurational Structures in Japanese.
Ph.D. dissertation, University of Washington.
Hornstein, Norbert ( 1 995) Logical Form: From OB to Minimalism. Blackwell, Oxford.
Jackendoff, Ray ( 1 972) Semantic Interpretation in Generative Grammar. MIT Press,
Cambridge, Mass.
Jackendoff, Ray ( 1 983) Semantics and Cognition. MIT Press, Cambridge, Mass.
lackendoff, Ray ( 1 987) 'The status of thematic relations in linguistic theory ' , Linguistic
Inquiry 18, 369-4 1 2.
lackendoff, Ray ( 1 990a) 'On Larson 's treatment of the double object construction' , Linguistic
Inquiry 21 , 427-455.
lackendoff, Ray ( 1 990b) Semantic Structures. MIT Press, Cambridge, Mass.
Jelinek, Eloise ( 1 984) 'Empty categories, case, and configurationality', Natural Language and
Linguistic Theory 2, 39-76.
Johns, Alana ( 1 984) 'Dative "movement" in Eskimo', in: D. Testen, V. Mishra and J. Drogo
(eds.), Papers from the Parasession on Lexical Semantics. Chicago Linguistic Society,
Chicago.
Johns, Alana ( 1 992) 'Deriving ergativity', Linguistic Inquiry 23, 57-88.
Kamp, Hans and Uwe Reyle ( 1 993) From Discourse to Logic. Kluwer, Dordrecht.
Kayne, Richard ( 1 984) Connectedness and Binary Branching. Foris, Dordrecht.
Kiparsky, Paul ( 1 987) Morphology and Grammatical Relations, ms., Stanford University.
Larson. Richard ( 1 988) ' On the double object construction' . Linguistic Inquiry 19, 335-
392.
Larson, Richard ( 1 990) 'Double objects revisited: A reply to Jackendoff' , Linguistic Inquiry
21 , 589-632.
Lev in, Beth ( 1 983) On the Nature of Ergativity. Ph.D. dissertation, MIT.
Levin, Beth ( 1 993) English Verb Classes and A lternations. University of Chicago Press,
Chicago.
Levin, Beth and Maika Rappaport ( 1 986) 'The fonnation of adjectival passives', Linguistic
Inquiry 17, 623-66 1 .
Levin, Beth and Maika Rappaport-Hovav ( 1 995) Unaccusativity: At the Syntax-Lexical
Semantics Interface. MIT Press, Cambridge, Mass.
Li, Yafei ( 1 990) 'On V-V compounds in Chinese' , Natural Language and Linguistic Theory
8, 1 77-208.
Machobane, Malillo ( 1 989) Some Restrictions on the Sesotho Transitivizing Morphemes. Ph.D.
dissertation, McGill University.
Mahajan, Anoop ( 1 990) The AlA' Distinction and Movement Theory. Ph.D. dissertation,
MIT, Cambridge, Mass.
Marantz, Alec ( 1 984) On the Nature of Grammatical Relations. MIT Press, Cambridge, Mass.
McClure, William ( 1 994) Syntactic Projections of the Semantics of Aspect. Ph.D. disserta
tion, Cornell lIPiversity.
McKay, G. R ( 1 9 15) Rembarnga: A Language of Central Arnhem Land. Ph.D. dissertation,
Australian National University.
Mithun, Marianne ( 1 984) 'How to avoid subordination', in Berkeley Linguistics Society,
pp. 493-509. Berkeley Linguistics Society, Berkeley, 1984.
Miyagawa, S. ( 1 989) Syntax and Semantics 22: Structure and Case Marking in Japanese.
Academic Press, San Diego.
Murasugi, Kumiko ( 1 992) NP-movement and the Ergative Parameter. Ph.D. dissertation,
MIT.
Nakamura, Masanori ( 1 993) 'An economy account of wh-extraction in Tagalog ' , in:
1 36 M A R K C. BAKER
Proceedings of the West Coast Conference on Formal Linguistics, CSLI, S tanford, Calif.,
1 993.
Nedjalkov, V. P. ( 1 976) ' Di athesen und satzstruktur i m Tschuktschischen ' , in: Ronald
Lotsch and Rudolf Ruzicka (eds.), Satzstruktur und genus verbi. Akademie Verlag, Berlin,
pp. 1 8 1 -2 1 1 .
Oerh1e, R ichard ( 1 97 5 ) The Grammatical Status of the English Dative A llernation. Ph.D.
dissertation, MIT.
Perlmutter, David ( 1 978) 'Impersonal passives and the Unaccusative Hypothesis', in Papers
from the Fifth Annual Meeting of The Berkeley Linguistics Society. Berkeley Linguistics
Society, Berkeley, Calif.
Pesetsky, David ( 1 987) ' B i nd ing problems with experiencer verbs ' , Linguistic Inquiry 18,
1 26- 1 40.
Pesetsky, Dav id ( 1 995) Zero Syntax. MIT Press, Cambridge, Mass.
Pinker, S tephen ( 1 989) Learnability and Cognition: The Acquisition of Argument Structure .
MIT Press, Cambridge, Mass.
Posta], Pau l ( 1 979) Some Syntactic Rules of Mohawk. Garland, New York.
Rappaport, MaIka ( 1 983) ' On the nature of derived nominals ' , in: Lori Levin, MaIka Rappaport
and Anne Zaenen (eds.), Papers in Lexical Functional Grammar. Indiana University
Linguistics Club, Bloomington, Ind.
Rappaport, MaIka and Beth Levi n ( 1 985) A case study in lexical analysis: The locative
�
Wasow, Thomas ( 1 977) 'Transformations and the Lexicon' , in: Peter Culicover, Thomas
Wasow and Adrian Akmajian (eds.), Formal Syntax. Academic Press, New York, pp.
327-360.
Webelhuth, Gert ( 1 992) Principles and Parameters of Syntactic Saturation. Oxford University
Press, New York.
Wil1iams, Edwin ( 1 980) 'Predication' , Linguistic Inquiry 1 1 , 203-238.
Zaenen, Anne ( 1 993) 'Unaccusativity in Dutch: integrating syntax and lexical semantics"
in: James Pustejovsky (ed.), Semantics and the Lexicon. Kluwer, Dordrecht, pp. 1 29- 1 6 1 .
Zu shi, Mihoko ( 1 992) 'The syntax of dative constructions in Japanese ' , ms., McGill
University.