Lexical Matters

LXC^L
M^TTPS
Lb!
Lecture Notes
No. 24
LEXCAL
MATTERS
COICO O_
Vd ^. SagdO ^nnaSaboIcsI
CSLI
LLJLh LhJLDJL1Y
L LFLLFLL
F11LhNF11L
Copyright 1992
Center for the Study of Language and Information
Leland Stanford Junior University
Printed in the United States
L data and other information appear at the end of the book
Contents
Introduction vii
Contributors xix
1 The Aspectual Interface Hypothesis l
CAROL TENNY
2 Thematic Relations as Links between Nominal Reference
and Temporal Constitution 29
MANFRED KRIFKA
3 Complex Predicates and Morpholexical Relatedness:
Locative Alternation in Hungarian 55
FARRELL ACKERMAN
4 On Obviation 85
DONKA F. FARKAS
5 Blocking of Phrasal Constructions by Lexical
Items 111
WILLIAM J. POSER
6 The Stress and Structure of Modifed Noun Phrases in
English 131
MARK LIBERMAN AND RICHARD SPROAT
7 Hungarian Derivational Morphology, Semantic
Complexity, and Semantic Markedness 183
FERENC KIEFER
8 Focus-Based Inferences in Sentence
Comprehension 209
GYORGY GERGELY
9 Combinatory Grammar and Projection from the
Lexicon 241
ANNA SZABOLCSI
vi / CONTENTS
10 The Lexical Entailment Theory of Control and the
Tough-Construction 269
PAULINE JACOBSON
1 1 A Lexical Analysis of Icelandic Case 301
IVAN SAG, LAURI KARTTUNEN, AND JEFFREY GOLDBERG
Author Index 319
Subject Index 323
Introduction
What is a lexical entry? What information does it include? How are
lexicons organized? In the early days of generative grammar, such questions
were given relatively simple answers. A lexicon was a list of entries, each of
which specifed a phonological shape, some representation of its meaning, a
grammatical category, and a specifcation of exceptionality. The excitement
concerning the lexicon was minimal, precisely because the grammatical
action, so to speak, wa elsewhere: lexical entries provided just enough
information to satisf the needs of elaborate syntactic rules.
In the thirty years that have passed since this era, grammatical theories
have changed in profound ways, the most striking of which may well be the
enhanced role of the lexicon. There is now considerable agreement within
the feld that lexical representations are highly structued and play an
essential role in grammatical description, as many burdens once borne by
transformations and other devices are now carried by lexical representations
or lexical rules of one sort or another.
In spite of this fundamental agreement, no real conensus ha yet been
achieved concerning the content of lexical entries, the precise nature of lexi
cal representations, the scope of the lexicon and lexical analyses in general,
or the matter of how the lexicon should be structured. It is evident that
lexical matters such as these will be of central concern well into the next
century, not just within the field of linguistics, but also in the neighor
ing disciplines where the study of language has assumed an increaingly
important role.
The contributions to the present volume, which (with the exception of
Gergely's) fall squarely within linguistics proper, illustrate very well the ex
treme diversity of empirical problems in this domain, as well as the variety
of techniques currently employed to investigate lexical matters. Despite
their diversity, the papers cluster naturally around four central themes,
which we have selected a the organizing foci of the volume:
Lexical Matters. Ivan A. Sag and Anna Szabolcsi, eds.
Copyright 1992, Stanford University.
viii I LEXICAL MATTERS
1. Argument Structure and the Nature of Thematic Roles
2. Blocking and the Boundaries of the Lexicon
3. The Nature of Lexical Meaning
4. Lexical Alternatives to Syntactic Analyses
1 Argument Structure and the Nature of Thematic Roles
Thematic roles have long been the standard devices for organizing lexi
cal argument structures. Substantial theories of thematic roles have been
baed on the assumption that each individual argument type can be given
a grammatical characterization that determines its syntactic behavior. But
evidence against this view has been mounting for some time now. It ha
been observed, for example, that it is difcult to extend the rigorous treat
ment of thematic roles beyond the core set of Agent, Patient, Experiencer,
Stimulus, etc. It is perhaps surprising that certain combinations of these
roles, for example, Agent and Patient, are typical, whereas others, like
Agent and Stimulus, are not attested at all. Thematic roles, moreover,
may be impossible to defne in semantic terms (a point already recognized
by the grammarians of ancient India and carefully documented by Dowty
(1991)) and many important generalizations are statable only with refer
ence to semantically coherent clases of predicates.
It i presumably due to this growing dissatisfaction with the standard
notion of thematic roles that Dowty's proposal to take a fesh look at
the issue was received with such enthusiasm. Among other things, Dowty
suggested to shift the focus to verb meanings and to recognize at most a
very limited set of proto-roles. The content of these roles is, to a great
extent, characterizable in cognitive, rather than purely grammatical terms.
For instance, the Patient proto-role is characterized by (a) undergoing a
change of state, (b) being an incremental theme, (c) being causally afected
by the event, (d) being stationary, and (e) having a referent whose existence
may depend on the action denoted by the verb.
C
AROL TENNY'S contribution ( "The Aspectual Interface Hypothesis" )
and that of
M
ANFRED
K
RIFKA ( "Thematic Relations as Links Between
Nominal Reference and Temporal Constitution" ) have a close relation.
They provide, in fact, an interesting example of how scholars belonging
to two quite diferent research traditions may achieve converging results
and insights.
It is well-known that the distinction between cumulative and quantized
reference in the domain of nominal interpretation (e. g. , apples vs. fve apples
parallels the contrast between telic and atelic interpretation in the verbal
domain (cf. run vs. run a mile and, furthermore, that the two interact in
grammar. Exactly when and how the properties of the nominal argument
afect those of the complex predicate is the subject of both Tenny's and
Krifa's contributions.
INTRODUCTION I ix
Using the background assumptions of the MIT Lexicon Project, Tenny
proposes the Aspectual Interface Hypothesis, according to which ( 1) the
mapping between thematic structure and syntactic structure is governed by
aspectual properties and (2) only the aspectual part of thematic structure
is visible to syntax. The crucial constraint she proposes is that the internal
argument of a simple verb is constrained so that it either undergoes no
change or motion, or it undergoes change or motion which measures out
the event over time.
The precise interpretation of "measuring out an event over time" may
be explicated, at least in a signifcant subset of the caes, in the terms of
fered by Krifa, whose work proceeds fom the same fundamental insight.
Krifa develops a semantic system baed on the particular lattice-theoretic
properties he assigns to three sorts of entities: objects, events, and times.
In his system he is able to give a precise interpretation to notions like cumu
lative, singular, quantized, and atomic reference. Defning telic predicates
a event predicates all of whose subevents have the same terminal point,
Krifa goes on to show that all quantized event predicates, and only these,
have the property of telicity. The crucial technique that Krifa then intro
duces to explicate the notion of "measuring out an event over time" is that
of a homomorphic mapping fom objects to events which preserves lattice
structure. Such a mapping exists in the case of drinking wine, for example,
but not in the case of drinking a glass of wine, and thus provides a precise
model-theoretic construction of the intuitive notion whose empirical utility
is argued for by Tenny.
On the grammatical side, Krifa's insights also parallel Tenny's, in that
he conceives of cumulativity, uniqueness of events, and graduality a prop
erties of thematic relations. He uses these properties to clasif the Pa
tient relations of diferent verbs and to distinguish them, for example,
fom Stimulus. The empirical scope of Tenny's informal proposal is sig
nifcantly broader than Krifa's, however. It extends to objects undergoing
change that does not afect the existence of their parts but rather, their
color or location (cf. redden and push (the cart) to New York). Krifa,
on the other hand, explores applications of his theory to such problems
a Finnish progressivity (marked by the partitive case of the argument)
and the interaction of perfectivity and the quantizedness of arguments in
Slavic.
Krif's and Tenny's contributions, both agreeing with Dowty's work
in their Davidsonian spirit and their attitude towards thematic roles, ad
dress the general question of why and how certain internal arguments
determine the apectual properties of the containing verb phrase.
F
AR
RELL
A
CKERMAN, directly infuenced by Dowty's work on roles, explores
a set of particularly challenging empirical problems in his paper "Com
plex Predicates and Morpholexical Relatedness: Locative Alternation in
Hungarian" .
} / LEXICAL MATTERS
Ackerman's study is couched in the famework of Lexical Mapping The
ory, developed by Bresnan and Kanerva ( 1989) and Bresnan and Zaenen
(1990) , inter alia, which seeks to explicate the relation between seman
tic (or thematic) roles and grammatical relations in terms of a mapping
fom the former to the latter that obeys certain principles. These prin
ciples make reference to a featural decomposition of grammatical rela
tions (subject is decomposed as [-objective] and [-restricted] ; objects are
all [+objective] and may be (semantically) restricted or not; and oblique
dependents are [-objective] and [+restricted] ) . Among the principles that
constrain the role-relation mapping in Lexical Mapping Theory are certain
intrinsic feature classifcations, for example, the classifcation of agents as
[-objective] (this prevents agents fom ever being realized a direct objects,
for example) .
Ackerman's examination of the systematic correlations between the
meanings and grammatical relations associated with various perfective
and resultative preverbs in Hungarian leads him to a number of conclu
sions. First, the intrinsic clasifcations (following Zaenen ( 1990)), should
be baed on Dowty's notion of Proto-Role properties. Second, a clear dis
tinction should be drawn between two kinds of operations: morpho lexical
(those which afect the lexical semantics of predicates) and morphosyntac
tic (those which don't) . The basic analysis of Hungarian preverbs, Acker
man argues, should all be in terms of morpholexical relatedness (expressed
within the word formation component of the grammar), for they involve not
only the Proto-Role properties, but also the very kinds of aspectual proper
ties, for example, telicity, explicated in the work of Tenny and Krifa. His
paper thus illustrates a new domain (in a third research paradigm) where
these basic semantic properties are able to contribute to the explanation
of why lexical argument structures are the way they are.
2 Blocking and the Boundaries of the Lexicon
Though syntactic and semantic rules are in general asumed to be inde
pendent in their application, there are a variety of cases that indicate that
certain grammatical rules must "know about each other." The most strik
ing of these is the phenomenon of blocking, whereby the availability of a
better-suited form renders a less specifc one ungrammatical. In general,
these are all cases which can be conceptualized as instances of the Else
where Principle (an insight also due originally to the grammarian of an
cient India), which states in essence that the existence of particular rules or
exception blocks the application of more general rules. Although standard
examples of this principle seem purely morpho lexical in nature (e. g. , the
existence of went blocks the application of the regular past tense sufxation
rule (*goed)), various researchers (Householder, McCawley, Horn, Dowty,
Levinson) have studied similar efects in other domains, relating blocking
INTRODUCTION / xi
phenomena, for example, to independently motivated aspects of Gricean
pragmatic theory. This raises the important theoretical issue of clarifying
the nature of the domains in which the Elsewhere Condition applies.
D
ONKA
F
ARKAS, in her contribution "On Obviation," examines classic
examples of obviation, including the impossibility of coreference between
matrix subject and pronoun in French examples like Pierre veut qu'il parte
'Pierre wants that he leave' . Accounts of this efect have been ofered in
terms of Chomsky's "avoid pronoun" principle or else Binding Theory. In
the latter cae, the domain in which the embedded subject is not to be
bound is extended to include the matrix clause. Farkas argues, however,
on the bais of data fom Western Romance-as well as new data from
Romanian, Hungarian, and Serbo-Croatian, that these approaches to ob
viation are unsatisfactory. The alternative she proposes is based on the
notion of blocking, thereby tying this phenomenon to her earlier results
concerning subjunctives and control.
According to Farkas' proposal, the coreferential reading in the relevant
caes is blocked by the availability of a rival infnitival form in the lan
guage. This immediately explains why Romanian has no obviation: it
minimally difers fom Western Romance in lacking rival infnitives com
pletely. A wealth of further, often subtle contrasts are shown to follow fom
this blocking analysis. For example, coreference of the sort just illustrated
is blocked precisely when the infnitive and the subjunctive complement
clause both conform to the canonical case of control, that is, cases in which
both the matrix and the complement subject bear a responsibility relation
to the situation denoted by the complement. Farkas argues that such fne
distinctions in verb meanings are in general refected in degrees of obvia
tion, a point she attempts to justif on cross-linguistic grounds.
The mechanism underlying Farkas' analysis, despite its larger efect on
coreference relations, is lexical blocking of the familiar sort. Such an anal
ysis is possible precisely because the "competition" in question is between
two lexical forms, each subcategorizing for a distinct type of complement.
Yet whether or not blocking must be a purely lexical matter is precisely
the point taken up in
W
ILLIAM
P
OSER
'
S paper "Blocking of Phrasal Con
structions by Lexical Items."
Poser points out three unusual cases where lexical forms block phrasal
constructions. The frst is that of Japanese "incorporated" periphrastic
verbs formed with suru ' do' . There is phonological and syntactic evidence
that, contrary to the usual assumption, these are phrasal. The fact that
these periphrastics cannot be formed directly fom deverbal nouns is ar
gued to be a blocking efect, due to the existence of the corresponding lex
ical verb. The second case is English comparatives and superlatives. The
phrasal status of the more and most forms is clear, yet they are blocked
by the availability of the lexical forms with -er and -est. Third, synthetic
expression of tense in Basque blocks periphrastic expression with the defec-
xii / LEXICAL MATTERS
tive verb ari. The dilemma that these data indicate is this: either blocking
is not just lexical, or the lexicon needs to be "extended" to encompas the
sorts of phraal construction discussed.
The frst possibility is the pragmatic approach used, for example, by
Householder ( 1971), to explain why the existence of the simpler pink blocks
more complex pale red. This explanation does not extend to morphology,
for a number of reasons: ( 1 ) oxen, for example, is not simpler than oxes; (2)
the blocking efect in morphology is much stronger than the unacceptability
of pale red; (3) blocking requires that the two forms be morphologically
related; pure semantic relatedness is not enough.
The second possibility is to "extend the lexicon." Poser proposes that
a morphological category is one that is potentially instantiated by a word
formation rule. Phrasal constructs involving XO categories are allowed to
instantiate morphological categories. A periphrastic construction is, then,
a in grammatical tradition, one in which morphological categories that are
typically instantiated lexically are instead instantiated at a phraal level.
This, in turn, allows one to say that if instantiation by a lexical verb is
available, phrasal instantiation is blocked.
Poser's notion of phraal constructs involving XO categories may seem
unfamiliar to readers steeped in the tradition of modern syntactic theory,
yet
M
ARK
L
IBERMAN and
R
ICHARD
S
PROAT appeal to just such a notion
in their contribution to the present volume, "The Stress and Structure of
Modifed Noun Phraes in English. "
In standard treatments of English modifed phraes, the combination of
adjective and noun is presumed to be the result of a syntactic rule (build
ing Nl structures) , wherea the combination of noun and noun is taken to
be lexical (NO) in nature. And correlating with this diference in mode of
combination, is assumed to be a diference in prosody: in adjective-noun
collocations, the noun i stressed, in accordance with the Nuclear Stress
Rule; in noun-noun combinations the frst noun is stressed, in accordance
with the Compound Stress Rule. Standard examples of this contrat in
clude detective novel versus defective novel, bank check versus blank check,
and so forth.
On the basis of a detailed examination of a large variety of modifed
phraes, Liberman and Sproat challenge this standard wisdom. For ex
ample, sof spot and wad man are adjective-noun combinations with com
pound stress, and steel plate and rice pudding are noun-noun combinations
with nuclear stress. These facts, a the authors show at length, cannot be
explained by appeal to discourse factors (diferences in inherent contrat)
or a distinction between "fxed" expressions and others (e.g., hard liquor
and blue moon are lexicalized adjective-noun combinations bearing nuclear
stress) . Accent, on Liberman and Sproat's view, is only partly predictable
on grounds of category or meaning.
These observations lead in turn to the view that all combinations of
INTRODUCTION I xiii
modifer-noun in English are in principle ambiguous between NO and N1
structures-a surprising conclusion that raises new questions, only some
of which Liberman and Sproat address in detail, about the exact relation
among such notions as being lexicalized, being an NO expression, and being
a fxed expression (Le. , having a special meaning).
3 The Nature of Lexical Meaning
The analysis of fxed expressions and the treatment of argument structure
B,re of course only two of the many issues that arise in the area of lexical
semantics. Two other notions that have been asumed to play an important
role in the organization of the lexicon, in predicting acquisition order, and
also with respect to predicting blocking efects of the sort we have already
discussed, are semantic complexity and semantic markedness. Complexity
may be thought of simply a the amount of semantic material contained in
a lexical item, for example, sell is more complex than give. Markedness, on
the other hand, has to do with how restricted a term is in its utility-baic
level terms like pine are held to be less restricted, thus more natural and
less marked, than either more specifc or less specifc terms in the same
semantic domain.
Since complexity is usually considered with respect to verbs, and basic
level predictions (markedness) with respect to category names, the relation
between these two notions i seldom examined. In his paper "Hungarian
Derivational Morphology, Semantic Complexity, and Semantic Marked
ness," FERENC KIEFER investigates precisely this problem. Kiefer seeks to
develop an appropriate notion of markedness for verbs and, with that in
hand, to examine the interesting circumstance that arises when complexity
and markedness make conficting predictions.
The semantic feld he investigates is defned by the rules of Hungarian
derivational morphology, involving the formation of causative, pasive, fe
quentative and perfective verb forms (with or without idiosyncratic mean
ings) . Morphological, syntactic and semantic complexity are carefully dis
tinguished fom the corresponding kinds of markedness, and the relation
ship of semantic complexity to syntactic and morphological complexity is
examined. What Kiefer fnds is that, with an increase of semantic com
plexity, neither morphological nor syntactic complexity decreases. Fur
thermore, the semantic complexity relations among derivationally related
lexical items which are semantically opaque are the same as those between
transparent lexical items.
Kiefer then shows that semantic markedness may conict with com
plexity considerations. For instance, verbs with a perfective prefx are less
marked than the corresponding iterative/generic verbs, despite the fact
that they are semantically more complex. And, similarly, verbs with a
directional prefx are less marked than corresponding generic/directionless
xiv / LEXICAL MATTERS
verbs. In this way, Kiefer argues, the facts of acquisition can be explained:
acquisition order is determined by markedness, rather than by complexity
(though, if no special markedness relations obtain, semantic markedness
may be equated with semantic complexity) .
Kiefer's study thus brings rather subtle and indirect evidence t o bear
on questions having to do with the psychological reality of lexical informa
tion. The methodology of GYORGY GERGELY, by contrast, is more direct.
In his contribution, "Focus-based Inferences in Sentence Comprehension,"
Gergely is concerned with verifing experimentally the claim that massive
incremental integration of diverse types of information takes place rapidly
in real-time language processing.
Recent work in sentence comprehension has established that listeners
use a variety of cues interactively in the construction of discourse interpre
tational structures, even before the full syntactic and semantic processing of
the sentence is complete. Working within this research paradigm, Gergely
investigates the precise on-line efects of some specifc cues that have pre
viously only been examined indirectly, using of-line measures. These are:
( 1 ) topic-focus (TF) structure, (2) inter-clausal semantic relations, and (3)
script-based knowledge.
The initial though in sentences like Though Daddy praised his daughter,
she was still not happy invites script-based inferences concerning cause
efect relations, one of which will be denied in the second clause. Experi
ments with English speakers have indicated that the pertinent cause-efect
relation can only be identifed in retrospect, during the processing of the
second clause. Gergely's paper is based on experiments with Hungarian
speakers, due in large part to the fact that in Hungarian, the TF structure
is marked unambiguously by word order and intonation. Gergely hypothe
sized that listeners can rely on the TF structure of the initial though clause
to restrict the set of expectable consequences to those that arise fom the
focused part of the clause. For example, the example just cited is felicitous
when the verb praised, rather than, for instance, the object his daughter,
is focused. The reason for this is that praising Iuight be expected to cause
happiness.
Gergely's results support this hypothesis. First, naming times for a
word probe, whose meaning is in adversative relation to the focus-based
consequence of the initial though clause, were signifcantly decreased, even
when the probe appeared before the end of the clause. This indicates that
before the main clause is processed, subjects have generated predictive
inferences about its adversative content. The selective nature of facilita
tion demonstrates that subjects inferred only those consequences that were
based on the focused part of the clause. Such facilitation was not present
when the initial clause contained no connective with comparable seman
tics, suggesting that when focus-based inferences cannot serve a predictive
function, they are discarded fom working memory at the end of the clause.
INTRODUCTION / xv
Gergely's experiments thus provide key insight into the semantic nature
of certain kinds of connectives, as well a the highly integrative nature of
language processing in general. In addition, his experiments showed that
focus-based inferences directly access pragmatic knowledge structures such
a scripts, thus raising the question of whether the probe facilitation here
is a further case of lexical priming. In contrast to the automatic and short
lived nature of lexical priming, however, the selective nature of the observed
priming efects and their survival even over a clause boundary indicate, as
Gergely shows, that they are more like the robust higher-order facilitation
of script-asociated words.
4 Lexical Alternatives to Syntactic Analyses
As we noted at the outset, a striking trend in contemporary syntactic the
ory is the lexical analysis of phenomena previously treated in terms of
transformational rules or other syntactic devices. With this trend comes,
quite naturally, an uncertainty as to which linguistic phenomena should in
fact receive lexical treatment. Several of the papers in this volume propose
lexically baed solutions to problems that have ofen been assumed to be
within the scope of syntactic movement theory, binding theory, control the
ory or case theory. The reason, they argue, is that the standard accounts
are descriptively inadequate and/or have theoretically undesirable proper
ties. We have already seen an argument of this type in Farka's lexical
analysis of obviation.
ANNA
S
ZABOLCSI, in her contribution "Combinatory Grammar and
Projection fom the Lexicon," explores the possibility of articulating a
sumptions about the lexicon in such a way that they derive a set of gen
eralizations that are traditionally conceived of a purely syntactic. For
example, Szabo1csi proposes to eliminate the need for principles pertain
ing to empty categories and binding by exploring the consequences of the
central asumption of categorial grammar, namely, that lexical items with
subcategorization fames are to be looked upon a functions.
Her paper frst demonstrates the explanatory value of the concept of
lexical items as functions, by providing a detailed comparison of the treat
ments of unbounded dependencies in Government-Binding Theory (G B) ,
Head-driven Phrase Structure Grammar (HPSG), and Combinatory Cat
egorial Grammar (CCG) . Taking this concept fully seriously, CCG goes
beyond fnctional application and introduces a set of functional opera
tions, that is, combinators, into the grammar. In this way it can dispense
with bound variables corresponding to gaps, and turn the empirical gener
alizations that GB and HPSG express in terms of constructs like the Empty
Category Principle, the Subcategorization Principle, the Nonlocal Feature
Principle, etc. , into constructive procedures.
The main part of Szabo1csi's paper is devoted to issues that tradition-
xvi / LEXICAL MATTERS
ally belong to Binding Theory. These matters pose a special problem for a
grammar baed on combinatory logic since this logic does not merely allow
one to dispense with bound variables, it has no variable binding machinery
at all. Szabolcsi argues that this restriction is in fact a virtue of CCG,
in that it forces the correct efects to be derived directly fom the lexical
meanings of anaphors and bound pronouns (crucially involving the dupli
cator combinator), without invoking the apparatus of a syntactic Binding
Theory.
In the last part of her paper, she extends her proposal to VP-ellipsis.
The fact that even quantifcational antecedents support the strict read
ing of VP-ellipsis (cf. Every man mentioned his merits before Mary did
. . . before Mary mentioned his merits') ha been taken to argue against
Reinhart-like approaches like Szabolcsi's and to motivate the introduction
of a new logical device, co-parametrization (see Gawron and Peters 1990) .
Szabolcsi challenges this conclusion, arguing that composition and duplica
tion, the two combinators whose presence in CCG accounts for extraction
and anaphora in the general cae, are in fct sufcient to derive the strict
readings here a well as in examples involving anaphors and gaps.
In her paper "The Lexical Entailment Theory of Control and the Touh
Construction,"
P
AULINE
J
ACOBSON presents another purely lexical analysis
that challenges conventional syntactic wisdom. Sentences of the type That
rock is hard for me to move exhibit a set of seemingly contradictory prop
erties. The matrix subject appears to be selected by the infnitive, a in
a raising construction; however, the gap following the transitive infnitive
resembles wh-traces more than NP-traces. One solution to this puzzle, sug
gested by Chomsky, is to asume that the subject/gap relation is mediated
by a null operator.
Building on a number of earlier results (including her own), Jacobson
presents a careful examination of data pertaining to these points and to fr
ther properties of the construction, and concludes that there is no syntactic
relationship between the subject of the tough-adjective and the gap. One
crucial fact is that there is no syntactic category matching requirement (cf.
That language is learable is hard for any theory to capture versus This
theory captures that language is learable). On the other hand, the con
struction allows for null complement anaphora, which is characteristic of
control constructions.
Jacobson argues for the following specifc asumptions about the lexical,
syntactic and semantic properties of the tough-adjective: (1) it subcatego
rizes for an infnitival complement that lacks an object; and (2) its subject
is understood a flling the gap position purely in view of some entailment
that is part of the lexical meaning of the adjective. This account is modeled
after the lexcal entailment theory of control, defended by Dowty, Chier
chia and others, which holds that want, for instance, denotes a relation
between individuals and properties, such that for any individual 1 which
INTRODUCTION / xvii
stands in the want-relation to a property P, x ha the property P in the
"want-world" of 1. Showing that the same kind of explanation can han
dle the puzzling relation between the tough-subject and the gap, Jacobson
suggests an interesting extension of this inuential theory.
The fnal contribution to the volume, "A Lexical Analysis of Ice
landic Case," by
I
VAN
S
AG,
L
AURI
K
ARTTUNEN, and
J
EFFREY GOLD
BERG (SKG), provides an essentially lexical treatment of the long-standing
problem of "quirky" cae in Icelandic complementation.
Although subjects in Icelandic are in general marked with nominative
case, and objects are in general accusative, there are three well-studied
clases of verbs which are exceptional in requiring (respectively) dative,
accusative or genitive case on their subjects. The interesting problem that
arises with respect to these quirky-subject verbs is their behavior in raising
constructions. The generalizations can be stated simply as follows: (1) the
raised subject of a subject-raising verb (e.g., viroist 'seems') is nominative
if its infnitival complement is headed by a non-quirky verb; (2) the object
of an object-raising verb (e. g., telur ' believes' ) is accusative if its infnitival
complement is headed by a non-quirky verb; and (3) a raised argument
must bear the appropriate quirky cae if the head of the complement is a
quirky-subject verb.
These facts, well-known in the literature on Icelandic syntax, strongly
suggest a two-stage analysis-one where quirky cae is lexically assigned
in underlying syntactic structures, but structural case (nominative for sub
jects; accusative for objects) is assigned to very superfcial syntactic repre
sentations, but only to those nominals that did not receive (quirky) lexical
case at an underlying level of syntactic representation. The notions of
before transformations have applied (for quirky cae) and afer transforma
tions have applied (for structural case) appear to be indispensable for an
account of these facts.
SKG, motivated by their desire to formulate all of linguistic knowledge
a declarative constraints on possible structures, set forth a treatment of
these facts that makes no reference to "before" or "after" stages of a trans
formational derivation. Their analysis posits two attributes-CASE (re
alized case) and DCASE (default case) , and allows the possibility that a
particular nominal may bear distinct values for these features. On SKG's
analysis, a quirky-subject verb specifes only the CASE value of its subject,
but a non-quirky verb specifes only that its subject's CASE value matches
its DCASE value. These minimal specifcations, together with the indepen
dently motivated treatment of realized subjects as [DCASE nom] and real
ized objects a [DCASE acc], lead to an account of the data just described
that is nonprocedural, monotonic, and constraint-based, as intended, and
yet expresses all the relevant generalizations. The SKG analysis is thus
another example (like those proposed by Farkas, Szabolcsi, and Jacobson)
of how complex data that have been taken to motivate transformational
xviii / LEXICAL MATTERS
analyses, can be dealt with in alternative terms, without sacrifcing either
explanation or the design properties that are motivated by processing con
cerns.
The strong Hungarian presence in the volume at hand is no accident.
Several of the papers in this volume were originally presented at the Sym
posium on Lexical Semantics, held during the 1987 LSA Linguistic Institute
at Stanford University, which was attended by a delegation of Hungarian
linguists (including one of the editors) and a number of American linguists
of Hungarian descent including the other editor) . The editors both regret
the various unforeseen delays, in consequence of which the contributions of
Tenny, Krita, Farkas, Kiefer, Gergely, and Szabolcsi, which were complete
in 1989, are only now seeing the light of day.
Neither the 1987 symposium nor the present volume would have been
possible without the generous support of both the Soros Foundation and the
International Research Exchanges Board (IREX) (formerly sponsored by
the American Council of Learned Societies and the Social Science Research
Council) , whose assistance we gratefully acknowledge. In addition, we wish
to thank Ferenc Kiefer and Andras Kornai for their help in organizing
the symposium, a well as Dikran Karagueuzian, Jonathan Ginzburg, and
Jongbok Kim for their asistance in the preparation of this volume.
Ivan A. Sag and Anna Szabolcsi
References
Bresnan, Joan, and Jonni M. Kanerva. 1989. Locative Inversion in Chichewa: A
Cae Study of Factorization in Grammar. Linguistic Inquiry 20: 1-50.
Bresnan, Joan, and Annie Zaenen. 1990. Deep Unaccusativity in LFG. In Grm
matical Relations: A Cross- Theoretical Perspective, ed. Katarzyna Dziwirek,
Patrick Farrell, and Errapel Mejias-Bikandi, 45-57. Stanford: CSLI Publica
tions.
Dowty, David. 1991. Thematic Proto-Roles and Argument Selection. Language
67(3) : 547-619.
Gawron, Jean Mark, and Stanley Peters. 1990. Anaphora and Quantifcation in
Situation Semantics. CSLI Lecture Notes No. 19. Stanford: CSLI Publica
tions.
Zaenen, Annie. 1990. Unaccusativity in Dutch: an Integrated Approach. In
Semantics and the Lexicon, ed. James Pustejovsky. Dordrecht: Kluwer.
Contributors
F
ARRELL
A
CKERMAN, Assistant Professor of Linguistics at the University
of California at San Diego, is interested in the interaction between lexical
semantics, morphology and syntax with special attention to the Finno
U gric languages.
D
ONKA
F
.
F
ARKAS is Associate Professor of Linguistics at the University
of California at Santa Cruz. Her research interests include the interface
between syntax and semantics (in particular, the semantics of complemen
tation), and morphology.
G
Y
O
RGY
G
ERGELY received his Ph.D. in 1986 in experimental psychology
at Columbia University, and has taught in the department of psychology
at the University of Rochester. His current research interests are language
processing, causal and intentional attribution in infancy, and experimental
and psychodynamic approaches to the development of the self.
J
EFFREY
G
OLDBERG, now a resident of Hungary, is a graduate student in
the linguistics department at Stanford University. His research interests
include syntactic theory and computational linguistics.
P
AULINE
J
ACOBSON is Professor of Cognitive and Linguistic Sciences at
Brown University. Her research centers largely on the syntax/semantics
interface. She is especially interested in investigating the feaibility of direct
model-theoretic interpretation of (surface) syntactic structures.
L
AURI
K
ARTTUNEN is Principal Scientist at the Xerox Palo Alto Research
Center and Consulting Professor of Linguistics at Stanford University. His
linguistic publications include works on pragmatics and semantics. In com
putational linguistics, he is best known for his contributions to unifcation
baed grammars and fnite-state morphology.
F
ERENC
K
IEFER is Senior Research Fellow at the Research Institute for
Linguistics of the Hungarian Academy of Sciences. His main research in
terests are morphology, semantics and pragmatics.
Iexical Matters. Ivan A. Sag and Anna Szabolcsi, eds.
xx / LEXICAL MATTERS
MANFRED KRIFKA is currently Assistant Professor in the Department of
Linguistics at the University of Texas at Austin. His research interests in
clude the semantics of mas terms, count terms and aspectual classes, po
larity phenomena, and focus. He has also worked on the syntax of Swahili.
MARK LIBERMAN obtained a doctorate in linguistics fom the Masachu
setts Institute of Technology in 1975. Fom 1975 to 1990, he worked at
AT&T Bell Laboratories, for the lat three years as head of the linguis
tics research department. He is now Tustee Professor of Phonetics at the
University of Pennsylvania. His current research interests include prosody,
phonetics, and computational linguistics, especially the problem of linguis
tic inference.
WILLIAM POSER is Associate Professor of Linguistics at Stanford Unver
sity. His interests include the phonetics/phonology interface, the formal
properties of phonological rules, phrasal phonology, and Japanese.
IVAN A. SAG is Professor of Linguistics and Symbolic Systems at Stanford
University and Vice Chair of the Department of Linguistics. He ha worked
mostly in the area of syntax, semantics and language processing.
RICHARD SPROAT holds a doctorate in linguistics from the Massachusetts
Institute of Technology. His research interests include morphology, com
putational morphology, text-to-speech synthesis, Chinese computational
linguistics, phonetics, and Celtic syntax. He is a member of the technical
staf in the Linguistics Research Department at AT&T Bell Laboratories.
ANNA SZABOLCSI has worked on performatives and focus in Montague
grammar, on the structure of the noun phrase in government and binding
theory, and on binding and extraction in combinatory categorial grammar.
She is currently Associate Professor of Linguistics at the University of Cal
ifornia at Los Angeles.
CAROL TENNY is Assistant Professor of Linguistics at the University of
Pittsburgh. Her research interests include theoretical syntax, the intersec
tion of syntax and semantics, and the representation of time in natural
language.
I
The Aspectual Interface Hypothesis
CAROL TENNY
1.1 Introduction: The Linking Problem
The problem of the linking of thematic roles to syntactic argument positions
is of fundamental importance because it is part of the larger problem of the
connection between syntax and semantics. Theories of generative grammar
have adopted, to a greater or lesser extent, the view of an autonomous syn
tax operating independently fom the semantics of a language. Up to a
certain point, this view does seem to characterize the behavior of natural
language. However, the problem remains that, in spite of the evidence for
the autonomy of syntax and semantics, there are strong generalizations to
be made about correspondences between meaning and syntactic structure.
There are familiar facts that cannot be ignored; such the fact that in
general, agents are subjects, and themes or patients are objects. Lexical se
mantics and syntax clearly interact-the problem is how to constrain that
interaction. Various mechanisms that have been proposed to deal with
this, such lists of linking rules that connect particular thematic roles
with particular syntactic or confgurational argument positions, are some
what stipulative and not entirely satisfactory. More principled approaches
to the problem have been presented in the recent literature, in the form of
hypotheses that there are uniform and universal constraints on the map
ping between syntax and lexical semantics. Perlmutter and Postal (1984)
proposed the Universal Alignment Hypothesis (UAR) in the famework of
Relational Grammar:
This paper is a 1989 revision of material, some of which wa presented at LSA 1987
and NELS 1987 and appears in Prceedings of NELS 18, published by University of
Masahusetts at Amherst. am grateful to audiences at these meetings for their com
ments. have s benefted from comments by Jane Grimshaw, Beth Levin, James
Pustejovsky, Maika Rappaport, Anna Szabolcsi, and Robert Van Valin. This work ha
been supported by grants from the System Development Foundation and from the Kapor
Family Foundation. am grateful for their support.
Lexical Matters. Ivan A. Sag and Anna Szabo1csi, eds.
2 / CAROL TENNY
Universal Alignment Hypothesis (U
A
H): There exist principles of
universal grammar which predict the initial relation borne by each nominal
in a given clause fom the meaning of the clause.
And Baker ( 1985) proposed the Universal Theta Assignment Hypothesis
(UTAH) , in the famework of Government and Binding Theory:
Universal Theta Assignment Hypothesis (U
TA
H): Identical them
atic relationships between items are represented by identical structural
relationships between those items at the level of D-structure.
Both of these hypotheses propose a fundamental connection between
"meaning" and some level of syntactic representation. The UAH maintains
that general principles constraining the mapping fom lexical semantics
to syntax do exist, although it gives no indication of what they might
be. The UTAH claims that the mapping between thematic and structural
relationships is consistent, although it does not explain why the mapping
is the way it is. The UAH and the UTAH are elegant idea that explain
a variety of phenomena in a simple way. However, these hypotheses lack
an account of the central mechanism by which thematic structures and
syntactic structures are connected. The thesis of this paper is that certain
aspectual properties mediate between syntax and lexical semantics, and
provide a principled basis for hypotheses such a the UAH and the UTAH.
The Aspectual Interface Hypothesis (AIR) is proposed to supplant the UAH
and the UTAH:
Aspectual Interface Hypothesis: The mapping between thematic
structure and syntactic argument structure is governed by apectual prop
erties. A universal aspectual structure associated with internal (direct) ,
external and oblique arguments in syntactic structure constrains the kinds
of event participants that can occupy these positions. Only the aspectual
part of thematic structure is visible to the syntax.
This proposal takes a strong stand on the current controversy over whether
grammar needs to refer to thematic roles. Under this view, the syntax
proper does not need to "see" thematic roles. It only "sees" certain syn
tactic/
aspectual structures the thematic roles are asociated with.
This is also in the spirit of much recent work arguing that the thematic
content of thematic roles need not be referred to by syntactic operations.
See Belletti and Rizzi 1988; Rappaport and Levin 1988; Levin and Rap
paport 1986; Grimshaw 1987a; Rappaport, Laughren, and Levin 1987.)
This paper focuses on the central part of the aspectual structure pro
posed by the AIR: the correspondence between the argument in a semantic
representation which ha the special aspectual role of "meauring out the
event, " and the syntactic argument which may be characterized as the
verb's internal argument. The crucial constraint on this correspondence
may be stated as:
THE ASPECTUAL INTERFACE HYPOTHESIS / 3
( 1 ) The internal argument of a simple verb is constrained so that it either
undergoes no change or motion, or it undergoes change or motion
which "measures out the event" over time.
The statement above may be regarded as one clause in a fully articulated
set of principles of aspectual structure. The discussion in this paper is
concerned with simple nonstative verbs (although ( 1) is true of stative
verbs as well) . Verbs with more complicated argument structures, such
as propositional-argument-taking verbs, are unexplored in this paper and
must await further research.
Section 1. 2 outlines some assumptions about lexical and argument
structure on which the discussion in the paper depends. Then in Sec
tion 2, various lexical semantic phenomena are discussed which illustrate
the special aspectual role of the internal argument. These include afected
ness, unaccuativity, the locative alternation, and psych verbs. Section 3
addresses the question of what the AIR says about thematic roles. Section
4 considers the implications of the AIR for language acquisition, and points
out some constraints it places on the lexicon.
1. 2 Some Assumptions
The AIH as stated above adopts a certain view of argument structure and
lexical structure. It employs some ideas developed in the work of the MIT
Lexicon Project. The MIT Lexicon Project has been investigating what
information must be included in the lexical entry for a natural language
predicate; or (to look at it another way) what a native speaker must know
about predicates in his or her language, in order to be able to use them
correctly. (Hale and Keyser 1987; Rappaport and Levin 1986; Rappaport,
Laughren and Levin 1987. ) A structured lexical entry has been developed
which has two parts: a structural part and a conceptual part. For example,
Hale and Keyser (1988) show the lexical entry for cut as in Figure 1.
VP
.
V arg
T produces linear separation
I
in material integrity of y
by sharp edge coming into
contact with latter
Figure 1 Lexical Entry for English cut
The prose part of the entry expresses the conceptual part of the verb's
meaning, including the number of open arguments and the roles they play
4 / CAROL TENNY
in the event described by the verb. (Notice that thematic roles 6I S6 are
not mentioned. ) The structural part indicates how these roles are to be
mapped onto syntactic structures. The tree-like structure simply indicates
that the
u
g
argument must be the syntactic object of the verb. What is

crucial here is that the verbal entry ha two parts: a semantic/conceptual
part and a syntactic/structural part; and it indicates the nature of the
correspondence between them. The AIR drives that correspondence, and
it has a a consequence the requirement that the argument linked to the
verb's direct object position must be apectually constrained.
The view of argument structure asumed here includes a distinction
between external and internal arguments, and a distinction between direct
internal and indirect internal arguments. (See Williams 1980 on external
arguments. ) External or internal argumenthood i essentially a syntactic
property. Argument structures are asumed to map onto syntactic struc
tures in a straightforward way, so that internal and external argumenthood
may be equated with syntactic positions. The expression "syntactic argu
ment structure" is used here to refer to these arguments mapped onto their
syntactic positions. In this paper the syntactic consequences of the AIR
are cat in the famework of Government and Binding Theory, but this
is not essential. I believe they may also be expressed in other syntactic
fameworks.
Employing the tools of Government and Binding Theory, external ar
guments are generated outside the verb phrae and are not governed by
the verb at D-structure. Internal arguments are generated within the verb
phrae. Direct internal arguments are generated as structural objects of
the verb, and are asigned structural cae by the verb, while internal in
direct arguments (oblique arguments) do not receive structural cae fom
the verb. Instead they are asigned inherent case or are assigned case by a
preposition. These three argument-types-external, direct, and oblique-
are asumed to be asociated with these distinctive syntactic positions. I
believe these defnitions of external, direct and oblique arguments may be
translated into other fameworks. What is necessary is simply that there
be some representation of argument structure which includes aspectual in
formation, and which ha some clear mapping to syntax.
2.1 Aspect ad the Internal Argument
In the theory presented here, the crucial aspectual property is asociated
with the direct internal argument of a verb. The direct internal argument of
a verb of change "meaures out" over time the event described by the verb.
The verb's direct internal argument may be thought of as being converted
into a function of time at some level of semantic representation. This is an
apectual property, because apect refers to the internal temporal organiza
tion of an event. The term "measures out" is used here in an informal sense,
a a convenient metaphor for uniform and consistent change, such as change
along a scale. The idea of an object measuring out an event can be eluci
dated using the aspectual property of delimitedness. A delimited event is
one that the language encodes a having an endpoint in time. The apectual
distinction between delimited and nondelimited events has a long history of
discussion in the philosophical and linguistic literature, dating back to Aris
totle. (Particularly notable in the recent literature are Vendler 1967, Dowty
1979, Hinrichs 1985, and Verkuyl 1972, 1976, 1978, 1987ab.) The diference
between delimited and nondelimited events is illustrated in (2) and (3) .
(2) Delimited:
a. destroy the city (in an hour/* for an hour)
b. climb a tree (in an hour/?for an hour)
(3) Nondelimited:
a. like jazz ( *in a day/for a day)
b. push the car ( *in an hour/for an hour)
The events described by the verb phraes in (2a) and (2b) are delimited
events; those in (3a) and (3b) are nondelimited events. Temporal adverbial
expressions such as in an hour and for an hour are usefl in distinguishing
between a delimited and a nondelimited reading.
1
Now consider the verb phrases below. In their salient readings, these
describe delimited events: 2
(4) a. perform a play
b. translate a poem
Not only do these events have temporal bounds, but these bounds are
provided by the referent of the internal argument. In other words, it is the
object that delimits the event. When you perform a play, you perform act
one, then act two, and so on, until you come to the end of the play. The
end of the play is the end of the event. Likewise, in translating a poem,
one may translate the frst stanza, then the second stanza, then the third,
until the end of the poem is reached and the translation event is over.
A very coarse test for this idea is provided by certain adverbials such
as halfay, which make reference to a kind of measurement:
(5) a. perform a play halfay
b. perform half a play
c. translate a poem halfay
d. translate half a poem
e. destroy the city halfway
f. destroy half the city
IThese adverbial expresions, however, are sensitive to lexical subtleties that make
them imperfect diagnostics of delimited ness. The defnition of delimitedness must be
baed on the existence of an understood temporal bound to the event, rather than solely
on tests using adverbials.
2
Some speakers fnd these to be vague between a delimited and a nondelimited reading.
6 / CAROL TENNY
Although the two verb phrases in each couplet do not mean exactly the
same thing, in each pair the second verb phrase represents one possible way
to understand the frst. That is, halfway through the event may be equated
with half of the object.
In the examples above, the direct internal argument meaures out the
event through its spatial extent or volume. This makes these examples
particularly clear. However, the argument need not measure out the event
through its spatial extent or volume; some other property of the object
may measure out the event, as in (6) below:
(6) a. redden the photograph
b. ripen the fuit
In (6a) it is the redness of the object, and in (6b) it is the ripeness of the
object that measures out the event. In both cases the verb names that
property of its argument which is to do the measuring.
The direct internal arguments in the verb phrases in (3) through (5)
above measure out and delimit the event described by the verb. However,
under the theory proposed here, all direct internal arguments undergoing
change are constrained to measure out the event, whether or not it is a
delimited event.
Consider ( 7) below:
(7) a. push the cart ( *in an hour/for an hour)
b. push the cart to New York (in an hour/?for an hour)
In (7a) , the verb and its direct internal argument describe a nondelimited
event, rather than a delimited one. In (7b) however, with the addition
of a goal phrase (an indirect internal argument) the event becomes delim
ited. Nevertheless, the cart still measures out the event in both of these
expressions. The delimited ness of ( 7b) is achieved through reference to the
very property of the direct internal argument that is meauring out the
event: namely, its location. The indirect internal argument-or the goal
phrase-delimits the event by referring to that property of the direct ar
gument that is undergoing the central change in the event. Example (7)
above illutrates how the arguments of verbs that describe nondelimited
events may be unifed with arguments of verbs describing delimited events
such as those in (3), (4) and ( 5) . Verbs such as push may be subsumed
under the condition that the direct internal argument of the verb is con
strained to measure out the event through a change in a single property.
The change in the direct internal argument during the course of the event
must be describable as a change in a single property. The event may be
delimited linguistically through reference to that change or that property.
To summarize the discussion thus far, the following hypotheses have
been made. All direct internal arguments undergoing change in the event
described by the verb measure out the event. Direct arguments are distin
guished arguments in this respect. With some verbs ( perform, trnslate,
destroy, redden, ripen) the direct internal argument measures out and de
limits the event. With other verbs ( push) the direct internal argument
measures out the event and an indirect internal argument may delimit it.
Only internal arguments can delimit the event, and they can do this only
by reference to the measuring performed by the direct internal argument.
Various elements of the idea proposed in this paper-that an internal
argument measures out an event-may be found in Gruber 1976; Jackendof
1987; Dowty 1979, 1988; Krifa 1987; Verkuyl 1972, 1976, 1978, 1987ab;
Hinrichs 1985; and Pustejovsky 1988.3 However, none of these authors has
extended the idea to include internal arguments in general, or employed
the idea a a truly general principle of correspondence between syntactic
structure and semantics.
The idea that only the direct internal argument measures out the event
explains an asymmetry between internal and external arguments regarding
the interaction of the aspectual properties of NP's and VP's. There is an
interesting asymmetry between internal and external arguments in the in
teraction of mass noun arguments with delimitedness. There is a class of
verbs which translate count-ness of the internal argument into delimited
ness of the event.4 When the internal argument is a count noun the event
is delimited (8a) , and when it is a mas noun it is nondelimited (8b). The
internal arguments are underlined in (8) below:
(8) a. Charles drank a mug of beer (??for an hour/in an hour) . (delim
ited)
b. Charles drank beer (for an hour/*in an hour) . (nondelimited)
To the best of my knowledge there is no corresponding clas of verbs that
translate count-ness of the external argument into delimited-ness of the
event. Whether the external arguments underlined in (9) and ( 10) are
mas or count nouns has no efect on the delimited-ness of the event:
(9) a. The heater melted the candle. (delimited)
b. Heat melted the candle. (delimited)
( 10) a. Snow surrounds the house. (nondelimited)
b. Seven trees surround the house. (nondelimited)
3
Krif (1987) and Hinrichs (1985) have written on certain delimiting predicates which,
roughly speaking, express homomorphisms from their arguments to the events they
decribe. Dowty ( 1991) ha employed this idea a one contributing property for the
ProtoPatient Role, mapping "incremental themes" (such as house in build Q house)
onto the ProtoPatient role. (See Section 2.7.)
4
It u well known that bare plurals in either subject or object position afect the as
pectual properties of a sentence. A bare plural subject or object makes it possible for
the sentence to describe some event of indefnite duration. Bare plurals, which behave
like mas nouns in many repects, do not exhibit the subject/object asymmetry de
scribed above. They must be considered a separate phenomenon for the purpose of this
discussion.
8 / CAROL TENNY
Mas/count properties of internal arguments but not of external arguments
interact with the aspectual properties of verb phraes and sentences. This
aymmetry between internal and external arguments makes sense in light
of the requirement that only internal arguments measure out events. 5
2. 2 Afectedness
The property of afectedness ha been much discussed in the literature
lately. Afectedness is interesting for two reaons: (i) because it is a se
mantic property that ha been implicated in certain syntactic phenomena,
and (ii) because it is always associated with direct internal arguments.
An afected argument ha been generally described as an argument which
undergoes some change. Undergoing change is a temporal process. An
afected argument can be more adequately described in apectual terms, as
an argument which measures out and delimits the event described by the
verb. An afected argument meaures out the event by virtue of its being
a direct internal argument. What is special about it is that it also delimits
the event. The characterization of these arguments a delimiting the event
unifes a larger clas of verbs having the same syntactic behaviors than
does the idea of being afected by some action. Although this fact does not
explain why these syntactic behaviors correlate with apectual properties,
it does explain why the crucial aspectual property is asociated with the
direct internal argument.
The correlation of afectedness with delimited-ness can be illustrated
by considering some of the syntactic phenomena in which afectedness
ha been implicated. Among the syntactic phenomena that afectedness
is relevant to are middle formation and NP pasivization. (See Ander
son 1979, Roberts 1985, and Hale and Keyser 1987 on these topics. The
aspectual nature of the afectedness constraint on NP-pasivization has
also been noted by Fellbaum (1987) . ) Consider the NP-pasives in ( 1 1 )
and ( 12) .
( 1 1 ) a. The Mongols' destruction of the city.
b. The city's destruction by the Mongols.
c. The missionaries' conversion of the natives.
d. The natives' conversion by the missionaries.
( 12) a. John's avoidance of Bill.
b.*Bill's avoidance by John.
c. Sally's pursuit of the cat.
d. *The cat's pursuit by Sally.
( M. Anderson 1979 ( =( 51) , (53) ) )
In ( 1 1 ) the arguments are afected and NP pasivization is possible. In ( 1 2 ) ,
where the arguments are not afected, NP pasivization does not yield good
5
A full acount of this phenomenon must be cat in a sound theory of mas- and
count-ness, and is beyond the scope of this paper.
expressions. The possibility of NP-passivization correlates closely with the
delimited or nondelimited-ness of the event. This is shown in ( 13) by the
applicability of temporal adverbials:
( 13) a. destroy the city in a day /*for a day
(the city's destruction in a day/*for a day)
b. convert the natives in a month/??for a month
(the native's conversion by the missionaries in a month/*for a
month)
c. avoid Bill *in an hour/for an hour.
d. pursue the cat * in an hour/for an hour.
Not only do (13a) and ( 13b) describe delimited events, but they describe
events delimited by the change in the argument. The argument meaures
out and delimits the events in these examples.
The property that governs NP-passivization is actually an aspectual
property. This explains why some verbs like perform take NP-pasives
when they do not actually seem to afect their arguments, a in ( 14) .
( 14) a. The company's performance of the play.
b. The play's performance by the company.
(M.
Anderson 1979 (=(48) ) )
c. John's translation of the poem.
d. The poem's translation by John.
An apectual defnition of an afected argument as a delimiting argument
unifes a wider range of relevant data than a defnition based on the notion
of being afected by some action.
Now consider the middles in ( 15) and ( 16) .
( 15) a. This door opens eaily (by pulling on the handle) .
b. This cinch tightens easily (if you give it a good yank) .
(after Hale and Keyser 1987)
c. This door opens easily in a minute/?for a minute.
d. This cinch tightens eaily in a minute/?for a minute.
(16) .*The trafc jam avoids easily.
b. *Fleeing buglars pursue easily.
c. avoid the trafc jam *in a minute/for a minute
d. pursue the burglar *in an hour/for an hour
In ( 15) the arguments are afected arguments, and middles are possible.
In (16) they are unafected and middles are not possible. Again, in ( 15)
the possibility of middle formation correlates with the delimited-ness of the
event, and in ( 16) the impossibility of middle formation correlates with the
nondelimited-ness of the event, as shown by the adverbials. The arguments
in ( 15) are delimiting and those in ( 16) are nondelimiting.
The distinction between delimiting and nondelimiting arguments ex
plains the diference in judgments speakers assign to the sentences in ( 17) .
10 / CAROL TENNY
( 17) a. The desert crosses easily for settlers with large wagons.
b.*The desert wanders easily for settlers with large wagons.
c. The globe circumnavigates in a day/easily with Pan Am
d. *The globe travels in a day/easily with Pan Am.
e. The enemy battalion infltrated surprisingly easily for the guerrilla
soldiers.
f.*The enemy battalion pursued surprisingly easily for the guerrilla
soldiers.
Although judgments about middles are notoriously murky, most speakers
agree with the relative grammaticality judgments illustrated here.6
The examples above show that afected arguments are delimiting argu
ments, and the notion of delimiting argument is more appropriate than that
of afected argument as a criterion for syntactic behavior. It is well-known
that afected arguments are always direct internal arguments of the verb.
This is predicted by a theory in which direct internal arguments measure
out events. Afected arguments are a subset of the class of direct internal
arguments; they are those arguments that not only meaure out but also
delimit the event.
To summarize, the aspectual constraints on internal arguments give H a
handle on the semantic property of afectedness, predict that afected argu-
6
The delimiting argument constraint on middles and NP-pasivization predicts that
the same clas of verbs will enter into both of thee constructions. This prediction is by
and large borne out. Verbs that form middles also form NP-pasives:
(i) the desert's crossing by settlers with large wagons
(ii)*the desert's wandering by settlers with large wagons
(iii) the globe's circumnavigation by Pan Am
(iv)*the globe's traveling by Pan Am
(v) the enemy battalion's infltration by the guerrilla soldiers
(vi)*the enemy battalion's pursuit by the guerrilla soldiers
Verbs that form NP-pasives (i-iv below) form relatively good middles compared to
verbs which do not form NP-pasives (v and vi below) , but middle formation is not a
productive a NP-passivization:
(i) ?Shakespeare's Othello performs more eaily with a good company than with a
bad one.
(ii) Haiku poetry does not translate at all.
(iii) ??European explorers found that native convert with the sword more eaily than
with the word.
(iv) (??)(*)The munitions research industry predicts that major cities will destroy
quickly and efortlessly with the weapons of the future.
(v)*Policemen avoid eaily in a crowd.
(vi)*Burglars pursue eaily on foot.
Examples (i), (ii), and (iv) above are poor to unaceptable for most speakers but they
improve with exposure. Middle formation appears to be slightly more constrained than
NP-pasivization. Since speakers' judgments about middles show s much variability,
it is likely that there are additional subtle factors at work with middles. It is also
possible that middle formation is more of a lexical phenomenon than NP pasivization.
Nevertheless, the delimiting constraint picks out by and large the correct clas of verbs
for both middle formation and NP-pasivization.
THE ASPECTUAL INTERFACE HYPOTHESIS / 1 1
ments should be internal arguments, and substitute the notion of delimiting
argument for afected argument as a criterion for certain kinds of syntactic
behavior. And although the reason for the interaction of delimited-ness
with middle-formation and NP-passivization is still obscure, recognition of
the fact that delimited-ness has syntactic properties is a step towards a
deeper understanding of the syntax of these constructions.
2. 3 Unaccusative and Unergative Verbs
Following the Unaccusative Hypothesis, unaccusative verbs are verbs whose
sole argument is an internal direct argument, while unergative verbs have
an external argument as their sole argument. (The Unaccusative Hypothe
sis and the representation of unaccusativity in syntactic theory is discussed
by Perlmutter and Postal 1984 and Burzio 1986. For contrasting views and
problems, see Van Valin 1987 and Zaenen 1987. For a useful overview of is
sues see Grimshaw 1987b.) These two classes of verbs constitute a minimal
pair with which to investigate the properties of internal and external ar
guments. The internal or external status of the arguments of unaccusative
and unergative verbs is a syntactic fact. However, it is well known that
there are strong general semantic tendencies associated with this syntactic
distinction. Unergative verbs are usually verbs in which the argument en
gages in some kind of volitional activity, while unaccusative verbs describe
situations in which the argument undergoes some kind of change. Unac
cusatives usually assign a patient or theme thematic role to their argument,
while unergatives more ofen than not assign agent thematic roles. These
familiar facts are illustrated by the following examples:
(18) a. unergatives: (external arguments; agent)
run, dance, whisper, bicycle, study
b. unaccusatives/ergatives: (internal arguments; patient, theme)
melt, feeze, evaporate, fall, open, collapse
In a model of grammar in which syntax and semantics are strictly au
tonomous, such a correlation of semantic properties with syntactic struc
tures is difcult to express. In a model in which a wide range of semantic
properties is visible to the syntax, the connection between syntax and se
mantics is difcult to constrain. The semantic generalizations about these
verbs are too strong to ignore. The proposal that the internal argument is
constrained to measure out the event provides a simple explanation. There
are difering apectual constraints on internal and external arguments, and
these aspectual constraints afect what types of thematic roles may occupy
those positions. Applying the adverbial halfway brings out a diference
between unaccusatives and unergatives:
(19) a.*Martha danced halfway.
*Thoma ate halfway.
* Jack whispered halfway.
12 / CAROL TENNY
b. The lake froze halfway.
The candle melted halfway.
The barn collapsed halfway.
Halfay applies naturally to the unaccUsatives in ( 19b) but not to the
unergatives in ( 19a) . 7
Adverbial phrases such a a little bit at a time, or a lot at once provide
additional tests. These are illustrated in (20) and ( 21) .
(20) a. The dancer danced slowly/*a little bit at a time.
The announcer talked slowly /* a little bit at a time.
The actress whispered slowly /*a little bit at a time.
b. The candle melted slowly / a little bit at a time.
The lake froze slowly / a little bit at a time.
The barn collapsed slowly/a little bit at a time.
( 21) a. ??Martha danced quickly, a lot of her dancing at once.
??The announcer talked quickly, a lot of him talking at once.
??The actress whispered quickly, a lot of her whispering at once.
b. The candle melted quickly, a lot of it melting at once.
The lake foze quickly, a lot of it feezing at once.
The barn collapsed quickly, a lot of it collapsing at once.
These expressions are awkward when used in conjunction with an unerga
tive verb, but quite natural when used with unaccusative verbs.s
The proposal that the internal argument is constrained to meaure out
the event through some property it possesses, while the external argument
is not so constrained, explains the semantic distribution of verb mean
ings across unergative and unaccusative verb clases. Those verb meanings
which become unaccusative verbs describe exactly those event types in
which the event participant may be construed a meauring out the event.
Verb meanings in which the event participant may not be construed as mea
suring out the event must become unergative verbs. And fnally there will
be a clas of verb meanings that may be construed in either way, and these
will be unaccusative in some languages and unergative in other languages.
This is the nature of the correlation that is attested cross-linguistically.
( Merlan 1985 examines this question in detail. ) The view advanced here
of the connection between lexical semantics and syntactic argument struc
ture predicts this. It maintains that there is no absolute mapping fom
verb meanings to syntactic argument structure. Rther, there is a kind of
7
There is a special sense of Martha danced halfay which is acceptable-namely that in
which Martha dances halfway to some destination. But in this cae the property which
is meauring out the event (location), although it u changing in the external argument
(Martha) , can only be expressed in its pure form through an internal argument, a in
Martha danced half the distance (home).
8
There are some verbs to which these adverbials do not eaily apply. Additional tests
need to be devised. The reader is reminded that the adverbials serve only to elucidate
the apectual propertie of internal arguments. They are not diagnostics.
event template asociated with argument structure that acts as a flter on
verb meanings. This event template consists of a set of universal aspectual
principles of argument structure, including the principle that the internal
argument measures out the event. 9
Van Valin ( 1987) showed that the aspectual properties of verbs can be
used to determine their class membership in the unaccusative or unerga
tive verb classes, and that Dowty'S aspectual calculus for lexical decompo
sition is an efective language for doing so. The theory proposed here has
much in common with Van Valin's proposal, but it difers in two crucial
ways. In both theories there is a mapping fom aspectual properties to
syntax, directly or indirectly. But according to Van Valin, the mapping
fom apectual properties of verbs to syntax is constrained only in that
apectual properties must be statable in a Dowty-style calculus. These
constraints may be language-particular. Under the view advanced here
there is one overriding and universal aspectual constraint; that the internal
argument, associated with the D-structural object position, measures out
the event. The internal argument must be convertible into a (nontrivial)
function of time. This overriding aspectual constraint unifes unaccusativ
ity with other verbal phenomena. The two views are not incompatible,
however, since within the limits of this general aspectual constraint, there
is room for a variety of mappings from Dowty-style aspectual properties
into syntax. Secondly, under the theory proposed here (which adopts the
Unaccusative Hypothesis) aspectual properties are directly correlated with
syntactic structures. This view says that certain apectual properties are
in fact syntactic properties, and predicts that further repercussions of these
properties may be found in syntax.
The aspectual properties of internal and external arguments also pro
vide a reason for the familiar fact that agents are never direct internal
arguments. lO Agents are a type of event participant which are inherently
90ne might say, taking this view, that there is a mapping from verb meanings to as
pectual structure. However, without a pre-apectual theory of verb meanings it would
be impossible to describe this as an absolute mapping. Even with such a theory it
may not be possible, since factors as imprecise as religious or cultural views may in
fuence what type of aspectual properties certain kinds of events are perceived by lan
guage speakers to have. It must be emphasized that this theory does not provide the
tools for a perfect and absolute mapping from pre-apectual verb meanings to syntac
tic structures. In a certain sense it preserves the essential autonomy of syntax and
semantics.
lOTwo well-known potential counterexamples to this generalization are syntactically
ergative languages, and certain lexical causatives. There is some doubt among re
searchers on Australian and Eskimo languages that syntactically ergative languages
do exist. (Jane Simpson (personal communication) and Alana Johns (personal com
munication), respectively. Also see Heath 1979 on Dyirbal. ) However, if there are
true syntatically ergative languages, they pose an interesting problem for the AIH. In
most general terms, the AIH requires that there be a distinction between arguments
that meaure out the event and those that do not, and that this distinction corre
lates with syntactic structure. Furthermore, the argument that meaures out the event
14 / CAROL TENNY
unsuited to meauring out an event on a scale. An agent by its very na
ture is fee to act in unspecifed ways to efect something. The actions
required of an agent are underspecifed and also not necessarily consistent
throughout an event, and so are not naturally construable as a scale on
which something could be measured.
2.4 The Locative Alternation
Next consider the locative alternation, investigated by Schwartz-Norman
(1976) , S. Anderson (1977) , and Rappaport and Levin ( 1986). The verbs
that participate in this alternation are verbs which take a goal and a theme
argument, either of which can be the direct argument of the verb. For
example:
(22) a. spray paint on the wall
b. spray the wall with paint
Either the goal, wall, or the theme, paint, can be the direct argument of
the verb spry. Only a very particular class of verbs show this alternation.
Not all verbs that have a goal and a theme argument can undergo this al
ternation. The verb push also takes a goal and a theme, but the alternation
is not possible in (23) :
(23) a. push a cart to San Fancisco
b.*push San Fancisco with a cart
The possibility of the locative alternation seems to depend on a very narrow
semantic property; the goal must be a fat surface or container, and the
theme must be a material which is removed or applied to that surface or
container. The verbs in (24 ) illustrate these properties:
(24) a. load hay on the wagon
b. load the wagon with hay
c. clear dishes fom the table
d. clear the table of dishes
e. cram pencils into the jar
f. cram the jar with pencils
Furthermore, this generalization holds across a wide variety of unrelated
languages. Examples (25) and (26) illustrate the alternation in Dutch and
Japanese:
must be closer syntactically to the verb than arguments that do not. If there are lan
guages in which the agent is closer to the verb than the theme (under deep syntactic
analysis) then this requirement is violated. If such languages exist, the details of this
mapping from apectual structure into syntactic structure would have to be revised.
The second cae is that of lexical causatives such a gal/op in John gal/oped the horse,
where the internal argument, the horse, ha agentive powers. Nevertheless, in this
sentence the horse's change of location is meauring out the event. (Consider John
gal/oped the horse to the edge of the feld. ) This example underscores the fat that
traditional thematic role labels may be too coarse to express the relevant apectual
properties.
(25) Dutch:
a. Jan plant bomen in de tuin.
John plants trees in the garden.
b. Jan be-plant de tuin met bomen.
John be-plants the garden with trees.
(De Groot 1984, his source Dik 1980)
(26) Japanese:
a. kabe n penki D nuru
wall on paint-ACC paint(VERB)
smear paint on the wall
b. kabe D penki de nuru
wall ACC paint-with paint(VERB)
smear the wall with paint
(Fukui, Miyagawa, Tenny 1985)
The locative alternation is an alternation in syntactic argument structure
that depends on certain characteristics of the theme and the goal. Since
we know syntax and semantics to be autonomous to such a large extent,
we must ak why such seeIningly trivial semantic properties should have
syntactic repercussions such as these? The proposal that the direct internal
argument is constrained to meaure out the event described by the verb
provides a simple explanation. These verbs describe exactly those events
in which the goal may be construed a meauring out the event. If you
apply some material to a fat surface, or fll a container with something,
then the material spreads out across the surface or rises up to the top of
the container in a consistent and uniform way, and in so doing may be seen
to meaure out the event.
The apectual constraints on internal arguments predict certain other
faIniliar facts about the locative alternation. Consider (27) below:
(27) a. spray the paint in the hole
b. spray the hole with paint
Phrae (27a) means that the paint is directed into the hole but not nec
essarily spread around on the surface or wall of the hole. (27b) , however,
is not an accurate paraphrae of (27a) because it means that the paint is
spread around on the surface or wall of the hole. This is because when
the goal, the hole, is the direct argument of the verb it is constrained to
meaure out the event.
The locative alternation is impossible when an instrument thematic role
is substituted for the material thematic role. Both material and instrument
are possible as indirect objects:
(28) a. spray the wall with water
b. spray the wall with a hose
But only the material can be a direct argument:
16 / CAROL TENNY
(29) a. spray water on the wall
b.*spray a hose on the wall
This is because a material is consumed a little at a time until it is gone,
thus meauring out the event, but an instrument does not undergo change
in such a way that it can be construed as meauring out the event. In this
cae it is precisely the apectual properties of the material and instrument
thematic roles which determine how they are mapped into the syntax.
Finally, this approach to the locative alternation can be extended to
verbs that take "path" arguments. For example:
(30) a. walk the bridge
b. walk across the bridge
c. walk to the bridge
In (30a) the brdge is a path argument which meaures out the event by
virtue of being the verb's direct argument. Although both (30b) and (30c)
are acceptable expressions, only (30b) may be a paraphrae of (30a) . In
(30a) and (30b) the walking event is understood to traverse the bridge, so
that a certain amount of bridge correlates with a certain amount of walking.
The bridge meaures out the walking.
Under this view, it is the apectual properties of thematic roles that
are relevant for syntactic argument structure. The aspectual diferences
between materials and instruments, and the apectual properties of cer
tain types of goals, together with the principle that the internal argument
meaures out the event, are what make the locative alternation possible or
impossible.
1 1
2.5 Psych Verbs
Next consider psych verbs. These are verbs that have an argument bearing
an experiencer thematic role. (The syntax of psych verbs ha been discussed
by Grimshaw ( 1987a), Pesetsky ( 1987), and Belletti and Rizzi ( 1988) ,
among others. ) I n the sentences below, John i s an experiencer. The expe
riencer is an external argument in (31a) and an internal argument in (31b):
(31) a. John fears ghosts.
b. Ghosts fighten John.
ll
The apectual properties of thematic roles determine the possibility of the locative
alternation a allowed by a universal semantics. Language-particular lexicalization may
allow only a subset of these universal possibilities. We see this for example with fll:
(i) fi ll the cup
(ii) *fill the cofee
However, the prediction is that these "universal possibilities" allow fll to be re-analyzed
a taking the locative alternation:
(iii) ?fill the cofee up to the rim of the cup
Compare this with (23b ) repeated below:
(iv)*push San Francisco with a cart
THE ASPECTUAL INTERFACE HYPOTHESIS / 1 7
Psych verbs present interesting minimal pairs in which similar (or identi
cal) verbs assign the same thematic role--experiencer-to the internal or
external argument position. At frst glance there seems to be no diference
in meaning between the members of a pair like fear and frighten. However,
there is an important diference in meaning between verbs with internal
argument experiencers and external argument exeriencers, and this is the
diference predicted by an aspectual theory of argument structure. When
the experiencer is an internal argument it measures out the event, and
when it is an external argument it does not. 12 This can be illustrated by
adding delimiting expressions to the sentence:
(32) a.* Johni feared the truth into drinkingi '
b. The truth fightened Joh into drinkingi .
(33) a. ?? John feared the trut" into innocuousnessi.
b.* The trut fightened John into innocuousnessi '
Sentence (33a) is problematic because fear is a stative verb, and the role
of the internal argument a the measure of change is only apparent with
verbs of change. (33a) is only acceptable if fear is interpreted (or perhaps
re-analyzed) as a verb of change. If this is possible (and it may be for a
very few speakers) then (33a) means "John made the truth innocuous by
fearing it" or "John feared the truth so much it became innocuous for him."
This re-analysis is easier with other experiencer-external psych verbs:
(34) a. John resented his neighbor so much, he resented him right into
the hospital (by attacking him with a bat) .
b. Mary admires her brother to pieces.
These expressions, facetious and idiomatic as they are, serve to illustrate
that if these verbs can be understood a verbs of change, the universal
aspectual structure of verbs of change-in which the internal argument
measures out the event-is available. This aspectual structure is brought
out by delimiting expressions that refer only to the internal argument. The
delimiting expressions in (32) , (33) and (34) ( into drinking, into innocu
ousness, into the hospital, to pieces) , are actually resultative secondary
predicates. They refer to the central property of the internal argument
which is changing and measuring out the event. They delimit the event
by picking out some point in that ongoing change at which the event ter
minates. (The expression to pieces does so metaphorically if not literally. )
When the experiencer is the external argument, the event may not be de-
1
2
Belletti and Rizzi (1988) propose a structure for psych verbs with external argument
experiencers in which both arguments are generated under the verb phrae, with the
experiencer as the highest NP governed by the VP. This is not inconsistent with the
theory proposed here since what is important is the correlation of apectual and syntactic
properties. Belletti and Rizzi are correct, then it is the highest NP generated under
the VP which meaures out the event.
18 / CAROL TENNY
limited by referring to a property of that experiencer. When it is an internal
argument, it is quite natural to do so.
The apectual constraints on internal arguments provide an explana
tion for the familiar fact that secondary resultative predicates may refer to
the object but not the subject of the sentence. Resultatives are syntactic
constructions that depend on the aspectual structure built up by the com
position of the verb and its direct argument. Since the external argument is
excluded fom this apectual structure, resultatives may not be predicated
of external arguments. Consistent with this i the fact that a resultative
secondary predicate is constrained not just to refer to the object but to
refer to the property of the object that is meauring out the event. The
resultative and the internal argument must agree on what property that is.
H the delimiting expression in (32b) is changed to a locational goal phrae,
the sentence is bad unless it can be interpreted a expressing motion (i. e. ,
change in location) :
(35) The truth frightened John to Ohio.
Sentence (35) is incoherent if it is John's becoming afaid that meaures
out the event. But it is a good sentence if it is John's changing in location
that meaures out the event. In the latter cae the sentence would be
understood to mean that the truth made John go to Ohio by fightening
him. Note that this view also unifes resultatives and goal phrases under
one aspectual structure. It has been noted (by Simpson (1983), among
others) that resultatives and goal phrases pattern together in many ways,
and that they do not co-occur. This is not surprising given that they fulfll
the same role in the aspectual structure of the sentence.
The case of these psych-verb pairs is particularly telling for a number
of reasons. First of all, the thematic role of the arguments in internal and
external position is exactly the same. (See Pesetsky 1987 for another view.)
The aspectual diference between ' fear' and 'frighten' cannot be traced to
diferences in thematic structure. They derive only fom the diference in
syntactic structure.
Secondly, the object-orientation of resultatives is a syntactic fact of
great generality. The connection of this fact with the aspectual properties
of internal arguments underscores the idea that these properties are syn
tactically rather than thematically baed. The tight connection between
aspect and syntax suggests that we may expect to fnd futher repercus
sions in syntax, of the aspectual properties of syntactic argument structure.
And conversely, aspect may be used as a tool in syntactic investigations.
Finally, psych verbs have posed a problem for theories that tightly con
strain the mapping of lexical semantics into syntax, because there seems to
be no diference in meaning between comparable verbs with experiencers in
internal and external argument position. However, there is a diference in
meaning that follows fom the diferent aspectual roles of the experiencer
in internal and external argument position. The thematic role of experi
encer happens to be adaptable to either of the aspectual representations of
internal or external argumenthood. In other words, the mapping of lexi
cal semantics into syntax may be tightly constrained only if it is based on
apectual properties, rather than on thematic roles.
2. 6 Other Verbs
Unaccusative verbs, locative alternation verbs, and psych verbs provide
various kinds of evidence for the AIR. The aspectual properties of argu
ment structure (and particularly of direct internal arguments) are easily
apparent with these verbs. There are a few other classes of verbs in which
this apectual structure is less apparent. These include achievement verbs,
verbs with goal agents, and verbs with source agents. These verbs may not
argue for the AIH in a strong way, but they are not inconsistent with it.
Achievement verbs (e. g. , explode, die) describe events that happen in
stantaneously (or nearly so) . They have internal arguments which measure
out the event, even though the measuring has no extended duration:
(36) The terrorist exploded the bomb.
It is not the duration of the meauring that is important; it is the fact that
the internal argument registers the temporal end of the event. This is true
in (36 ) since it is the bomb that undergoes the change that delimits the
event, not the terrorist.
Verbs with goal agents (e. g. , acquire, buy, receive) are another inter
esting clas. Although the external argument, the agent, seems to be a
kind of locational endpoint for the event, it is the internal argument that is
actually meauring out the event over time. These verbs behave with rate
adverbials in the same way as do unergatives:
(37) a. John acquired the property slowly, acquiring a little of it at a time.
b. John acquired the property slowly, *a little of him acquiring it at
a time.
Verbs with source agents (e. g. , send, give, mail) also describe events in
which the external argument seems to mark a signifcant point of time in
the event:
(38) Christopher sent the package to Bill.
Christopher's act of sending and the package's arrival at Bill's location
both seem to mark a potential endpoint of the event. Nevertheless, it is
the change undergone by the package and not by Christopher, which is
referred to by the delimiting expression (the goal phrase, to Bill). In this
respect, send is similar to push, discussed in Section 2. 1. 13
The discussion up to this point has focused on verbs that describe events
13
Double object constructions based on these verbs pose interesting questions for the
AIH which will not adress here. Se Tenny 1987.
20 / CAROL TENNY
in which the internal argument undergoes some change or motion. It has
been argued that with all of these verbs, the change or motion in the
internal argument is such that it "meaures out the event," or is a function
of time. There are many verbs which are not statives, but which describe
events in which the internal argument undergoes no change:
(39) a. George pounded the wall.
b. Mary embraced her sister.
c. Carla studied the elk population.
The internal arguments of the verbs above do not "measure out the event"
in the sense described above. We can state the generalization about internal
arguments of verbs of change a follows:
(40) The internal argument of a simple verb is constrained so that it either
undergoes no change or motion in the event described by the verb,
or it undergoes change or motion which "measures out the event."
Even with verbs where the internal argument undergoes no change or mo
tion, the apectual structure in which the internal argument meaures out
the event is sometimes latent. A resultative can be predicated only of the
direct internal argument of a verb of contact. A direct internal argument
(41a) , but not an external argument (41b) or indirect internal argument
(41c) , can be forced to be understood as undergoing a change that delimits
the event:
(41) a. George pounded the walli to piecesi '
b.*Georgei pounded the wall to exhaustioni.
c.*George pounded on the walli to piecesi '
Thus the principle that internal arguments can measure out events has
wider generality than what might at frst appear to be the cae. This
suggest a bolder hypothesis as a corollary to (40) :
(42) All verbs of change (simple nonstative verbs) have latent in them
the apectual structure in which an internal argument can measure
out the event.
Principle (42) is proposed here a a direction for research. Whether this
universal potential aspectual structure is instantiated or not likely depends
on various language-particular conditions. Assuming modular principles
of grammar, aspectual structure imposes its own constraints, which then
interact with an assortment of other constraints, both in the core and the
periphery of the grammar.
2. 7 Summay
This section of the paper has argued that there is a strong correlation
between aspectual properties and syntactic structures, which constrains
the mapping of thematic roles into syntax. The aspect/syntax correlation
centers on the verb's internal arguments. With verbs in which the internal
argument undergoes any change or motion, all and only direct internal
arguments, or D-structural objects of the verb, measure out events. The
event is delimited linguistically within the verb phrase, through reference to
that "measuring out." These principles constitute the universal apectual
structure of verbs of change, or simple nonstative verbs. This aspectual
structure is not always fully realized with all verbs of change, but it is
latent in them and potentially realizable.
The correlation between aspectual properties and syntactic argument
structure exerts some constraint on the mapping of lexical semantics into
syntax, because only certain types of thematic roles are compatible with
the requirement that they measure out an event. This theory amounts to a
proposal for a certain kind of rigorous connection between lexical semantics,
or verb meaning, and syntactic argument structure. However, under this
view the essential autonomy of syntax and semantics is preserved, because
they interact only through a restricted aspectual vocabulary.
The aspectual theory of argument structure introduces a fundamental
semantic asymmetry paralleling the familiar syntactic asymmetry between
subjects and objects, or between external and internal arguments. In a
sense, this is a rather conservative proposal.
Finally, it must be noted that this apectual theory of argument struc
ture introduces an unusual view of apect. Aspectual properties have usu
ally been associated with predicates and predicative expressions, such as
verbs, verb phrases, or clauses. But with this approach, we can also talk
about the aspectual properties that arguments and noun phrases receive
fom their governing predicates.
3 The AIH and Thematic Roles
This paper argues for the existence of an aspectual interface between the
matic structure and syntactic structure. Under this view, thematic roles
are not directly visible to syntactic structure. This does not mean that the
matic roles do not exist, or do not need to be characterized in some fashion.
What implications does the AIR have for the nature of thematic roles?
The idea that it is the compatibility of various thematic roles with
certain aspectual constraints that governs their mapping into syntactic ar
gument positions imparts a view of thematic roles like that propounded
in Dowty 1991. Dowty argues that linguistic theory needs to recognize
two proto-roles, a Proto-Agent and a Proto-Patient, instead of a set of dis
crete thematic roles; and these proto-roles are "higher order generalizations
about meanings" (Dowty 1991). Associated with these proto-roles are col
lections of agent-like and patient-like characteristics which are the criteria
by which argument selection takes place (fom Dowty 1991):
Argument Selection Principle: The argument of a predicate having
the greatest number of Proto-Agent properties entailed by the meaning
22 / CAROL TENNY
of the predicate will, all else being equal, be lexicalized a the subject of
the predicate; the argument having the greatest number of Proto-Patient
properties will, all else being equal, be lexicalized as the direct object of
the predicate.
Corollay 1: If two arguments of a relation have (approximately) equal
numbers of entailed Proto-agent and Proto-patient properties, then either
may be lexicalized a the subject (and similarly for objects).
Under this approach thematic roles are essentially generalizations about
elements of meaning which map to internal or external argument positions.
This view of thematic roles is highly compatible with the AIR.
Dowty lists the properties which contribute to Proto-Agent and Proto
Patient roles a:
Contributing Properties for the Agent ProtoRole:
a. volition
b. sentience
c. cause event
d. movement
(e. referent exists independent of action of verb)
Contributing Properties for the Patient ProtoRole:
a. change of state (including coming-to-being, going-out-of-being)
b. incremental theme (Le. , determinant of aspect)
c. causally afected by event
d. stationary (relative to movement of Proto-Agent)
(e. referent may not exist independent of action of verb, or may not exist
at all)
These properties are aligned according to whether or not they can "measure
out the event. " The Proto-Patient properties are conducive to doing this;
the Proto-Agent properties are not. The AIR unifes changes of state,
afected theme and incremental theme verbs under one system. In fact, the
AIR derives the distinction between Proto-Agent and Proto-Patient roles,
which otherwise is unmotivated. Dowty'S theory of thematic roles is not
only compatible with the AIR, it is motivated by it.
4 Implications for Language Acquisition
The claim that predicates of natural language characterize events in a very
particular and constrained fashion has implications for cognitive science in
general, and for language acquisition in particular. I will close with some
brief speculations about issues relating to language acquisition.
The Aspectual Interface Hypothesis has three specifc consequences for
language acquisition; it underscores the necessity of syntactic information
i n the acquisition of verbs, i t predicts certain interactions between cogni
tive and linguistic development, and it makes predictions about what are
impossible verbs.
The importance of syntactic information to verb meanings is refected
in the practice of many linguists, of including within the lexical entry of
a verb some indication of how the verb' s arguments are to be realized
syntactically. Because the syntactic information cannot be derived from
the core (pre-apectual) meaning of a verb by any perfect mapping, it
must be independent information that has to be acquired and listed in the
lexicon. Among researchers on language acquisition, the claim has been
made that syntactic information is necessary for verb acquisition. The
claim ha been made in its general form by Landau and Gleitman ( 1985)
and as a more particular thesis by Gropen, Pinker, and Goldberg ( 1987) .
Gropen et al. report on experiments which indicate that children rely on a
principle which asociates afectedness (an apectual property) with direct
object-hood (a syntactic property) , in learning the meanings of the locative
alternation verbs.
The importance of the acquisition of syntactic information does not
preclude the need for acquiring conceptual information or machinery as
well. The AIR predicts that cognitive development will enter into language
acquisition in very particular ways. The AIH constrains to some extent the
interaction of children's cognitive growth with their linguistic capabilities.
Children's understanding of the nature of events and change, and particu
larly their ability to recognize properties that "meaure out events" should
afect the difculty with which they learn various predicates.
A theory proposing a correlation between apectual and syntactic struc
ture makes strong predictions about possible and impossible verbs. If the
view advanced in this paper is correct, verbs that violate the aspectual con
straints on argument structure should be impossible verbs. They should
not be found in any language, and they should be impossible for children
to learn through normal processes of language acquisition. In the next few
paragraphs a few "impossible verbs" are described which violate apectual
principles of argument structure.
Verbs with two or more direct interal arguments: Verbs of change
describe events measurable on a single scale, which can only be associated
with a direct internal argument. A verb with two or more direct internal
arguments, both meauring out the event, would violate this principle.
Such a verb would describe an event meaured out on two independent
scales. An example would be a verb meaning "A melts and B feezes,"
such that the melting of A and the feezing of B independently defne or
measure out the event. This event comes to an end when A has melted and
B has fozen. Such a meaning i expressible through conjunction, although
it cannot be lexicalized in a single verb. Note that a verb meaning "A
melts and feezes" would be a possible verb. The scale on which this event
24 / CAROL TENNY
is meaured out would be a little peculiar to characterize, but it would be
a single scale.
Verbs with two or more delimiting oblique arguments: An event de
scribed by a verb may be delimited only once. A verb with more than
one delimiting argument would violate this principle. This would be a
verb taking two or more arguments indicating endpoints to the event
for example, two or more goal arguments. Such a verb would be used
a follows: Push the cart to San Frncisco to Mexico, meaning push
the cart to San Fancisco and to Mexico. Again, the constraint on this
structure is not pragmatic because the meaning can be expressed through
conjunction.
Verbs with no interal arguments ( including implicit or default argu
ments): If all verbs of change describe events meaurable on some scale,
then all verbs of change have potential internal arguments. A verb violat
ing this principle would be a verb which indicates no scale on which the
event is meaured out. There are many verbs which occur with an exter
nal argument and without an overt internal argument, but these verbs can
generally take some kind of internal argument such as a cognate object
or a refexive. However, the impossible verb being invented here would
have no way of taking any internal argument or direct object. There is a
strong claim here; that there may be verbs that can never take external
arguments but there are no verbs that can never take internal arguments
(including cognate objects and refexives) . This claim may be too strong,
but the prediction would be that where default internal arguments are im
possible, it is because of stylistic, pragmatic or other extragrammatical or
language-particular considerations.
Verbs with two or more exteral arguments: Although the AIH provides
no explanation of why a verb can only have one external argument, it does
characterize external arguments aspectually. A verb taking two or more
external arguments has not been attested, as far as I know. If there were
such a verb it would be a verb which takes two arguments, neither of which
measures out the event, or participates in the internal aspectual structure
of the event. An example would be a verb that means "some indefnite
or incompletely specifed change is happening with argument A and some
indefnite or incompletely specifed change is happening with argument B. "
This verb would have as part of its meaning a notion similar to "while. "
An example would be a verb with arguments A and B, meaning:
A scratches and B dances
A sings while B jumps
This does not rule out a verb that means something like "A makes B dance
by scratching," in which case B's traversal of the dance event is measuring
out the event; and the scratching belongs in a manner clause, without
aspectual constraints. (Dance Bill up the stairs (by scratching).) (This
is putting aside the special properties of syntactic causative constructions,
which are not lexicalized events. )
A verb describing an event in which one argument measures out the
event and the other does not is always possible, even if the change or
activity of the external argument is seemingly unrelated to the meauring
out change. A verb meaning something like "A scratches and B gets wetter"
is in principle a possible verb, if one of the arguments has the properties
of an external argument, and the other of an internal argument. It is only
necessary for such a concept to be useful to the speakers of the language
for such a verb to be created. Such a verb might mean something like "A
scratchingly wets B."
Several kinds of aspectually "impossible verbs" have been imagined and
described above. This kind of prediction about possible or impossible verbs
may be testable in language acquisition experiments. I know of no exper
imental work that sheds light on them, and the appropriate tests may
be difcult to design. Nevertheless, the aspectual properties of argument
structure introduced by the AIR do make strong, and in principle falsifable,
predictions for natural language acquisition.
References
Anderson, Mona. 1979. Noun Phrase Structure. Doctoral dissertation, University
of Connecticut.
Anderson, Stephen R. 1977. Comments on the paper by Wasow. In Formal
Syntax, ed. Peter Culicover, Thomas Wasow, and Adrian Akmajian, 361-
378. New York: Academic Press.
Baker, Mark. 1988. Incorportion: A Theory of Grammatical Function Changing.
Chicago: University of Chicago Press.
Belletti, Adriana, and Luigi Rizzi. 1988. Psych-Verbs and B-Theory. Natural
Language and Linguistic Theory 6(3): 291-352.
Burzio, Luigi. 1986. Italian Syntax. Dordrecht: Reidel.
de Groot, Caper. 1984. Totally Afected Aspect and Three-Place Predicates in
Hungarian. In Aspect Bound: A Voyage Into the Realm of Germanic Slavonic
and Finno- Ugrian Aspectology, ed. Casper de Groot and Hannu Tommola,
133-151. Dordrecht: Foris.
Dik, Simon. 1980. Studies in Functional Grammar. London and New York:
Academic Press.
Dowty, David. 1979. Word Meaning and Montague Grammar. Dordrecht: Reidel.
67(3):547-619.
Fellbaum, Christiane. 1987. On Nominals with Preposed Themes. In Proceedings
of NELS 1 7. Amherst: University of Massachusetts, GLSA.
Fukui, Naoki, Shigeru Miyagawa, and Carol Tenny. 1985. Verb Classes in English
and Japanese: A Case Study in the Interaction of Syntax, Morphology and
Semantics. Lexicon Project Working Papers 3. Cambridge, Mass. : MIT
Center for Cognitive Science.
26 / CAROL TENNY
Grimshaw, Jane. 1987a. Psych Verbs and the Structure of Argument Structure.
Manuscript, Brandeis University.
Grimshaw, Jane. 1987b. Unaccusatives-An Overview. In Proceedings of
NELS 1 7. Amherst: University of Masachusetts, GLSA.
Gropen, Jess, Steven Pinker, and R. Goldberg. 1987. Constrained Productivity
in the Acquisition of Locative Forms. Talk delivered at Boston University
Conference on Language Acquisition, 1987.
Gruber, Jefrey S. 1976. Lexical Structures in Syntax and Semantics. Amsterdam:
North Holland. Includes 1965 MIT PhD thesis, Studies in Lexical Relations;
reissued by Indiana University Linguistics Club (Bloomington, 1970) .
Guerssel, Mohamed. 1986. On Berber Verbs of Change: A Study of Tansitivity
Alternations. Lexicon Project Working Papers 9. Cambridge, Mas. : MIT
Hale, Kenneth, and S. Jay Keyser. 1986. Some Tansitivity Alternations in En
glish. Lexicon Project Working Papers 7. Cambridge, Mas. : MIT Center for
Cognitive Science.
Hale, Kenneth, and S. Jay Keyser. 1987. A View from the Middle. Lexicon
Project Working Papers 10. Cambridge, Mas. : MIT Center for Cognitive
Science.
Hale, Kenneth, and S. Jay Keyser. 1988. Explaining and Constraining the En
glish Middle. In Studies in Generative Approaches to Aspect, ed. Carol Tenny.
Lexicon Project Working Papers 24. Cambridge, Mass. : MIT Center for Cog
nitive Science.
Heath, Jefrey. 1979. Is Dyirbal Ergative? Linguistics 17:401-463.
Hinrichs, Erhard. 1985. A Compositional Semantics for Aktionsarten and NP
Reference in English. Doctoral dissertation, Ohio State University.
Jackendof, Ray. 1972. Semantic Interpretation in Genertive Grmmar. Cam
bridge, Mas. : MIT Press.
Jackendof, Ray. 1983. Semantics and Cognition. Cambridge, Mas.: MIT Press.
Jackendof, Ray. 1986. Parts and Boundaries. Talk delivered at MIT Lexicon
Seminar.
Jackendof, Ray. 1987. The Status of Thematic Relations in Linguistic Theory.
Linguistic Inquiry 18(3): 369-41 1 .
Krifa, Manfred. 1987. Nominal Reference and Temporal Constitution: Towards
a Semantics of Quantity. Manuscript, Universitat Tiibingen.
Landau, Barbara, and Lila Gleitman. 1985. Language and Experience: Evidence
frm the Blind Child. Cambridge, Mass. : Harvard University Press.
Levin, Beth, and MaIka Rappaport. 1986. The Formation of Adjectival Passives.
Linguistic Inquiry 17:623-661.
Merlan, Fancesca. 1985. Split Intransitivity: Functional Oppositions in Intransi
tive Infection. In Grammar Inside and Outside the Clause: Some Approaches
to Theory from the Field, ed. Johana Nichols and Anthony Woodbury, 324-
362. New York: Cambridge University Press.
Nwachukwu, P. Akujuoobi. 1987. The Argument Structure ofIgbo Verbs. Lexicon
Project Working Papers 18. Cambridge, Mas.: MIT Center for Cognitive
Science.
Perlmutter, David, and Paul Postal. 1984. The I-Advancement Exclusiveness
Law. In Studies in Relational Grammar, ed. David Perlmutter and Carol
Rosen, 81-125. Chicago: University of Chicago Press.
Pesetsky, David. 1987. Binding Problems with Experiencer Verbs. Linguistic
Inquiry 18: 126-140.
Pesetsky, David. 1988. Psych Predicates, Universal Alignment, and Lexical De
composition. Manuscript, University of Massachusetts at Amherst.
Pustejovsky, James. 1988. The Geometry of Events. In Studies in Generative
Approached to Aspect, ed. Carol Tenny. Lexicon Project Working Papers 24.
Cambridge, Mass. : MIT Center for Cognitive Science.
Rappaport, MaIka, Mary Laughren, and Beth Levin. 1988. Levels of Lexical
Representation. Lexicon Project Working Papers 20. Cambridge, Mass. :
MIT Center for Cognitive Science.
Rappaport, MaIka, and Beth Levin. 1988. What to Do with Thematic Roles. In
Syntax and Semantics: Thematic Relations, ed. Wendy Wilkins. New York:
Academic Press.
Roberts, Ian. 1 985. The Reprsentation of Implicit and Dethematized Subjects.
Doctoral dissertation, University of Southern California.
Schwartz-Norman, Linda. 1976. The Grammar of 'Content' and 'Container' .
Joural of Linguistics 12: 279-287.
Simpson, Jane. 1983. Aspects of Warlpiri Morphology and Syntax. Doctoral
dissertation, Masachusetts Institute of Technology.
Tenny, Carol. 1987. Grammaticalizing Aspect and Afectedness. Doctoral disser
tation, Massachusetts Institute of Technology.
van Valin, Robert. 1987. The Unaccusative Hypothesis vs. Lexical Semantics:
Syntactic vs. Semantic Approaches to Verb Clasifcation. In Proceedings of
NELS 1 7. Amherst: University of Masachusetts, GLSA.
Vendler, Zeno. 1957. Verbs and Times. Philosophical Review 66: 143-160. (Also
in Vendler 1967, pp. 97-121. ) .
Vendler, Zeno. 1967. Linguistics i n Philosophy. Ithaca: Cornell University Press.
Verkuyl, Henk J. 1972. On the Compositional Nature of the Aspects. Dordrecht:
Reidel.
Verkuyl, Henk J. 1976. Interpretive Rules and the Description of the Aspects.
Foundations of Language 14: 471-503.
Verkuyl, Henk J. 1978. Thematic Relations and the Semantic Representation of
Verbs Expressing Change. Studies in Language 2: 199-233.
Verkuyl, Henk J. 1987a. Aspectual Asymmetry and Quantifcation. In Tempo
ralitit, ed. Veronika Ehrich and Heinz Vater. Heidelberg Workshop on Tense
and Aspect (Feb. 1986).
Verkuyl, Henk J. 1987b. Nondurative Closure of Events. In Studies in Discourse
Representation Theory and the Theory of Generalized Quantifers, ed. Jeroen
Groenendijk, Martin Stokhof, and Fank Veltman. Dordrecht: Foris.
Williams, Edwin. 1980. Predication. Linguistic Inquiry 1 1( 1) : 203-238.
Zaenen, Annie. 1987. Are There Unaccusative Verbs in Dutch? In Proceedings
of NELS 1 7. Amherst: Univeristy of Masachusetts, GLSA.
Z
Thematic Relations as Links between
Nominal Reference and Temporal
Constitution
MANFRED KRIFKA
This paper treats the correspondence between the reference type of NPs
(i.e. , mas nouns, count nouns, meaure constructions, plurals) and the
temporal constitution of verbal predicates (i. e. , activities, accomplish
ments) . A theory will be developed that handles the well known infuence
of the reference type of NPs in argument positions on the temporal consti
tution of the verbal expressions, asuming an event semantics with lattice
structures and thematic roles a primitive relations between events and ob
jects. Some consequences for the theory of thematic roles will be discussed,
and the efect of partitive cae marking on the verbal aspect, as in Finnish,
and of apectual marking on the defniteness of NPs, like in Slavic, will be
explained.
1 Introduction
It ha been observed for some time that semantic distinctions in the nominal
domain and in the verbal domain show certain resemblances to each other,
namely the distinction between mass and count terms on the one hand and
the distinction between "aspectual classes" or "aktionsarten" on the other.
Concerning the nominal domain, I think that one should not contrast
mas nouns like wine to count nouns like apple directly, because they dif-
The thery developed in this paper represents parts of my doctoral thesis at the Uni
versity of Munich (Krifa 1986). I learned a lot from my supervisors, Theo Vennemann
and Godehard Link. And there are many more people who helped to clarif my idea
and their preentation: Riner Bauerle, David Dowty, Hana Filip, Franz Guenthner,
Sebatian Lobner, Jan Tore L!nning, Uwe Monnich, Barbara Partee, Craige Roberts,
Arnim von Stechow, Henk Zeevat, Barbara Zimmermann, Ede Zimmermann, and surely
more. Thanks to them all.
30 J MANFRED KRIFKA
fer in their syntactic distribution and in their semantic type; the frst can
serve a an NP, wherea the second cannot. One should contrat instead
expressions like wine and an apple, or apples and fve apples, or wine and
a glass of wine. The frst element in each of those pairs ha the property
of referring cumulatively (cf. Quine 1960): whenever there are two entities
to which wine applies, this predicate applies to their collection as well.
The second member in each pair does not have this property: whenever
there are two (diferent) entities to which an apple applies, this predicate
does not apply to their collection. Let us subsume these properties under
the heading of nominal reference. Predicates like wine will be called cu
mulative, and predicates like fve apples will be called quantized. As for
"aspectual classes" or "aktionsarten, " I would like to use another name
for this concept, because these terms were originally coined for related,
but quite diferent phenomena in the morphology of the Slavic and Ger
manic languages. I will call the notion we are afer temporal constitution,
which was invented as the German term "Zeitkonstitution" by Fam;ois
( 1985) and covers a concept which was treated perhaps most prominently
by Vendler (
1
957) . I will concentrate here on what Vendler calls activities
and accomplishments, which I call atelic and telic expressions, following
Garey ( 1957) . To give a preliminary defnition: A verbal expression is
atelic if its denotation has no set terminal point (e.g. , run), and it is telic
if it includes a terminal point (e. g. , run a mi/e). This well-known semantic
distinction is supported by a battery of tests (cf. Dowty 1979) . For exam
ple, in ordinary, e. g. , non-iterative interpretations, atelic expressions allow
for durative adverbials like for an hour, but do not allow for time-span
adverbials like in an hour, whereas with telic expressions the situation is
reversed.
( 1) a. John ran (for an hour)/( *in an hour).
b. John ran a mile ( *for an hour)/(in an hour) .
That nominal reference and temporal constitution are related became clear
in two ways. First, the two concepts are felt to be semantically similar. For
example, a quantized NP like an apple denotes an object with precise limits,
just as run a mile denotes an event with precise limits. On the other hand,
a cumulative NP like wine denotes something without clear limitation, just
like what run denotes also has no clear limitation. Second, it wa observed
that the reference types of verbal arguments often determine the temporal
constitution of complex verbal expressions, insofar as a quantized argument
yields a telic verbal predicate, and a cumulative argument yields an atelic
verbal predicate:
(2) a. John drank wine (for an hour)/(*in an hour).
b. John drank a glass of wine ( *for an hour)/(in an hour) .
However, we cannot observe this efect with any verbal predicate, as the
following examples show:
THEMATIC RELATIONS / 31
(3) a. John saw a zebra (for an hour) j( *in an hour) .
b. John saw zebra (for an hour) j( *in an hour) .
This suggests that the lexical semantics of the verb plays a crucial role in
the way the nominal reference type of the arguments afects the temporal
constitution of the complex expression. More specifcally, it seems that the
thematic role of the argument is responsible for this efect; for example,
we fnd it with arguments which can be described as "consumed object"
a in (2) , but not with arguments which can be described as "observed
objects" as in (3). Therefore, a theory which explains this efect will have
consequences for the theory of thematic roles.
Some historical remarks: The similarity between nominal and verbal
distinctions wa observed already by Leisi ( 1953). The efect of verbal
arguments wa investigated frst by Verkuyl (1972) in his work on apec
tual composition, who dealt with features like [+SPECIFIED QUANTITY)
that are projected fom the argument to the verb phrase. Another ap
proach relying on feature projection is Platzack 1979. Dowty (1979) criti
cized these feature-based approaches, as they merely describe the facts and
do not really explain them. Dowty himself, a well as Hoepelman (1976)
and Hoepelman and Rohrer ( 1980) , developed theories in the paradigm of
model-theoretic semantics to capture the facts in a more explanatory way.
See Krifa 1986 for a detailed criticism of their approaches. It seems to
me that the general insight of the feature-baed approach, that the argu
ments and the complex expression have something in common, is lost in
these model-theoretic approaches. The theory presented here and in Krifa
1986, 1989 is more in the spirit of ter Meulen 1984 and Bach 1986, who
tried to characterize the similarities of noun denotations and verb denota
tions by model-theoretic means, but they remain at a rather informal level.
There is one explicit model-theoretic approach which looks similar to the
one developed here, namely Hinrichs 1985. But Hinrichs' theory crucially
depends on the notion of a stage of an individual, which complicates his
formalizations and ha some unintuitive side efects. The theory presented
here comes most closely to Dowty 1987, 1989, and Link 1987.
2 The Semantic Representation Language and
Its Interpretation
In this section, I will introduce the semantic representation language and
the basic facts about the model structure of its interpretations. I asume
a type-theoretic language with function symbols and identity. For reasons
of simplicity, it is assumed to be extensional.
To handle the semantics of cumulative and quantized reference, we must
provide for the semantic operation of joining two individuals to a new in
dividual. This means that our model structure must be of the form of a
lattice (cf. Link 1983) . Here, I can simplify Link's approach to apects
32 / MANFRED KRIFKA
relevant to my argument; for example, I will not distinguish between indi
vidual entities and quantities of matter. But I will extend Link's approach
to cover event predicates as well.
Assume that we have two non-overlapping sorts of entities, objects
(characterized by a predicate C), events (characterized by a predicate 8) ,
and times (characterized by a predicate T) . The extensions of C, and
T have the structure of a complete join semi-lattice without a bottom ele
ment. Let U be a two-place operation (join), and , L, ~ two-place relations
(part, proper part, overlap) . Then the following postulates must hold for
admissible interpretations of the semantic representation language:
(PI) 'x, y, z[ (x U y z)
[
O
(x)
1 O
(
y
) 1 O
(z)]
Y [(x) / (y
) 1 (
z
)] Y
[T(x) / T(y)
1
T(z)]]
(restriction to C,, T)
(P2) 'x, y3z[x U y z] (completeness)
(P3) 'x, y[x U y =
y ~
x] (commutativity)
(P4) 'x[x U x x]
(idempotency)
(P5) 'x, y, z[x U [y U z] = [x U y] U
z
] (associativity)
(P6) 'x, y[x Y f x U Y y]
(part)
(P7) -x'
y[x y]
(no J element)
(P8) 'x, y[x g Y f X Y 1 - x =
y] (proper part)
(P9) 'x, y[x '
Y f 3z[z x 1 z ylJ (overlap)
We can generalize the join operation to the fusion operation, which maps
a set P to its lowest upper bound:
(PIO) 'x, P[[(P
O Y
P Y
PT) FU(P) = x]
f 'y[P(y
)
y x] / 'z[
V
y[P(y
) y z]
-
x zlJ (fusion)
We now defne some higher-order predicates and relations to characterize
diferent reference types.
(Pl 1 ) 'P[ CUM(P)
'x, y[P(x) 1 P(y
)
-
P(x U y)lJ
(cumulative reference)
(PI2) 'P[SNG(P) f 3x[P(x) / 'y[P(y)
-
x = yJJJ (singular reference)
(PI3) 'P[SCUM(P) * CUM(P) 1 -SNG(P)]
(strictly cumulative reference)
(P14) 'P[ QUA(P) * 'x, y[P(x) 1 P(y
)
- - y L xlJ (quantized reference)
(PI5) 'P[SQUA
(P)
-
QUA(P) 1 'x[P(x)
- 3y[y L xJJJ
(strictly quantized reference)
(PI6) 'x, P[ ATOM(x, P) T P(x) 1 -3y[
y L x 1 P(y)l J
(x is a P-atom)
(PI7) 'P[ ATM(P)
'x[P(x)
3y[y x 1 ATOM(y, P)J J J

(P has atomic reference)
Postulate (PI7) says that if P is atomic, then every x which is P contains
a P-atom. The following theorems hold, as can be easily checked:
( TI) VP[SNG(P) QUA(P))
(T2) VP[SNG(P) CUM(P))
(T3) VP[ QUA(P)
..SCUM(P)
)
(T4) VP[ QUA(P) ATM(P)
)
We have to postulate some structure for events and times. First, we
assume that the time lattice is atomic, that is, ATM(1) , with T as the set
of atoms ( time points) . (I leave the question open as to whether objects and
events are atomic as well) . Second, we assume a temporal order relation ::,
which is a linear order for time points. With these notions, we can defne
convex times, or time intervals. (In the following, I will use t, t' , etc., as
variables for times, and e, e
'
, etc. , a variables for events.)
(PI8) ATM(T) / Vt[T(t) .. ATOM(t, 1)|
(T is the predicate of time points)
(PI
g
)
"
t, t' , t"[T(t) / T(t') / T(t") [t ::t / [t ::t'
/ t' ::t
i l
t ::t"
)
/ [t ::t
' Y
t' ::tj / [t ::t'
/ t' ::t t = t')))
(:: is a linear order for time points)
(P20)
"t, t' [t :: t' ..Vt", t
il
' [t
il
t / t
il
' t'
t
il
::t
i l
'])
(extension of :: to times in general)
(P2I) Vt[CONV(t) ..Vt' , t", t"' [t' t / t
il
t / t' ::t
il
' :: t
i l
t
il
' t))
(convex times, or intervals)
Third, we assume a function
T
from the extension of to the extension of 1,
the temporal trace function; this function maps an event to its "run time,"
or temporal trace. It is a homomorphism relative to the join operation:
(P22) "e, e' [T(e) U T
(e' )
T(e U e' )
)
That is, the join of the temporal traces of two events equals the temporal
trace of the join of these events.
3 Cumulativity and Quantization for Object and
Event Predicates
In this section, we will apply the notions we have developed so far to
the semantic description of certain predicate types. First, we will look at
predicates on objects, and then at predicates on events.
Characterizing object predicates like wine versus a glass of wine, or ap
ples versus fve apples, is straightforward. If we represent these expressions
by predicates in the semantic representation languages, we have:
(4) a. wine 0 / CUM( wine)
b. a.glass. of. wine 0 / QUA(a. glass. of. wine)
c. apples 0 / CUM( apples)
d. fve.apples 0 / QUA(fve. apples)
(4a) says that wine is a predicate on objects (note that we make no dis
tinction between stuf and objects for reasons of simplicity) , and that it is
34 / MANFRED KRIFKA
cumulative. (4b) says t hat a.glass.of wine is also a predicate on objects,
but that this predicate is quantized. Similarly, apples is a cumulative ob
ject predicate, and fve. apples is a quantized object predicate. I will not go
into the semantic composition of predicates which are syntactically or mor
phologically complex, like a.glas.ofwine, apples, or fve.apples; see Krifa
1986, 1989 for a treatment.
Now look at expressions like run and run a mile. In the standard
treatment (5i) , one-place verbal predicates are reconstructed as applying
to objects, just as object predicates. For example, run is analyzed as
applying to every object that runs. However, there are good reasons to
assume that these predicates have also an argument place for events (cf.
Davidson 1967), as in (5ii), or even that they are predicates on events, and
that the participants are related to these events by thematic relations like
Agent, Theme, etc. (cf. Parsons 1980, Carlson 1984, Bauerle 1988), a in
(5iii):
(5) Mary runs. i. run(Mary)
. run(Mary, e)
iii. run (e)/ AG(e, Mary)
Obviously, if we want to model the temporal constitution of verbal expres
sions, we should choose either (5ii) or (5iii) a a representation format,
because the temporal constitution can most eaily be formulated with the
help of the event argument 6. Furthermore, it will turn out that the rules
can be more easily formulated in the format (5iii ) , which factorizes a verbal
predicate into an event property and the thematic information. So I will
base what follows on this representation format.
How can we characterize an atelic event predicate like run and a telic
event predicate like run.a. mile within our theoretical famework? We may
say that the frst is cumulative and the second is quantized: If we have
two events of running, then they form together an event of running; and if
we have an event of running a mile, then no proper part of it is an event
of running a mile. So we can reconstruct atelic and telic expressions by
cumulative and quantized event predicates, respectively:
|) a. run ^ CUM (run)
b. run.a.mile / QUA(run. a.mile)
We might ask how this characterization of telic and atelic predicates relates
to the traditional one, that telic predicates have a set terminal point and
atelic predicates lack such a set terminal point. There is, in fact, a close
relationship:
The notion of a "set terminal point" cannot be defned for bare events
or "event tokens" , but only for events with respect to a certain description,
event predicates, or "event types." For consider a concrete event of running
and a concrete event of running a mile; then surely both events have a
terminal point (both events might be even identical) . The diference is
that an event of running might be a part of another event of running which
has a later terminal point, whereas this is not possible for an event of
running a mile.
We can defne the notion of a telic event predicate like follows. First,
let H defne a function TP which maps events to the last time point in
their run time. Then we can defne the notion of event predicates which
have a set terminal point, STP.
(P23) 'e, t [ TP(e) = t T(t)
A
t r(e) A 't' [t' r(e) t' t
ll
(the terminal point of an event)
(P24) 'P[STP(P) 'e[P(e) 'e' [P(e') A e' e TP(e) = TP(e')
]
] ]
(event predicates with set terminal point)
An STP event predicate, then, applies to events such that all subevents
which fall under the predicate have the same terminal point. In a natural
interpretation of run and run.a.mile, we can assume the following proper
ties:
(7) a. run A
..STP( run)
b. run. a. mile A STP( run. a. mile)
That is, run. a. mile is a predicate with a set terminal point, as every
subevent of an event of running a mile ha the same terminal point. This
is diferent for run. In general, we may characterize telic predicates P as
STP(P) , and atelic predicates P a ..STP(P).
If we defned a mapping fom objects to spatial regions and defne the
notion of a border of regions, and hence, of objects, then we could character
ize nominal predicates like a. glass. of wine and wine similarly, as implying
a set border or a not implying a set border. In this way, we could capture
the similarity between expressions like 'run' and 'run a mile' with 'wine'
and 'a glass of wine' , respectively.
It turns out, however, that a good deal of the similarity can already be
covered by the notions of cumulative and quantized predicates. The reason
is that there is a relation between predicates with a set terminal point
and an interesting class of cumulative event predicates. This clas can be
defned as follows: With the exception of singular event predicates (that
refer to one event only) , event predicates in natural language typically have
the property that they apply to events which have diferent terminal points.
For example, a predicate like run, or run a mile, refers to events that end
at diferent times. Let us defne the notion of natural event predicates,
NEP, as an event predicates with that property:
(P25) 'P[NEP(P) P A
3e, e' [P(e) A
P(e') A
TP(e) = TP(e')]]
Now we can prove that cumulative natural event predicates cannot have a
set terminal point:
36 / MANFRED KRIFKA
(T5) VP[ CUM(P) / NEP(P) -STP(P)]
Proof: Assume an event predicate P with CUM(P) and NEP(P). As P is
natural, there are two events el , e2 such that P(el ) , P(e2) , and - TP( el )
TP(e2 ) . Assume that TP(el ) TP(e2 ) . As P is cumulative, it holds that
P(el U e2) . As r(el U e2) r(el ) U r(e2), we have - TP( el ) TP(el U e2) .
But i t holds that e
l ! el U e2 . Conequently, we have -STP(P).
This mean that, under the assumption that P is cumulative, CUM(P),
and not singular, -SNG(P), we can normally assume that P has no set
terminal point, -STP(P). That is, strictly cumulative event predicates
can safely be taken as atelic under the traditional defnition (lacking a set
terminal point) .
On the other hand, whenever we have a quantized event predicate P,
QUA(P), this will have a set terminal point, STP(P) . This is because when
QUA(P) and P(e), then e has no proper part; so all parts e' of e will have
the same end point, a e' and e are in fact identical. Therefore all quantized
event predicates will be telic, under the traditional defnition. But note that
there are predicates with set terminal points that fail to be quantized. One
example is 'walk to the station': If this predicate applies to an event e,
then it will also apply to the latter half of e; so it is not quantized.
In the following, I will view telic predicates simply as quantized event
predicates, and atelic predicates as strictly cumulative event predicates.
4 A Famework for Object and Event Reference
Before we turn to a formal description of the infuence of nominal arguments
to verbal predicates, I will sketch the syntactic and semantic famework I
am assuming by way of an example (see Krifa 1986 for a more explicit
treatment) .
I asume a categorial-like syntactic representation; this is, however, not
essential. Verb argument places come with features such as category (like
NP) , cae (like 8ubj, obj) , and theta-roles (like ag, pat) . The expressions
that fll these arguments must have the same values for these features.
The value of the theta feature is interpreted semantically by corresponding
thematic relations. In the derivation tree in (8), I specify the expression,
its syntactic category, and its semantic interpretation. The general syn
tactic operation is concatenation, and the semantic operation is fnctional
application.
A verb is interpreted as a one-place predicate of events; the syntactic
argument slots have no counterpart in its semantic representation, but
only in its syntactic categorization. The theta-role information, which is
specifed with the argument slots in syntax, is passed to the subcategorized
NPs, where it is realized as a part of the semantic representation of the
determiners (e. g. , pat is realized as PAT(e,x) ) . With free adjuncts like in 0
pen, the thematic relation are specifed within the adjunct. Here I assume
that the theta role is specifed in the preposition; the NP governed by the
preposition only has a dummy theta feature "empty" that is not realized
in the interpretation of the determiner.
(8) drank ; S/NP[subj, ag] , NP[obj, pat] ; -e[ drink(e)]
water ; N ; water
; NPlobj, patl/N ; ,P' ,P>e3x
IP(e) A PAT(e, x) A P' (x)]
water ; NP[obj, pat] ; -P-e3x[P(e) / PAT(e, x) / water(x)]
drnk water ; S/NP[subj, ag] ;
-P-e3x[P(e) / PAT(e, x) / water(x)] (-e[drink(e)] ) =
-e3x[ drink(e) / PAT(e, x) / water(x)]
in ; [S/S]/NP[obj, empty] ; IN
pen ; N ; pen

; NPlobj, emptyl/N ; ,P' MP>e3xIP( e) A R( e, x)
A p'
(x)
]
a pen ; NP[obj, empty] ; -
R-P-e3x[P(e) / R
(e, x) / pen (x)]
in a pen ; SIS ; -P-e3x[P(e) / IN(e, x) / pen(x)]
drank water in a pen ; S/NP[subj, ag] ;
-P-e3x[P(e) 1 IN(e, x) / pen(x)] (-e3x[drink(e) / PAT(e, x)
/ water( x)] ) =
-e3x, y[ drink (e) / PAT(e, x) / water(x) / IN(e, y) / pen(y)]
pig ; N ; pig

; NP
I.
ubj, a
.
] /N ;
,'
,P3x
IP(e) A AG(e, x) A P' (x)]
a pig ; NP[subj, ag] ; .P3x[P(e) / AG(e, x) / pig(x)]
a pig drank water in a pen ; S ;
-P-e3x[P(e) 1 AG(e, x) / pig(x)] (-e3x, y[ drink(e) / PAT(e, x)
/
water(x) / IN(e, y) / pen(y)] )
=
-e3x, y, z[ drink (e) / AG(e, x) / pig(x) / PAT(e, y) / water(y)
/ IN(e, z) / pen(z)]
After all fee variables are bound, we obtain a predicate on events with
out fee variables, the sentence radical ( S) . This can be transformed to
a sentence (S') by the application of a sentence mood operator, e. g. , the
declarative operator, which simply binds the event variable with an exis
tential quantifer.
38 / MANFRED KRIFKA
(9)
. ; D
/S
, .P3eP(e)
A pig drnk water in a pe
n. ; Sf ;
3e3x, y, z[ drink(e) / AG(e
, x) / pig(x) / PAT(e, y) / water(y)
/
IN(e, X)
/ pen(z)]
This representation of declarative sentences thus conforms to the truth
scheme of Austin ( 1961) , who asumed that a declarative sentence consists
of two baic semantic constituents, namely a specifcation of an event type
and a reference to a specifc event, which is claimed to be of the specifed
type. Types of events I capture by event predicates, and the reference to
a specifc event by the existential quantifer. Surely, both reconstructions
will turn out to be too simple, but they sufce for the present purpose,
and the analysis to be developed hopefully can be recat in more complex
representations.
5 The Impact of Arguments
In this section, which repeats part of Krita 1989, I will show how the
impact of the nominal reference of arguments on the temporal constitution
of verbal predicates can be captured formally.
The baic idea is that, with certain thematic relations, the reference
properties of the syntactic arguments carry over to the reference properties
of the complex construction. There is a way to visualize this transfer of
reference types, namely space-time diagrams. In these diagrams, space is
represented by one axis, and time by the other. Objects, with their spatial
extension, can be represented as lines, and events can be mapped to the
time axis. Now consider 6 the event of drinking a quantity of wine w
(which is gradually disappearing during the drinking):
( 10)
I
By this picture the intuitive notion that the object is subjected to the
event in a gradual manner should become clear. Consider two possible
descriptions of w and, consequently, e. First, let w be described as wine,
and hence e a drink wine. As wine is cumulative, it is normally the cae
that it can also be applied to proper parts of w, like w' . But then it should
be possible to apply the predicate drink wine to the corresponding proper
part of e, namely e', a well. Secondly, let w be described a a glass of
wine, and e consequently as drink a glass of wine. As a glass of wine is
quantized, no proper part of U can be described as a glass of wine. But
then no proper part of e can be described as drink a glass of wine either.
Technically speaking, we have to assume a homomorphism fom objects
to events which preserves the lattice structure. This should follow fom the
properties of the thematic relation that mediates between event and object.
To characterize these properties, I assume the following notions:
(P26) 'R[SUM(R) 'e, e' , x, x' [R(e, x) / R(e' , x')
R(e U e' , x U x')]]
(summativity)
(P27) 'R[ UNI- O(R) 'e, x, x' [R(e, x) / R(e, x'
) x = x']]
(uniqueness of objects)
(P28) ' R[ UNI-E(R) 'e, e', x[R( e, x) / R( e' , x) e e'
ll
(uniqueness of events)
(P29) 'R[MAP- O(R) 'e, e
' , x[R(e, x) /
e' ! e 3x' [x' !
x /R(e'
, x' )
)))
(mapping to objects)
(P30) 'R[MAP-E(R) 'e, x, x' [R(e, x) /x' ! x 3e' [e' !
e / R(e', x'))))
(mapping to events)
Summativity (that is, cumulativity for two-place relations) provides the
basic connection between thematic relations and the join operation U . For
example, two (distinct) events of drinking a glas of wine yield an event of
drinking two glasses of wine. Uniqueness of objects captues the fact that
an event is related to a specifc object, for example, a drinking of a glas of
wine is related via the patient role to this glas of wine, and to nothing else.
Uniqueness of events says that there is only one event related to the object
by the thematic relation; for example, for a specic glas of wine there can
be only one drinking event. Mapping to objects can be exemplifed by our
example a follows: every part of a drinking of a glass of wine corresponds
to a part of the glas of wine. And mapping to events says in the example
at hand that every part of the glas of wine being drunk corresponds to a
part of the drinking event. These are just informal characterizations of the
properties of thematic relations; they will be discussed in more detail below.
The following postulate covers the notion of iterativity. It is a relation
between an event e, an object x and a thematic relation R saying that at
leat one part of x is subjected to at least two diferent parts of e. This
notion applies to, e.g. , the reading of a book if at least one part of the book
is read twice. It is more permissive than the usual notion of iterativity,
40 / MANFRED KRIFKA
which would claim in the example at hand that the whole book must have
been read at leat twice.
(P31) 'e, x, R[ITER(e, x, R) - R(e, x) 1 3e' , e", x' [e' e l e" e
/
e'
= e" / x' x 1 R( e', x' ) / R( e", x' )
l}
(iterativity)
Which properties must we assume for thematic relations to derive their
homomorphism properties? Let us translate an expression like read a letter
by predicates ,
( 1 1 ) = Ae3x[a(e) 1 8(x) 1 O
(
e, x)]
where O represents the verbal predicate ( read), 8 represents the nominal
predicate ( a letter) , and 0 represents a thematic relation (here, a specifc
patient relation) . In the following, I will examine the efects of some prop
erties of 8 and 0 on . The verbal predicate a will be considered to be
cumulative throughout.
We start with the question: What are the conditions for to be cu
mulative? One set of conditions is: is cumulative if 8 is cumulative and
o is summative (an example is read letters) . Proof: Assume el , e2 (not
necessarily distinct) with (el ) , (e2). According to the defnition of ,
there are two objects Xl , X2 with a( el ) ' 8(X
I
) ' 6(el , x
I
) and a( e2) ' 8(X2) '
O(e2, X2 ) . Because a and 8 are cumulative, it holds that a(el U e2 ) and
8(X
I
U X2) , and because 0 i summative, it holds that O(el U e2, X
I
U X2).
Hence (el U e2), that is, is cumulative:
(T6) 'P, Q, R[ CUM(P) / CUM(Q) 1 SUM(R) CUM( Ae3x[P(e) /
Q(x) P R(e, x)])]
As singular predicates (e. g. , the letter) are cumulative a well, albeit in a
somewhat pathological way, this result holds for them, too. Consider the
following example, with the. letter as a predicate with singular reference
applying to the letter.
( 12) read the letter
Ae3x[ read(e) / PAT(e, x) 1 the. letter(x)]
But if we want to understand read the letter as atelic, a in he read the
letter for an hour, then we clearly have to asume either a partitive reading
or an iterative reading. Partitive readings will be treated in Section 7. As
for the iterative reading, it can be shown that if is strictly cumulative,
o is summative, and 8 has singular reference, then we get an iterative
interpretation.
Proof: If is strictly cumulative, then we have two distinct el , e2 with
(el ) , (e2) . According to the defnition of , there are two objects Xr ,
X2 with 8(X
I
) ' 6( er , xr ) and 8(X2) ' O(e2, X2) . Because 0 i s summative, it
holds that O(el U e2, Xl U X2), and because 8 has singular reference, it holds
that Xl X2 With O(el U e2 , X
I
) , O(el , x
I
) , O(e2, xr ) and - el e2, the
conditions for iterativity (P31) are met, as Xl is subjected to two diferent
parts of the event el Ue2, namely e
l
and e2. So the following theorem holds:
THEMATIC RELATIONS I 41
(T7) VP, R, e, x[SNG(P) / SUM(R) 1
SCUM( -e3x[P(x) / R(e, x)] ) -
ITER(e, x, R)]
If we exclude the iterative interpretation and retain singular reference of 6
and summativity of (, then it follows that cannot be strictly cumulative:
(T8) VP, R, e, x[SNG(P) / SUM(R) / -ITER(e, x, R) -
-SCUM(-dx[
P(x) / R(e, x)] )]
Hence a verbal predicate like read the letter, under a non-iterative and non
partitive interpretation, cannot be strictly cumulative, and hence cannot
be atelic.
Sometimes the iterative interpretation is excluded in the frst place,
namely with efected or consumed objects, as in write the letter or drink the
wine. The reason is that an object can be subjected to an event of drinking
or writing a maximum of one time in its career. Therefore, uniqueness of
events should be postulated for the respective thematic relations. And this
excludes an iterative interpretation. Proof: Assume to the contrary that (
is unique for events, ((eo, xo) and ITER(eo, xo, (). Because of iterativity,
it follows that there are el , e2, Xl with el eO, e2 eo, - el = e2 and
Xl Xo for which it holds that ((el , x
I
) and ((e2 , xd. But this contradicts
uniqueness of events.
(T9) VR, e, x[R(e, x)
1
UNI-E(R) - -ITER(e, x, R)]
Let us now investigate the infuence of quantized nominal predicates like a
letter. Under which conditions can we asume that they cause the complex
verbal predicate to be quantized as well? One set of conditions is that the
thematic role ( must satisf uniqueness of objects and mapping to objects,
and that iterative interpretations are excluded. Proof: We assume to the
contrary that ( is quantized, (e l ), (e2) and e2 e el . Then there are
Xl , X2 with 6(XI)
'
((el , xI ) and 6(X2) ' ((e2, x2), according to the defnition
of . Because e2 L el and ( satisfes mapping to objects, there is an X3
such that X3 Xl and ((e2, X3) . Because uniqueness of objects, it holds
that X3
= X2, and therefore X2 Xl . As we have (( e2, x2) , e2 e el and
-ITER(e
I
, XI , () , we can infer that -X
I
X2 . With X2 Xl , this yields
X2 L Xl , but this contradicts the assumption that 6 is quantized. Hence
there are no el , e2 as asumed above, and that means that is quantized.
(TID) VP, R, e, x[ QUA(P) / UNI- O(R) / MAP- O(R) / -ITER(e, x, R)
QUA(-e3x[P(x) / R(e, x)] )]
As a special case of (TID), we have the following theorem for thematic re
lations which satisf uniqueness of events (e. g. , efected and consumed ob
jects), a this property excludes an iterative interpretation in the frst place:
( TIl) VP, R[ QUA(P) / UNI- O(R) / MAP- O(R) / UNI-E(O) -
QUA(-e3x[P(x) I R(e, x)] )]
Even in the iterative cae it holds that examples like ' read a letter' are
atomic. The conditions for thematic relations which are relevant for this
42 / MANFRED KRIFKA
result are that they satisf uniqueness of objects and mapping to events.
We have to asume 6 not only to be quantized, but to be strictly quantized,
which is not a substantial limitation. Proof: Assume an el with <(ed,
hence an Xl with 6(Xl ) and B(e l , Xl ) ' Because 6 is strictly quantized, Xl
contains a proper part X2 , that is, X2 L Xl , with ,6(X2) ' Because of
mapping to events, there is an e2 with e2 !; el and ((e2, x2). Because of
uniqueness of objects, X2 is the only object with this property. Hence there
is no X with 6 (x) and B( e2, x). But then '<( e2) holds, and this means that
e l contains a <-atom, e2. As we made no special assumption for el l it
follows that < is atomic.
(TI2) VP, R[SQUA(P) " MAP-E(R) " UNI- O(R)
ATM(Ae3x[P(x)
" R(e, x)] )]
In Krifa 1986, 1989, I have shown how we can explain with the help of
these results why durative adverbials like for an hour select for atelic verbal
predicates, wherea time-span adverbials like in an hour select for telic
predicates. The underlying reaon is that durative adverbials presuppose
that the verbal predicate they are applied to is strictly cumulative, and
time-span adverbials presuppose that the verbal predicate they are applied
to is atomic. Now, quantized verbal predicates telic predicates) are
atomic and not strictly cumulative; hence their distribution with respect
to those adverbials is explained. Strictly cumulative verbal predicates
atelic predicates) can be combined with durative adverbials, and they can
also be combined with time-span adverbials under the presupposition that
they are atomic. Normally, this presupposition is not warranted, and hence
the combination with time-span adverbials sounds odd.
To conclude this section, let us use the properties of thematic relations
to clasif the patient relations of diferent verbs. It is usefl to introduce a
new notion that says that the object is subjected to the event in a gradual
manner, as visualized by the space-time diagram ( 10) . I call this gaduality;
it comprises uniqueness of objects, mapping to objects, and mapping to
events.
(P32) VP[ GRAD(P) UNI- O(P) " MAP- O(P) " MAP-E(P)] (graduality)
The criteria for the clasifcation of thematic roles can be applied to tran
sitive verbs. This yields at leat three interesting classes; two classes can
be further subdivided for independent reasons.
( 13) example summa- gradu- uniqu. label
tivity ality events
write a letter gradual efected patient
eat an apple gradual consumed patient
read a letter gradual patient
touch a cat afected patient
see a horse stimulus
I think the conditions of summativity, graduality and uniqueness of events
are intuitively plausible for the respective patient relations. In the next sec
tion, I will discuss the transfer properties for thematic relations in greater
detail.
6 Some Consequences for the Theory of Thematic
Relations
In this section, we will discuss some consequences which follow fom the
asumption of properties of thematic relations, a discussed in the last
section.
The most general property is 8ummativity, which obtains for all patient
relations, and probably for all thematic relations whatsoever. This means
that thematic relations are not sensitive to the "size" of the entities they
relate to each other.
One thing which summativity can buy us is a simpler and intuitively
more appealing treatment of cumulative reading than the one ofered in
Scha 1981. For example, if there are two events, one to be described with
(14a) , the other with ( 14b) ,
(14) a. John saw three zebra.
b. Mary saw four zebra.
and if the zebra John and Mary saw do not overlap, then the sentence John
and Mary saw seven zebras can be derived if one asumes summativity for
the experiencer relation and the stimulus relation, and that the count noun
relation contains extensive meaure fnctions compatible with the object
lattice, a notion introduced in Krifa 1986, 1989. EXP and STI should
represent the experiencer and stimulus relation, and zebra(x, n) says that x
and H are zebra.
(15) see(ed ^ EXP(e l > John) / STI(el , xt} ^ zebr(xl
'
3)
see(e2) ^ EXP(e2, Mary) ^ STI(e2, X2) ^ zebr(x2
'
4)
. Xl ~ X2
see(el U e2) / EXP(el U e2, John U Mary) / STI(el U e2, Xl U X2)
^ zebr(XI U X2, 7)
Note that the derived sentence ha rather weak truth conditions, a it re
mains unspecifed how the zebra relate to John and Mary individually.
This is a it should be, a the diferent possibilities are not diferent "read
ings" of the sentence. In contrat to other theories of cumulative predica
tion, for example the one by Gillon (1987) , this i a natural outcome of a
very simple rule and need not be stated by a complicated rule involving
quantifcation over partitions of sets and the like.
An objection against this treatment might be that a sentence like John
and Mary saw seven zebras are understood saying that they saw exactly
seven zebra, a reading which is not captured by 3e, x[ see( e )IEXP( e, JohnU
Mary) ^ STI(e, x) ^ zebra(x, 7)], a this representation allows for John and
44 / MANFRED KRIFKA
Mary to have seen more than seven zebras. But this problem can be han
dled if we assume a pragmatic rule that enforces maximally informative
readings, as the sentence x saw H zebras is more informative than x saw H
t
zebrs, if H ` H
t
(cf. Krifa 1986, 1989) .
To cover collective readings, a e. g. , John and Mary (jointly) own three
houses, we need of course a diferent representation, which will not be
developed here. And it should be clear that distributive readings can be
treated in this famework a well (cf. Link 1983 for distributivity in lattice
model structures) .
Uniqueness of objects has been discussed by several authors. For exam
ple, it corresponds to "thematic uniqueness" in Carlson 1984 and "unique
ness of role-bearers" in Dowty 1987, and it is a requirement for the treat
ment of thematic relations as functions, as e. g. , in Link 1987. Furthermore,
Carlson suggests that thematic roles may serve to discriminate events fom
one another on the basis of this property and the discrimination of ob
jects involved in the events: If 0 is unique for objects, then we can infer
fom O(e l , xd
A O(e2, x2) A "Xl X2 that . el e2 But note that I
do not assume uniqueness of objects for every thematic relation, as these
authors seem to do. Obviously, it does not hold for the stimulus rela
tion, a e. g. , I can see a zebra and, with the same event of seeing, see
the mane of the zebra as well. And it does not obtain with afected ob
jects, as e. g. , I can touch a shoulder and a person with the same event of
touching.
Next, consider mapping to events and mapping to objects, the two re
lations which constitute the core of the construction of the homomorphism
fom objects to events. They seem to be sound asumptions for gradual
patient relations. Take as an example the reading of a book; every part of
the book corresponds to a part of the reading and vice versa. With other
thematic relations, these properties normally do not obtain; for example,
there is no correspondence between parts of the person that is reading and
the reading event. But note that as we can have sum individuals, it is pos
sible that mapping to events and mapping to objects (a well as uniqueness
of objects) apply to other thematic relations in certain circumstances as
well. As an example, consider see seven zebrs. Even if a single experiencer
is involved, this predicate can be applied to events with diferent temporal
structues, for example to events where seven zebras are seen simultane
ously, or to the sum of seven consecutive events, in each of which a single
zebra was seen. Now, in the second case, it does make sense to speak of
mapping to events and mapping to objects, as for every part of the complex
seeing event (down to the observings of single zebras) there is a part of the
sum individual of the zebras which is seen in this event. Note that in cases
like this one, predicates as 'see seven zebras' can be understood as telic;
for example, ( 16) can be understood to say that the seven zebras were not
seen simultaneously.
( 16) John saw seven zebras in an hour.
Time-span adverbials like in an hour select for atomic verbal predicates
(cf. Krifa 1986, 1989). The simplest way to get an atomic reading of see
seven zebras is that in the relevant event, the zebras were seen in some
temporal succession.
The fact that the object roles of verbs like see sometimes have the
same mapping properties as the object roles of verbs like eat indicates
that the properties we have discussed so far are not "hard-wired" in the
thematic relations, but follow from other knowledge sources. Consequently,
we should asume that even the object role of verbs like eat does not exhibit
graduality as some grammatical feature, but simply because the normal
way of eating enforces the graduality properties.
However, there are some problems with the mapping properties. With
mapping to events, it is ofen the cae that only a certain class of parts of
the object are relevant. As an example, consider eat the apple and peel the
apple; in the frst cae, all the parts of the apple are involved, whereas in the
second cae, only the surface parts are. Another example is read the book
and bur the book; surely, there are parts of the book which are relevant in
the second cae (e.g. , the cover of the book) which do not count as parts
of the book in the frst case. To handle these phenomena, we may asume
that the verb selects specifc aspects of an object (e.g. , only its surface).
Perhaps more problematic is mapping to objects. As an example, con
sider build the house. There are surely parts of the event of building a
house which cannot be mapped to parts of the house. An example is the
erection of the scafold, which is clearly part of building the house, but the
scafold is not a part of the house, and even vanishes when the house is
fnished. Therefore, mapping to objects does not hold in a strict sense for
complex events.
This problem can be solved if we assume that predicates like build the
house refer to events consisting of events which themselves fall under difer
ent quantized predicates. A list of such predicates may be called a scenario,
after Link 1987. For example, the building of a house consists in raising
a loan, buying a place, and so on. This can be captured by a predicate
i ,e3el ' " en[il ( el ) / / in(en) / e e l | | en] , where all the
ii are quantized and disjoint from each other. It can be shown that i is
then quantized a well. Proof: Assume to the contrary that i(el)
'
i(e2)
and el g e2; then there is at least one ii , 1 : i : H, and e
3
, e4 such that
ii ( e
3
) , ii (e4) and e3 g e4, which contradicts the assumption that ii is
quantized.
However, an objection to this solution may be raised, as many events
lack a standard scenario (cf. Link 1987) . For example, with the building of
a house, there need not be an erection of a scafold. Therefore, we have to
assume that a complex event of a certain type ha to be related to some
46 / MANFRED KRIFKA
scenario of quantized subpredicates which need not be exactly specifed,
but which at least must qualify as being quantized, and this is all we need.
Uniqueness of events, fnally, characterizes those patient relations which
describe the coming into being and disappearing of objects, because there
can be only one such event for every object. This is another property that
should not be considered as a grammatical feature, but as an external fact
about the world.
Note that with many verba efciendi, we fnd a certain ambiguity: They
can be either token-oriented and type-oriented, so to speak. For example,
it is possible to write the same letter more than once, if one refers to the
letter type, not to the letter token. Such type-oriented verbs were called
"performance verbs" by Verkuyl ( 1972) . The approach outlined here can
be extended to type reference. Types may be conidered a abstract enti
ties with a part relation that corresponds to the part relation for concrete
entities we considered so far. For example, if y, y' are types and y' y,
and i f a concrete object or event T realizes the type y, then there should
be an object or event T
t
, T
t
T, that realizes type y'. The specifc pa
tient relation of performance verbs then describes the realization of a type.
Verbs like play and compose (as in play/compose a sonata) have patient
relations relating an event to types, whereas write can be analyzed as either
token-oriented or type-oriented. We cannot assume uniqueness of events for
the type-oriented patient relation of write and play, and therefore we can
understand a predicate as write a letter as atelic in its performance read
ing and under an iterative interpretation, as e.g. , in ( 17a). On the other
hand, with the patient relation of compose, we should asume uniqueness
of events, which explains why ( 17b) is bad.
( 1 7) a. The secretary wrote this letter for three years.
b. *Scarlatti composed this sonata for three years.
In this paper, I cannot go into the semantics of types, or kinds, and
their relation to tokens; see Krita 1986 for a more elaborate treatment.
7 Progressive and Partitive
The famework developed so far can be extended in many diferent direc
tions and applied to interesting problems. Here, I will treat two topics,
namely the marking of progressives in Finnish and German and an inter
action between aspect and defniteness in Slavic languages.
I start with progressives. There are two ways to mark progressives in
natural language. Most often, it is marked by verbal morphology, or by a
periphrastic verbal construction, as in Czech or English. Sometimes, it is
marked by some special prepositional or partitive case marking of an NP,
as in German or Finnish (cf. Moravcsik 1978 for the meaning of diferent
object markings in general and Heinamaki 1984 for the partitive objects in
Finnish) :
( 18) a. John snedl rybu.
b. John jedl rybu kdyz Mary vztoupila.
( 19) a. John ate a fsh.
b. John was eating a fsh (when Mary came in) .
(20) a. John aB einen Fisch.
b. John aB an einem Fisch (als Mary hereinkam) .
( 21) a. John soi kalan.
b. John soi kalaa (kun Mary tuli sisaan).
Progressivity normally is considered to be a verb-oriented category. How is
it possible, then, that it is marked on an argument of the verb? The theory
developed here provides an answer, because it predicts that a change of the
reference type of the nominal predicate will afect the temporal constitution
of the complete construction. This I want to show more precisely.
Although progressivity seems to elude a satisfing model-theoretic se
mantic description, it is clear since Bennett and Partee 1972 that some
notion of partiality is involved in it. As a frst approximation, which suf
fces for our purposes, we can consider a predicate like be drinking a glass
of wine a applying to events which are parts of events to which drink a
glass of wine applies. That is, progressivity is associated with the following
operator:
(22) PROG = APAe' 3e[P(e) 1 e' e]
On the other hand, one can assume that partitivity can be associated with
a similar operator (cf. Bach 1986) . For example, the partitive of 'fsh' can
be analyzed a referring to parts of a fsh.
(23) PART APAX' 3x[P(x) 1 x' x]
Although the partitive may be analyzed like this in languages as Finnish
which have a clear partitive cae marking, the German case probably ha
to be handled diferently because partitive objects like an einem Fisch
have a rather limited distribution. They should instead be analyzed as
prepositional objects governed by the verb. We have to assume a lexical
restructuring rule which takes verbs like (24a) with an accusative object
and a patient theta role and transforms them into verbs like (24b) with
a prepositional object and a "partitive" patient relation. The partitive
patient relation is related to the normal patient relation as in (24c) :
(24) a. essen S/NP[nom, ag] , NP[acc, pat]
b. essen S/NP[nom, ag] , NP[ an-obj, pan-pat]
c. 'e, x[ PART-PAT(e, x) 3x' [PAT(e, x' ) 1 x' x
l l
Now, consider the following two expressions, (25a) representing a verbal
progressive (English style, e.g. , be eating a fsh), and (25b) representing a
nominal progressive (Finnish or German style, e. g. , an einem Fisch essen) ,
with ( as verbal predicate ( eat) , 8 as nominal predicate ( a fsh) , and ( as
the specifc thematic relation.
48 / MANFRED KRIFKA
(25) a. <v -e
'
3e, x[a(e) A t
(x) A 8(e, x) A e
' I;
e]
b. <n -e3x, x' [a(e) A t
(x) A
8(e, x' ) A x' I; x]
We asume that 8 is gradual and unique for events. At least in German,
the progressive marking by prepositional phrase is possible only with verbs
like drink or write, marginally possible with read, but impossible with see
or pat:
(26) a. Hans schrieb/?la an einem Brief.
b.*Hans sah/streichelte an einer Katze.
Furthermore, the verbal predicate a should be divisive, that is, if it applies
to an event, it applies to every part of it as well. Even if this is not exactly
true, we can asume it in the general cae.
A fnal point is worth mentioning. In (25) , I used the general part
relation instead of the proper part relation. I think that this captures the
semantics of progressivity, but pragmatically one can infer from the use of
the progressive form, which is more complex than the corresponding simple
form, that the proper part relation holds. Therefore, we have to show that
<v is similar to <n using the proper part relation.
Proof: First I show that for all e, <v(e) <n (e). Let <v(e2) , then there
is an el with a(el ) and e2 L eb and an Xl with
t
(Xl ) and 8(el
'
Xl ) . Because
O is divisive, it holds that a(e2 ) . With mapping to objects, uniqueness for
objects and uniqueness for events, there is an X2 with X2 g Xl and 8(e2 , x2 ) .
But then <n(e2) holds, too. Secondly I show that for all e, <n (e) <v (e).
To do this, we have to make an additional asumption, namely that with
all nominal progressives, the whole object is eventually subjected to the
event (this means ignoring the problems of the imperfective paradox). Let
<n(e2 ) , then a(e2) holds, and there are X2 , Xl with 6(Xl ) , 8(e2 , x2 ) and
X2 L Xl . Now the additional asumption is that there is an el with a(el )
and 8(el , xl ) . Because of mapping t o events, there is an 3 with 8(e3, x2 )
and e3 I; el . Because of uniqueness of objects, X2 is the only X for which
8(e3, x) holds, and because of uniqueness of objects, e3 is the only e for
which 8(e, X2) holds, hence e3 e2 and ..e2 = el (because ..Xl X2) , and
therefore e2 e el . But then it holds that <v(e2 ) .
By this method, it can be explained how it i s that a marking on the
noun can serve to mark an essentially verbal category. Note that in Finnish
the partitive is used in many more caes; it serves to express the progres
sive even with nouns like read and buy, and it may be employed to mark
irresultative verbs, a e.g., to shoot and wound versus to shoot dead. This
can be explained by an analogical extension of this type of marking to
conceptually similar caes. The tertium comparationis of this extension is
that the expression with a partitive object denotes an event which is not
a complete a an event denoted by the corresponding expression with an
accusative object.
k
8 Perfective and Defniteness
Let us now look at defniteness in Slavic. We have seen how a nominal
predicate operator can have an efect that is similar to a verbal predicate
operator. As the transfer of reference properties works in both directions,
we should not be surprised to fnd the converse case as well, that is, a
verbal predicate operator afecting the meaning of a nominal predicate.
We observe this most clearly in Slavic languages. The observation and
data in this section are based on Wierzbicka 1968 for Polish and Filip 1985
for Czech.
As it is well known, Slavic languages mark perfective apect (or ak
tionsart; the diference does not matter here) , whereas they do not mark
defniteness of the NP. For example, the Czech NP vino can mean either
'wine' or 'the wine', hruska can mean either ' a pear' or 'the pear', and
hrusky can mean either ' pears' or 'the pears' .
(27) a. vno i. Ax[wine(x)]
ii. AX[X FU(wine) / wine(x)]
b. hruska i. Ax[pear(x, I)]
ii. AX[X FU(Axpear(x, 1)) / pear(x, 1)]
c. hrusky - Ax[pears(x)]
. AX[X FU(pears) / pears(x)]
I represent defnite NPs on the bais of a predicate P a predicates ap
plying to the fusion of all P-elements, given that the predicate P applies
to the fusion. For example, the wine will apply to the fusion of all wine
quantities (which is a wine quantity as well, a wine is cumulative). Sim
ilarly, the pears will apply to the fusion of all pears. And the pear will
apply to one pear if there is only one; otherwise, the fusion of the ob
jects which fall under the predicate (aJ pear would not fall under that
predicate.
According to this interpretation, hruska i s a quantized predicate in both
readings, wherea the two readings of vino and hrusky difer in their refer
ence type: in the defnite reading, they are quantized (as they have singular
reference), wherea in the indefnite reading, they are cumulative.
Now consider the following examples:
(28) a. Ota pil vino.
'Ota dran wine/?the wine' (imperfective)
b. Ota vypil vino.
' Ota drank the wine/*wine' (perfective)
(29) a. Jedl hruSku.
' He ate a pear/?the pear' (imperfective)
b. Snedl hrusku.
'He ate a pear/the pear' (perfective)
50 / MANFRED KRIFKA
(30) a. J edl hrusky.
'He ate pears/?the pears' (imperfective)
b. Snedl hrusky.
'He ate the pears/*pears' (perfective)
These data can be interpreted as follows: (28) and (30) show that aspect
marking can distinguish between the indefnite and the defnite reading of
mass nouns and bare plurals, as the perfective aspect is compatible only
with the defnite interpretation of the object. (29) shows that a verb in the
perfective aspect may have an indefnite object if this object is quantized.
Hence it is not defniteness, but quantization which is required by the
perfective aspect marking.
The data can be explained as follows: Let H assume that the perfective
operator has scope over the complex verbal predicate. One of its meaning
components is that the predicate it applies to is quantized. That is, at
leat part of the meaning of the perfective can be captured by the mod
ifer >P>e[P(e) / QUA(P)] . This follows fom the usual assumptions for
perfectivity which says that it conveys the meaning that an event is "com
pleted. " This makes sense only for events which are quantized (or have a
set terminal point), as events under a cumulative description have no set
terminal point and hence cannot be said to be completed.
If we assume the normal transfer properties for the object role of verbs
like eat and drink, then we see that only with a quantized object the
complex verbal predicate will be quantized as well. If the perfective as
pect forces a quantized interpretation of the complex verbal predicate, the
complex verbal predicate will again force a quantized interpretation of the
object NP. This means in the case of (28b) and (30b) the defnite interpre
tation of the object, as this is the ony quantized interpretation (note that
all singular predicates are quantized). In the case of (29b), we can assume
both the defnite and the indefnite interpretation, as the latter one will
also yield a quantized object.
In a similar way, the imperfective aspect may force a non-quantized
interpretation of the verbal predicate, which consequently enforces a non
quantized interpretation of the object NP. However, this requirement seems
to be much weaker.
Note that the treatment of the Slavic defniteness marking proposed
here is essentially compositional, although the phenomenon itself seems
not to be compositional at frst sight, as the interpretation of vino depends
on other constituents. It is simply that the unwelcome reading is excluded
by general principles, just as in rob the bank the unwelcome readings of
bank are excluded by the lexical meaning of rob.
It should be stressed here that perfectivity is not just an expression
of quantization. If it were just this, we could not explain why languages
typically use a variety of perfectivity markers, even for the same verb, with
slight meaning diferences. But quantization is at least a component of the
meaning of perfectivity. There seem to be interesting variations between
languages in what has been called "perfectivity." Singh ( 1991) , for example,
argues that perfectives in Hindi are not related to quantization, but to
atOInicity of the basic predicate.
9 Some Final Remarks
To summarize, I hope to have made it clear that a semantic representation
is feasible in which the intuitive similarities between the reference type of
noun phrases and the temporal constitution of verbal expressions is cap
tured in a simple way. I have shown how the reference type of a noun
phrase can afect the temporal constitution of a verbal expression and vice
versa. I have discussed the properties of thematic relations that allow this
transfer of reference properties. Finally, I have applied these insights to
explain the marking of progressives by the case of NPs, and the efects of
aspect to the defniteness of NPs.
In Krifa 1989, I have shown how the theory can be extended to cover
quantifcation and negation. Furthermore, I have explained why durative
adverbials like for an hour and time-span adverbials like in an hour se
lect for cumulative and quantized event predicates, respectively. In Krifa
1986, I also treated the infuence of locative and directional adverbials on
the temporal constitution of a complex verbal expression with examples
such as walk to the school in/
*
for an hour versus walk towards the school
*
in/for an hour. In Krifa ( 1990), I have shown how measure expressions
in nominal constituents can express a measure on events. For example,
the sentence '4000 ships passed through the lock last year' has a reading
in which it does not imply that there are 4000 ships that passed through
the lock, but that there were 4000 events of passing of a single ship. An
other area of application is the semantics of the frequentative aktionsart;
in many languages, it remains unspecifed whether a sentence with a fe
quentative predicate claims that there is more than one event or more than
one participant in an event, an ambiguity which can be rendered easily in
our semantic representation.
References
Austin, John. 1961. Tuth. In Philosophical Papers, 117-133. Oxford: Oxford
University Press.
Bach, Emmon. 1986. The Algebra of Events. Linguistics and Philosophy 9: 5-16.
Bauerle, Rainer. 1979. Temporale Deixis, Temporale Frage. Tiibingen: Gunter
Narr.
Bennett, Michael, and Barbara Partee. 1972. Toward the Logic of Tense and
Aspect in English. Bloomington: Indiana University Linguistics Club.
Carlson, Gregory N. 1984. Thematic Roles and their Role in Semantic Interpre
tation. Linguistics 22:259-279.
52 I MANFRED KRIFKA
Davidson, David. 1967. The Logical Form of Action Sentences. In The Logic of
Decision and Action, ed. Nichola Rescher, 81-95. Pittsburgh: Pittsburgh
University Press.
Dowty, David. 1987. Events, Aspects, and NP Semantics. Paper presented at the
Conference on Logic and Linguistics, July 1987, Stanford University.
Dowty, David. 1989. On the Semantic Content of the Notion "Thematic Role" .
In Properties, Tpes and Meaning, Vol. 2: Semantic Issues, ed. Gennara
Chierchia, Barbara Partee, and Raymond Turner. Dordrecht: Kluwer.
Fauconnier, Gilles. 1978. Implication Reversal in a Natural Language. In Formal
Semantics and Pragmatics for Natural Language, ed. Fanz Guenthner and
S. J. Schmidt, 289-301. Dordrecht: Reidel.
Filip, Hana. 1985. Der Verbalaspekt und die Aktionsarten dargelegt am Beispiel
des Tschechischen. Master's thesis, Universitat Miinchen.
Fillmore, Charles. 1968. The Case for Cae. In Universals in Linguistic Theory,
ed. Emmon Bach and Robert T. Harms, 1-90. New York: Holt, Rinehart
and Winston.
Fancois, Jean. 1985. Aktionsart, Aspekt und Zeitkonstitution. In Handbuch
der Lexikologie, ed. Christopher Schwarze and Dieter Wunderlich, 229-249.
Kronberg: Athenaeum.
Garey, Howard B. 1957. Verbal Aspects in Fench. Language 33:91-1 10.
Gillon, Brendan S. 1987. The Readings of Plural Noun Phrases in English.
Linguistics and Philosophy 10: 199-220.
Gruber, Jefrey S. 1976. Lexical Structures in Syntax and Semantics. Amsterdam:
North Holland. Includes 1965 MIT PhD thesis, Studies in Lexical Relations;
reissued by Indiana University Linguistics Club (Bloomington, 1970).
Heinamaki, Orvokki. 1984. Aspect in Finnish. In Aspect Bound: A Voyage into
the Realm of Germanic, Slavonic and Finno- Ugrian Aspectology, ed. Casper
de Groot and Hannu Tommola, 153-178. Dordrecht: Foris.
Hinrichs, Erhard. 1985. A Compositional Semantics for Aktionsarten and NP
Reference in English. Doctoral dissertation, Ohio State University.
Hoepelman, Jakob. 1976. Mas Nouns and Aspects, or: Why We Can't Eat
Gingercake in an Hour. In Amsterdam Papers in Formal Grammar 1, ed.
Jeroen Groenendijk and Martin Stokhof, 132-154. Universiteit van Amster
dam. Centrale Interfaculteit.
Hoepelan, Jakob, and Christian Rohrer. 1980. On the Mas-Count Distinction
and the French Imparfait and Passe Simple. In Time, Tense, and Quantifers,
ed. Christian Rohrer. Tiibingen: Niemeyer.
Krifa, Manfred. 1986. Nominalreferenz und Zeitkonstitution. Zur Semantik von
Massentermen, Plurlteren und Aspektklassen. Doctoral dissertation, Uni
versitat Miinchen. Published 1989, Miinchen: Fink.
Krifa, Manfred. 1989. Nominal Reference, Temporal Constitution and Quan
tifcation in Event Semantics. In Semantics and Contextual Exressions, ed.
Renate Bartsch, Johan van Benthem, and Peter van Emde Boas, 75-115.
Foris: Dordrecht.
Krifa, Manfred. 1990. Four Thousand Ships Passed Through the Lock: Object
Induced Measure Functions on Events. Linguistics and Philosophy 13:487-
520.
Leisi, Ernst. 1953. Der Wortinhalt. Seine Struktur im Deutschen und Englischen.
Heidelberg: Quelle und Mayer.
Link, Godehard. 1983. The Logical Analysis of Plurals and Mass Terms: A
Lattice-Theoretical Approach. In Meaning, Use, and Interpretation of Lan
guage, ed. Rainer Bauerle, Christopher Schwarze, and Arnim von Stechow,
302-323. Berlin: Mouton.
Link, Godehard. 1987. Algebraic Semantics for Event Structures. In Proceedings
of the 6th Amsterdam Colloquium, ed. Jeroen Groenendijk, Martin Stokhof,
and Fank Veltman, 243-262. University of Amsterdam. Institute for Lan
guage, Logic, and Information.
L0nning, Jan Tore. 1987. Mas Terms and Quantifcation. Linguistics and
Philosophy 10: 1-52.
Moravcsik, Edith. 1978. On the Cae Marking of Objects. In Universals of Human
Languages, Vol. 4, ed. Joseph H. Greenberg, 249-289. Stanford University
Press.
Parsons, Terence. 1980. Modifiers and Quantifiers in Natural Language. Cana
dian Joural of Philosophy (Suppl) 4:29-60.
Platzack, Christer. 1979. The Semantic Interpretation of Aspect and Aktion
sarten: A Study of Interal Time Reference in Swedish. Dordrecht: Foris.
Quine, Willard V. 1960. Word and Object. Cambridge, Mass.: MIT Press.
Scha, Rero. 1981. Distributive, Collective, and Cumulative Quantification. In
Formal Methods in the Study of Language, Vol. 2, ed. Jeroen Groenendijk,
Theo Janssen, and Martin Stokhof, 483-512. Amsterdam: Mathematisch
Centrum.
Singh, Mona. 1991. The Perfective Paradox, or: How to Eat a Cake and Have
It, Too. In Proceedings of the 1 7th Annual Meeting of the Berkeley Linguistic
Society. University of California at Berkeley.
ter Meulen, Alice. 1984. Events, Quantities, and Individuals. In Varieties of For
mal Semantics, ed. Federick Landman and Frank Veltman, 259-280. Dor
drecht: Foris.
Vendler, Zeno. 1957. Verbs and Times. Philosophical Review 66: 143-160. (Also
in Vndler 1967, pp. 97-121. ).
Vendler, Zeno. 1967. Linguistics in Philosophy. Ithaca: Cornell University Press.
Verkuyl, Henk J. 1972. On the Compositional Nature of the Aspects. Dordrecht:
Reidel.
Wierzbicka, Anna. 1967. On the Semantics of the Verbal Aspect in Polish. In To
Honor Roman Jakobson: Essays on the Occasion of his Seventieth Birthday,
Vol. 2, 2231-2249. Den Haag: Mouton.
Complex Predicates and

Morpholexical Relatedness: Locative
Alternation in Hungarian
FARRELL ACKERMAN
1 Introduction
This paper examines Hungarian locative alternation predicates as well as
restrictions on pre nominal deverbal adjective formation in this language.
Sentences exhibiting Locative Alteration are familiar in the linguistic
literature:
(1) a. The peasant loaded (the) hay onto the wagon.
OBJ
th
eme
OBL
goal
b. The peaant loaded the wagon with (the) hay.
OBJ
goal
OBL
th
eme
Accounts of these constructions commonly address two issues. First, there
is an observable alternation of grammatical functions (GFs) for the os
tensible theme and goal arguments of participating predicates: the theme
alternates between
OBJ
and
OBL
while the goal alternates between
OBL
and
OBJ
. Second, there are semantic (or, aspectual) diferences associated
with each variant of this construction: there is argued to be a holistic ef
fect associated with the variant in which the goal appears as
OBJ
or as
SUBJ
depending on the transitivity of the predicate. 1 Both the recurrence
The present analysis of locative alternation follows the general outlines of a joint presen
tation of this material by Ackerman and Koml6sy at the Stanford LSA Lexical Semantics
Workshop 1987. The present analysis supersedes the analysis found in Ackerman 1990.
The discussion of deverbal adjectives is based on Laczk6 and Ackerman 1984-Cf. Lazk6
and SzaboJcsi 1988 for further discussion.
l
Cf. Fillmore 1968, Anderson 1971, Salkof 1983, Pinker 1989, Levin and Rappaport
1986, Rappaport and Levin 1988, among others.
L
e
xical Matters. Ivan A. Sag and Anna Szabo1csi, eds.
56 / FARRELL ACKERMAN
of this phenomenon and its arresting cross-linguistic similarities make it a
perennial issue withn the lexical semantic literature.
Rappaport and Levin ( 1988) argue quite cogently that an adequate
account of these phenomena in English should be formulated in terms of
Predicate Argument Structures (PAS) . This level of representation is inter
preted as a projection of the syntactically relevant aspects of the composi
tional semantic representations typifed by the lexical conceptual structures
of Jackendof (1983) and the MIT lexicon project. They argue explicitly
against the viability of a feature baed analysis.
Contrary to their claims, I propose that a feature based analysis does
extend explanatorily to the phenomena at hand. 2 In particular, I demon
strate that a particular interpretation of the feature baed mapping theory
of Lexical Functional Grammar yields an instructive insight into the orga
nization of the lexicon. Moreover, this model of the lexicon provides lexical
entries for both simple and complex predicates with precisely the right
sorts of information to account for relatedness between predicates, difer
ences between related predicates, correct grammatical function assignment
to arguments and constraints on prenominal deverbal adjective formation.
This model does not appeal to underlying levels of syntactic representa
tion. The baic proposal constitutes a synthesis of the views of mapping
fom semantic proto-role features to grammatical function features found
in Zaenen 1990 and the monostratal proposal to capture unaccusativity
efects found in Bresnan and Zaenen 1990.
I suggest that the architecture of mapping theory yields a natural parti
tioning of lexical operations within the lexicon. The frst type of operation,
referred to a Morpholexical, afects the lexical semantics of predicates by
altering the semantic properties (SP) associated with predicates: locative
alternation and causativization are two phenomena efected by morpholex
ical operations. This type of operation has consequences for the assign
ment of the intrinsic features or ICs within lexical mapping theory.3 It will
be shown that, in determining intrinsic (IC) feature clasifcation, lexical
semantic properties indirectly constrain the Grammatical Function ( GF)
realization of arguments. Locative alternation predicates will be related by
morpholexical rule: in efect, a morpholexical rule is a function which takes
one verbal form and associated information a an argument and yields an
other verbal form and asociated information as a value. The second type
of operation, referred to a Morphosyntactic, assigns features supplemental
2
1 leave to another forum the question a to how the present proposal explains the
English data: a will be seen, the distributions of data in English and Hungarian are
sufciently similar to entertain an extension of the present analysis to English.
3
Morpholexical operations correspond to what Simpson (1983) refers to a semantic
redundancy rules, while Morphosyntactic operations correpond to her relation-changing
rules. Both types of rules are lexical rules, i.e they apply within the lexicon, on her
account U well a on the account developed here.
COMPLEX PREDICATES AND MORPHOLEXICAL RELATEDNESS / 57
SP. . . SP . . .
\ \
^,)
h h
h h

Or, Or,
Morpholexical
Rule
Intrinsic Class
Morphosyntactic Rule
Defaults
SP. . . SP . . .
\ \
P' (^' ,^' ))

h h
h h

Or' ,Or')
Figure 1 Model of (Modifed) Mapping Theory
to those supplied by IC assignment: these operations can efect the fnal
GF assignments to arguments but cannot afect the lexical semantics or
consequent IC feature assignment of predicates: passive and locative in
version are two phenomena afected by morphosyntactic operations. These
operations will play no role in the present paper. In general, the bifurca
tion of lexical operations stipulated in a theory such as that proposed by
Zubizarreta ( 1987) follows naturally on the present account. The model of
mapping advanced here has the shape shown in Figure 1.
The immediately relevant information in this diagram concerns the
functions performed by morpholexical and morphosyntactic rules. Mor
pholexical rules manipulate semantic properties (SP) of arguments, i. e. ,
these are roughly the proto-role features of Dowty 1991. Arguments are
interpreted as hierarchically organized valence slots represented here as (.
Assemblages of semantic properties associated with arguments determine
the intrinsic (or, IC) feature assignments of arguments, i.e. , the values for
h. Morphosyntactic rules and defaults further determine the feature ma
trices of individual arguments, Le. , the values for 1, within the limits de
fned by monotonicity: this principle basically insures that the operations
responsible for complementing the IC features of arguments will neither
alter IC features nor align Ie features with incompatible and contradictory
features. The entire feature composition of arguments determines their GF
status: grammatical functions are not atomic, they are feature matrices.
The treatment of locative alternation phenomena as well as the well
known unaccusative efects exhibited by constraints on prenominal deverbal
adjectives, e. g. , the yellowed leaf vs. * the limped man, will be derived from
the operation of morpholexical rules. On the present account, locative al
ternation alters semantic information associated with lexical entries with
consequences for grammatical function realization, while de verbal adjective
formation appeals to the refexes of this information contained in lexical
entries: syntactic efects are attributed to manipulation of featural infor-
mation in lexical entries.4 The information accessed concerns the lexical
apect ( [TELIC] ) and the IC feature composition (specifcally, the [-0]
vs. [-r] status) of lexical entries. 5
The paper is organized as follows. Section 1 introduces the Hungarian
locative alternation constructions. Section 2 proposes a version of LFG
mapping theory appropriate for treating locative alternation phenomena
in terms of complex predicate formation. As will be seen, Hungarian loca
tive alternation constructions contain complex predicates consisting of a
Pre Verb (PV) and a verb stem: the relevance of PVs for establishing lex
ical aspect and for IC feature determination will be discussed. Section 3
examines constraints on pre nominal deverbal adjective formation in Hun
garian and explains the observed unaccusative efects associated with these
constraints in terms of the interaction between lexical aspect and the IC
feature composition of arguments motivated in Section 2.
2 Locative Alternation and Complex Predicates
Hungarian exhibits a productive means of predicate formation via prefx
at ion: a prefxal preverb
6
is juxtaposed to a verbal stem and-modulo the
(semantic class of the) preverb and the (semantic class of the) verb-the
combination results in an alteration of lexical meaning, changes in valence
and/or case government and/or changes in the GF status of arguments, as
contrasted with these properties of the simple verb stem. The constitutive
pieces of such morpho lexical compositions, i.e. , the prefxal preverb and
verbal stem, are separable under specifable syntactic conditions. 7
I turn now to a restricted domain of complex predicate formation,
namely, the domain of locative alternation. The Hungarian variants exhibit
all the characteristics of locative alternation constructions mentioned pre
viously: grammatical functions change, while the semantics of the variants
remains basically similar but difers along the so-called holistic dimension.
Consider the Hungarian constructions presented below. 8
4
The treatment of deep unaccusativity in Bresnan and Zaenen 1990 informs the present
speculations in this domain: they propose a way in which a monostratal thery with
rich lexical asumptions can accommodate the types of efects generally attributed to
underlying syntactic relations.
5
The present proposal parallels the position propounded in Zanen 1990 and was pro
posed in previous versions of the present paper. For the most, I will rely on Zaenen
'
s
formulations here.
6
1n Ackerman 1987 I distinguish between two basic types of PVs, specifically, those
which bear no synchronic syntactic relation to the verbal stem, i.e., prefixal preverbs,
and those that do bear some synchronic syntactic relation to the stem, i.e., argumen
tal preverbs interpretable a incorporated elements. For the most part this paper will
concentrate on prefxal preverbs.
'This syntactic behavior of ver bal prefxes explains the earlier reference to juxtaposition
rather than afxation.
B
lntransitive variants will be ignored in the present work although the basic analysis
proposed here extends without complication to them as well. The following abbreviations
(2) a. a paraszt (ra= )rakta a szenat
(3)
(4)
the peasant (onto)loaded-3sg/DEF the hay-ACC
a szekerre
the wagon-SUBL
The peasant loaded the hay onto the wagon.
b.*a paraszt (ra=)rakta a szekeret
the peasant (onto)loaded-3sg/DEF the wagon-ACC
szenaval
hay-INSTR
'The peasant loaded the wagon with hay. '
a. a parazt meg=rakta a szekeret
the peasant PERF-Ioaded-3sg/DEF the wagon-ACC
(szenaval)
(hay-INSTR)
'The peasant loaded the wagon (with hay) . '
b.*a paraszt meg=rakta a szenat
the peasant PERF-Ioaded-3sg/DEF the hay-ACC
a szekerre
the wagon-SUBL
'The peasant loaded the hay onto the wagon. '
a. a paraszt tele=rakta a szekeret
the peasant full-Ioaded-3sg/DEF the wagon-ACC
'The peasant loaded the wagon full with hay' .
b. * a paraszt tele=rakta a szenat
the peasant full-Ioaded-3sg/DEF the hay-ACC
a szekerre
the wagon-SUBL
'The peasant loaded the hay full onto the wagon' .
szenaval
hay-INSTR
In contrat to their English analogues, the Hungarian locative alternation
predicates are complex, i. e. , they consist of a Pre Verb (PV) and a verb stem.
The preceding sentences illustrate the types of complements permitted to
co-occur with three representative verb forms: (2) contains verb forms with
an optionally present directional preverb, (3) contains verb forms with the
Perfective preverb meg, and (4) contains verb forms with the Resultative
preverb tele.
The presence or absence of prefxes correlates with diferences in case
government and function assignment for nominal complements as well as
and conventions are employed in the Hungarian examples: ACC " accusative, SUBL =
sublative, INSTR = instrumental, DEL = delative, ABL " ablative, IN = inessive, PERF
" perfective, DEF = defnite conjugation, and '=' is employed a a juncture marker
between PV (preverbs) and verbal stems.
with aspectual diferences in the meaning associated with alternative pat
terns. In particular, wherea hay must be marked in the ACC and fnction
a the OBJ complement of the simple verb or complex verb with direc
tional prefx r 'onto' ) , it must bear INSTR marking and function as an
OBL when it co-occurs with complex predicates containing either the PERF
( meg) or resultative prefx ( tete ' full' ) . Conversely, wagon must appear
with a locational cae marker and fnction a an OBL with simple verbs or
complex predicates containing directional prefxes, while it must bear ACC
marking and function as an OBJ when it co-occurs with complex predicates
containing either a PERF or resultative prefx.
There is evident systematicity in the correlation of prefx, case-marking
and grammatical function: nominals fnctioning a loc arguments are al
ways marked with LOC cae and bear the OBL relation in conjunction
with directional preverbs, while these nominals are marked ACC and fnc
tion as an OBJ when they co-occur with PERF or resultative preverbs.
In a similar fahion, nominals functioning a th arguments are marked
with the ACC case and bear the OBJ relation when they co-occur with
directional preverbs, while these nominals are marked with the INSTR
cae and function as OBLs when they co-occur with PERF or resultative
preverbs.
It remains to be observed that the juxtaposition of pre verbs and verb
stems in Hungarian is traditionally asumed to alter the aspectual status
of the afected predicate. In particular, the presence of a preverb generally
confers perfective or completive aspect on the predicate. There are certain
preverbs such a meg which appear devoid of lexical semantic meaning and
seem to function almost exclusively a aspectual markers, while there are
other preverbs such a tele which simultaneously contribute lexical semantic
and apectual values to the predicate. For present purposes I will asume
that the presence of a preverb correlates with alteration of the lexical apect
of a predicate and that this is registered in the lexical entry for a predicate.
Following Zaenen 1990 I will assume that the preverb makes the afected
predicate [+TELICj .
Distribution of time adverbials
There is standard evidence which can be adduced for the claim that the
presence of a pre verb correlates with the aspect of a predicate. Follow
ing Vendler 1967 and Dowty 1979 (among others) there is a rich tradi
tion devoted to the interaction between predicates and temporal adjuncts.
Consider the tabular representation of predicate and adverb interactions
presented below:9
9It should be observed that Hungarian possesses a construction in which a directional
PV follows a verbal stem and conveys prgressivity. Example (5a) below reflects this
reading, e.g. raxta ra a szerat 'wa loading the hay onto'.
(5) egy onin at egy ora alatt
' for an hour' 'in an hour'
a. rakta (ra) a szenat +
'loaded the hay'
b. ra=rakta a szenat +
'onto=loaded the hay'
c. tele=rakta a szekeret + .
' full-loaded wagon'
d. meg=rakta a szekeret +
'PERF-loaded wagon'
Fom the present perspective it is important to observe that the simple verb
stem (or verb stem with post posed directional prefx, cf. preceding footnote)
and prefxed verb stems display diferent co-occurrence options. The par
ticular correlation of adverbial and verb is signifcant for the present thesis:
prefxed verbal stems co-occur with an adverbial generally acknowledged to
correlate with perfective apect, while the simple stem (or, stem and post
posed directional prefx) co-occur with adverbials generally acknowledged
to correlate with progressive apect. IO
In summary, the assumption that prefxes are asociated with aspectu
ality is borne out in the co-occurrence constraints of time adverbials and
certain predicates. As previously mentioned, I will asume with Zaenen
( 1990) that the relevant apectual value is [+TELIC] : Dutch verbs contain
ing separable preverbs of direction are like Hungarian verbs in this respect.
Having seen the importance of pre verbs for determining lexical aspect, se
mantic interpretation of the variants and the grammatical function status
of arguments, it is necessary to investigate how complex predicate forma
tion should fgure in our theory of grammar.
3 A Version of Mapping Theory
In the version of mapping theory advanced here the slots representing the
valence of a predicate, i. e. , expediently expressed here in terms of thematic
role labels, are associated with Semantic Properties and these semantic
properties are partially determinative of the Grammatical Fnction status
of these arguments. All of this information is contained in lexical entries
a will be amply illustrated below.
The lexical mapping theory developed in Bresnan and Kanerva 1989
decomposes grammatical functions (GFs) into (sometimes underspecifed)
feature bundles. Two binary-valued features are postulated, namely, [o]
for Objective and [r] for (Semantically) Restricted and these feature bun-
l
O
Difrences betwen perfective vs. progressive apect will also be relevant in the dis
cussion of constraints on deverbaI adjective formation in Section 4.
dIes defne natural classes of functions. Both the feature composition of
GFs and the natural classes defned by these features are listed below.
l
1
Feature Composition of Grammatical Functions
SUBJ [-0] OBJ [-r] OBJe [+0] OBLe [-0]
[-r] [+0] [+r] [+r]
Natural Classes Defned by Features
[-r] SUBJ, OBJ [+r] OBLe , OBJe
[-0] SUBJ, OBLe
[+0] OBJ, OBJe
The basic question with respect to these features concerns the principles
which are responsible for associating them with the valence positions of a
predicate. One claim advanced in Bresnan and Kanerva is that a certain
class of features are determined by lexical semantic properties of the pred
icate. This is formulated by them in terms of Intrinsic Classifcation or IC
feature classifcation: certain features are assigned to arguments by virtue
of the thematic role of those arguments. This assignment is presented
below:
I (ntrinsic) C (lassifcation)
agent [-0] , patient/theme [-r] or [+0]12, other [-0]
3. 1 Partial Lexical Entries
In the present paper IC feature clasifcation will also be keyed to lexi
cal semantic properties of elements in the predicate's valence structure.
However, I will be assuming (along with Zaenen 1990) that thematic roles
are not atomic, but are rather abbreviations for constellations of semantic
properties as on the analysis of Dowty 1991. Dowty proposes that verb
meanings entail the existence of arguments defned schematically a con
stellations of a small set of recurring semantic properties. 13 The semantic
properties employed here follow those suggested by Dowty: 14
11 For the intuitive motivations underlying the postulation of these features see Bresnan
and Kanerva 1989.
1 2The optional asignment of [+0] to th is proposed for Chichewa double object con
structions in Alsina and Mchombo 1989. Since Hungarian does not contain double object
constructions this option will not play a role here.
1
3
Cf. Ladusaw and Dowty 1988 and Culicover 1989 for further discussion.
14The particular properties germane to the present discussion will be causally afected,
change of state and incremental theme so I will, for the most part, take no stand on the
precise repertoire of properties listed above. The delimitation of a reliable and explana
tory set of semantic properties, a Dowty acknowledges, is as fraught with vagueness a
the defense of particular collections of (atomic) thematic or semantic roles: in the present
paper I have nothing illuminating to say about these issues. For present purposes it is
sufcient to suggest that there is a natural partition of operations and that morpholex
ical operations a defned here are driven by lexicasemantic properties of predicates
irrespective of whether these be atomic thematic roles or features constitutive of such
arguments.
Proto-Role Properties
Proto-Agent Proto-Patient
volition change of state
sentience incremental theme
causes event causally afected
movement stationary
referential possibly referential
On such a view, traditional thematic role labels are (misleading) abbre
viations for particular constellations of properties. That is, thematic roles
are interpretable as feature bundles. 1 5 On Dowty's account, the algorithm
for the grammatical function realization of feature ensembles is as follows:
Argument Selection Principle
The argument of a predicate having the greatest number of Proto
Agent properties entailed by the meaning of the predicate will, all else
being equal, be lexicalized as the subject of the predicate: the argu
ment having the greatest number of Proto-Patient properties will, all
else being equal, be lexicalized as the direct object of the predicate.
A corollary relevant to locative alternation, among other constructions con
taining OBL(ique) complements, is formulated as follows:
Corollary
With a three place predicate, the non-subject argument having a
greater number of entailed Proto-Patient properties will be lexical
ized as the direct object, the non-subject argument having fewer en
tailed Proto-Patient properties will be lexicalized as an oblique or
prepositional object . . .
In sum, the system of mapping proposed by Dowty directly associates an
unweighted inventory of semantic properties with specifc syntactic roles
via Argument Selection PrinciplesI6 and various corollaries.
On the present account, as in Zaenen 1990, I will be assuming that
(possibly weighted) constellations of properties determine the value for the
Intrinsic Classifcation feature of arguments: as will become clear below,
this means that semantic properties indirectly constrain the GF realization
of arguments.
The previous assumptions will yield partial lexical entries for predicates
as shown in Figure 2.
A constellation of semantic properties is associated with a valence po
sition, i. e. , B, and determines the values for the IC features, i. e. , h.
1
5
That the entities reified into role labels are organized into hierarchies will be asumed
in Section 3. 5 below.
1
6
Cf. Fillmore 1977a, 1977b and Gawron 1983 for discussion of SUBJ and OBJ selection
principles which are keyed to hierarchically arranged lexica-semantic features.
SP . . . SP . . .
\ \
P (, . . . ,)

L
Figure 2 Schematic Partial Lexical Entry
3. 2 Basic lexical entries: unmarked IC assignment
The principles for determining specifc Ie feature asignments for baic
or underived predicates, i. e. , for simple verb stems without PVs, roughly
follow those proposed by Zaenen (1990).
Principles of IC Feature Assignment
(1) The argument with the most or most heavily weighted proto-patient
properties is intrinsically clasifed a [-r] .
(2) The argument with the most or most heavily weighted proto-agent
properties is intrinsically clasifed a [-0] .
(3) All other arguments are intrinsically clasifed as [-0] .
These feature specifcation principles yield the following partial lexical en
try for the baic or underived predicate rk ' load' : the frst argument ha
a preponderance of proto-agent properties and is therefore asigned [-0],
the second argument ha a preponderance of proto-patient properties and is
therefore asigned [-r] , while the third argument has fewer of the preceding
properties or diferent properties and is therefore asigned [-0] P
(6) rak ( ag, th,
[-0] [-r]
loc)
[-0] Intrinsic Classifcation (Ie)
I will asume that the arguments expediently labelled a th and loc in
the preceding examples share the proto-patient property causally afected.
The proposed Ie feature asignment principles obtain, by hypothesis, for
basic predicates. The crucial characteristic of Hungarian locative alterna
tion predicates, however, is that they contain derived complex predicates.
As mentioned previously, the presence of a preverb correlates with the al
teration of inherent lexical apect and depending on the preverb and the
(clas of the) verb stem, with an alteration of the lexical semantics, cae
government and grammatical function asignment to arguments a well.
'' will employ thematic role labels for convenience, but they should be understood a
abbreviations of semantic properties in the same way that grammatical function names
should be considered a abbreviations for feature matrices.
COMPLEX PREDICATES AND MORPHOLEXICAL RELATEDNESS I 65
3.3 Complex lexical entries: marked IC assignment
I will asume that complex predicate formation with preverbs is a mor
pholexical operation. This means that the combination of a preverb and
verb stem can alter the lexical semantics of a basic predicate: this alter
ation is registered in the lexical entry of the complex predicate. Since
IC feature assignment is keyed to lexical semantic properties of predicates,
the application of such an operation can alter IC feature assignment. These
morpholexical operations are a way of relating predicates, i.e. , of account
ing for observed similarities, while also accounting for whatever diferences
they may display: the diferences are attributable to divergent inventories
of semantic properties and whatever refexes of these may be exhibited as
a function of diferent IC feature assignments. We have already seen that
the presence of a preverb alters lexical apect: the correlate of this will
be that the relevant complex predicate will be associated with the feature
[+TELICl . In addition, each preverb will confer certain other semantic
properties on arguments of the complex predicate. In general, morpholex
ical operation responsible for complex predicate formation will lead to a
marked IC feature assignment viz. the principles proposed for feature as
signment with basic predicates.
There are two marked feature assignments associated with preverbs
which are relevant for locative alternation. The preverbs will be divided
into two types. There are (uses of) preverbs which contribute a specifcation
of directionality to the action denoted by the verb stem: these PVs will be
subscripted with the diacritic [+dirl . This type i exemplifed by the PVs
r 'onto', be 'into', ki 'out of', etc. There are also (uses of) preverbs which
do not contribute any directional sense to the action denoted by the verb
stem. This type is exemplifed by the PVs meg PERF' and tele ' full'.
Whereas both types of preverb determine that the lexical aspect of the
complex predicate is [+TELIC] , each type yields a diferent distribution of
lexico-semantic properties to the arguments of the complex predicate.
3. 3. 1
PV
[+dirj
PV
[+dirj
S establish motional and topological constraints on the loe argument
of the complex predicate. The motional notions and topological constraints
are refected in the repertoire of surface case and post positions which can
be employed to encode the loe complement. These considerations become
evident with respect to locative alternation in the contrast between the
lexical entries for complex predicates baed on the verbal stem rak 'load'
when composed with either the directional preverb ki 'out of' or be 'into'.
Each of these PVs require their loe complement to be expressed by cases
or adpositions which encode certain features. For example, the PV ki re
quires its loe complement to be expressed by a morpheme which has the
features [-goal, +containmentl . 18 Two representative markers compati-
18
The cae features utilized here are intended to be heuristic: their theoretical status
ble with this requirement are the cae marker -b61 'from out of' and the
postposition al61 ' from under' . Similarly, the PV be constrains its loc com
plement to be expressed by any member of a set of morphemes which bear
the features [+goal, +containment] . Representative markers fom this set
are the case marker ba 'into' and the postposition ala 'to under' . The
presence of these PVs also determines the telicity value of the complex
predicate. The relevant lexical entries might be represented roughly as fol
lows, where the annotation for the function O
B
L is intended to convey the
requirement that this function must be encoded by markers that bear the
stated features:
ki=rak V ' unload' (ag, th, loc)
O
B
L
case
= O [-goal]
[+containment]
TELIC = +
be=rak V 'into load' (ag, th, loc)
O
B
L
case
= O [+goaIJ
[-containment
TELIC = +
In summary, PV
[+
dir
]
's alter the lexical aspect of a predicate and specif
the motional and topological features of loc arguments. This results in a
complex predicate with a lexical entry that contains the aspectual feature
[+TELIC1. The lexical entry for the complex predicate will difer minimally
from that for the basic or simple predicate: whereas the former will contain
a negative telicity feature value, the latter will have a positive value for this
feature. The partial representation of ni-rk 'onto-load' is presented below:
r1-rak V 'onto-load' ( ag, th, loc)
[-oj [-r] [+r]
TELIC = +
3.
4 PV[-dir]
I will assume here that PV
[
-
d
i
r
]
S correlate both with the association of
[+TELIC] and with a distribution of semantic properties among argu
ments. The resultative or purely aspectual [-dir] preverbs associate the
property chang
e of state with an argument containing the feature causally
afected which is associated with [-0] in basic predicates. The argument
now exhibiting a preponderance of proto-patient properties will be assigned
the [-r
J
IC feature value, while the argument with fewer (or less heavily
weighted) proto-patient values will receive [-0] .
The motivation for assigning [-oj to the th is based on three interrelated
is both arguable and irrelevant to the point at issue. [+goal] is intended to designate
centripetal or aferent motion, [-goal] designates centrifugal or eferent motion, [+con
tainment] designates that the location is construable U denoting a bounded space with
an interior.
COMPLEX PREDICATES AND MORPHOLEXICAL RELATEDNESS J 67
considerations: ( 1) [-0] is the feature assignment which follows fom the
principles of IC feature assignment presented previously, (2) Hungarian is a
language whose predicates can only have a single [-r] intrinsically classifed
argument (following the parametric diference between languages proposed
in Alsina and Mchombo 1989, and (3) the assignment of [-0] avoids a
violation of the constraint against multiple [-r] IC arguments, while yield
ing the correct function assignments for both transitive and intransitive19
variants of the locative alternation predicates.
In the predicates at issue, these assignments mean that the argument
which is roughly correlative with the loc role will receive [-r] , while the
argument roughly correlative with the th will receive [-0] . The PV
dir
also
determines the telicity value for the complex predicate as +. The resulting
partial lexical entry is presented below.
meg-rak V 'onto-load' ( ag, th, loc)
[-0] [-0] [-r]
TELIC +
3. 5 Full Lexical Entries
Since grammatical fnctions, on the present interpretation, are tiered fea
ture bundles, there must be additional ways to complement the IC feature
composition of arguments. Within mapping theory there are essentially
two ways to do this: there are morphosyntactic operations such as pas
sive which determine the argument inventory which will be syntactically
expressed and there are default asignments of features which are depen
dent on the IC features of arguments. These alternative ways of asigning
features are presented below:
Morphosyntactic Operations
e.g., Passive 0)
I
o i. e. , suppress the highest argument
The efect of passive is to "suppress" the highest argument (cf. below) .
This means that this argument will not be expressed syntactically. The
suppression of this argument would lead to the ostensible th playing the
role of SUBJ, given the well-formedness conditions stated below.
The basic default asignments are presented below:
Defaults (a) Subject default 0
|
[-r]
(b) Elsewhere default (
I
[+r]
19Though \ignore here the treatment of intransitive variants of this construction, there
is one apect of their analysis with respect to the alignment of th with [-oj which will
be addressed after the presentation of default asignments below.
These defaults function as instruction to associate arguments with par
ticular features. The arguments are organized in terms of a hierarchy. In
particular, the argument with the most proto-agent properties is higher
than the argument with the most proto-patient properties. Other argu
ments are arrayed within a hierarchy such a that presented below. The
hierarchy can be stated in terms of thematic role labels:
Argument hierarchy (partial)
ag < exp < ben/goal < th/pat <
loc
The highest argument will be referred to as 0. The SUBJ default asociates
the highest argument with the feature [-r] , while the (b) default asso
ciates other arguments with [+r], as long a this doesn't conict with the
argument S IC feature asignment (cf. below for monotonicity) . This pro
cedure leaves the argument with [-r] IC underspecifed and therefore able
to assume either the SUBJ or OBJ function (cf. below for well-formedness
conditions) .
The assignment of features to complement IC features is constrained by
Monotonicity:
Monotonicity Constraint on Feature Assignment
Feature assignment is additive, i. e. , neither rules nor defaults can
alter feature assignments nor result in contradictory assignments of
features for arguments.
Once the assemblages of features have been determined, there are two well
formedness conditions on fnal lexical entries for predicates:
Well-formedness Conditions on Lexical Forms
(1) Functional Uniqueness: Every grammatical function and every func
tional feature must have a unique value.
(2) SUBJ condition: Every lexical form must contain a SUBJ function.
2
o
The manner in which features are associated with arguments and result in
asignments of GFs can be illustrated in the cae of the basic Hungarian
verb rk 'load' .
(7) rak (ag,
[-0]
[-r]
th,
[-r]
SUBJ SUBJ/OBJ
SUBJ OBJ
TELIC "
loc)
[-0]
[+r]
OBL
OBL
Intrinsic Classifcation (IC)
Default
Bi-Uniqueness Constraint
The verb rk 'load' is asociated with a thematic argument inventory con
sisting of an 0g, th and loco The arguments are associated with features
2
0
A lexical form is the lexical representation that results from the mapping of arguments
into grammatical functions. The SUBJ condition is possibly parametric: there may be
some languages for which it does not obtain.
following the algorithm for semantic determination of IC feature assign
ment. Specifically, the ag argument is associated with a [-0] IC since it
exhibits the most proto-agent properties, the th argument is associated
with a [-r] IC since it contains the most proto-patient properties, while
the loe argument is associated with the IC [-0] since it is the remaining
argument in the argument structure. Since there are no (morphosyntac
tic) operations such as passive which apply here, the defaults then apply:
the SUBJ default requires that the highest argument is associated with the
feature [-r] , while the elsewhere default, states that all other arguments
are asociated with the feature [+r] as long as this assignment does not
violate the monotonicity contraint. In the case of 'load', the highest ar
gument is ag which is associated with [-r] , while the loe is associated with
the feature [+r] . The th cannot be associated with [+r] without violating
monotonicity, so it remains underspecifed. The object created by the as
signment of features to arguments, namely, a Lexical Form is constrained
to conform to the two well-formedness conditions (1) functional uniqueness
and (2) the SUBJ condition. The result of assigning features to ' load' is a
representation in which the ag is fully specifed and must appear as SUBJ,
the th is underspecifed and can, consequently, appear as either a SUBJ or
an OBJ, while the loe is fully specifed and must appear as an OBL. The
SUBJ condition, which obtains in Hungarian, applies and determines that
the th is OBJ. The fnal GF assignments are indicated in the text: 0g is
mapped into SUBJ, th mapped into OBJ and loe into OBL.
The assignment of features to the basic verb 'load' is straightforward:
ag receives [-0] IC, th the [-r] IC, and loc the [-0] IC, then defaults apply
fully specifying the ag with [-r] by the SUBJ default, and fully specifing
the loe with [+r] by the elsewhere default, and leaving the th underspecifed.
These assignments result in the ag as SUBJ, th as OBJ and loe as OBL.
In the present instance, as in all others, the interaction between in
trinsic classifcation, rules and default assignments lead to an increasing
specifcation of the features associated with semantic roles and this results
in GF realizations for arguments.
The lexical entries for complex predicates containing PVs are derived
in the following way. Since both tele and H6g are [ -dir] the loe argument
is associated with [-r] IC, while the th argument is associated with [-0]
IC. The ag argument receives the [-oj IC. The SUBJ default applies to the
ag assigning it [-r] , while the elsewhere default applies to the th argument
assigning it [+r] : the loe argument remains underspecifed. Since the ag
is fully specified as [-0] and [-r] it surfaces as the SUBJ. Likewise the
th is fully specifed as [ -0] and [+r] and therefore surfaces as an OBLo
The loe is underspecifed and in principle could assume either the SUBJ or
OBJ function. However, a violation of functional uniqueness would ensue
if the th appeared as a SUBJ, so it assumes the OBJ function. The entire
predicate is, of course, associated with the feature TELIC ~ +.
Composition of Complex Predicates
ra=rak ' onto-load' ( ag, th, loc )
[-0] [-r] [+r] IC
[ -r] [-0] Default
SUBJ OBJ OBL
TELIC = +
tele=rak 'full=load' ( ag, th, loc )
[-oj [ -oj [-rJ IC
[-r] [+r] Default
SUBJ OBLo OBJ
TELIC +
meg=rak 'PERF-load' ( ag, (th) , loc )
[-0] ( [-0] ) [-r] IC
[ -
r] ( [+r] ) Default
SUBJ (OBLo) OBJ
TELIC ~ +
The derivation is somewhat diferent for complex predicates containing
[+dir] PVs. The th argument in predicates containing a [+dir] PV is asso
ciated with [ -rJ , while the loc argument receives the [ -oj IC feature. After
the defaults have applied the ag i fully specifed as [-0] and [-rJ , the th is
underspecifed as [-rJ, and the loc is fully specifed as [ -0] and [+r]. These
feature specifcations, in conjunction with the well-formedness conditions,
result in GFs assignments whereby the ag is SUBJ, the th is OBJ and the
loc is an OBLo .
21
2
1 The atute reader may have calculated that the conventions for default feature assign
ment might lead to incorrect function asignments for intransitive locative alternation
predicates. The apparent problem can be illustrated with reference to the intransitive
locative alternation predicate tele=folyik 'full fow' which is associated with the following
argument structure and presumable IC feature assignments:
tele=folyik 'full fow (th, loc)'
[-oj [-r
J
The straightforward application of the SUBJECT and elsewhere defaults would asociate
[-rJ with the highest argument, e.g., th, while the loc argument would remain underspec
Hied. This would yield function asignments where the th would be the SUBJECT and
the loc would be the OBJECT: locative alternation, in this instance, would appear to
be a transitivizing operation. These are, unfortunately, the wrong asignments: the
intransitive locative alternation predicates remain intransitive with the th functioning
a an OBL and the loc functioning a the SUBJECT. This situation might be explained,
a suggested by an anonymous reader, if we follow the spirit of Bresnan and Zaenen's
attempt to yield an external versus internal argument distinction within lexical mapping
theory. If we asume that an external argument 0 must bear a preponderance of proto
agent properties, while arguments with a preponderance of proto-patient properties are
always internal irrespective of the IC features they bear, then the SUBJECT default can
be interpreted a an instruction to asociate an external argument with [-rJ . For the
cae at hand, the predicate tele=folyik 'full=fow' does not have an external argument,
In this section we have seen that the conventions for composing [dir]
PVs with predicates act in conjunction with various assumptions of map
ping theory to yield the correct assignment of GFs to arguments of transi
tive predicates. I turn now to an investigation of how this analysis interacts
with observed constraints on deverbal adjective formation.
4 Deverbal Adjective Formation
The morpho lexical relatedness analysis of locative alternation constructions
refers criterially to the efects of composing complex predicates in the lex
icon: this composition process determines the apectuality of predicates
in terms of the feature [+TELIC] as well a the distribution of semantic
properties determinative of IC feature assignment. In this fnal section I
will demonstrate that reference to these features (or more accurately, to
the properties they denote) is central for explaining otherwise puzzling
constraints on deverbal adjective formation in Hungarian. In the course
of this presentation I will argue that proposals baed on thematic roles,
grammatical fnction (surface or deep), or morphology, i.e. , the presence
or absence of PVs, are inadequate.
Consider the following examples of deverbal adjectives formed fom
locative alternation verbs:22
(8) a.
b.
a rarakott
the onto-loaded
a telerakott
the full-loaded
,
/*
k '
szenath
sze erloc
hay/wagon
k'
/*
'
sze erloc szenath
wagon/hay
The examination of restrictions on perfect adjective formation in English
have led to the hypothesis that there is a thematic constancy constraint
underlying a similar distribution of judgments in that language: 2
3
roughly,
a verb can be adjectivalized only if it contains a th( ere) in its argument
array. The th argument functions as the head (or externalized argument)
of the NP in which the derived attributive adjective appears. This proposal
ha been challenged in Levin and Rappaport 1986 where it is argued that
data such as:
(9) the stuf ed pillO
Wloc vs. *the stufed featherst
h
Le. , it's th argument is still internal despite its [-0] IC feature. The elsewhere default
applies to the th providing a [+r] feature and insuring that it assumes an OBL func
tion, while the lac argument must assume the SUBJECT function in order to satisfy the
SUBJECT condition. It can do this since it is the underspecifed argument asociated
with a [-r] IC. The proposal to distinguish betwen external and internal arguments
on the bais of semantic properties and independent, in some sense, of the IC features
borne by the relevant arguments leads to various empirical consequences for phenomena
other than locative alternation. I will not explore these consequences here.
22The prefixal preverb in these examples is italicized and the diacritic for juxtaposition,
i. e. , '=', ha been omitted since in these adjectival forms the prefix is not separable from
the nominal stem.
2
3
Cf. Wasow 1977, Bresnan 1982.
run afoul of the such a hypothesis: among other problems, certain loc ar
guments can inexplicably head nominals containing deverbal adjectives,
while certain th arguments inexplicably cannot. Levin and Rappaport
propose that this distribution can be explained by the Sole Complement
Generalization:
Sole Complement Generalization
An argument that may stand as a sole NP complement to a verb can
be externalized by APF (= Adjectival Passive Formation)
This generalization attempts to relate the observation that acceptable ad
jectival passive constructions correlate with acceptable clauses containing
a single internal argument. In other words, we fnd clauses of the following
sort correlative with the previously cited j udgments for adjectival passive
forms:
(10) Henry stufed the pillow/
oc vs. * Henry stufed the feathersth
Two questions arise with respect to this sort of explanation for the observed
judgments: ( 1) if such a generalization actually obtains then it represents
a mystery eaily a arresting a the constraint on deverbal adjectivalization
which it is supposed to explain, and (2) it is not clear how or whether this
sort of "explanation" can be extended to account for similar constraints in
languages with null complements, i.e. , languages in which the notion "sole
complement" is difcult to defne.
In the following discussion I will argue that featural distinctions pre
viously referred to in order to account for locative alternation predicates
are likewise crucial for the explanation of constraints on deverbal adjecti
valization in Hungarian. In efect, we will see that neither thematic con
stancy nor the sole complement constraint provide explanations for these
data: in particular, the thematic constancy proposal, arguably, founders
on data such a those presented by Levin and Rappaport ( 1986) while the
baic efects of the sole complement generalization as well as constraints
on adjective formation follow from a feat ural treatment of the relevant
data.
The. Hungarian constructions (8a) and (8b) conict with the thematic
constancy condition in much the same fashion as the English examples
cited above: the loc argument is head of the NP, while the th cannot be
the head of this NP. In Hungarian, unlike in English, complements of the
verb may be omitted in surface syntax:
( 1 1 ) a paraszt ra=rakta
the peasant onto=loaded-3sg/DEF
'The peaant loaded (something determinate) (somewhere
determinate) ' .
( 12) a paraszt tele=rakta
the peaant full=loaded-3sgjDEF
'The peaant full-loaded (something determinate) (with something
determinate) . '
Wherea the notion of sole complement is difcult to defne in such a lan
guage, the notion of argument asociated with a [-r] IC feature is quite
transparent. Specifcally, the complex verbs presented above have [-r] as
sociated with the th argument of verbs containing a directional prefxal
preverb and with the loc argument for complex verbs containing tete or
meg. In other words, appeal to the feature [-r] explains the present distri
bution of grammaticality judgments. This treatment can be formulated as
follows:
[-r] Condition on Adjectivalization (Preliminary)
A verb O a is related to an adjective f if the argument structure of
O a contains an argument annotated with [-r] .
This condition claims that the presence of [-r] constitutes a necessary
condition on deverbal adjectivalization in Hungarian. 24
Given the fact that the arguments associated with [-r] in the preceding
examples are realized a ACC marked complements, it might be claimed
that surface OBJ(ecthood (correlative with accusative marking in the
present instances) is the determinant of adjectivalization rather than the
presence of [-r] . The inadequacy of this line of speculation becomes appar
ent by examining two types of evidence. First of all, there are verbs which
govern the ACC cae for their OBJ complements but which cannot undergo
the relevant adjectivalization process. Examples of this sort suggest that
neither cae marking nor the surface relation OBJ provide sufcient condi
tions for the formulation of a constraint on adjective formation. Second,
there are some intransitive verbs whose nominative marked SUBJ argument
appears a the modifed head in an NP containing its deverbal adjectival
form.2
5
These examples suggest that surface OBJecthood is not a neces
sary condition on the formulation of a constraint on adjective formation. I
examine each of these caes in turn below.
There are numerous simple (Le. , unprefxed) transitive verbs which gov
ern the ACC cae for their OBJ th arguments. Consider the two represen
tative lexical entries below:
2
4We will se below that restrictions on perfect adjective formation for one of the in
transitive variants of the locative alternation predicates precludes a stronger formulation,
i.e., one in terms of both necessary and sufcient conditions.
2
5
This latter phenomenon, obviously, raises the question about unacusativity or un
derlying objecthood in Hungarian. We will see below that the presumable unaccusative
verbs in Hungarian are those that bear an apectual feature [+TELIC] and contain a
[-r] argument.
( 13) keszit 'to prepare' (ag, th)
[-0] [-r]
( 14) sargit 'to yellow' (ag, th)
[-0] [-r]
The account of deverbal adjective formation presented thus far predicts
that verbal forms such as these should undergo perfect adjectivalization:
they contain arguments which have a [-r] . The judgments associated with
the following constructions are therefore surprising.
( 15) *a keszitett ruha
the prepared clothes
( 16) *a sargitott level
the yellowed leaf
The presence of [-r] in the lexical entries for these predicates make it
mysterious why such forms are prohibited. However, it turns out that
predicates related to those presented above can participate in adjective
formation of the relevant sort. Consider the following partial lexical entries
and nominal constructions in this regard:
( 17) el=keszit 'to prepare' (ag, th)
[-0] [-r]
(18) meg=sargit 'to yellow' (ag, th)
[-0] [-r]
( 19) az elkeszitett ruha
the PERF-prepared clothes
'the prepared clothes'
(20) a megsargitott level
the PERF-yellowed leaf
'the yellowed leaf'
In these constructions the [-dir] PV confers [+TELIC] lexical apect on
the complex predicate. In addition, the complex predicate contains an
argument with a [-r] IC feature. A revised version of the constraint on
adjective formation might be formulated a follows:
Condition on Adjectivalization (Revised)
A predicate L is related to an adjective 3if (i) the argument structure
of L contains an argument annotated with a [-r] IC feature and (ii)
the predicate bears the feature [+TELIC] .
The preceding discussion clearly demonstrates that a constraint on adjec
tive formation keyed to ACC case marking and/or the surface OBJ function
status of an argument alone is inadequate. The inadequacy of such a formu
lation becomes even more evident upon examining the behavior of certain
intransitive verbs with respect to perfect adjective formation.
Consider the following partial lexical entries for the intransitive variants
of the transitive verbs presented in ( 1 7) and ( 18).
(21) a. kesziil 'to get prepared' (th)
b. el=kesziil
(22) a. sargul
b. meg=sargul
'to yellow' (th)
[-r]
(th)
[-r]
[-r]
(th)
[-r]
TELIC = +
TELIC = +
The revised analysis predicts that only adjectives based on the (b) vari
ants will be well-formed since only these forms contain both the aspec
tual feature [+TELIC] and an argument annotated with the [-r] IC fea
ture. The correctness of this prediction is exemplifed by the following
constructions.26
(23) a. *a kesziilt ruha
the prepared clothes
b. az elkesziilt ruha
the PERF-prepared clothes
'the prepared clothes'
(24) a. * a sargult level
the yellowed leaf
b. a megsargult level
the PERF-yellowed leaf
'the yellowed leaf'
The preceding distributions not only conform to the predictions of the
revised condition on adjective formation, but they illustrate clearly that
surface OBJecthood is not a determinant of candidacy for adjectivalization.
Some notion of "deep" or underlying OBJecthood would seem to be at issue:
such phenomena are ordinarily interpreted as indicative of unaccusativity.
2
6
It should be observed that there are other ways to save the ill-formed constructions
found in the text other than by providing simple verb stems with prefxes. In particular,
a observed in Laczk6 and Ackerman 1984 it is often possible to provide the simple verbal
stem with an adverbial; a balm keszilt ruha, 'the clothes prepared for a ball' (lit. the for
a ball prepared clothes) or, az e/sokent ftott fu, 'the boy who ran and came in frst' (lit.
'the frst run boy' ). In addition, there are certain constructions which are acceptable
without either prefxes or adverbials of the sort alluded to above. For example, ji stolt
hus 'smoked meat' and vagdalt hus 'ground meat' exemplif such constructions. These
forms raise questions which cannot be addressed here.
This is the nature of the analysis proposed by Rappaport and Levin (1988)
for somewhat similar distributions of data in English. The present analysis
yields the observed unaccusative efects by appealing to the presence of the
IC feature [-r] in the condition on adjective formation. Bresnan and Zaenen
(1990) argue for independent reasons that "deep" unaccusativity efects
can be captured by appealing to the IC feature [-r] . The present analysis
adduces the present distributions as further support for their proposal.
As on their analysis, there is no need in the present case to appeal to a
distinction between underlying vs. surface grammatical relations.
Given the determinative role attributed to aspectual considerations in
the analysis, one might ask whether aspectual considerations are relevant
for the other deverbal formations a well. The relevance of aspect becomes
clear when we see that forms such as kbziil and sargul can function as
progressive active adjectives, while their prefxed partner cannot: 27
(25) a. a keszillo ruha
the preparing clothes
' the clothes being prepared'
b. * az elkesziilo ruha
the PERF-preparing clothes
(26) a. a sargul6 level
the yellowing leaf
'the yellowing leaf'
b.*a megsargul6 level
the PERF-yellowing leaf
It is likely that constraints on the formation of deverbal adjectives such as
those listed above are keyed to aspectual properties of the base predicate.
For example, the base predicate arguably is asociated with the feature
[ -TELIC] . For present purposes it is sufcient to suggest that aspectual
properties of predicates are, at least partially, determinative of constraints
on word-formation.
In this section we have seen that constraints on perfect adjective for
mation rely on the joint presence of [+TELIC] and a [-r] IC feature: only
if these features are co-present is it possible to form perfect adjectives.
4. 1 Preverbs and Tansitivity
The preceding section argued for the inadequacy of IC feat ural information
alone to explain the observed constraints on adjective formation. Aspec
tual considerations appear to play a pivotal role. Since we have seen the
importance of aspectual information in constraining adjective formation it
2
7
The present examples illustrate the adjectives in their attributive use. These same ad
jectives can be employed predicatively with attendant progressive aspectuai distinctions,
e.g., a rha keszu/oben van 'the clothes are being prepared' (lit. the clothes in-preparing
are) vs. * a ruha meg=keszu/oben van.
COMPLEX PREDICATES AND MORPiOLEXICAL RELATEDNESS / 77
might well be asked whether aspectual information alone could explain the
relevant distributions. In the present section I investigate another class
of cases that contain complex predicates but which exhibit a restriction
on perfect adjective formation: perfect adjective formation is permitted
with complex predicates containing [-dir] PVs but prohibited with those
containing [ +dir] PVs: in both instances we are dealing with [+TELIC]
predicates but we find diferences in grammaticality along the line pre
dicted by the hypothesis that both apectual and IC feature information
play a central constraining role.
There are clusters of verbs in Hungarian where the PV
[-dir]
correlates
with a change in transitivity for an asociated verbal stem: the simple stem
or complex predicate with PV[+
dir
]
are intransitive, while the complex pred
icate containing a PV
[-di
r]
is transitive. These predicates generally denote
contrasts between activities vs. accomplishments/achievements following
the Vendler /Dowty typology of aspectual types. These pairs are exempli
fed by the constructions below:
(27) a. a paraszt (ni= )celzott a madarra
the peasant (at- )aimed-3sg the bird-SUBL
'The peasant aimed at the bird. '
b. * a paraszt (ra= ) celozta a madarat
the peasant (at-)aim-3sg/DEF the bird-ACC
c. a paraszt meg=celozta a madarat
the peasant PERF-aim-3sg/DEF the bird-ACC
'The peasant aimed at the bird. '
d.*a paraszt meg=celzott a madarra
the peasant PERF=aim-3sg the bird-SUBL
(28) a. a paraszt (ra= )nevetett a filra
the peasant (at-)laughed-3sg the boy-SUBL
'The peasant laughed at the boy.'
b. *a paraszt (ra= )nevette a flt
the peasant (at- )laughed-3sg/DEF the boy-ACC
c. a paraszt ki=nevette a filt
the peasant PERF=laughed-3sg/DEF the boy-ACC
'The peasant made fun of the boy. '
d. *a paraszt ki=nevetett a flira
the peasant PERF=laughed-3sg the boy-SUBL
The verbs in the preceding constructions yield the following distribution
with respect to perfect adjective formation:
(29) a.*a racelzott madar
the at-aimed bird
b. a megcelzott madar
the aimed bird
'the aimed at bird'
(30) a.*a ranevetett f
the at-laughed boy
b. a kinevetett fu
the PERF-laughed boy
'the made fun of boy'
This is precisely the expected scatter of grammaticality judgments, given a
condition which contains reference both to telicity and the Ie feature [-rJ.
It is not the distribution to be expected if reference to apectual properties
alone was a sufcient condition on adjectivalization. Consider the lexical
entries for the relevant verbs:
(31) (ra
[
+dir
)
=)celoz 'aim at' (ag, loc)
[-oj [-oj
(32)
(33)
(34)
meg
[
-dir)
=celoz
(ra
[+dir
)
= ) nevet
ki
f
-dir)
=nevet
'aim' (ag, loc)
[-oj [-rJ
' laugh at' (ag, loc)
[-oj [-oj
' make fun of' (ag, loc)
[-oj [-r
J
The argument inventories of the preceding verbs resemble the inventories
of locative alternation predicates with one crucial diference: they do not
contain a tho Recall the conventions for asociating features with arguments
of locative alternation predicates: [-dir
J PVs specif the association of
[-r] with loc, while for [ +dir] PVs [-r
J
is asociated with the th, i. e. , the
argument with most proto-patient properties. In the present instances,
these conventions will asSure that the loc arguments of (32) and (34) are
associated with [-rJ , while [-rJ will not be asociated with any argument
in (31) and (33) since these latter arguments are goals.
2
8
This contrats
with meg=celoz 'shoot down' and ki=nevet 'make fun of' : the prefxes
meg and ki determine the [-rJ Ie features for the original loc argument.
Since the argument structure of these verbs contain the requisite aspectual
and Ie feature specifcations they can serve as a base for perfect adjective
formation.
In summary, the analysis of the present data set suggests that a condi
tion relying on the mere presence of aspectual specifcations in complex
predicates is not sufcient to yield the observed constraints on perfect
adjective formation. Instead, the analysis proposed for locative alterna
tion constructions extends explanatorily into this domain. The empir-
2
8
Given our feature asignment conventions loe argument of the baic verb will asigned
a [-oj IC feature, while the variant with a PV [+dirJ will be asigned a [-oj IC feature.
ical predictions of an analysis baed on the mere presence of preverbs
vs. an analysis baed on the presence of both aspect and a [-rJ IC fea
ture are further distinguished in the two classes of predicates examined
below.
4.2 Relatedness Between Manner of Motion and Change
of Location Predicates
Hungarian contains many intransitive verb clusters related by morpholexi
cal rule. In these complex predicates a [+dir] PV is composed with a verb
denoting motion. The preverb specifes the directionality of the motion and
sets constraints on the topological nature of a lac complement in the same
fahion a it does for locative alternation predicates. This phenomenon is
illustrated in the following constructions utilizing the manner of motion
predicate Jut 'run' .
(35) a. a paraszt futott a szobaban
the peaant ran the room-IN
'The peasant ran (around) in the room. '
b. a parazt be=ftott a szobaba/*szobab61
the peaant into-ran the room-ILL/*room-EL
'The peaant ran into the room.'
c. a parazt ki=futott a szobabOl/*szobaba
the peaant out-ran the room-EL/*room-ILL
'The peaant ran out of the room.'
The presumable lexical entries for these verbs are a follows:
(36)
(37)
(38)
fut 'run'
be
[+dir]
=ft
ki
[
+
dir]
=fut
(ag, (loc))
[-oj [-oj
'run-into'
'run-out'
(ag, locg
o
al)
[-oj [-oj
(ag, loc
source
)
[-oj [-oj
As in all previous instances, the directional preverb adds the apectual
feature [+TELICJ
to the lexical entry of the complex predicate. Both be
'into' and ki 'out of' are [+dir
J (in their use with the present class of motion
predicates) so that they correlate with the asignment of [-oj IC to their lac
arguments. The absence of [-rJ in the lexical entries of these predicates
predicts that they cannot serve as a base for perfect adjective formation.
Consider the following constructions containing the perfect adjective forms
baed on the predicate Jt 'run'.
(39) a.*a futott szoba
the run room
'the run-in room'
b. *a be-futott paraszt/szoba
the into-run room
'the peasant who ran in/the run-into room'
c. *a ki-futott paraszt/szoba
the out-run room
'the peasant who ran out/the run-out-of room'
As predicted, these prefxed manner of motion predicates cannot serve
a bases for perfect adjective formation. Once again, it is clear that the
presence of a preverb does not constitute a sufcient condition for the
application of deverbal perfect adjective formation.
The behavior of a fnal class of intransitive dyadic predicates likewise
suggests that the mere presence of a preverb is insufcient to license perfect
adjective formation. There are numerous unprefxed dyadic intransitive
verbs in Hungarian which govern an oblique cae for their OBL complement.
Their government pattern and GF selection requirements remain unaltered
when they compose with a [-dir] PV. This clas is exemplifed in the
following constructions containing the verb (meg=)art 'harm, hurt' which
governs the DAT case for its th argument.
(40) a. a nedvesseg (meg=) artott
the dampness (PERF= )damaged-3sg
'The dampness damaged the paper. '
b.* (meg)artott papir
'the damaged paper'
a papirnak
the paper-DAT
As can be seen, the mere presence of the [-dir] PV is insufcient to license
perfect adjective formation.
2
9
5 Conclusions
In this paper I have argued for a mapping theory which yields a natural
partitioning between operations which afect the lexical semantics of pred
icates, that i, Morpholexical vs. those which don't, that is, Morphosyn
tactic. I have argued that Hungarian locative alternation constitutes a
case of Morpholexical Relatedness. The clusters of predicates which par
ticipate in this alternation are created in the word formation component
of the grammar. The semantic properties which characterize the argu
ments asociated with these predicates are determinative of the Intrinsic
Classifcation features assigned to these arguments. The particular seman
tic properties conferred on arguments a well the value for the aspectual
feature [TELIC] , are correlative with the presence of directional uses of
Pre Verbs.
2
91 will not speculate here about the relation between lexical semantics, IC feature
asignment and invariant cae government with these predicates. It is not to clear
at this time whether the interaction between these phenomena is principled or simply
idiosyncratic.
Finally, the basic ingredients for the thesis of morpho lexical relatedness
between locative alternation predicates is extended explanatorily to the
treatment of constraints on perfect adjective formation in Hungarian. In
particular, the independently required necessity to refer to telicity values
of predicates as well as the specifc way in which PV types help deter
mine IC feature values for arguments appears to be required for explaining
observed constraints on deverbal adjective formation. This reliance is en
coded in the proposed condition on adjectivalization. In sum, this analysis
relies on the interaction between semantic properties of arguments, aspec
tual values of predicates and the feature decomposition of grammatical
functions.
References
Abraham, Werner, and Sjaak de Meij . 1986. Topic, Focus, and Confgurationality.
Amsterdam: John Benjamins.
Ackerman, Farrell. 1987. Miscreant Morphemes: Phrasal Predicates in Ugric.
Doctoral dissertation, University of California at Berkeley.
Ackerman, Farrell. 1990. Locative Alternation vs. Locative Inversion. In Pro
ceedings of the 9th Annual West Coast Conference on Formal Linguistics, ed.
Aaron L. Halpern. Stanford: CSLI Publications.
Ackerman, Farrell, and Andras K6mlosy. 1987. Relatedness in Hungarian
Spray /Load Verb Clusters. Presented at LSA Lexical Semantics Workshop,
Stanford University.
Alsina, Alex, and Sam Mchombo. 1989. Object Asymmetries in Chichewa.
Manuscript, Stanford University and University of California at Berkeley.
Anderson, Stephen R. 1971. On the Roles of Deep Structure in Semantic Inter
pretations. Foundations of Language 6: 197-219.
Anderson, Stephen R. 1977. Comments on the paper by Wasow. In Formal
Syntax, ed. Peter Culicover, Thomas Wasow, and Adrian Akmajian, 361-
378. New York: Academic Press.
Brassai, Samuel. 1860, 1863-65. The Hungarian Sentence. Magyar Akademiai
E
rtesito. 1 :279-399, 3: 3-128, 173-409 (in Hungarian).
Bresnan, Joan. 1982. Control and Complementation. In The Mental Represen
tation of Grammatical Relations, ed. Joan Bresnan. Cambridge, Mass. : MIT
Press.
Bresnan, Joan, and Jonni M. Kanerva. 1989. Locative Inversion in Chichewa: A
Cae Study of Factorization in Grammar. Linguistic Inquiry 20: 1-50.
Bresnan, Joan, and Lioba Moshi. 1989. Applicative Constructions in Ki-Chaga.
Manuscript, Stanford University and University of Georgia.
Bresnan, Joan, and Annie Zaenen. 1990. Deep Unaccusativity in LFG. In Gram
matical Relations: A Cross- Theoretical Perspective, ed. Katarzyna Dziwirek,
Patrick Farrell, and Errapel Mejia-Bikandi, 45-57. Stanford: CSLI Publica
tions.
Channon, Robert. 1980. On Place Advancements in English and Russian. In
Morphosyntax in Slavic, ed. Catherine V. Chvany and Richard D. Brecht.
Columbus: Slavica Publishers.
Culicover, Peter. 1989. On Thematic Relations. MIT Working Papers in Linguis
tics 10. Massachusetts Institute of Technology: Department of Linguistics.
Dowty, David. 1989. On the Semantic Content of the Notion "Thematic Role" .
In Properties, Types and Meaning, Vol. 2: Semantic Issues, ed. Gennara
Chierchia, Barbara Partee, and Raymond TUrner. Dordrecht: Kluwer.
67(3): 547-619.
Fillmore, Charles. 1968. The Case for Case. In Universals in Linguistic Theory,
ed. Emmon Bach and Robert T. Harms, 1-90. New York: Holt, Rinehart
and Winston.
Fillmore, Charles. 1977a. Topics in Lexical Semantics. In Current Issues in
Linguistic Theory, ed. Roger W. Cole. Bloomington: Indiana University
Press.
Fillmore, Charles J. 1977b. The Case for Case Re-opened. In Syntax and Se
mantics, Volume 8, ed. Peter Cole. New York: Academic Press.
Foley, William, and Robert Van Valin. 1984. Functional Syntax and Universal
Grammar. New York: Cambridge University Press.
Fkui, Naoki, Shigeru Miyagawa, and Carol Tenny. 1985. Verb Clases in English
and Japanese: A Case Study in the Interaction of Syntax, Morphology and
Semantics. Lexicon Project Working Papers 3. Cambridge, Mass.: MIT
Gawron, J. Mark. 1983. Lexical Semantics and the Semantics of Complementa
tion. Doctoral dissertation, University of California at Berkeley. Also appear
ing in Garland Outstanding Dissertations in Linguistics, 1989.
Gawron, J. Mark. 1986. Valence Structure Preservation and Demotion. In Papers
from the 22nd Regional Meeting of the Chicago Linguistic Society. University
of Chicago.
Jackendof, Ray. 1983. Semantics and Cognition. Cambridge, Mas. : MIT Press.
Jackendof, Ray. 1985. Multiple Subcategorization and the 8-criterion. Natural
Jazay, Laszlo, and Laszlo TOth. 1987. Az orosz igeaspekusr61 magyar szemmel.
Budapest: Tnkonykiado.
Kiefer, Ferenc. 1982. The Aspectual System of Hungarian. In Hungarian Lin
guistics, ed. Ferenc Kiefer. Amsterdam: John Benjamins.
Kiss, Katalin E. 1987. Confgurationality in Hungarian. Budapest: Academic
Press and D. Reidel Publishing Company.
Laczko, Tibor, and Farrell Ackerman. 1 984. Nominal Passives and Thematic
Roles. Manuscript, Hungarian Academy of Sciences, Budapest.
Laczko, Tibor, and Anna Szabolcsi. 1988. A fnovi csoport szerkezete (The Struc
ture of Noun Phrases). In Magyar mondattan, ed. Ferenc Kiefer. Akademiai
Kiado.
Ladusaw, William, and David Dowty. 1988. Toward a Nongrammatical Account
of Thematic Roles. In Syntax and Semantics: Thematic Relations, ed. Wendy
Wilkins, 61-73. New York: Academic Press.
Levin, Beth, and Maika Rappaport. 1986. The Formation of Adjectival Pasives.
Linguistic Inquiry 17:623-661 .
.
Levin, Lori. 1986. Operations on Lexical Forms. Doctoral dissertation, Mas
sachusetts Institute of Technology.
Pinker, Steven. 1989. Learability and Cognition: the Acquisition of Argument
Structure. Cambridge, Mas.: MIT Press.
Pollard, Carl, and Ivan A. Sag. 1987. Information-Based Syntax and Semantics,
Vol. I: Fundamentals. CSLI Lecture Notes No. 13. Stanford: CSLI Publica
tions.
Rappaport, Maika, and Beth Levin. 1988. What to Do with Thematic Roles. In
Syntax and Semantics: Thematic Relations, ed. Wendy Wilkins. New York:
Academic Press.
Salkof, Morris. 1983. Bees are Swarming in the Garden. Language 59: 288-346.
Simpson, Jane. 1983. Aspects of Warlpiri Morhology and Syntax. Doctoral
Vendler, Zeno. 1967. Linguistics in Philosophy. Ithaca: Cornell University Press.
Wasow, Thoma. 1977. Tansformations and the Lexicon. In Formal Syntax, ed.
Peter Culicover, Thomas Waow, and Adrian Akmajian. New York: Aca
demic Press.
Waow, Thoma. 1981 . Major and Minor Rules in Lexical Grammar. In Lexical
Grammar, ed. Teun Hoekstra et al. Dordrecht: Foris.
Zaenen, Annie. 1990. Unaccusativity in Dutch: an Integrated Approach. In
Semantics and the Lexicon, ed. James Pustejovsky. Dordrecht: Kluwer.
Zubizaretta, Maria-Luisa. 1987. Levels of Representation in the Lexicon and
Syntax. Dordrecht: Foris.
4
On Obviation
DONKA F. FARKAS
1 Introduction
This paper addresses the question of why, in many languages, the subject of
certain subordinate clauses must be disjoint in reference from certain NPs
in the immediately higher clause, as illustrated by the following French
example.
( 1) Pierrei veut qu'il.i
/
j parte.
P. wants that he leave.SUBJ(unctive)
In what follows, the complements whose subjects obey this restriction will
be called obviative.
This problem has been discussed in recent Government and Binding
literature especially in connection with Romance languages (see Picallo
1984, 1985; Meireles and Raposo 1983; Kempchinsky 1985; Bouchard 1982,
1983; among others) . Common to all these works is the attempt to account
for obviative complements within Binding Theory (BT) . In Section 2 I
will review previous solutions and then argue that the BT approach is
empirically inadequate. In Section 3 I will propose a new treatment, one
in which the restriction on the interpretation of the subject pronoun in
,1 ) appears as a consequence rather than as a cause. The solution I will
argue for rests on Kiparsky's notion of blocking applied in the realm of
subcategorization, as well as on lexico-semantic considerations involving
the semantic relations that obtain between the arguments of a verb in virtue
of its meaning. Under the analysis presented in Section 3 the phenomenon
of obviation is crucially connected to control.
I wish to thank Larry Horn, Nirit Kadmon, John Goldsmith, Nicolas Ruwet, and Draga
Zec for useful comments and criticisms of earlier drafts. Earlier versions of this paper
were presented at Stanford University, University of Illinois at Urbana-Champaign, Yale
University, and at the 1988 winter LSA meeting.
86 / DONKA F. FARKAS
2 Previous solutions
2.1 The Pragmatic Approach
The simplest possible way of disposing of the problem illustrated in ( 1)
would be to invoke the Chomskian "avoid pronoun" commandment (or the
more general Gricean "be as concise a possible" maxim) , given that the
infnitive is used instead of the subjunctive in examples like ( 1 ) in Fench,
as well a in Spanish, Catalan and Portuguese. (Throughout this paper I
will use Fench for illustrative purposes. )
(2) Pierre veut partir.
P. wants to leave
As noted by Picallo ( 1985) , the immediate problem with a simple-minded
"avoid pronoun" approach is that the infnitive does not always block the
use of a fnite form in these languages. The verb promettre 'to promise' ,
for instance, may take both infnitive and indicative complements:
(3)
{ qu'ili
/
j part ira.
Pierrei
promis
that he
f
j will leave.
Pi proI
l
sed
de partIr.
Prep. leave.INF
Since in Romance languages all obviative complements are in the subjunc
tive, one could dispute the relevance of (3) by claiming that the efect of the
"avoid pronoun" commandment is somehow restricted to subjunctive com
plements. As shown in (4), however, there are subjunctive complements
which are nonobviative although an infnitive alternant is possible as well:
(4)
a.
{qU'ili s'en aille.
J' ai propose i Jeani
that hei leave.SUB
I have su
ggested to J'i
de s'en aller.
Prep. leave.INF
b.
{que je n'aie pas pu te voir.
Je regrette
that I not have.SUBJ been able to see you
I regret
de ne pas avoir pu te voir.
Prep. not have.INF been able to see you
I will nonetheless argue below that the ungrammaticality of ( 1 ) is indeed
connected to the grammaticality of ( 2) ; the correct solution has to account
both for this connection and for the data in (3) and (4). (A more complex
"avoid pronoun" solution will be discussed below. )
Finally, note that, although (5b) i s slightly worse than (5a),
(5) a. * Je veux que je parte.
I want that I leave. SUB
b.*Ili veut qu'ili parte.
Hei wants that hei leave. SUB
ON OBVIATION / 87
one cannot explain the problem away by invoking the Gricean "avoid am
biguity" maxim either, since there is no danger of ambiguity in (5a).
2. 2 The BT Approach
Researchers working in the GB famework have been exploring the attrac
tive possibility of assimilating the phenomenon illustrated in ( 1) to the
well-known case in (6) , covered by principle B of Binding Theory, which
forbids a pronominal fom selecting an antecedent within its binding do
main.
(6) * Johni saw himi .
In order to implement this idea, one has to analyze obviative subjunctive
clauses as not constituting a binding domain as far a their subjects are
concerned. Since obviative clauses in Romance languages are a subset of
subjunctive complements, one ha to fnd some special characteristic of
this subset which distinguishes its members from nonobviative subjunctive
complements as well as from indicative ones, and the boundaries of binding
domains have to be made sensitive to this characteristic. If the analysis is
correct, the characteristic in question should be connected to obviation in
other languages as well.
PicaBo (1984, 1985) , Meireles and Raposo ( 1983) , and Salamanca ( 1981 )
suggest that the distinctive feature of obviative clauses is located i n the
Tense component of their I(nfection) node. (AGR, the other component of
I cannot be responsible for obviation because its properties are the same
for all subjunctive clauses, and therefore it cannot be used to isolate any
subset thereof. ) The main idea behind all of these proposals i that the
Tense of obviative clauses is somehow dependent on the Tense feature of
the immediately higher clause, and it is this dependency that the defnition
of binding domains is sensitive to.
In order for any solution based on this idea to work, the set of obvia
tive clauses must turn out to be coextensive with the set of clauses whose
Tense is dependent. Suner and Padilla-Rivera ( 1984) and Zaring (1985,
chap. II) , convincingly argue that there is no notion of "dependent" Tense
which isolates the right class of clauses for Spanish, Fench, and Brazilian
Portuguese. Sequence of tense restrictions, for instance, hold not only in
obviative clauses but also in some nonobviative subjunctive clauses and
in certain indicatives. Strengthening the requirement to subsequent time
reference does not work for the same reason. Both complement clauses in
(7) have subsequent time reference and yet they are nonobviative.
(7) a. Mariei promet a Paul qu'ellei partira tot.
M' i promises P. that shei will leave soon
b. J'ai propose au professeur que je fasse l'experience moi-meme.
I proposed to the professor that I perform. SUB the experiment
myself.
The independently detectable properties of Tense in obviative clauses in
Romance languages do not distinguish them fom nonobviative ones and
therefore obviation cannot be a consequence of any of these properties.
Before abandoning this approach I will discuss two further empirical
challenges to it. The frst comes fom Romanian, where subjunctive clauses
have semantic and morphological properties similar to those in other Ro
mance languages (except for the fact that the subjunctive is used in a wider
range of complements, a point we will return to below). Crucially, however,
subjunctive clauses are not obviative:
(8) Ion; vrea [ei
/
j sa pIece] .
Ioni wants [Proi
/
j SUBJ leave]
(The 'e' above marks the dropped pronominal subject. )
l
Now the Romanian facts come as a complete surprise under any BT ap
proach relying on Tense because there is no independently justifed reason
for treating Tense in (8) as essentially diferent fom the Tense in ( 1) .
The next problem i s brought up in Ruwet 1984, where it is noted that
the strength of obviation in Fench depends on the semantics of both the
matrix and the obviative clauses. The relevant generalization is that obvi
ation is weakened if the degree of "agentivity" of the subordinate or main
clause subject decreases:
(9) a. Je veux que je puisse partir.
I want that I can. SUB leave
b. Je veux que je sois autorise a partir tot.
I want that I be.SUB authorized Prep. leave early
c. Je veux que je guerisse aussi vite que possible.
I want that I recover. SUB as soon as possible
d. Je voudrais bien que je parte tot.
I would like it well that I leave.SUB early
of agentivity is the same as the notion of responsibility, (RESP) , which,
as argued in Farkas 1988, is involved in controller choice with verbs like
promise and persuade. Following Farkas 1988, I will asume that the RESP
relation obtains between an individual and a situation if the individual
brings the situation about. The sentences in (9a)-(9c) difer fom ( 1) in
that the embedded subject is understood as not being responsible over the
situation denoted by the complement. The switch to the conditional and
the addition of the adverb in (9d) afect the responsibility of the matrix
IThe fact that an overt pronoun is not possible in (8) under the coreferential reading
is not specifc to subjunctive clauses:
(i) Ioni tie ca ei
j j
/el
j
f
*i e detept.
Ion knows that he is clever.
The problem of obviation U defned above is connected to the alternation between null
and overt pronouns but the two issues are distinct and I will not deal with the latter
here. For discussion, see Montalbetti 1984 and Levinson 1987.
ON OBVIATION / 89
subject over the complement situation. Thus, one would say (9d) rather
than ( 1) just in case it is not up to the speaker whether he leaves or not.
Ruwet notes that, while judgments difer with respect to the acceptability
of examples like those in (9) , they are all signifcantly better than ( 1 ) .
He further notes that by removing responsibility from both subjects, one
arrives at practically perfect sentences, like the example in ( 10) .
( 10) Je voudrais bien que je puisse partir des aujourd'hui.
I would like well that I can. SUB leave aleady today
Ruwet's data are incompatible with any solution based on Tense since there
is no plausible way of making the contents of the Tense of the embedded
clause depend on the factors discussed above. As Ruwet correctly argues,
these data cannot be handled by any purely confgurational account.
An altogether diferent BT solution is proposed in Kempchinsky 1985.
She assumes that subjunctive clauses contain a subjunctive operator which
moves to their COMP at LF just in case the matrix verb subcategorizes for
the subjunctive clause. In Spanish or Fench this means that the whole
I node moves to COMPo As a result, the embedded clause is no longer a
binding domain for its subject (because it no longer contains a governor of
the subject) . Kempchinsky notes that the Romanian data are explained
under the assumption that the subjunctive operator in this language is not
under I and therefore its movement to COMP does not afect the binding
domain of the subject, which is governed by I .
The crucial ingredient of this analysis is the movement of the subjunc
tive operator to COMP at LF. This movement is supposed to be trig
gered by the fact that subcategorization requirements must be met at LF
and therefore, if a verb subcategorizes for a subjunctive complement, the
head of the complement must fulfll the subcategorization requirement.
Assuming that COMP is the head of S' , this means that the COMP of sub
categorized subjunctive complements must contain a subjunctive element
at LF.
The main shortcoming of this solution is that it incorrectly predicts
that the class of obviative complements is coextensive with the class of sub
categorized subjunctive complements in languages like Spanish or French.
There are, however, plenty of subcategorized subjunctive complements in
these languages which are nonobviative, such as the complements of neg
ative epistemic verbs ( nier 'deny' , douter 'doubt' ) or of factive-emotives
( tre heureux, 'be happy', regretter 'regret' ) . Note also that Ruwet's data
constitute just as much of a problem for this analysis as for the one dis
cussed previously. I therefore conclude that the BT solutions proposed
so far are inadequate because they fail to predict the class of obviative
complements in Romance languages and because they fail to account for
the role the semantic factors discussed by Ruwet play in obviation. These
shortcomings are shared by the pragmatic approach as well.
Yet another approach is proposed in Bouchard 1982, 1983. Bouchard
takes obviation to be the manifestation of his own version of "avoid pro
noun," suggestively called the Elsewhere Principle. According to it, a pro
noun cannot be used in a position where an anaphor is possible, i. e. , in
a position where the pronoun will be interpreted as coreferential with an
NP that govern it.
2
Hermon ( 1985) , in discussing obviation in Imbabura
Quechua, proposes a version of Bouchard's Elsewhere Principle where the
presence or absence of [AGR] is the crucial factor in determining whether
PRO is an anaphor or a pronominal. In order to explain why only subjunc
tive clauses are obviative, Bouchard makes reference to the fact that both
subjunctives and infnitives have "unrealized tense" (see Bresnan 1972) . It
is not made clear, however, how and why this property comes to play a role
in obviation, and therefore there is no account of why promettre ' promise',
for instance, takes nonobviative complements, although their tense is also
unrealized.
Picallo (1985) correctly notes that Bouchard's solution fails to account
for the fact that in Fench, as well a in Catalan, obviation is BU(bject)
oriented, i. e. , its efect is weakened if the matrix argument involved is not
the Bu, a in (4a) or ( 1 1 ) :
( 1 1 ) ?Marie a convaincu Pauli qU'ili s'en aille. /de s'en aller.
M. ha convinced P' i that hei leave.SUBJ jPrep. leave.INF
Finally, note that this solution does not fare any better than the previ
ous ones with respect to the data discussed in Ruwet 1984. Hermon's ver
sion, if extended to Romance languages, would have the additional problem
of positing an abstract and unmotivated [AGR] diference between obviative
and nonobviative subjunctive clauses.
3 Towards a Solution
3. 1 Data To Be Accounted For
Below I will summarize the main facts discussed so far and bring in
some new data fom two non-Romance languages, Hungarian and Serbo
Croatian.
Within Romance
(i) We have seen above that in those Romance languages that have them,
obviative clauses are a subset of subjunctive complements. We therefore
have to account for the fact that indicative clauses are nonobviative (recall
the promettre 'promise' / vouloir 'want' contrast) . We also have to dis
tinguish obviative clauses fom nonobviative subjunctives (recall the nier
'deny' , regretter 'regret' jvouloir 'want' contrat a well as the data dis
cussed by Ruwet) .
2
For Bouchard PRO is an anaphor when referentially dependent on a governing NP and
a pronominal otherwise.
ON OBVIATION / 91
(ii) We also have to give an account of the diference between Roma
nian and the other Romance languages. Such an account should link the
non-existence of obviation to some other feature of Romanian, which dis
tinguishes it fom its better known Romance sisters.
(iii) Next, we have to account for the fact that obviation tends to be Su
oriented, Le. , the argument that the complement subject must be disjoint in
reference fom tends to be the matrix subject. Recall the contrast between
( 1 ) and (4a) or ( 11) . 3
(iv) Finally, one has to predict the locality of the phenomenon, i. e. , one
ha to predict that the two NPs that must be disjoint in reference are in
adjacent clauses:
( 12) Mariei souhaite que Jeannej veuille qu'ellei
/
*j parte.
Marie wishes that Jeanne want.SUBJ that she leave.SUBJ
Beyond Romance
The correct account of obviation for Romance languages should extend to
other languages a well. Below, we will consider the relevant data in Serbo
Croatian, which is a Balkan language, and Hungarian, which i neither
Balkan nor Romance.
In what follows, the complements of verbs like convince, try, or begin,
whose Su arguments are necessarily referentially dependent on a matrix NP,
will be called dependent Su clauses. The term accidental Su dependency
will be used for caes where the subject of a complement happens to be
referentially dependent on a matrix argument.
Hungarian
Hungarian ha both infnitive and subjunctive complements. The infnitive
i the only possible complement form in dependent Su clauses when the
controller is the matrix subject:
( 13)
{ elkezdett beszel-ni.
Janos
started talk-INF
J.
igyekszik kedves len-ni.
tries nice be-INF
When the controller is a non- Su, verbs subcategorize for either infnitive
or subjunctive complements:
( 14) a.
{ hogy ej
/
*i, *k men-jen vele.
Janosi
eggyozte Maritj ,
that (she) go-SUB with him.4
J. convInced M.
*men-ni vele.
go-INF with him
3
Subject orientation is only a tendency, rather than a clear cut phenomenon because
judgments concerning fnite clauses whose subjects are coreferential with a non-Stl ma
trix argument and which have an infnitive counterpart show considerable variation from
speaker to speaker and from verb to verb.
b.
{ vasarol-ni.
Janos elkuldte Mariti
shopINF
J. sent M.
*hogy ei vasarol-jon. 5
that (she) shop-SUBJ
In cae of accidental Su dependency, the complement is either in the infni
tive or in the subjunctive and is obviative, i. e. , the subject of the subjunc
tive complement is disjoint in reference fom the matrix subject:
( 15) Janos moziba akar men-ni.
J. movie-into wants go-INF
Janosi azt akarja, hogy e.i
f
j moziba men-jen.
J.
it-ACC wants that (he/she) movies-into go-SUB
Not all subjunctive complements are obviative, however. The verb kovetel
'demand' , for instance, is like akar 'want' except that it takes only sub
j unctive complements. As shown in (16), its complement is nonobviative:
( 16)
{ k6veteli, hogy ei
f
j kapjon t6bb etelt.
Janosi
demands that (he/she) get-SUB more food
J.
*k6vetel kap-ni t6bb etelt.
demands get-INF more food
Obviation in Hungarian is infuenced by the following factors. First, in
certain contexts in which the whole complement is focused an otherwise
obviative clause becomes nonobviative:
( 1 7) Ha az-t akarod, hogy velunk gyere,
if that-ACC want that us-with come. SUBJ,
viselked-j szepen.
behave-SUBJ well
If you want to come with us, behave well.6
In this cae the demonstrative pronoun az-t 'that-ACC' , whose antecedent
is the clause, has to be used. The exact syntactic analysis of the construc
tion exemplifed above is immaterial for present purposes. What matters
is that this construction is not possible with infnitival complements:
4
Hungarian is a Su pro drop language too.
5
Examples like ( 14b) were brought to my attention by A. Szabolcsi. It is not clear,
however, that the relation between the infnitive and the direct object of the matrix
in ( 14b) is the same a in obligatory control caes but I will not pursue this issue any
further here. The subjunctive in ( 14b) is grammatical under an interpretation in which
the complement is understood a a purpose clause.
6
Thanks again to A. Szabolcsi for bringing this type of examples to my attention.
Exatly which contexts allow a focused nonobviative subjunctive clause u not clear to
me at present. The mere presence of the pronoun UZ 'that' appears to be insufcient,
since (i) is bad, or at leat signifcantly worse than ( 17).
(i) * Janosi a-t akarja, hogy ei jojjon veliink.
J. that-ACC wants that (hei) come.SUBJ with us.
( 18) * Janos azt akarja velunk jon-ni.
J. that wants with us come-INF
ON OBVIATION / 93
Obviative complements become nonobviative also if the complement sub
ject is focused. In this case the subject must occur immediately before the
subjunctive verb and it receives special stress, indicated by capitals.
(19) JanOSi (azt) akarja, hogy
O
i menjen Marival (es nem Laszlo).
Ji (it) wants that hei go.SUB M. with (and not L. )
Serbo-Croatian
The relevant facts in Serbo-Croatian (Belgrade dialect) can be summarized
as follows. In case the complement subject is obligatorily controlled by the
matrix subject or direct object, the complement form is a so-called "da2
complement" whose subject must be null. (Serbo-Croatian is a Su pro drop
language as well. )
(20) a. Petari je pokusao da ei
/
*j dodje.
b.
P. Aux tried Comp (he) come
*Petari je pokusao da oni dodje.
P. Aux tried Comp he comes
Ana je naterala Marijui
Comp (she)
.
come
{ da ei
/
*k dodje
A. Aux forced M i
*da onai dodJe.
Comp shei come
"Da2 complements" in Serbo-Croatian are not formally distinguished fom
so-called "dal complements" with respect to either the complementizer or
verbal infection. However, da2 complements are special in that they obey
tense restrictions and have modal connotations characteristic of subjunctive
complements (see Zec 1987 and references quoted therein).
Now in case the subject of a da2 complement is accidentally dependent
on some NP in the matrix, it must necessarily be null:
(21)
{ da ei
f
j dodje.
Petari je zeleo
Comp (hei
f
j)
.
come
Pi Aux wants
da Onj
/
*i dodJe.
Comp hej
/
*i come
Obviation in Serbo-Croatian reduces then to the null/overt pronoun choice
in da2 complements. In case the complement subject is accidentally depen
dent in reference the null pronoun must be chosen, i. e. , the overt pronoun
is obviative.
7
The data below come from Zec 1987.
SIn cae the matrix verb is an auxiliary, the infnitive is used. The use of a finite verb
form in obligatorily controlled complements is a Balkan trait that the Belgrade dialect
of SerboCroatian shares with Romanian.
In Serbo-Croatian, just like in Hungarian, the obviation efect disap
pears in cae the complement subject is emphatically stressed: the exam
ples in ( 20b) and (21) lose their stars if there is emphatic stress on the
complement subject.
Since the facts from Hungarian and Serbo-Croatian are strongly remi
niscent of the Romance obviation data, we should seek a unitary account.
The main new facts concern the efect of emphasis on obviation and the
fact that in Serbo-Croatian obviation reduces to the contrast between a
null and an overt pronoun.
3. 2 The "Blocking" Approach
I suggest that we should view obviation as the manifestation of a constraint
on the use of certain complement forms rather than as a problem pertaining
to pronoun use or interpretation.9 According to this claim, in Romance
languages and in Hungarian obviation is due to the old rivalry between the
subjunctive and the infnitive. Under this proposal, ( 1) is ungrammatical
not because a pronoun selects its antecedent in the wrong domain but
rather because a subjunctive complement is used instead of an infnitive
one. This claim is diferent from Bouchard's in that it does not refer to
the choice between a pronominal and an anaphor directly. Recall now
that in Serbo-Croatian the two rival complement forms difer only with
respect to whether the subject is obligatorily null or not. What is common
to the innitive in Romance and Hungarian, and da2 complements with
null subjects in Serbo-Croatian is that they are the complement forms
used in dependent Su clauses. We can then start from the preliminary
generalization in (I):
(I) The complement form used to mark dependent Su clauses blocks the
use of the subjunctive under certain circumstances. 10
The term 'blocking' is used here in the sense used in lexical semantics and
morphology. Thus, Horn (1984) , in discussing semantic relationships be
tween items in the lexicon, notes that "the meaning, use, or very existence
of a given word or expression is afected by the existence and range of a
related and more baic or specifc entry in the lexicon" (p. 111) . (For rel
evant work in lexical semantics, see McCawley 1977 and Horn 1978, and
in morphology, Aronof 1976 and Kiparsky 1983. ) The existence of a more
specifc lexical item or construction is said to block the use of the more
general one in cases where a choice is theoretically possible. My claim
is that the same phenomenon is responsible for the obviation efect dis
cussed here, where the two competing expressions are complement forms.
In case the more specifc complement form, the one that marks Su depen-
9This suggestion is in the spirit of the pragmatic approah outlined in Section 2. 1. The
diferences will become clear below. Under the present proposal the term 'obviation' u a
misnomer for the phenomenon exemplifed by (1) but I will continue to use it nonetheless.
lOThe term 'subjunctive' is used loosely here. It will be eliminated below.
ON OBVIATION / 95
dence, can be used the use of the less specifc form, the subjunctive, is not
allowed.
Below I will frst point out the immediate advantages of this approach; I
will then briefy comment on the subjunctive/infinitive rivaly, and fnally,
I will turn to the mysterious "certain circumstances" in (I).
Viewing obviation a one complement form (the infnitive) blocking
another ( the subjunctive) allows us to correctly predict that subjunctive
clauses will be obviative only in contexts in which an infnitive is also
possible, i. e. , that obviative clauses are a subset of those subjunctive clauses
which can in principle alternate with innitives. The following facts are
thus accounted for:
(i) Romanian subjunctive clauses are nonobviative because in this lan
guage the complement form used to mark Su dependency is the sub
j unctive.
(22) loni a incercat ei
/
*j sa pIece.
I. tried SUBJ leavel l
loni a fortat-o pe Mariaj ej
/
*i, *k sa pIece.
I. forced M. SUBJ leave
The diference between Romanian and the other Romance languages con
cerning obviation is thus connected to the diference in the role played
by the infnitive. The relevant diference between Romanian and Serbo
Croatian is that in Romanian the null pronoun is more widely used than
in the Belgrade dialect of Serba-Croatian and therefore there is no spe
cial complement form indicating Su dependency and no signifcant difer
ence between indicative and subjunctive complements with respect to the
null/ overt subject pronoun contrat. 12
(ii) The complement of kovetel 'demand' in Hungarian is nonobvia
tive because this verb, unlike akar 'want', takes only subjunctive comple
ments. Subjunctive complements controlled by non-subjects in Hungarian
are nonobviative for the same reason.
(iii) The disappearance of obviation under emphasis in Hungarian and
Serba-Croatian is also explained. In Hungarian, in certain contrative con
texts illustrated in (17) in which the complement may not be infnitive, the
subjunctive is nonobviative. Note also that in both Hungarian and Serbo
Croatian emphatic subjects are nonobviative because they may not be null.
So far, the blocking approach shares advantages with the "avoid pro
noun" solution, except that the former can account for obviation in Serbo
Croatian while the latter cannot. The problem raised by Serbo-Croatian
1
1 For details see Joseph 1983 and Farka 1985b.
1
2
Comorovski (1985) finds that the "avoid pronoun" principle seems stronger in sub
junctive complements than in indicative ones. She connects this diference to the fact
that the use of an overt subject in subjunctive complements triggers the use of the
complementizer U well.
for the "avoid pronoun" approach is to explain why the force of the com
mandment is so much stronger in obviative clauses than in nonobviative
ones. Of couse, we have not accounted for the Serbo-Croatian data ei
ther until the term 'subjunctive' is eliminated in (I) and until the "certain
circumstances" are specifed.
The fact that indicative clauses are nonobviative is stipulated in (I) .
Below I will ofer some justifcation for treating infnitive and subjunctive
complements in a clas that contrasts with indicative complements.
3.3 On the Infnitive/Subjunctive Rivalry
It wa proposed above that obviation in Romance and Hungarian should
be accounted for by claiming that the possibility of using an infnitive com
plement blocks the use of a subjunctive one under certain circumstances.
We have also seen that the infnitive does not block the indicative in these
languages. I will now attempt to justify this contrast by arguing that cer
tain subjunctive complements share with infnitive complements semantic
characteristics that set them apart fom indicative complements. Under
the assumption that we have the same type of blocking here as lexical
blocking, which requires the two forms involved to be semantically close,
we have some explanation for why the infnitive blocks the subjunctive but
not the indicative.
In what follows, I will assume that sentences, as well as infnitival
complements, can be used to denote situations, i. e. , individuals having
properties and standing in various relations at various spatio-temporal
locations. 1
3
I will also assume that verbs are lexically associated with
predicate-argument structures, where the arguments identif participants
in the situations denoted by clauses involving the predicate in question.
The arguments of the predicate end up being associated with syntactic con
stituents. The details of this association are irrelevant for present purposes.
Now in the case of complex sentences, one of the arguments of the
matrix predicate, namely the one linked to the complement, denotes a
situation. The situation denoted by the complement (se) is a constituent
of the situation denoted by the matrix (Sm) . The matrix predicate may
impose various restrictions on Se as part of its lexical meaning since Se is one
of its arguments. These restrictions may involve the immediate constituent
parts of Se, i. e. , the participants in Se, as well as its spatiotemporal location.
The more such restrictions are imposed on Se, the more dependent it is on
S
m. The parameters of semantic dependency that are most consequential
for linguistic phenomena are (i) Su dependency and (ii) world dependency.
The most important consequence of semantic dependency is its correlation
with the use of moods. 14 Below, I will briefy discuss each dependency
1
3
The term 'situation' is borrowed from Barwise and Perry 1983.
1
4Noonan (1985) and Ransom ( 1986) discuss the same question from diferent perspec-
ON OBVIATION / 97
parameter and in the process, I will argue that a third parameter, time
reference dependency is less basic.
A. Su Dependency
There are predicates which require that the "external" argument of 8, i.e. ,
the argument that would be linked to the subject of the complement, be
necessarily referentially dependent on some argument of the matrix pred
icate. This means that these arguments are referentially dependent on
some participant in @. 1 5 Complements realizing such arguments have
been called here dependent Su complements. Examples of verbs requiring
dependent Su complements are "aspectuals," such as begin, start, try, and
'directives', such as persuade or convince.
Note now that it is frequent for languages to use a special comple
ment form to indicate Su dependency, a form in which the Su is not ex
pressed. In languages that have infnitival complements, they are crucially
connected to Su dependency. These languages use infnitival complements
in at least some dependent Su complements. Concerning the languages
discussed here, in Western Romance the verb form in dependent Su com
plements is the infnitive and there are no infnitival complements whose
Su argument is fee in reference. The situation is similar in Hungarian, the
major diference being that in Hungarian the infnitive is exclusively used
only in Su dependent clauses controlled by a subject; 16 recall that if the
controller is a non-subject the subjunctive may be the only possibility (cf.
( 14a) ) . 1 7 Note also that the time reference of dependent Su complements is
also dependent on that of the matrix but there are complements whose Su
reference is fee and whose time reference is dependent, such as the com
plements of promise, or foresee. The claim that the infnitive is associated
with Su dependency rather than time reference dependency is justifed by
the fact that these verbs may always take indicative complements. There
tives. They both treat time-reference dependency as basic and do not discuss world
reference dependency at all.
1
5
The two arguments are nevertheless diferent qua arguments, each being constrained
by its own predicate.
1
6The diference is actually less sharp: in dependent Su complements where the con
troller is a non-subject the subjunctive is not ruled out in Western Rmance either, a
illustrated in (i):
(i)
?J'ai c
nvaincu Marie
that she come.SUBJ
{ qu'elle vienne.
I convInced M.
de venir.
Prep. come.INF
1
7
Nothing in what follows depends crucially on the syntactic status of complements and
therefore the remarks below are compatible with a syntactic treatment in which at leat
some of these complements are VPs, a well a with one in which they are all sentential.
An important question which I will not address here is how to ensure that only subjects
(i.e., external arguments) can be controlled.
are no languages that use the infnitive just in cae Se is dependent in time
reference and fee in Su reference.
B. World Dependency
In order to introduce the notion of world dependency two more general
assumptions have to be stated. First, unlike in "situation semantics," I
asume that collections of situations, called "worlds" play a central role
in the interpretation of sentences. A sentence is true in a world w if the
situation it denotes is part of w. Second, the possible worlds framework
asumed here difers fom most in that partial worlds are admitted. (For a
discussion of partiality, see Landman 1986.) Thus, a sentence may be true,
false or undecided in a particular world.
In the case of verbs taking situation-denoting complements, it often is
the case that the world in which the complement is evaluated is diferent
fom the world in which the matrix is evaluated. Thus, it may well be the
case that (23a) is true in the real world (WR) , while (23b) is not.
(23) a. John thinks that Mary is home.
b. Mary is home.
What is required, however, is for (23b) to be true as far as John is con
cerned, i. e. , to be true in what John takes Reality to be. If know is substi
tuted for think, (23b) has to be true in both worlds.
What world the complement is interpreted in depends on the lexical
meaning of the matrix predicate, which is said to "introduce" (or "create" )
this world. In what follows the introduced world will be called "the new
world," while the world in which the matrix sentence is evaluated will be
called "the base world. " In the cases we will consider here the base world
will always be WR. The type of semantic dependency under discussion has
to do with the properties of the new world.
I will briefy argue now that there are two distinct ways of introducing
a world. One is to introduce a world whose situations are independent
of the base-world or any other world in the sense that the new world is
not asumed to inherit any parts fom any world. I will call such worlds
independent. Another way of introducing a world is to, in a sense, modif
another world, i. e. , introduce a world U
t
that is just like some other world
U, except that a particular situation, Se holds in w' . I will say that in this
case U
t
inherits fom w all those situations that do not confict with Se and
I will say that w' is dependent on w. One gets fom w to w' by adding Se
to w and performing all the accommodating changes required to maintain
consistency. I will call 8 in such cases a highlighted situation for which U
t
is a background world.
Independent Worlds
As an example of independent worlds, consider the worlds introduced by
"cognitive" verbs such as think, believe, and know, exemplifed in (23a).
ON OBVIATION / 99
(The former two are taken here in their "strong" sense, in which the referent
of their Su takes the complement to be true. ) The base world in this cae is
WR and the new world is one that represents Reality as John sees it, a world
that I will denote by WR(J). In what follows, I will call the referent of the
(active) subjects of predicates introducing worlds anchors of those worlds.
If instead of think we had know, the world introduced would be part of
WR(J). The point that is of interest here is that the world introduced by
these verbs is independent of the base-world or any other world, in the
sense that all we know about the new world on the basis of sentences like
(23) i that Se holds in them. There are no situations in the base world, or
any other world, that the new world inherits. Other verbs that introduce
independent worlds are dream, imagine, fantasize. The world introduced
by them is not Reality anchored to the referent of their subject but is a
world that plays that role under limited circumstances. The point again is
that these worlds are independent. No situations fom WR or wR(a), where
a is the anchor, are inherited by these worlds. 18
Dependent Worlds
Things are quite diferent in the cae of desideratives like want, wish, and
desire, exemplifed in (24) .
(24) John wants Mary to leave.
As noted in McCawley 1977, one cannot speak of a single buletic alterna
tive introduced by want since one may hold contradictory wishes simul
taneously without thereby being guilty of logical inconsistency. (This, of
course, i not so with cognitives. ) I will therefore assume that desidera
tives introduce a set of buletic alternatives anchored to the referent of their
subject, and that their complement holds D some world of this set. The
point that is of relevance here is that in this world Se is a highlighted sit
uation whose background is wR(a), i.e. , the world in which Se holds is one
where Se is added to wR(a). Viewing things this way helps us account for
several semantic properties of desideratives, which distinguish them fom
cognitives.
First, note that (25a) does not entail (25b) even if John is flly aware
that being away fom one's wife entails having a wife.
(25) a. John wants to be away fom his wife.
b. John wants to have a wife.
The highlighted situations are diferent in the two caes and therefore the
two worlds are diferent. In (25a) the situation added to WR(J) is one in
which John is away fom his wife, while the situation added to WR(J) in
the case of (25b) is one in which John has a wife.
l
8
The question of inheritance is independent from the pragmatic asumption, mentioned
in McCawley 1977 and discussed at length in Fauconnier 1985, according to which worlds
are not supposed to difer in relevant respects, unless explicitly stated.
Second, having a background world and a highlighted situation helps
us account for the role Se plays in specifying the buletic alternative intro
duced by the verb. Thus, note that the buletic alternative introduced by a
desiderative cannot be just any world in which Se holds, i. e. , the realization
of Se cannot count as a sufcient condition for a world to be a buletic al
ternative of the anchor's. Presumably, a world in which John cannot bear
to be separated from his wife and in which he is away fom her is not a
buletic alternative of John's even if (25a) is true. The buletic alternative
introduced in (25a) is one where John's marriage is the way John thinks it
is in Reality. As mentioned before, realizing Se cannot count as a necessary
condition for a world to be a buletic alternative of the anchor's because one
may hold two contradictory wishes simultaneously without wanting them
to be realized simultaneously.
Thus, if (26) is true of John, there is a buletic alternative of John's in
which he is a lawyer and not a marine, and another, in which he is a marine
and not a lawyer.
(26) John wants to become a lawyer but if he can't get into a good school,
he wants to join the marines.
The two alternatives have diferent background worlds. The background
world of the frst is WR( J) , while that of the latter is one in which the
situation denoted by the if clause has been added to WR(J) .
I will end the discussion of desideratives by drawing attention to a
diference between want and wish concerning the nature of Se. Consider
the contrat in (27):
(27) a. I want it to be Friday.
b. I wish it were Fiday.
said by an ordinary person in our ordinary world. I suggest that (28a) is
strange because want imposes a further restriction on Se, namely that it
should be possible, for all that the anchor knows, for Se to be realized and
therefore Se cannot be counterfactual as far as the anchor is concerned. The
fact that it cannot be factual either follows fom pragmatic considerations,
since if it were, the conditions would be met for the use of an emotive
factive. The verb wish does not impose this requirement on Se. Note
that this diference correlates with time reference restrictions, since one
can know how things turned out in the past but usually one can only
conjecture about the future. The complement of want is acceptable with
past time reference in English, at leat for some people, if the anchor does
not know whether the situation it denotes is factual or not.
There is cross-linguistic variation with respect to the details of the con
straints particular desideratives impose on Se. For Fench, Ruwet ( 1984)
claims that the use of vouloir 'want' in the indicative implies that the an
chor is ready to do what it takes in order to bring Se about, while souhaiter
ON OBVIATION / 101
'wish' implies that the anchor's attitude towards the fulfllment of his wish
is passive.
"Directives," such as ask, order, demand, request can be analyzed as
similar to want except that the person responsible for bringing about Sc
is the individual denoted by the internal argument of the verb (the direct
object or the object of the preposition of or from) .
With respect t o time dependency, we have seen that it is associated
with Su dependency and that it may accompany world dependency as
well, as a result of a constraint on the realizability of Sc. We have already
seen that not all world-dependent complements have dependent time ref
erence (cf. wish) . Conversely, there are complements whose time reference
is dependent but which are neither world nor Su dependent. Consider, for
instance, the complements of promise and foresee. Their time reference
is dependent: it must necessarily be non-anterior to the time reference of
the matrix. As noted above, they are not, however, dependent Su com
plements and they do not introduce a new world, and therefore certainly
not a dependent one. Their complements are supposed to hold in the base
world at some non-anterior time. Promise difers from the other two in
that it imposes the realizability condition on the referent of its (active)
subject and sc, and therefore these two arguments must be in the RESP
relation.
Going back to the main point of all this, namely the subjunctive/in
finitive rivalry, note that world dependency, just like Su dependency, and
unlike time reference dependency, is associated with the use of a special
"non-indicative" mood. Thus, languages that have mood distinctions use
a non-indicative mood in world-dependent complements. In languages
that have a subjunctive, this is the mood used to mark world depen
dency, except, of course, under special circumstances, when the infnitive
blocks it. 19
Common to the infinitive and the subjunctive then, is that they mark
semantic dependency. The "rivalry" between them manifests itself in the
tendency of one "dependent" mood to take over the rightful domain of
the other. The existence of the rivalry is not surprising in view of the
fact that the domains of the two moods often overlap, as for instance in
the case of a verb like try or in the case of directives that take depen
dent Su complements, like persuade. Thus, we have already noted that in
Balkan languages the subjunctive has supplanted the infinitive to a large
extent, and therefore the subjunctive has come to be used to mark not only
world dependency but subject dependency as well. In English the reverse
has happened: the subjunctive has almost disappeared, its place having
19
Marking world dependency is only one of the uses the subjunctive may be put to.
Thus, in Fench it is used to mark subordination, in preposed complements, as well a
non-asertiveness, in emotive-fatives (cf. Palmer 1986).
been taken by the infnitive. Consequently, we have infnitive complements
whose subject reference is fee, as in (28):
(28) John wants ( for) Mary to leave.
Recall also that in Hungarian and Western Romance the situation is mixed,
the infnitive marking Su dependency and the subjunctive marking world
dependency. In cases of overlap, the infnitive tends to win in cae of
Su control, while the subjunctive has at least a chance in case of non
Su control. (In Hungarian in these cases verbs select either the sub
junctive or the infnitive, while in Western Romance both complement
types are possible. ) The cae of Serbo-Croatian is diferent since there
only Su dependency is formally marked (by an obligatorily null subject
pronoun) ; recall that world-dependent complements ( da2 complements)
are only semantically distinguished fom independent ones ( dal comple
ments).
We are ready now to return to obviation. Languages which have ob
viative complements are languages which distinguish dependent Su com
plements fom world-dependent ones. The battleground between the two
forms is limited to those cases where the complement meets the semantic
requirements for both Su and world dependency, i. e. , Sc is world dependent
and the participant connected to its subject is referentially dependent on
some participant in Sm. The generalization in (I) can then be reformulated
a in (II) , which now covers Serbo-Croatian a well.
(II) In certain world-dependent complements, the form used to mark
subject dependency blocks the form used in world-dependent com
plements.
For Fench, for instance, (II) requires the choice of the infnitive over
the subjunctive in certain world-dependent complements, while in Serbo
Croatian the choice will be between an obligatorily null pronoun over a
potentially overt one. This gives the disjoint reference efect in both lan
guages. Blocking, of course, presupposes the existence of two complement
forms, and therefore no obviation efect exists in Romanian.
Note that calling this preference for the dependent Su construction
"blocking" is appropriate because this construction is semantically more
specifc than the world-dependent one. Recall that obviation manifests
itself in case of accidental Su dependency. Now the fact that the comple
ment is world dependent is predictable fom the meaning of the matrix
verb but the fact that its Su is referentially dependent is not. The use of
the dependent Su form in this case is therefore more informative than the
use of the world-dependent one. The less informative complement form
may be chosen only in case the conditions are not met for the use of the
more informative one, which results in the disjoint reference efect under
discussion here. Let us turn now to a less mysterious characterization of
the "certain" complements involved.
ON OBVIATION / 103
3.4 The Semantic Role of the Controller and of the
Controlled Argument
So far we have narrowed the range of obviation to world-dependent com
plements whose subjects are accidentally dependent on a matrix argument.
Obviation wa treated here a the result of one complement form blocking
another in these caes. Note, however, that nothing we have said so far
accounts for Ruwet's weakening factors. Recall that based on examples
such a (10), repeated here as (29),
(29) Je voudrais bien que je puisse partir des aujourd'huis.
I would like well that I can.SUBJ leave already today
Ruwet concludes that the semantic role of the two arguments involved
afects obviation. The remarks that follow are elaborations of ideas in
Sul er and Padilla-Rivera 1984, Zaring 1985 and Ruwet 1984.
We have seen above (cf. examples (9a)-(9c)) that obviation is weak
ened in cae the complement is passivized, or if it contains a modal or a
"non-agentive" main verb. Common to all these cases is the fact that the
controlled argument (the complement subject) is not seen as bringing Se
about, i.e., it is not in the RESP relation with the situation of which it is a
participant. The tendency of obviation to disappear when the complement
is pasivized, or when it contains a modal or a "non-agentive" main verb
leads me to conclude that the controlled argument must bear the RESP rela
tion to the situation denoted by the complement. This means that the con
trolled argument is the initiator of the situation in which it is a participant.
It wa also noted above that obviation is weakened if the controller is
not seen a being D the RESP relation with Se (cf. example (9d) ) . Relevant
here is Ruwet's observation that obviation is weaker in the complement of
souhaiter 'wih' than in that of vouloir 'want' . As mentioned in Section
3.3 above, Ruwet also notes that the Su argument of the latter is more
likely to be interpreted a being in the RESP relation with Se than the Su
argument of the former. Ruwet also notes that the complement of exiger
'demand' is not obviative. The relevant characteristic of this verb is that its
Su argument may not be in the RESP relation with Se. I conclude therefore
that the obviation efect is most robust if both controller and controlled
argument are in the RESP relation with Se, the former as a result of the
characteristics of Sm, the latter a a result of the characteristics of Se. (We
have here, of course, two distinct arguments of two distinct predicates even
though they happen to be referentially dependent. )
This situation is reminiscent of facts pertaining to controller choice,
discussed in Farkas 1988. It wa argued there that the unmarked controller
choice for infnitival complements falls on the matrix argument that bears
the RESP relation to Se, when the controlled argument bears this relation
to Be a well. This suggests that there is a correlation between the use
of dependent Su constructions and the RESP relation, which is the link
104 / DONJA F. FARKAS
between obviation and control. I suggest that there is a "canonical control
case," canonical for the use of a dependent Su complement. The semantic
characteristic of the canonical control case is that both the complement Su
and the matrix argument it is referentially dependent on bear the RESP
relation to sc. We expect then that if a language has a complement form
marking Su dependency, it will use it in complements conforming to this
characteristic and that if the characteristic is not met we will fnd cases in
which the dependent Su construction is not used. A case in point is the
Hungarian equivalent of exiger, kovetelni, which, as seen in (16) above, may
not take infnitival complements but only subjunctive ones. This property,
I claim, is not accidental but rather, it is the result of the fact that the
controller may not be in the RESP relation with sc.
lt is therefore not surprising that obviation is strongest in case the se
mantic characteristics of the "canonical control cae" are met, in view of
the fact that obviation is the result of preferring the dependent Su comple
ment form over some other complement form. The generalization in (II)
can now be rendered more precise:
(III) In world-dependent complements that conform to the canonical con
trol cae, the form used to mark Su dependency blocks the form used
for world dependency.
The generalization as formulated now accounts for Ruwet's weakening fac
tors. The proposal made here elaborates the suggestions in Ruwet 1984.
It difers fom both Suier and Padilla-Rivera 1984, and Zaring 1985. Ac
cording to Suner and Padilla-Rivera 1984 obviation is connected to the
semantic feature [+WILL] of the matrix predicate. In discussing this pro
posal, Zaring 1985 correctly notes that the role of the complement subject
is also relevant and proposes that a complement is obviative when the ma
trix subject is an "instigator" and the complement subject, an "instigatee."
Zaring does not, however, elaborate on the meaning of these terms. In par
ticular, it is not clear how to postulate a semantic relationship between the
matrix subject and the complement subject since they are not arguments
of the same predicate. In the analysis suggested above, the two subjects
are related by both bearing the RESP relation to the same situation, the
situation denoted by the complement.
The question to be addressed next is that of why the RESP relation
should play such a prominent role in control and obviation. The answer
is suggested by Ruwet ( 1984), who notes that when a matrix argument is
in what we call here the RESP relation to Sc there is a crucial diference
between the case when it is coreferential with the complement argument
bearing the RESP relation to the same situation and the case when it is
not. Let us call the matrix argument im (the matrix initiator) and the
complement argument ic (the complement initiator). If the two initiators
are referentially independent, im is indirectly responsible for Sc in that it
ON OBVIATION / 105
has to act on ie, who brings Se about. If the two initiators are referentially
dependent, on the other hand, the realization of Se depends on a single
individual and the ie has direct responsibility over Se. It is therefore not
surprising that languages will tend to mark this diference. What we have
called obviation here is one way of marking it. The generalization in (III)
then can be seen as a rule that limits the feedom allowed by subcatego
rization restrictions whose efect, besides limiting the number of ways one
can say the same thing, is to mark a relevant semantic diference.
3. 5 Locality and Su Orientation
There are two more characteristics of obviation that have to be accounted
for, locality and Su orientation.
Locality
Recall that the disjoint reference efect characteristic of obviation always
involves an argument of a matrix verb and the subject of its immediate
complement. This follows in our analysis fom the fact that obviation is
dependent on semantic characteristics of a matrix argument and Se, char
acteristics that are imposed by the lexical meaning of the matrix verb. It
follows therefore that Se must be an argument of that V since only the
immediate arguments of a verb may be afected by the verb's meaning.
Now imposing semantic constraints on a situation may involve imposing
semantic constraints on its participants. Further details of the situation in
question are unavailable.
Su Orientation
As noted before, Su orientation refers to the strong cross-linguistic ten
dency of obviation to disappear in cae the controller is not the subject
of the matrix clause. Recall that in Fench, vouloir 'want' contrasts with
persuader 'persuade' or convaincre 'convince' in that the complements of
the latter are nonobviative, (or less obviative) as shown in ( 11) , repeated
here as (30).
(30)
{qU'ili s'en aille.
?Marie a con
:
aincu Pauli
that hei leave.SUBJ
M. has convInced Pi
de s'en aller.
Prep. leave.lNF
The same appears to hold in Serbo-Croatian, as seen in ( 31).
(31)
da ei/j ostane na veceri.
Petari je molio Marijuj
P.i Au asked M.j
Comp (he/she) stay at dinner
*
da oni ostane na veceri.
Comp hei stay at dinner
?
*
da onaj ostane na veceri.
Comp shej stay at dinner
I suggest that this property of obviation is connected to the general ten
dency of languages to formally mark coreference with a subject more
explicitly than coreference with a non-subject. Thus, note that cross
linguistically, refexive pronouns may always be bound by a subject but
not always by a non-subject. Second, in "switch reference" languages it
is again coreference with the subject that matters. Lat, note that there
is a cross-linguistic tendency for the infnitive to act a a "same subject"
construction. Recall that in Fench and Hungarian, for instance, depen
dent Su complements may be in the subjunctive only if controlled by a
non-subject .
We can account for these facts as well as for the Su orientation of
obviation by restricting the notion of canonical control cae to caes of Su
control. The defnition of canonical control case will then have a semantic
side, pertaining to the semantic role of the controller and the controlled
argument, and a syntactic side, pertaining to the grammatical role of the
controller. The blocking efect we have discussed here is strongest when
both requirements are met.
4 Conclusion
It ha been argued here that obviation i the efect of a blocking mecha
nism that limits the choice between two complement types subcategorized
for by a verb. The range of the phenomenon is limited to semantically
dependent complements and is most robust in canonical control situations.
The disjoint reference efect discussed here thus reduces to the requirement
of using a control construction, if possible, in canonical control situations.
The blocking mechanism we have invoked is the same as lexical blocking,
discussed in McCawley 1977, and Horn 1978, 1984, where it is argued that
the existence and interpretation of a lexical item may be afected by the
existence and interpretation of another. In these works it is shown that
the choice of a formally more complex and semantically less specifc form is
restricted to cases where the use of a formally less marked and semantically
more specifc form is not possible. What we have seen here is that, not sur
prisingly, the same is true with respect to the choice of complement forms.
The generalization in (III) can thus be seen as a principle operating in the
lexicon, which limits the choice between two subcategorized complement
forms, a principle driven by the same mechanism as lexical, morphological
or phonological blocking, expressed by some adequately generalized version
of the Elsewhere Principle.
Note that our proposal difers fom that in Bouchard 1983 in that no
mention is made of anaphors and pronominals here. As a result, we lost
the direct connection between obviation and the efects of principle B of
Binding Theory, a connection that is present in all the analyses of obviation
formulated within Government and Binding theory. Note, however, that
ON OBVIATION / 107
if one adopts an analysis of the coreference properties of anaphors and
pronominals along the lines proposed by Bouchard or Reinhart 1983, the
connection is regained. The fundamental reason for the ungrammaticality
of ( 1 ) and (6) , repeated here as (32a) and (32b),
(32) a. Pierrei veut qU'il.i/j parte.
P. wants that he leave.SUBJ
b.
*
Johni saw himi .
would then be the same: a less specifc form has been chosen when a more
specifc one was available, and a result, one has failed to formally mark
coreference with a subject. The diference in range between obviation and
the anaphor/pronominal choice results from the fact that the former con
cerns the choice between subcategorized forms, a choice sensitive to the
semantic characteristics of the two situations involved. The principle gov
erning the pronominal/anaphor choice and our generalization (III) are then
members of a family of constraints on choice, whose common denominator
is the Elsewhere Principle.
References
Aronof, Mark. 1976. Word Formation in Generative Grammar. Cambridge,
Mas.: MIT Press.
Barwise, Jon, and John Perry. 1983. Situations and Attitudes. Cambridge, Mass. :
MIT Press.
Bouchard, Denis. 1982. On the Content of Empty Categories. Doctoral disserta
tion, Masachusetts Institute of Technology.
Bouchard, Denis. 1983. The Avoid Pronoun Principle and the Elsewhere Princi
ple. In Proceedings of NELS 13, 29-36. Amherst: University of Massachusetts,
GLSA.
Bresnan, Joan. 1972. Theory of Complementation in English Syntax. Doctoral
Comorovski, Ileana. 1985. Control and Obviation in Romanian. In Proceedings
of the 2nd Annual Easter States Conference on Linguistics.
Farkas, Donka. 1985a. Intensional Descriptions and the Romance Subjective
Mood. New York: Garland Publishers.
Farkas, Donka. 1985b. Obligatory Controlled Subjects in Romanian. In Pa
pers from the 21st Annual Meeting of the Chicago Linguistic Society, 90-100.
University of Chicago.
Farkas, Donka. 1988. On Obligatory Control. Linguistics and Philosophy 1 1 : 27-
58.
Fauconnier, Gilles. 1985. Mental Spaces: Aspects of Meaning Construction in
Naturl Language. Cambridge, Mas. : MIT Press.
Hermon, Gabriela. 1985. Syntactic Modularity. Dordrecht: Foris.
Horn, Laurence R. 1978. Lexical Incorporation, Implicature, and the Least
Efort Hypothesis. In Proceedings of the Parasession on the Lexicon, ed.
Donka Farkas, Wesley M. Jacobsen, and Karol W. Todrys, 196-209. Chicago:
Chicago Linguistic Society.
Horn, Laurence R. 1984. Ambiguity, Negation, and the London School of Par
simony. In Proceedings fm NELS 14, ed. Charles Jones and Peter Sells,
108-31 . Amherst: University of Massachusetts, GLSA.
Joseph, Brian. 1983. The Synchrony and Diachrony of the Balkan Infnitive. New
York: Cambridge University Press.
Kempchinsky, Paula. 1985. The Subjunctive Disjoint Reference Efect. In Pro
ceedings of the 1 5th Symposium on Romance Languages.
Kiparsky, Paul. 1982. Word Formation and the Lexicon. In Proceedings of the
1982 Mid-America Linguistics Conference, ed. Frances Ingemann, 3-29, 47-
78. Lawrence: University of Kansas.
Landman, Federick. 1986. Towards a Theory of Information. In GRASS 6.
Dordrecht: Foris.
Levinson, Stephen. 1987. Pragmatics and the Grammar of Anaphora: a Par
tial Pragmatic Reduction of Binding and Control Phenomena. Joural of
Linguistics 23:379-434.
McCawley, James. 1977. Conversational Implicature and the Lexicon. In Syntax
and Semantics
g
: Prgmatics, ed. Peter Cole, 245-259. New York: Academic
Press.
Meireles, J. , and Eduardo Raposo. 1983. Subjunctives and Disjoint Reference
in Portuguese: Some Implications for Binding Theory. Paper presented
at the Colloqui International de Linguistica Teorica i Llengues Romanicas,
Barcelona.
Montalbetti, Mario. 1984. Afer Binding: On the Interretation of Pronouns.
Doctoral dissertation, Massachusetts Institute of Technology.
Noonan, Michael. 1985. Complementation. In Language Typology and Syntactic
Description Vol. J, ed. Timothy Shopen, 42-140. New York: Cambridge
University Press.
Palmer, Fank. 1986. Mood and Modality. New York: Cambridge University
Press.
Picallo, Carme. 1984. The InH-node and the Null Subject Parameter. Linguistic
Inquiry 15: 75-102.
Picallo, Carme. 1985. Opaque Domains. Doctoral dissertation, City University
of New York.
Ransom, Evelyn. 1986. Complementizers: How Do They Mean. In Proceedings
of the 2nd Annual Meeting of the Pacifc Linguistics Conference, ed. Scott
DeLancey and Russell S. Tomlin. University of Oregon.
Reinhart, Tanya. 1983. Coreference and Bound Anaphora: A Restatement of
the Anaphora Questions. Linguistics and Philosophy 6: 47-88.
Ruwet, Nicolas. 1984. Je veux partir /* Je veux que je parte. In Cahiers de
grammaire, chapter 7. Toulouse: Le Mirail.
Salamanca, D. 1981. Subjunctive Syntax. Manuscript, Massachusetts Institute
of Technology.
ON OBVIATION / 109
Suier, Margarita, and J. Padilla-Rivera. 1984. On the Subjunctive and the Role
of the Features of INFL: Evidence from a Null Subject Language. Manuscript,
Cornell University.
Zaring, L. 1985. The Syntactic Role of Verbal Infection in French and Brazilian
Portuguese. Doctoral dissertation, Cornell University.
Zec, Draga. 1987. On Obligatory Control in Clausal Complements. In Working
Papers in Grammatical Theory and Discourse Structure, ed. Masayo !ida,
Stephen Wechsler, and Draga Zec, 136-168. Chicago: University of Chicago
Press.
Blocking of Phrasal Constructions by

Lexical Items
WILLIAM J. POSER
Introduction
Blocking is the widely observed phenomenon where the existence of one
form prevents the creation and use of another form that would otherwise
be expected to occur.
1
Perhaps the most prominent cases are those in which
the existence of an irregular form prevents the corresponding regular form
fom being used. In English for example, the existence of the irregular
plural men for man is said to block the regular form
*
mans.
Since the inception of the idea ( Paul 1896), and in virtually all work
on morphology over the pat sesquidecennium, blocking has been taken to
be restricted to the word-formation component ( Esau 1973, Aronof 1976,
Allen 1978, Clark and Clark 1979, Toman 1980, Zwanenburg 1981) . Cases
of non-lexical blocking are not cited as examples of the phenomenon.2
Moreover, the theoretical proposals that have been put forward account
only for blocking of one lexical form by another. For example, Miyagawa
( 1980) proposes that the morphological categories of a language defne a
set of slots in the lexicon, each of which, in the normal case, may be in
stantiated only once, so that if a more specifc rule (of which the extreme
cae is the existence of an irregular form) applies to instantiate a category,
a more general rule may not apply. Similarly, Kiparsky ( 1982a) proposes
that morphological rules are subject to the Elsewhere Condition ( Anderson
This is a revised version of a paper presented on 27 April 1986 at the Southern Califor
nia Conference on General Linguistics, University of California, San Diego, California.
Thanks to Steve Anderson, Joe Emonds, Paul Kay, Paul Kiparsky, Alec Marantz, Mariko
Saiki, and Michael Wescoat for discussion of these issues and comments on the draft.
l
The term blocking is due to Aronof ( 1976, 41) who defnes it U . . . the nonoccurrence
of one form due to the simple existence of another."
2
Exceptions are Di Sciullo and Williams 1987 and Hualde 1988, to which we will refer
later.
Lexical Matters . Ivan A. Sag and Anna Szabo1csi, eds.
112 / WILLIAM J. POSER
1969, Kiparsky 1973) so that a more specifc rule instantiating a complex
morphological category will apply disjunctively with a more general rule in
stantiating the same category, thus producing the blocking efect. On both
accounts, blocking is expected only internal to the lexicon, under Miya
gawa's proposal because the set of morphological categories is simply the
structured part of the lexicon, and under Kiparsky's proposal because mor
phological rules are by defnition lexical. I argue here on the bais of three
examples that the blocking phenomenon is not restricted to the lexicon,
that is, that it is possible for lexical forms to block phraal constructions.
1 The Existence of Phrasal Blocking
In this section I present three examples in which lexical items block phraal
constructs, in each case ofering evidence for the lexicality of the blocker,
the phrasal status of the blocked item, and the claim that the relationship
between the two is that of blocking.
Example I: Japanese Periphrastic Verbs
The Periphrastic Construction
Japanese has a very large number of periphrastic verbs consisting of a
nominal part followed by the verb suru 'do'. The majority of such verbs
are based on loans fom Chinese, a are the examples in ( 1) .
( 1) denwa suru t o telephone
sanpo suru to take a walk
kenkyuu suru to do research
This periphrastic construction i the usual way of borrowing verbs from
foreign languages; rather than adapting a foreign verb directly to Japanese
verbal morphology Japanese normally borrows a nominal form and then
creates a periphratic verb. Some examples of periphratics based on loans
fom English are given in (2).
(2) doraibu suru to drive
nokku suru to knock
Periphratics based on native nouns, such as those in (3), also exist but are
relatively uncommon; one reaon for this will be discussed below.
(3) tatigare suru to be blighted
tatiuti suru to cross swords
The nouns that appear in periphratic verbs are never restricted to the
periphrastic context; they invariably may appear in Noun Phraes in other
contexts.3 For example, the noun kenkyuu that forms the bae of the
periphratic verb kenkyuu suru 'to study, to do research' may also appear
3
There are a number of superfcial exceptions to this statement, but these are all
examples of the historically related but synchronically quite distinct lexical pseudo
periphratics, described in detail in Poser 1991.
B
LOCKING OF PHRASAL CONSTRUCTIONS BY LEXICAL ITEMS / 113
a the head of the subject
NP of a sentence, as i n (4) , or as the head of a
relative clause serving as a predicate nominal, as in ( 5) .
( 4) Sono kenkyuu-ga taisetu-da.
that research-N important-be
That research is important.
(5) Sore-wa Tanaka-san-ga site-iru kenkyuu-da.
that-T Tanaka-Mr. -N doing-be research-be
That is the research that Mr. Tanaka is doing.
Periphrastic verbs come in two forms, referred to as the "incorporated" and
"unincorporated" forms. In the incorporated form, the nominal component
is not case-marked, as in ( 1) , (2) and (3) , while in the unincorporated form
it bears accusative case, as in (6) .
(6) denwa-o suru to telephone
sanpo-o suru to take a walk
tatigare-o suru to cross swords
More generally, in the unincorporated form the nominal behaves like an
ordinary direct object NP, so that if the unincorporated periphrastic has
a logical object, the logical object appears a a genitive modifer of the
nominal component of the periphrastic, as in (7). In contrast, the logical
object of an incorporated periphratic, like other direct objects in Japanese
bears accusative or dative case, as in (8) , and cannot be modifed.
( 7) Eigo-no benkyoo-o site-iru.
English-G study-A doing-be
( He) is studying English.
(8) Eigo-o benkyoo site-iru.
English-A study doing-be
( He) is studying English.
The unincorporated periphrastics are unequivocally phrasal but the incor
porated periphrastics have been treated in most of the literature as single
words, whether lexically derived (Inoue 1976, Poser 1980, Miyagawa 1987,
1990, Grimshaw and Mester 1988) , or derived by incorporation in the syn
tax ( Kageyama 1977). However, there is considerable evidence that they
too are phraal in character ( Hasegawa 1979, Poser 1989). The evidence
for phrasal status may be briefy summarized as follows:
(a) Periphrastic verbs are accented like phrases rather than like any other
sort of verb ( Poser 1989).
(b) Reduplication afects only the suru component of the periphrastic
( Kageyama 1977) .
(c) Periphrastic verbs do not undergo even highly productive lexical nom
inalizations (Poser 1989) .
(d) Sentence-internally periphrastics are analyzable into the nominal and
verbal portions, in that the nominal may be omitted in whether
constructions, which require repetition of the verb (Poser 1989).
( e) It i s possible to Right Node Raise the suru portion alone ( Poser 1989) .
(f) It is possible to delete the verbal noun in the second conjunct of a
pair of conjoined sentences ( Kageyama 1977).
(g) Periphrastics are analyzable across sentence-boundary in that the
nominal part may be omitted in too-clauses, in which the verb of the
frst sentence is repeated in the second sentence ( Poser 1989) .
(h) Periphrastic verbs are analyzable at the discourse level across speak
ers into the nominal and suru, in that the nominal part may be
omitted in responses to yes-no questions ( Poser 1989).
'ue phrasal periphrastic verbs contrast i n these properties as well as a
number of others with historically related forms that have now been lexi
calized ( Poser 1991 ) .4
Given that periphrastic verbs are phrasal constructions, we do not ex
pect it to be possible to block them, but in fact there is reason to believe
that such blocking takes place.
Deverbal Noun Formation
Japanese has a fairly productive process of simple deverbal noun forma
tion.5 The deverbal noun is segmentally identical to the verb stem if the
verb is a vowel-stem verb, and consists of the verb stem followed by the
vowel Iii if the verb is a consonant-stem verb.6 Some examples are given
below in (9).
(9) Simple Deverbal Nouns
Verb Stem Derived Noun Gloss
ir iri parching
kari kari borrowing
mamor mamori protection
oyog oyogi swimming
sabak sabaki judgment
4
The atute reader may fnd it peculiar t hat \ attribute the hypothesis that incorpo
rated periphratics have a phraal structure to Hasegawa ( 1979) but in addition to my
own work cite only Kageyama ( 1977), who treats periphrastics as single words, for ar
guments to this efect. The reason for this apparent paradox is that Haegawa, whose
paper constitutes a reply to Kageyama's arguments for incorporation in the syntax, con
tributed no new arguments for phrasal status but rather argued, contra Kageyama, that
incorporation never takes place at all.
5
1 use the term simple to refer to the least specialized kind of deverbal noun formation
in Japanese. There are others, for example the manner nominals derived by sufxation
of -kata.
6
Although treated in much past work a a sufx, Yoshiba (1981) proposes that this
Iii be inserted by a morphological epenthesis rule, and Poser (1984) proposes that it be
inserted by a phonological epenthesis rule.
BLOCKING OF PHRASAL CONSTRUCTIONS BY LEXICAL ITEMS / 115
This kind of nominalization appears to be a lexical process. The pre
cise meaning taken on by the noun varies considerably, fom the abstract
'act of V-ing' through the agent noun, as illustrated by the examples
in ( 10).
(10) Thematic Types of renyoomeisi
Verb Gloss Noun Gloss Thematic Type
lru parch iri parching action
kariru borrow kari borrowing action
kumoru become cloudy kumori cloudiness result
moru serve, dish up mori a serving theme
oou cover ooi a cover instrument
tetudau help tetudai helper, help agent
tumu load tumi shipment, load theme
Moreover, in a number of cases deverbal nouns are accented irregularly,
further indicating they are lexical. In general, de verbal nouns are accented
on the ultima if the verb stem is accented and non-compound. Otherwise,
they are unaccented (Kawakami 1973). But a the examples in ( 11) show,
there are a number of exceptional cases in which the accent falls on the
frst syllable rather than on the ultima. Moreover, as the examples in (12)
show, there are also caes in which a deverbal noun that ought, if regular,
to be accented, is in fact unaccented.
( 1 1 ) Initial Accented Deverbal Nouns
domoru stammer domori stammering, a stammerer
hanareru separate hanare isolation
nagau sing fom door to door nagasi strolling musician
orosu sell at wholesale orosi wholesale trade
sabaku judge sabaki judgment
sawagu make noise sawagi noise, hubbub
seku dam up seki dam
suru pick pockets Surl pickpocket
tanomu request,ask tanomi a request, favor
tataru curse tatari curse
tomu become rich tomi riches
( 12) Unaccented Nouns Derived fom Accented Verbs
takuramu scheme, plan takurami a design, a trick
takuwaeru store, lay in takuwae store, hoard
todoroku roar, peal todoroki a roar, a peal
tumugu spin tumugi pongee
tutusimu be discreet tutusimi discretion
Blocking of Periphrastic Verbs
We might expect that we would be able to take a native verb, derive fom
it a noun, and form a periphratic with this deverbal noun as its nominal
bae, yielding forms like those in (13). However, this turn out to be
impossible; with rare exceptions, incorporated periphratics may not be
formed directly fom deverbal noun.
( 13) Periphrastics Baed on Root Deverbal Nouns
*
iri suru
*
mamori suru
*
oyogi suru
*
sabaki suru
*
uketori suru
parch
protect
swim
judge
receive
Note that the claim here is that the incorporated periphratics are im
possible. As I note below, the same is not true of their unincorporated
counterparts. This means that one must be careful to distinguish between
true incorporated periphratics and unincorporated periphrastics in which
the accusative cae particle has been elided by the process referred to as
O-Ellipsis. s In the transitive case this can readily be determined by the
cae-marking of the object: accusative in the case of a true incorporated
periphrastic but genitive in the case of an unincorporated periphratic that
has undergone O-Ellipsis. Another diagnostic is whether the form with a
case less nominal bae is considered appropriate in writing and in formal
speech, since O-Ellipsis is permissible only in caual speech.
I propose that this is a blocking efect, that is, that the periphrastic
forms are unacceptable because a corresponding lexical verb already exists,
as also suggested by Kageyama ( 1982) .
One alternative explanation that we mut consider is that we have here
only an ordering efect, that is, t hat periphratic verbs based on deverbal
nouns are impossible simply because periphrastic verb formation occurs at
a point at which the deverbal nouns have not yet been created. But there is
good reaon not to accept this proposal. One argument is theory-internal.
This is the fact that deverbal noun formation is lexical and periphrastic
7There are a small number of exceptions, real and apparent, to the blocking of pe
riphratic verbs by their lexical counterparts. An apparent exception is sakadati suru
'stand on end, stand on one's hea', which ha the non-periphratic counterpart saka
datu. In this cae the two forms have diferent meanings. The periphrastic form cannot
be used to refer to inanimate things, such a hair, standing on end, while the lexical
form ha precisely this use. Real exceptions include tatigare suru 'be blighted' which
coexists with its lexical counterpart tatigarr, tabenokosi suru 'leave food behind' <
tabenokosu, fmitaosi suru 'cheat' < fmitaosu, torisimari suru 'check on', < torisi
maru, and norikae suru 'change trains' < norikaeru. Saiki { 1987} observes that the
blocking efect sems to be weakening among younger speakers.
B
In caual speech it is possible for the accusative cae-marker U to be omitted. This is
known a O-Ellipsis.
BLOCKING OF PHRASAL CONSTRUCTIONS BY LEXICAL ITEMS / 1 17
formation is post-lexical. Since lexical rules precede post-lexical rules, de
verbal noun formation must precede periphratic formation, and hence the
deverbal nouns must be available. Secondly, there is direct evidence that
the deverbal nouns are available at the point at which periphrastic forma
tion applies. In ( 14) we have a number of examples of periphrastic verbs
whose nominal component is a compound the second member of which is
deverbal.
( 14) Periphrastic Verbs with Compound Deverbal Nominal Component
Periphratic Gloss Analysis
amamori suru leak rain rain + leak
amayadori suru take shelter fom the rain rain + take shelter
aibumi suru stamp foot + tread on
atomodori suru retreat afer + turn back
atozusari suru finch afer + withdraw
hitobarai suru clear a room of people person + sweep
hitomisiri suru be bahful person +see+ know
hitoriaruki suru be independent one-person + walk
hitotigai suru mistake a person for another person + mistake
iede suru leave home house + leave
igui suru live in idleness live + eat
kantigai suru misjudge perception + difer
kimayoi suru waver spirit + be confused
maeoki suru make introductory remarks font + put
mizuarai suru wah with water water + wash
nebumi suru appraise price + evaluate
senobi suru straighten one's back back + straighten
tatiuti suru cross swords sword + strike
tukimi suru engage in moon-viewing moon + see
yukimi suru engage in snow-viewing snow + see
In addition to the many more-or-less random examples of this type, certain
frst members are quite common. For example, periphrastics meaning 'do V
in advance' are created rather feely from nominals formed by compounding
the noun mae ' front, before' with the de verbal noun. Some examples are
given in ( 15).
( 15) Periphrastic Verbs with Nominal Component Containing mae
Periphratic Gloss
Analysis
maebarai suru prepay font + pay
maegari suru draw (money) in advance font + borrow
maegasi suru advance ( money) font + lend
maeuri suru sell in advance font + sell
Similarly, periphratics meaning 'do V a little' are formed fairly produc
tively by compounding the deverbal noun with the number 'one' , hito.9
( 16) Periphrastic Verbs with Nominal Component Containing hito
Periphrastic Gloss Analysis
hitohasiri suru go for a run one + run
hitomawari suru go round one + go around
hitonemuri suru nap one + sleep
hitooyogi suru swim a little bit one + swim
hitoyasumi suru take a short rest one + rest
The existence of such periphrastics sufces to demonstrate that de verbal
nouns must be available for periphratic verb formation. The reaon that
these are acceptable while other periphrastic verbs baed on deverbal nouns
are not, is that these have no corresponding lexical verbs. That is, there
are no verbs
*
amamoru,
*
amayadori ,
*
asibumu,
*
atomodoru,
*
atozusaru,
*
hitobarau,
*
hitohasiru,
*
hitomawaru,
*
hitomisiru,
*
hitonemuru,
*
hitoo
yogu,
*
hitoriaruku,
*
hitotigau,
*
hitoyasumu,
*
iederu, * iguu,
*
kantigau,
*
kimayou,
*
maebarau,
*
maegariru, maegasu,
*
maeoku,
*
maeuru,
*
ne
bumu,
*
senobiru,
*
tatiutu,
*
tukimiru, or
*
yukimiru. The great majority
of the nominals on which periphrastics of the type illustrated in ( 14), ( 15) ,
and (16) are baed are Noun-Noun compounds whose frst component is a
non-deverbal noun. 10 Although there are sporadic examples of the type,
Japanese does not productively generate verbs by compounding a noun
with a verb, so these nominals cannot be derived by nominalizing a verb
with this structure. Moreover, a shown in Poser ( 1984, 93), the accentua
tion of these nominals is consistent with a derivation in which the second
member is nominalized and then compounded with another noun, but not
with a derivation in which a compound verb i created and then nomi
nalized. Thus, periphrastics baed on nominals containing non-deverbal
nouns sufer no competition fom non-periphratic verbs, and hence, are
not blocked.
Periphratics may also be baed on deverbal nouns to which the sufx
bakari 'only' has been attached, as in ( 17) , or to which the topic-marking
sufx wa has been attached to focus the verb, as in (18).
( 17) Yomi-bakari site
Reading-only do-ing
I am only reading.
imasu.
be
am grateful to Yo Matsumoto for pointing out the relevance of the compounds of
hito.
100f the examples cited, the sole exception is igui 8U7 whose nominal bae, igui, is
a compound both of whose members are deverbal. (The first member is derived from
the verb iru, 'be located in a place, live' . ) However, even this example conforms to the
larger generalization, a there is no compound verb * iguu.
(18)
Yomi-wa site
Reading-TOP do-ing
I am reading.
imasu.
be
These too show that the deverbal noun must be available for periphrastic
formation. They are not blocked since their meaning difers fom that of
the simplex verbs fom which they are derived.
In sum, the acceptability of periphrastics based on compound nouns
with a deverbal member, and the acceptability of periphrastics based on
simplex deverbal nouns to which bakari and wa have been sufxed demon
strate that the ordering explanation is untenable, while the blocking hy
pothesis makes exactly the correct prediction.
Still another explanation might be baed on the fact that not every
noun can enter into the incorporated periphratic construction-the noun
must have a suitable argument structure and other syntactic properties.
We might suppose that deverbal noun formation creates nouns which, un
like the non-derived loans fom Chinese and English, lack the syntactic
properties necessary to enter into the periphratic construction. But this
is belied by the fact already demonstrated that deverbal nouns can form
the basis for periphratic verbs provided that there be no corresponding
non-periphrastic verb. Moreover, there is no syntactic property that the
non-native nominal baes of periphratic verbs exhibit that the native ones
do not, other than, of course, the ability to enter into the periphrastic
construction. Thus, just as non-native nominals can assign case in the
absence of the verb suru in certain constructions, such a purpose clauses
( 19), so can native deverbal nouns (20) . There is no evidence that the
properties of the deverbal nouns themselves are in any way distinct from
those of the non-deverbal nouns fom which periphrastic verbs may be
formed.
( 19) Hanako-wa eigo-o benkyoo-ni amerika-e itta.
Hanako-T English-A study-D America-AD went
Hanako went to America to study English.
(20) Taroo-wa Hanako-kara hon-o uketori-ni dekaketa.
Taroo-T Hanako-ABL book-A receiving-D went-out
Taroo went out to get a book fom Hanako.
In sum, periphrastic verbs based on deverbal nouns are not possible so long
as there is a corresponding lexical verb. Alternative explanations for this
behavior being untenable, this appears to be a case in which lexical items
block a phrasal construction.
1 1
l IThe blocking analysis proposed here eliminates the need for the Japanese analogue
of Do-Support proposed by Kuroda (1965). Kuroda proposed that sentences like ( 18)
have a deep structure in which the particle UQ appears to the right of the entire verb.
Example II: English Comparatives and Superlatives
Comparative and superlative adjectives in English may be formed in two
ways. There is a lexical construction, involving the sufxation of the mor
phemes -er and -est, and a periphrastic construction, in which the adverbs
more and most precede the adjective.
The lexicality of the afxation of -er and -est seems clear. They
exhibit no behavior inconsistent with lexical status: their phonology is
that of words, the afxes are bound, and nothing not itself a sufx may
intervene between the adjective and these afxes. Moreover, compara
tive and superlative adjectives are in certain caes formed by suppletion,
clear evidence of lexicality. Thus, we have better for * good+er and best
for * good+est , worse for * bad+er, worst for * bad+est . 1
2
-er and - est
also occur inside of compounds, a in surer-footed, fairer-minded, lightest
skinned.
Similarly, the phrasal status of the periphrastics in more and most
seems clear. more and most can be followed by arbitrary conjunctions of
adjectives, a in ( 21) , and it is possible to interpolate appositive material
between more and most and the following adjective, as in (22) .
( 21) Periphratic Comparatives and Superlatives of Conjoined Adjectives:
(a) more [curious and inquisitive]
(b) most [economical, efcient, and frugal]
The si-Insertion transformation (p. 63, ex. 121), which takes the form V-AUX-wa V
wa-si-AUX, moves wa to the right of the verb stem and inserts the verb suru. The
subsequent discussion suggests that he intended to generalize this rule to other particles
a well, though he does not state the more general form of the rule. Leaving aide the
open question of how si-Insertion is to be formulated in a more restrictive thtry than
clasical Transformational Grammar, the blocking account seems clearly superior to that
baed on si-Insertion. First, since the periphratic construction exists independently,
the rule of si-Insertion is simply unnecessary. Second, the blocking account provides a
complete account of when suru appears and when it does not, while it is an arbitrary
stipulation that certain morphemes trigger si-Insertion. Kuroda says that the purpose
of si-Insertion is to carry the AUX. However there is no explanation of why a verb
to which these particles are attached cannot itself bear the AUX. Third, the blocking
analysis explains why it is that the accentuation of the V is that of the verbal noun,
which is left unexplained on Kuroda's acount. Finally, the blocking analysis provide a
cleaner account of what wa and similar morphemes can attach to. These morphemes
baically attach to NP, though some of them can attach to S-bar complements and to
participles. They cannot attach to adjectives or adverbs. On the blocking analysis, this
distribution is maintained, a the form to which wa and its ilk attah in the periphratric
construction is a noun. On the si-Insertion analysis, these morphemes must be allowed to
attah underlyingly to tensed verbs. Note that the blocking issue would arise even if we
were to accept Kuroda's proposal. Given the independent existence of the periphratic
construction, we would still need an account of why some but not all peripheratics
baed on deverbal nouns are acceptable.
12
0ne conceivable, though unattractive alternative would be to claim that the non
suppletive forms were phraal and that only suppletive forms are lexical. In this cae we
would have a diferent argument for phraal blocking, since the suppletive forms block
the regular forms gaoder and * gaadest.
(22) Interpolation into Periphrastic Comparatives and Superlatives:
(a) This situation is more, I suppose the term is delicate, than I
had thought.
(b) Watson, this is the most, how shall I say, curious case that I
have ever seen.
This is not true of the lexical forms. In (23) we see that each term in
a conj unct requires its own comparative or superlative sufx. When the
comparative sufx falls on the lat term the sentence is grammatical, but
the comparison is restricted to the lat adjective. When it is on a non-fnal
term the example is simply ungrammatical. 13
(23) Interpolation into Periphrastic Comparatives and Superlatives:
(a) He is taller, slimmer, and handsomer than John.
(b) He is tall, slim, and handsomer than John.
(c)
*
He is taller, slim, and handsome than John.
(d)
*
He is tall, slimmer, and handsome than John.
The lexical forms are possible only in a fairly small range of cases,
determined in a way not fully understood by the length or stress pattern
of the stem. Generally speaking, lexical comparatives and superlatives
of adjectives with mono- and di-syllabic stems are perfect, while lexical
forms derived fom adjectives with longer stems are unacceptable. On the
other hand, while it is always possible to form periphrastic comparatives
and superlatives fom adjectives with longer stems, periphrastic forms of
adjectives with mono- and di-syllabic stems are generally unacceptable. 14
The crucial observation is, then, that whatever the nature of the principles
governing the well-formedness of lexical comparatives and superlatives, the
acceptability of the periphratic forms is inversely related to that of the
lexical forms, as illustrated in (24) . As far a I can see, the only plauible
explanation for this is that the periphrastic forms are blocked by the lexical
forms.
1
5
1
3
It ualso true that interpolation between the adjective and the comparative or superla
tive sufx is imposible, in the sense that there are no acceptable examples of it, but
the un acceptability of these examples is not of great probative value since I have been
able to fnd no circumstances in which, were the the comparative and superlative afxes
c1itics or even independent words, we would expect the interpolation to be acceptable.
l40ne systematic exception to this generalization occurs in such metalinguistic con
structions a It's mor big than good, where the periphratic form is not only acceptable
but obligatory (cf. * It's bigger than better). What distinguishes such sentences from or
dinary comparatives is that the nature of the comparison is diferent from that conveyed
by lexical comparatives. In the usual cae, saying that A is Adj-er than means that
on some scale of Adj-ness, A lies farther from the reference point than B. In contrat,
when we say A is more Adjl than Adh we mean something like "It is more accurate
or appropriate to say that A m Adj
l
than to say that A is Adj
2
." In other words, the
comparison u here between the appropriateness of two utterances rather than between
two situations in the world.
1
5
Di Sciullo and Williams ( 1987) have also noted this example.
(24) English Comparatives
Bae Lexical Periphrastic
big bigger
*
more big
small smaller
*
more small
good better
*
more good
fun funnier
*
more funny
silly sillier
*
more silly
childish
*
childisher more childish
regal
*
regaler more regal
damaging
*
damaginger more damaging
symmetric
*
symmetricer more symmetric
vivacious
*
vivaciouser more vivacious
Example III: Basque
A third example of blocking of a phraal construction by lexical items
is found in Basque, as described by Hualde ( 1988, 38-41). In Baque,
progressive aspect is normally expressed by means of periphrasis with the
defective verb ari . Thus, the progressive counterpart of (25) is (26) , in
which ari appears between the verb stem and the auxiliary.
(25) Jon abiatzen da.
John leave- imf AUX-intr
John leaves.
(26) Jon abiatzen ari da.
John leave- imf ari AUX- intr
John is leaving.
However, a handfl of verbs have synthetic (non-periphrastic) present and
past tense forms, and these lack periphrastic progressive forms in ari.
Hualde provides the examples in (27), where the grammatical, synthetic
form in the frst column contrasts with the ungrammatical but expected
periphrastic form in the third column.
(27) daki he knows
*
jakiten ari da
doa he goes
*
joaten ari da
dakar he brings
*
ekartzen ari da
dabil he walks
*
ibiltzen ari da
dator he comes
*
etortzen ari da
dauka he possesses
*
edukitzen ari da
The existence of the lexical tense-aspect forms of these verbs apparently
blocks the corresponding periphrastic forms.
As evidence that the periphrastic forms are indeed phrasal, that is, that
ari is not lexically attached to the main verb, Hualde ofers the fact that
in negative constructions ari need not be adjacent to the verb stem at all.
In (28) ari precedes the verb kantatzen and is separated fom it by the
direct object madrigalak. This appears to be compelling evidence for the
phraal status of ari periphratics, and hence for the claim that we have
here another cae of blocking of a phrasal construction by lexical items.
(28) Jon ez da ari madrigalak kantatzen
John NEG AUX ari madrigals sing- imf
John is not singing madrigals.
In sum, Japanese, English, and Basque appear to provide real examples of
blocking of phrasal constructs by lexical items. 16
2 Implications
The three examples of blocking of phrasal constructions by lexical items
presented here are problematic for the existing theory of blocking since
they cannot be accounted for in terms of unique instantiation of complex
morphological categories, at leat if we take such morphological categories
to be those flled by the word-formation component. Either we must fnd
some other account of blocking, one under which we do not expect blocking
to be restricted to the lexicon, or we must in some way "extend" the lexi
con to encompas the sorts of phraal construction that we have discussed
here.
2. 1 The Pragmatic Approach
There is, in fact, one proposal in the literature that does not predict the
restriction of blocking to the lexicon. This is the proposal, due to House
holder ( 1971) and McCawley ( 1977), that efects very much like those that
are usually acribed to blocking are to be attributed to pervasive Gricean
principles, to wit the principle that ceteris paribus the speaker expends as
little efort a possible to say what he wants to say and therefore chooses
the simplest available form. For example, Householder (1971) observes that
it is awkward to say pale red. He proposes that the reason for this is that
English has a simpler way of expressing the same notion, namely the word
pink. Insofar as pink and pale red express the same meaning, the speaker
will minimize his expenditure of efort and choose the former. pale red
will be chosen only when pink is for some reason inappropriate, as when
the hearer does not know the meaning of pink and pale red is given as a
defnition.
Neither Householder nor McCawley uses the term blocking or discusses
the usual caes of morphological blocking, nor are their works cited in the
1
6
Di Sciullo and Williams ( 1987) cite the relationship between the synthetic and pe
riphratic forms of the Latin pasive a an example of lexical blocking of a phraal
construct, which it may well be, though to be sure it is necessary to ofer evidence of
the phraal character of the periphratic construction and to rule out alternative ex
planations of the relationship. In general, paradigms containing both synthetic and
periphratic forms are good candidates for instances of lexical blocking of phraal
constructs.
literature on morphological blocking. Thus it does not seem that their
proposal was intended to extend to these cases, nor ha it been so inter
preted. However, Di Sciullo and Williams (1987) suggest such an approach
to blocking, without citing Householder or McCawley, or going into any
detail.
This proposal ha the advantage that it is not restricted to the lexicon.
Thus, we must entertain the possibility that the unique instantiation ac
count of blocking is incorrect, and that it is rather a pragmatic efect due
to miniInization of efort.
Although this proposal is attractive, it does not solve our problem. To
begin with, it fails to account for a number of the clasical examples of
morphological blocking, since it predicts that form A will block form B
only if form A involves a lesser expenditure of efort. This means, other
t hings being equal, that we expect blocking only if the blocker contains
less phonological material than the blocked form. But this prediction is
incorrect. For example, we cannot appeal to the pragmatic proposal to
explain the blocking of English * oxes by the irregular oxen, since both
forms are of equal phonological and morphological complexity. SiInilarly,
in Japanese, the verb kuru 'come' has the irregular present neutral nega
tive stem kona- in place of the regular * kina. The fact that kona- blocks
kina- is inexplicable on the pragmatic hypothesis, since the two stems
are of equal complexity. Even worse i the present neutral afrmative
form of the Japanese verb 'to do'. The irregular present neutral afrma
tive form suru is actually longer than the expected but incorrect * su.
1
7
A parallel example in English is the irregular plural children, which is
surely not simpler than the regular
*
childs. Examples such a these show
that the pragmatic hypothesis does not handle the traditional caes of
blocking.
A second problem with extending the pragmatic hypothesis to all cases
of blocking is the fact that in the typical cae of blocking the judgments
are much stronger than in Householder's pale red example. While it is
true that pink is generally preferable to pale red, it is still possible to use
the latter when there is sufcient motivation, a, for example, in defning
pink for someone who does not know its meaning. In contrast, we cannot
explain the irregular form men to a person learning English by equating it
with * mans. This latter form is not simply verbose; it is impossible.
The very property that makes the pragmatic hypothesis attractive,
namely that it predicts the existence of blocking outside the lexicon, also
provides an argument against it. Under the pragmatic hypothesis, it should
be possible for phraal constructs of any size to be blocked. But in point of
fact the examples of blocking of phraal constructs known to me all involve
17
The stem of 'do' throughout most of its paradigm is simply Is;' The addition of the
preent neutral afrmative sufx Irul to this stem will yield * U.
BLOCKING OF PHRASAL CONSTRUCTIONS BY LEXICAL ITEMS I 125
blocking of small phrases; there appear to be no examples of blocking of
large syntactic units. For example, the red book does not block the book
which is red.
A further difculty for the pragmatic approach arises fom the fact that
the correspondence between two forms depends only on their meaning-if
two potential forms have the same meaning, the simpler form should block
the other one whether or not they are morphologically related, as is indeed
the cae in Householder's example of pale red, which has only a semantic
relation to pink. IS
In general, there is no blocking efect when two forms are not related. 19
Thus,
(29) John is smarter than Tom.
blocks
(30) * John is more smart than Tom.
but not such synonymous but structurally unrelated sentences as:
(31) John's intelligence exceeds Tom's.
(32) John ha more intelligence than Tom.
or
(33) John ha greater intelligence than Tom.
The pragmatic approach therefore fails to provide a fully adequate account
of blocking, partly because it cannot account for the classical observation
that irregular forms block regular forms, and partly because it fails to
restrict blocking efects to structurally related forms where the blocked
form comprises a small syntactic unit.
2. 2 Morphological Constructions
These problems with the pragmatic hypothesis suggest that we ought to
try the other available route, namely fnding some way to characterize
certain phasal constructs as instantiating morphological categories in spite
of their non-lexical status. Roughly speaking, what we want to do is to
1
8
Another apparent counterexample is that of doublets of a type common in Japanese,
where there is both a native simplex verb and a periphratic verb baed on a loan
from Chinese, where on the pragmatic account we might expect the simplex verb to
block the periphratic. In many caes the members of these doublets appear to be
perfectly synonymous. An example is manabu 'study', the native counterpart to the
Sino-Japanese periphrastic benkyoo SQ7Q In most if not all of these caes, however,
the members of the pair belong to diferent stylistic registers. It is usually the Sino
Japanese periphratic that belongs to the higher register, but there are exceptions, such
a manabu, which is the higher register member of the pair. Insofar a the pragmatic
constraint is to use the simplest available form, the absence of blocking here is expected
if we take availability to be relative to the chosen register.
1
9
The difculty of determining which utterances count a relevant alternatives is dis
cussed in some detail by Horn (1978) in a critique of the proposal of McCawley'S of
which we here consider an adaptation.
extend the boundary of the lexicon, so that we can treat a class of phrase
formation rules as essentially morphological in character, in the sense that
they instantiate morphological categories.
The question that arises is how to instantiate t his idea. I will ten
tatively propose that we should distinguish between morphological rules,
by which I mean processes that instantiate morphological categories, and
word-formation rules, by which I mean the non-phonological rules that
operate within the lexicon. Since word-formation rules all instantiate mor
phological categories, all word-formation rules are morphological rules, but
the converse need not be the case. Insofar as there are syntactic rules that
instantiate morphological categories, these rules are morphological rules
but not word-formation rules. 2o This provides us with a reconstruction of
the traditional notion of periphrasis: a periphrastic construction is one in
which morphological categories which are typically instantiated lexically
are instead instantiated at a phrasal level.
The question then arises as to how to define a morphological category
other than by saying that it i something that is instantiated by a word
formation rule. Suppose that we say that a morphological category is a
category potentially instantiated by a word-formation rule. Then we would
say that the category of comparative adjectives is a morphological category
because in some languages it is instantiated by word-formation rules. The
fact that it may be instantiated by a phrasal construction as well, as in
English, does not afect the claim that this category is morphological in
nature.
This distinction between morphological rules and word-formation rules
permits a straightforward account of the English comparative adjective.
The morphological category of comparative adjective may potentially be
flled either by a lexical form or by a periphrastic form. If the lexical form
exists, the category is instantiated and so the periphratic form is blocked.
Similarly, in the case of the periphrastic verbs, the morphological cat
egory is the verb with the argument structure and other properties of the
related noun. If this category is instantiated by a lexical verb, the pe
riphrastic form is blocked, but if it is lef empty it may be instantiated by
a phrasal construction.
One question that arises is what kinds of phrasal constructs may in
stantiate morphological categories, and hence be blocked by lexical forms.
The three examples that we have seen both involve phrasal categories that
are in a certain sense "small." As I pointed out above, one defect of the
pragmatic proposal is that it predicts that it should be possible to block
any sort of phrasal construction. We should like to avoid the same problem
here. I conjecture that it is only what I will call small categories that can
instantiate morphological categories. By a small category I mean a cate-
20
A thery with this property is that of Anderson (1991).
gory that dominates only zero-level projections. The English periphrastic
comparatives and superlatives are presumably categories of type A
1
and
contain only categories of type ADYo and AO. The Japanese incorporated
periphrastics are "small" since, on the analysis of Poser 1989, they are
of category yO and contain only categories of the same level, namely yO
and NO.
This defnition helps us to explain a fact about the Japanese pe
riphrastics that would otherwise seem problematic. As I have noted,
although incorporated periphratics are blocked by corresponding lexical
verbs, their unincorporated counterparts are not. This contrast follows im
mediately once we recognize that the unincorporated periphrastics are not
"small" ; the nominal part of an unincorporated periphrastic is a full NP,
a illustrated in example ( 7) above a well as in (34) .
(34) Butyoo-ga suru yoo-ni merrel
Boss-NOM do so ordered
kenkyuu-o mada sinakatta.
research-ACC still do- neg-
p
ast
sita
He still hasn't done the research that his boss ordered him to do.
Since the unincorporated periphrastic contains a full NP it is not a "small"
category and so cannot instantiate a morphological category. 21
Whether i t i s possible to derive the restriction of phrasal blocking to
"small" phrases is unclear. It may well be that this restriction can be de
rived fom principles governing the distribution of morphological features,
but I am not at present prepared to defend this position.
The proposal that morphological rules be considered to be a superset
of the word-formation rules, with blocking applicable to morphological cat
egories, not to words, provides an account both of the classical blocking
caes and of the attested caes of phrasal blocking, without falsely predict
ing blocking to be a more general phenomenon than it is. In this sense,
it seems that the proposal is on the right track. However, as presented
here the proposal is also excessively vague, and its viability depends on
whether subsequent research provides an adequate theory of morphological
categories and how they are instantiated. If this approach is correct, it
21 Let me note briefy two inadequate alternatives that I considered. First, it is not
adequate to say that "small" categories are merely non-recursive. The incorporated
periphrastics, on the analysis given here, are recursive in that they contain another cat
egory of the same type (y
O
contains another y
O
) . Second, we might defne the "small"
categories as those that are monotonic in the sense that they contain no projections
higher than themselves. The incorporated perphrastics and the periphratic compara
tive and superlative adjectives satisfy this defnition since they contain only zerolevel
projections, but this defnition fails to distinguish between the incorporated perphra
tics and their unincorporated counterparts. Insofar a the latter are of category y
2
they
should count as "small' even if they dominate full NPs. This defnition might, however,
be tenable if the unincorporated periphratics were of category y
l
, a possibility that I
am not at present prepared to rule out conclusively.
provides reconstructions of the traditional notions of periphrasis and con
struction, notions used regularly for descriptive purposes, but which have
no home in current morphological and syntactic theory. 22
3 Summay
Although the current literature on blocking is restricted to the lexicon,
there appear to be cases of blocking of phrasal constructions by lexical
items. Three examples are presented here, namely the blocking of Japanese
periphrastic verbs by their lexical counterparts, the blocking of English
periphrastic comparative and superlative adjectives by lexical comparative
and superlative forms, and the blocking of Basque periphrastic progressive
verb forms by lexical progressives. These examples require a modifcation
of the theory of blocking. One possibility is a purely pragmatic account,
along the lines suggested by Householder (1971), McCawley ( 1977), and
Di Sciullo and Williams ( 1987). This, however, is subject to a number of
objections. Instead I propose an extension of existing lexical accounts of
blocking to encompass blocking of phraal constructions that instantiate
morphological categories.
References
Allen, Margaret R. 1978. Morphological Investigations. Doctoral dissertation,
University of Connecticut.
Anderson, Stephen R. 1969. West Scandinavian Vowel Systems and the Order
ing of Phonological Rules. Doctoral dissertation, Masachusetts Institute of
Technology.
Anderson, Stephen R. 1991. A-morphous Morhology. New York: Academic
Press.
Aronof, Mark. 1976. Word Formation in Generative Grammar. Cambridge,
Mass. : MIT Press.
Clark, Eve, and Herbert Clark. 1979. When Nouns Surface as Verbs. Language
55: 767-81 1 .
Di Sciullo, Anne-Marie, and Edwin Williams. 1987. On the Defnition of Word.
Cambridge, Mass. : MIT Press.
Esau, Helmut. 1973. Nominalization and Complementation in Moder German.
Amsterdam: North Holland.
Fillmore, Charles J. , Paul Kay, and Mary Catherine O'Connor. 1988. Regularity
and Idiomaticity in Grammatical Constructions. Language 64:501-538.
22
The work of Fillmore, Kay, and O'Connor ( 1988) on Construction Grammar is an
exception. Although one might entertain the idea that what can be blocked are con
structions, it appears that from the point of view of Construction Grammar every phraal
collocation is a construction-there is no distinction made between phraal collocations
that instantiate morphological categories and other phraal structures-so that their
notion of construction is too general to be useful for the purpoe of delimiting the scope
of blocking.
Grimshaw, Jane, and Armin Mester. 1988. Light Verbs and Theta-Marking.
Haegawa, Nobuko. 1979. A Comparison of Two Approaches to a Class of
Japanese Constructions. Manuscript, University of Wahington, Seattle.
Horn, Laurence R. 1978. Lexical Incorporation, Implicature, and the Least
Efort Hypothesis. In Proceedings of the Parsession on the Lexicon, ed.
Donka Farka, Wesley M. Jacobsen, and Karol W. Todrys, 196-209. Chicago:
Chicago Linguistic Society.
Householder, Frederick W. 1971. Linguistic Speculations. New York: Cambridge
University Press.
Hualde, Jose Ignacio. 1988. A Lexical Phonology of Basque. Doctoral dissertation,
University of Southern California.
Inoue, Kazuko. 1976. Henkei Bunpoo to Nihongo (Tansformational Grammar
and the Japanese Language). Tokyo: Taishukan.
Kageyama, Taroo. 1977. Incorporation and Sino-Japanese Verbs. Papers in
Japanese Linguistics 5: 1 17-156.
Kageyama, Taroo. 1982. Word Formation in Japanese. Lingua 57: 215-258.
Kawakami, Shin. 1973. Doosi kara no tensei meisi no akusento (Accentuation of
Deverbal Nouns) . Imaizumi Hakase Koki Kinen Kokugogaku Ronsoo 55-70.
Kiparsky, Paul. 1973. Elsewhere in Phonology. In A Festschrif for Morris Halle,
ed. Stephen R. Anderson and Paul Kiparsky. New York: Holt, Rinehart, and
Winston.
Kiparsky, Paul. 1982a. From Cyclic Phonology to Lexical Phonology. In The
Structure of Phonological Representations (Part I), ed. Harry van der Hulst
and Norval Smith, 131-175. Dordrecht: Foris.
Kiparsky, Paul. 1982b. Word Formation and the Lexicon. In Proceedings of
the 1982 Mid-America Lin9uistics Conference, ed. Frances Ingemann, 3-29,
47-78. Lawrence: University of Kansas.
Kuroda, Sige-yuki. 1965. Generative Grammatical Studies in the Japanese
Language. Doctoral dissertation, Massachusetts Institute of Technology.
Reprinted 1979 by Garland Publishers, New York.
McCawley, James. 1977. Conversational Implicature and the Lexicon. In Syntax
and Semantics 9: Pragmatics, ed. Peter Cole, 245-259. New York: Academic
Press.
Miyagawa, Shigeru. 1980. Complex Verbs and the Lexicon. In Coyote Papers, !
Thcson: Department of Linguistics.
Miyagawa, Shigeru. 1987. Lexical Categories in Japanese. Lingua 73: 29-51.
Miyagawa, Shigeru. 1990. Light Verbs and the Ergative Hypothesis. Linguistic
Inquiry 20:659-688.
Paul, Hermann. 1896. uber die Aufgaben der Wortbildungslehre. Sitzbungs
berichte der konigl. bayer. Akademie der Wissenschafen (phiIosophisch
philologische und historische Classe) 692-713.
Poser, William J. 1980. Periphrastic Verbs in Japanese. Manuscript, Mas
Poser, William J. 1984. The Phonetics and Phonology of Tone and Intonation in
Japanese. Doctoral dissertation, Massachusetts Institute of Technology.
Poser, William J. 1989. Japanese Periphrastic Verbs and Noun Incorporation.
To appear in Natural Language and Linguistic Theory.
Poser, William J. 1991. Lexical and Phrasal Periphrastic Verbs in Japanese.
Manuscript, Stanford University.
Saiki, Mariko. 1987. On the Manifestations of Grmmatical Functions in the
Syntax of Japanese Nominals. Doctoral dissertation, Stanford University.
Scalise, Sergio. 1984. Genertive Morphology. Dordrecht: Foris.
Szymanek, Bogdan. 1985. Disjunctive Rule Ordering in Word Formation. In
Papers and Studies in Contrastive Linguistics, Vol. 20, ed. Jacek Fisiak, 45-
64. Poznan: Adam Mickiewicz University.
Toman, Jindrich. 1980. Wortsyntax. Doctoral dissertation, University of Cologne.
Yoshiba, Hiroshi. 1981. The Mora Constraint in Japanese Phonology. Linguistic
Analysis 7:241-262.
Zwanenburg, Wiecher. 1981. Le principe du blocage dans la morphologie deriva
tionelle. In Linguistics in the Netherlands 1981, ed. Saskia Daalder and
Marinel Gerritsen, 65-72. Amsterdam: North Holland.
The Stress and Structure of

Modifed Noun Phrases in English
MARK LIBERMAN AND RICHARD SPROAT
1 Introduction
Our topic is the stress pattern of English noun phrases in which the head
noun is preceded by a sequence of modifers. l We assume a context of use
that is rhetorically stress-neutral; the phenomena of FOCUS, CONTRAST
and ANAPHoRA-henceforth FCA-are taken to be perturbations of the
patterns that we discuss. We attempt to establish the basic regularities
that shape the complex data in this area, against the background of a broad
(and thus complex) description. Our purpose is to establish an adequate
set of descriptive categories, able to support a formal model of the syntax,
semantics and prosody of complex nominals. We would like such a model
to be adequate for parsing and asigning stress to modifed noun phrases
in unconstrained English text. We start with a careful description of the
phenomena, followed by a more formal account of the proposed syntactic
analysis, and a sketch of the implications for parsing and stress-assignment
algorithms.
Many syntacticians (e. g. , Jackendof 1977) have noted the existence of
at least four distinct prenominal positions, arranged in a right-branching
structure:
( 1)
4 3 2 1 0
the three exotic chess boards
John's many large book bags
those few Chinese store owners
Position 4 is stereotypically occupied by articles, demonstratives, and pos
sessive phrases; position 3 by certain quantifers and numerals; position
I The authors would like to thank an anonymous reviewer for useful comments, and
Julia Hirschberg and Mats Rooth for some discussion.
Lexical Matters. Ivan A. Sag and Anna Szabo1csi, eds.
132 / MARK LIBERMAN AND RICHARD SPROAT
2 by adjectives; and position 1 by nouns making up the initial element
of a compound word. We will refer to items in positions 1 and 2 as
MODIFIERS.
Each of the positions in ( 1) may be occupied by an item that has inter
nal structure of its own: John's brother's (position 1) , distressingly few (3),
more exotic (2), liquor store (1). Items characteristic of positions 1 and 2
are often repeated, either in parallel or a layered fashion: powerful, luxuri
ous automobiles; powerful economic forces. Normally, a strict ordering of
these positions is required-phrases like
*
large many John's book bags are
about as wrong a arrangements of English words can be.
The left hand edge of English noun phrases is more complex than the
simple pattern ( 1) indicates. For instance, certain quantifers cannot substi
tute in position 3 of the pattern, co-occurring with articles, demonstratives
and possessive phrases only in so-called partitive constructions:
*
John's
all/some/any/each large book bags, all/some/any/each of John's large book
bags; see Jackendof 1977, pages 104f for some discussion. Other complex
ities arise in the handling of pronouns, such, defniteness, and so forth.
However, we will not discuss the intricacies of material to the left of posi
tion 2 in ( 1) ; our concern is with the complexities of the inner structure,
and their infuence on stress patterns.
1.1 A Sketch of Our Conclusions
We take a traditional view of modifers in positions 1 and 2 of ( 1) : a position
1 modifer combines with a noun to form a compound noun, whereas a
position 2 modifer forms a phraal category. Expressed in terms of X-bar
theory, position 1 is flled by modifers of
N
O
whereas position 2 contains
modifers of
N
1
. The structure of large exotic chess boards is thus:
(2) [
N
1
large [
N
1 exotic [
N
O chess boards]] ]
Position 1 modifers-CoMPOUND MODIFIERS-are thus adjunctions to
N
O
,
wherea position 2 modifers-PHRASAL MODIFIERS-are daughters of
N
1
,
modifing a right sister that is either
N
1
or
N
O
. 2
We shall agree with the traditional generative view (Chomsky and Halle
1968, Liberman and Prince 1977, Hayes 1980) that constructions involv
ing positions 1 and 2 in ( 1) are assigned diferent default stress patterns
due to the diference in category of their parent node. English stress is
normally assigned recursively to rightmost elements, but the stress rule
for nouns, simple or compound, will ignore a single non-complex element
at each level. This implies that
N
1
constructions are "right dominant"
i. e. , have main stress on the head noun-while
N
O
constructions are "lef
dominant"-i. e. , have main stress on the modifer-as long as the head is
2
We know of no distinction in types of modification that would motivate maintaining
these two alternative categories for the right sister of position 2 modifers, so one could
asume that such modifers are always adjoined to N
l
. Under that asumption, the
sister of exotic in (2) would be more correctly given a [N1 [NO chess boards]] .
STRESS AND STRUCTURE OF MODIFIED NOUN PHRASES / 133
a single word. We thus disagree with the view (Bolinger 1972, Ladd 1984)
that the diference between compound and phrasal stress is derivable fom
FCA efects.
Position 1 modifers are usually taken to be joined with their heads in a
separate lexical or morphological component, whereas position 2 modifers
are taken to be syntactic in origin; a recent expression of this view is given
in Di Sciullo and Williams 1987. We agree that constructions dominated by
NO are words, but we see no good evidence for assuming that they should
be considered anything other than syntactic constructions in English. We
shall return in a later section to a fuller discussion of the theoretical issues
in English noun phrae modifcation.3
We also agree with the traditional view that the two diferent types
of syntactic modifcation imply diferent sorts of semantic relationship as
well. However, a will become clear fom the discussion below, there does
not seem to be a single, clean semantic distinction such that all nominals
with left hand stress will fall into one semantic class and all nominals with
righthand stress will fall into the other.
The main novelty of our position is this: we argue that both position
1 and position 2 may be occupied by modifers of a wide variety of cat
egories. We show in particular that positions 1 and 2 may be flled by
both adjectives and nouns or phraes made up of these and other cate
gories, sometimes linked with the possessive 's. This implies that for any
modifer-noun sequence, both NO and N1 structures are always available in
principle. Thus the contextually appropriate parsing of such phraes may
require a judgment a to the relative plausibility of the semantic relations
implied by the structural choices. This is at least somewhat consistent
with Bolinger's ( 1972) dictum that "accent is predictable (if you're a mind
reader)" . However, we will suggest that there are some strategies that can
achieve fairly high accuracy without telepathic assistance.
1.2 The Problem
It is common in running text or speech in English to fnd PREMODIFIED
NOMINALS containing one or more modifers, such as the examples in ( 1 ) .
Depending on the style of the material and on the defnition of 'phrase',
something like 30 to 70 percent of all phrases can be expected to end in
such units. As a matter of practice, the location of the main stress in these
sequences is quite variable. Although the fnal noun is the commonest lo
cation, pre-fnal main stress i also quite fequent. Depending on the type
of material studied, somewhere between 10 and 60 percent of the premodi-
3
For those reaers familiar with the work of Abney 1987, we note that we shall use the
terms ' noun phrae' and ' NP' in their traditional sense for the bulk of this paper. Also,
we shall generally use 'Xp' (e.g., NP, AP, . . . ) to refer to the (contextually) maximal
projection of the category in question with no commitment to the bar level of this
projection.
fed noun constituents show pre-fnal main stress, with 30-40 percent being
typical of newspaper writing.
There are several sorts of reasons why the main stress in a noun group
might fall before the head. We distinguish caes where the preterminal
stress serves to underline an important word, or to avoid stressing a re
dundant one, fom cases where a preterminal stress pattern is normal or
natural for the phrase in neutral contexts of use. In the following example,
and henceforth in this paper, boldface is used, where needed, to mark the
word that bears the main stress:
3) a. Stress pattern "natural" for the phrae:
- At 9: 00, there is an important meeting.
. At 9:00, there is a staf meeting.
iii. John was wearing a red jacket.
iv. John was wearing a life jacket.
b. Stress pattern determined by "FCA" considerations
i. We're only concerned with solvable problems.
ii. He replaced his low-interest bonds with high-interest
bonds.
While FCA efects are not uncommon and cannot be ignored, it is nonethe
less true that FCA phenomena must in turn interact with other principles
such as the lexical stress pattern of polysyllabic words-to produce ob
served stress patterns; see Hirschberg 1990 for some recent discussion of
issues in modeling some FCA efects and their interaction with lexical stress
placement. In particular, a familiar generalization seems to underlie the
stress pattern of cases like those in 3a) , where [A N] phraes typically
show main stress on the head noun righthand stress) , while [N N] units
typically show stress on the noun in the lefthand position. These simple
generalizations hold true more often than not. We can provide a rough
experimental check by having someone read some text, and counting up
the stress patterns employed for constituents of the appropriate kinds. In
our experience, more than 90% of [A N] units will be read with right hand
main stress, while around 75% of [N N] units will show lefhand main
stress .
.
Although the contrary cases remain fairly fequent, we might suppose
that they represent the expected efect of the phenomena of FCA. Be
cause theories of FCA phenomena are not very well developed, it is hard
to check this notion with total assurance. However, one typically has some
feeling for the application of such analyses in particular cases, and we can
look for them in all the textual examples whose stress patterns are con
trary to the predictions of part of speech sequence. On this basis, FCA
type explanations do not seem to help very much, especially for the [N N]
anomalies.
The exact statistics depend very much on the style of the material
surveyed, but two sample caes will give the favor of the situation. In
one text (taken fom a book on computer vision) , 190 out of 214 [A N]
constituents showed right hand stress, while 76 out of 92 [N N] constituents
showed lef hand stress.4 A plausible explanation in terms of FCA could be
found for 18 of the 24 lef-stressed [A N] units, and for none of the 16
right-stressed [N N] units. Thus we are left with about 3% anomalous left
stressed [A NJ expressions and about 17% anomalous right-stressed [N N]
expressions. In a second text (several stories from the main section of the
New York Times) , 153 out of 169 [A N] constituents had the expected
right hand stress, while 102 out of 138 [N N] constituents had the expected
left hand stress. FCA-type explanations applied to 8 of the 16 left-stressed
[A N] units, and to none of the 36 [N N] units, leaving about 5% non
FCA lef-stressed [A N] expressions, and about 26% non-FCA right-stressed
[N N] expressions. In these two sample texts, only 60% of the "anomalous"
[A N] caes (and none of the anomalous [N N] cases) had a plausible FCA
explanation.5 This means that in 6. 7% of all [A N] sequences there was a
useful FCA explanation for the stress pattern (be it left or right) , and in
none of the [N N] caes.
So, while the stress pattern of an [X N] expression is strongly correlated
with the lexical category of the word preceding the head noun, there are
quite a few apparent exceptions, most of which cannot be attributed to
discourse efects. Theories which explain patterns of modifier-noun stress
mostly or solely on the basis of FCA factors (such as that of Ladd 1984) , are
unlikely to be correct. We shall return to this issue at subsequent points in
the paper. We devote the body of the paper to a survey of the "standard"
pattern, as well a the systematic classes of exceptions to it.
2 The Standard Pattern
Before we take up the minority cases of left-dominant [A N] expressions and
right-dominant [N N] expressions, it will be helpful to survey the standard
forms of these constructions. Nearly all of the material discussed in this
section is familiar fom the literature, but since the full range of issues
is rarely found discussed in one place it seems useful to review the facts
here.
4We require that both the head noun and its modifier be content words, thus disallowing
anaphors and excluding a few easily predictable [A N] FCA examples involving head
words like one.
5
Maidment ( 1989) ha noted that news reporters often tend to place right hand stress
on [N N] sequences that he feels should properly be left-stressed, and speculates (p. 187)
that "this feature of broadcat speech is due to a desire . " on the part of the broadcaster
to defer the intonational 'payof' for a long a possible in order to create suspense or to
make the news item sound more portentous." Indeed, Bolinger ( 1972, p. 643) notes that
similar efects can even interfere with normal lexical stress for non-compound words.
This possibility adds another possible dimension to FeA efects on which we will not
comment further.
2.1 Standard Compound Nouns: NO [N N] Expressions
It is a well-known fact that English permits the fee formation of compound
nouns, the commonest type being of the form [N N]. The orthographic con
ventions for encoding compounding are varied. Short, frequent or fossilized
compounds are often written a single words:
(4) drugstore, icewater, postman, gamecock, basketball, bathroom, ash
can, bartender, poolhall, earwax, lawnmower, fagpole, marshmal
low, wallpaper, keyboard, waveform
Sometimes (especially in attributive position) a hyphen is used, and some
times the two words are written separately: masthead, mast-head, mast
head. The correlation between typographical practice and the semantic
regularity of the resulting compound is at best imperfect. Some com
pounds that are commonly written as single words have a meaning that
seems fairly compositional (5a), while other cases that must be written a
two words have a special meaning that surely must be lexically listed (5b) :
(5) a. phonecall (a call on the phone)
b. overseas cap (a particular style of cap)
The fequency of the compound, and the length of its constituent words,
seem to be more important factors than semantic compositionality.
To some extent, the typography is a matter of style. Some writers prefer
words written solid, while others like hyphens, or fnd pleaure in spaces.
However, the typographical usage in actual text is quite variable. Even
individual writers are not always consistent-within a single chapter of a
textbook on the automobile electrical system, we have found spark plug,
spark-plug, and sparkplug. However they may be spelled, such compound
nouns are generally pronounced with primary stress on the frst element.
Indeed when the compound is run together or spelled with a hyphen, frst
element stress is almost inevitable.
Because noun compounding is an easy way to create terms of art, tech
nical writing, in the broad sense, is especially prone to such coinages. How
ever, ordinary life also provides plenty of examples-a few days of recording
those that came up in reading and listening produced over 5,000 examples,
a few of which are given below:
(6) drug abuse, line backer, tool cabinet, feat day, knife edge, crop
failure, dart game, body hair, shoe imports, phone jack, tea kettle,
heat lamp, utility man, node name, post ofce, soap pad, printer
queue, boat race, cocktail sauce, folk tale, land use, property value,
star wars, junk yard, combat zone
2. 2 Semantic Relations in NO Compound Nouns
The productive types of NO [N N] compounds fall into a number of diferent
categories. As an initial cut, we will distinguish two broad classes on the
bais of the semantic relation that holds between the two nouns. In the frst
type, a paraphrae of the meaning has the frst noun providing an argument
for a predicate associated with the second (head) noun; we will call this
type ARGUMENT-PREDICATE compounds. A typical example would be lion
tamer, paraphrased as 'one who tames lions' . Extensive work has been done
within theoretical linguistics on these so-called SYNTHETIC COMPOUNDS;
see for example, Roeper and Siegel 1 978, Selkirk 1982, Lieber 1 983, Sproat
1985, Levin and Rappaport 1992, Marantz 1989. All of this work treats the
relation between the lefthand member and the head in argument-predicate
compounds by analogy to the relation between verbs and their objects.
In the second type of compound, paraphrases of the meaning involve
a predicate not implicit in either word, with a meaning like POSSESSION,
PURPOSE, etc. , for which the two elements of the compound provide argu
ments. We will call this type ARGUMENT-ARGUMENT compounds. Some
examples are pie chart, paraphrased as 'a chart that is like a pie', and
keyhole saw, paraphrased as 'a saw used to make keyholes' . The seman
tics of these has been much less extensively studied, but Lees 1960 and
in particular Levi 1978 are two works which discuss these compounds at
length.
2. 3 Argument-Predicate Compounds
The head noun in the argument-predicate type of compound may be an
"agentive" -er nominal, a gerund, a derived nominal, or a noun without a
verbal counterpart whose meaning nevertheless seems to put it in this cat
egory. A generalization which holds over all of these cases is that when the
left hand member is assigned an internal thematic role-typically whatever
argument is normally assigned to the direct object of the verb fom which
the compound' s head is derived-the main stress is on the lefhand member
of the compound. The compound is therefore an NO by our assumptions.
Compounds with "agentive" heads. By these we mean, of course,
examples such as the following, where in each case the lefhand member is
interpreted as the internal argument of the verb fom which the right hand
member is morphologically derived via afxation of -er:
(7) shock absorber, torch bearer, syntax checker, car dealer, grain ex
porter, fre extinguisher, lens grinder, door knocker, deer hunter,
rocket launcher, steel maker, can opener, music publisher, paint re
mover, knife sharpener, opium taker, window washer
It is easy to fnd (or think of) hundreds of other natural and familiar
seeming examples. There are a number of cases where quite large sets of
objects will go nicely with a particular head. In the cases in (8) , X could
quite plausibly be any one of thousands of things:
(8) a. X maker: X anvil, arrow, battery, bobbin, buckle, button, car,
carriage, cheese . . .
b. X supplier: X asbestos, cable, copper, leather, lumber, paper,
pipe, steel . . .
c. X collector: X art, book, car, clock, coin, mirror, pottery,
stamp . . .
There are relatively few values of X for which it might not, under some
circumstances, be appropriate to speak of an X maker, an X supplier, or
an X collector.
All of the above agentive compounds have left hand main stress since
their left hand members function as internal arguments to the head. Need
less to say, any agentive nominal that can occur alone, can also occur in
constructions in which a preceding noun does not function as an object for
the agentive's underlying predicate, but instead has one of the other sorts
of relation possible in noun phrases. These constructions may show main
stress on either member, depending on their nature:
(9) a. Left-dominant cases: ghost writer (does not write ghosts), base
ball writer (does not write baseballs)
b. Right-dominant cases: Virginia creeper (does not creep Virginia),
ga drier (does not dry gas), girl swimmer (does not swim girls)
This is expected given that agentive nouns are (afer all) nouns, and should
therefore be able to function like non-agentive nouns no matter what other
properties they may have. So, parallel to girl swimmer there are other
right-dominant N
1
constructions with appositive interpretations, such as
boy athlete. Examples such a those in (9) cause problems for computa
tional analysis, since it i often hard to be sure that an argument-predicate
interpretation is wrong.
Agentive-headed argument-predicate compounds are not ordinarily well
formed if the left member is a meaure noun or other pseudo-object. Nor
is there normally any way to incorporate arguments whose expression nor
mally requires a preposition or particle ( though one fnds a few examples
like city-dweller, church-goer and looker-upper):
( 10)
*
This meeting looks like a day-laster.
*
Smoking-stoppers tend to be irritable.
*
a water-looker-for
*
a for-water-Iooker
*
a water-looker
The lefthand member of agentive compounds may be modifed in vari
ous ways, but it usually may not be quantifed, usually may not be plural,
and certainly may not have its own determiner or other noun phrase speci
fers. Absence of determiners and other specifers for modifers is generally
required in English, and this is a point we will take up again later. On its
face, this suggests that the modifer for agentive nominals may be either
NO or Nl, but no higher projection of N (and no projection of D) is allowed
(see Fabb 1984) :
( 11) a toxic waste dumper
*
a three dog owner
*
a dogs owner
*
a the Grand Canyon admirer
Compounds with gerunds as heads. English gerunds participate in a
variety of constructions; in the one that concerns us here, a gerund is the
head of a compound noun, whose lefthand member (usually) functions as
the object:
( 12) cattle breeding, ale brewing, carpet cleaning, number crunching, cost
cutting, tape dubbing
Note that these caes are sometimes difcult in practical contexts to dis
tinguish fom the phraal NP V-ing construction and it is certainly not
difcult to construct genuinely ambiguous examples:
( 13) The man eating shark was repulsive.
Sometimes an -ing nominal acquires an additional meaning that partly
or entirely supersedes the act/process one; this meaning often denotes the
result of the act or process-e. g. , dropping, building, writing-or some
materials or methods central to the process or action-caulking. Such
caes do not usually form compounds of the type exemplifed in ( 12), al
though (like all nouns) they participate feely in other sorts of compound
ing and modifcation. That is, a preceding noun will not serve as object
if a gerund head is used in the "result" sense. Thus a mouse dropping
is normally a dropping that comes fom a mouse. If punk pilots adopted
the practice of showering public gatherings with thousands of live mice,
mouse dropping would be a natural way to refer to this unnatural act, but
would not serve to denote its pitiful result, which we would have to call
dropped mice.
As in the case of agentive-headed compounds, the argument-predicate
compound constructions are typically ill-formed with pseudo-objects or
with arguments expressed by means of prepositions:
(14)
*
He is capable of day-waiting.
*
Cigarette-quitting is hard work.
*
Oil-drilling-for is chancy.
Argument-predicate compounds with derived nominals B heads.
As is well known (see Chomsky 1970, Thomason 1985, Sproat 1985, Safr
1987, among many others) , derived nominals can have meanings that relate
to an act or process associated with the related verb, or to the cause,
instrument, method, resulting state of such an act or process. The second
type of meaning-result, etc.-seems more erratic, while the frst type
act, process or event-is more regularly found and its gloss is more easily
predicted. We will follow such previous work by distinguishing these two
types of meanings as PROCESS NOMINALS and RESULT NOMINALS.
In a compound noun headed by a process-nominal, the lef member may
correspond to a noun phrase in various syntactic relations with the related
verb:
( 15) a. Subject: cell division, commando raid, sunrise, police action,
snake bite, proton decay, ether drif, moon glow
b. Object: nest construction, dream analysis, haircut, birth control,
dress design, steel production, ofce management, heart massage,
heat regulation, trash collection
The subject types ofen have (though clearly do not always have) right
dominant stress patterns:
( 16) enemy invaion, police intervention, staf attempts, faculty decision,
student inventions
Argument-predicate compounds without a deverbal head? There
are many examples of [N N) compounds whose head is not deverbal (and in
deed may not have any corresponding verb) but seem nevertheless to corre
spond semantically to a predicate that takes other nominal constituents as
arguments. Such words seem to form compounds of the argument-predicate
type. Sometimes there is a corresponding expression with a postnominal
PP expressing the argument, and sometimes not:
( 17) expert in ballistics ballistics expert
critic of music music critic
format of data data format
department of chemistry
?broker of commodities
*
bufof opera
*
thief of cars
chemistry department
commodities broker
opera buf
car thief
Of course, a we shall see in Section 2. 4, the range of argument-argument
compounds is so broad that it i hard to be sure that examples such as
those in ( 17) are not included in it.
Compounds whose lef member is self. Compounds headed by agen
tives and derived nominals can feely occur with the word self a lef hand
member. In all caes, the normal stress pattern is right dominant, for
the same reasons-presumably FCA reasons-that refexive pronouns are
generally deaccented in phraal contexts. Compounds headed by V +ing
nominals do not so eaily take self as a left member, although the corre
sponding compound adjectives are common:
( 18) self-starter, self-promotion, self-igniting
2.4 Argument-Argument NO Compounds
We now turn to NO compounds where a paraphrase links the two words in
the compound with a predicate not implicit in either one. We are limiting
this category to endocentric compounds, so that their English paraphrase
will be somethlng like ' an Nl N2 is an N2 relative-clause-containing-Nl ,
e. g. , 'an ankle bracelet is a bracelet that is worn on the ankle, ' or ' rub
bing alcohol is alcohol that is used for rubbing' . The range of predicates
implied by such paraphrases is very large. Since this type of compound
formation can be used for new coinages, any particular compound will
in principle be multiply ambiguous (or vague) among a set of possible
predicates.
Consider hair oil versus olive oil. Ordinarily, hair oil is oil for use
on hair, and olive oil is oil derived fom olives. But if the world were a
diferent way, olive oil might be a petroleum derivative used to shine olives
for added consumer appeal, and hair oil might be a lubricant produced by
recycling barbershop foor sweepings.
A coherent categorization is hard to fnd. In their everyday mean
ings, the compounds olive oil and hair oil resonate with many similar
examples:
( 19) a. Like olive oil
pattern 5OURCE-PLANT 5UB5TANCE-THEREFROM-DERlVED
peanut oil sesame oil safower oil soybean oil corn oil
palm oil cottonseed oil corn syrup bean paste wheat bran
wheat four chickpea four rice four wheat bran barley malt
corn starch mango pulp orange juice lemon juice apple juice
carnauba wax guar gum beet sugar cane sugar pine tar
b.
hand cream
eye drops
Like hair oil
pattern BODY-PART 5UB5TANCE-THEREON- U5ED
skin cream face powder foot powder
nail polish underarm deodorant hairspray
lip gloss
toothpaste
Such broad resonances have been the driving force behlnd clasifcatory
schemes for argument-argument compounds of the type most fully devel
oped in Levi 1978. A practical problem, discussed by Downing ( 1977) , is
that the set of patterns that would be required to achieve complete coverage
appears to be open ended.
A more fundamental problem is that there does not seem to be any
non-arbitrary way to decide on a single, coherent categorization of such
patterns. The olive oil type of pattern exemplifed in ( 19a) ,
SOURCE-PLANT SUBSTANCE-THEREFROM-DERIVED
covers substances such as oil, four, syrup, bran, juice and so on, which
are extracted by a variety of methods, use various portions of the source
plant, and so on. This much generalization seems unproblematic-almost
everyone would agree that these examples are instances of a type. However,
the set can be extended in many directions. Each extension seems natural,
but the resulting set is less and less coherent. One set of extensions (20a)
(20c) gradually relaxes the limitation of N1 to particular plants, allowing
for plants in general, to living things in general and fnally to inanimate
sources of materials; another (20d) allows N2 to be a count rather than a
mass noun:
(20) a. SOURCE-PLANT is a more general category:
fuit juice, vegetable oil, grain alcohol
b. N1 is an animal:
whale oil, fsh oil, chicken fat, horse hair, goose down
c. N1 is inanimate:
rock dust, river water, bread crumbs
d. N2 is a subpart that can be extracted or removed rather than a
derived substance:
peach pit, chicken wing, rose petal, pine cone, coca leaf, corn cob,
fsh scale, corn husk, peanut shell
Plausibly, a general pattern to cover all the cases in (20) and the
SOURCE-PLANT SUBSTANCE-THEREFROM-DERIVED
instances in ( 19) would be
SOURCE SUBSTANCE-OR-PIECE- WHICH-COMES-FROM-SOURCE,
in which COMES-FROM includes derivation of a "new" substance by press
ing, grinding, and so forth, of the source, and also by separation of a part
from the whole.
Now consider the common cases where N1 is the whole of which N2 is
a part:
( 21) tire rim, mountain peak, arrow head, door knob, bed post, piano
keys, shirt sleeve, table leg
These examples certainly are analogous to peanut shell or or fsh scale, since
both ft the rough schema N2 is part of Nl. The examples like olive oil,
in turn, are analogous to peanut shell, since both ft the pattern N2 comes
from Nl . But it seems less plausible that olive oil and mountain peak are
in the same category. Each example, taken as a nucleus of generalization,
yields several sets with which it shares some properties, but which may not
share any properties with one another. A related example will give us new
sets of neighbors, which overlap with the previous sets but are not exactly
the same. Because of arguments like this, we doubt that an approach
such as that proposed in Levi 1978 can be coherently and systematically
pursued.
The ' connected-with' theory. As several writers have noted-e. g. ;
Dowty ( 1979)-the facts are consistent with a linguistic rule of argument
argument compounding that contributes only a vague 'connected-with'
predicate, the more specifc meanings arising fom lexicalization and from
the usual contextual circumscription of linguistically vague expressions.
On this view, there is no well-defned hierarchical categorization of such
examples, since a given form may have analogical connections in many
directions.
Thus the core meaning of olive oil would be something like 'oil con
nected with olives' , and hair oil would be 'oil connected with hair' . In
the case of these particular expressions, it is likely that their more spe
cifc meanings should be lexically listed. Novel caes-e. g. , thistle oil or
moustache oil-cannot get more specifc meanings fom a direct dictionary
entry, but they can get a predisposition to go in one direction or another
from analogy with the patterns in ( 19), the meanings of their component
words, and common sense.
Words and fxed phraes easily acquire special meanings; utterances are
always interpreted in context; and analogy with fxed expressions is a pow
erful determinant of everyday phraeology. So the mechanisms required
by the connected-with theory are in any case available. By the nature of
this theory, it cannot be disproved by positive examples, since (by some
argument or another) it licenses any compound in any meaning. To show
that this theory is wrong, we must show that some meaning relations are
systematically excluded, and that some alternative hypothesis will distin
guish those that occu fom those that don' t. This is nearly impossible to
do without a precise account of such meaning relations,6 which no one can
at present provide.
A survey of argument-argument compounds. Our task is to distin
guish the [N N) constructions with left hand stress (NO) from those with
righthand (N1) stress. Since the part-of-speech categories give us no help,
any successful algorithm must rely in part on the semantic relationship
between the words. The argument-predicate compounds previously con
sidered were nearly all left-dominant. In this section, we will survey some
common types of argument-argument compounds, limiting consideration
to examples that are also left-dominant. Afer right-dominant expressions
of form [N N) have been treated in a later section, we will discuss possible
sources of the distinction.
Since any [N N) form could in principle have many meanings, we have
chosen examples that rely on the existence of a meaning that is fozen
through common use, that is implied by common sense given the word
meanings, or that is strongly preferred due to the existence of a "schema"
emerging fom the analogical force of many related forms. It is an open
question if such schemata achieve the technical status of linguistic "con
structions" or formation rules, whether of morphology, syntax, or seman
tics. In practice, new instances of such compounds may usually be coined
freely, whether by analogy or by rule. We divide the cases into categories
that are designed mainly for expository convenience, and that are defnitely
far fom complete. Fudge 1984 contains discussion of some of these classes.
6
Though see Downing 1977 for some discussion.
(22) pattern ELEMENT-IN-CAUSAL-CHAIN RESULT:
drug deaths, heat rash, job tension, snow blindness, food poisoning,
hay fever, tire track
For some reason, the head nouns in the pattern in (22) are usually nega
tive in connotation. Persuasive examples with positive heads are hard to
construct, and are generally right dominant, as in the example for your
listening pleasure.
(23) pattern TYPICAL-RESULT-OR-PRODUCT ELEMENT-IN-CAUSAL
CHAIN:
disease germ, polio virus, growth hormone, honey bee, silkworm,
song bird, oil well, sob story
(24) pattern SUBTYPE TYPE:
a. Nl a name, N2 virtually redundant (hence the construction is
almost exocentric):
pine tree, cactus plant, collie dog, lilac bush, ivy vine
b. Nl specifes species or type:
preacher man, tree shrew, girlfriend, man servant, sports activ
ities, tape measure, soldier ant
Note that polio virus is like pine tree, in that N2 is a generic biological
classifer and Nl is a name for a particular type; however, a polio virus
causes polio but is not a polio, while a pine tree does not cause pine, but
is a pine.
(25) pattern
Nl
THING-MADE-OUT-OF-
Nl
:
daisy chain, cable network, mountain range, grape arbor, chocolate
bar, snowball, sugar cube (but cf. (68))
(26) pattern POWER-SOURCE MECHANISM:
steam engine, water wheel, vacuum cleaner, air brake, cable car (but
cf. (75) )
(27) pattern INSTRUMENT ACTION or INSTRUMENT RESULT
head butt, sword thrust, tank attack, knife wound, pot roast, pan
fies, shovel cakes (but cf. (75))
(28) a. pattern PLACE- WHERE-
N2
-IS-FOUND
N2
:
feld mouse, mountain lion, desert rat, sea cruise, surface tension,
farm boy, marsh gas, feld trial, pond scum, house fy, street cop,
city folk, alley cat
b. pattern TIME- WHEN-
N2
-ocCURS
N2
:
morning sickness, night blindness, Easter bunny, birthday party,
Christmas present, morning coat, Sunday school (but cf. (70))
(29) pattern
Nl
SOMETHING-WITH-PURPOSE-ASSOCIATED-WITH-
Nl
:
horse doctor, nose drops, fy paper, test pattern, cooking utensils,
arms budget, plant food, face towel
(This is a large and diverse set that could be further subdivided. )
(30) a. pattern THING-CONTAINED CONTAINER:
picture book, photo album, gunboat, bear country
b. pattern CONTENTS CONTAINER:
parts bin, gin bottle, butter dish, olive jar, milk can, juice glass,
garbage can, water bucket, stock drawer, linen closet, cofee mug,
laundry basket, punch bowl, water tumbler, oil tank
c. pattern
Nl
SHELTER-OR-TRANSPORTATION-FOR-
Nl
:
dog kennel, horse barn, cow shed, guest room, hay loft, corn crib,
grain elevator, hay wagon
( 31) pattern SUBJECT-MATTER THING-FOR-WHICH-SUBJECT-MATTER
IS-RELEVANT: tax law, abortion vote, budget debate, adventure
story, love song, detective novel, oil crisis
(32) pattern
Nl
THING-LIKE-
Nl
:
catfsh, dragonfy, hermit crab, spider monkey, garter snake, kettle
drum, fogman, cat burglar
Types and subtypes could be multiplied. For the reasons that we gave
earlier, the categories are not disjoint and do not have sharp boundaries.
2. 5 Compound Nouns whose Heads are not Nouns
English can also make compound nouns from various combinations of verbs,
prepositions and nouns; see, for example, Selkirk 1982. Almost always,
the stress pattern of these constructions is left-dominant. Alo, most are
written with a hyphen or as a single typographical word. These cases
mostly fall outside the topic of this paper, but we list them briefy for
completeness.
Phrasal verbs (verbs that combine with a "particle" or intransitive
preposition) are nominalized feely. Common examples often have an id
iosyncratic meaning:
(33) fall-away, run-away, throw-away, tear-away, slow-down, melt-down,
come-down, splash-down, walk-out
Prepositions can also combine with verbs in the opposite order, although
this type of combination is less productive:
(34) downdraf, downfall, downpour, downturn, downtrend, outburst,
outbreak, outcast, outlook, outreach
Erratically, a verb may combine with a noun to form a noun. One
[ V N] pattern produces an exocentric compound whose referent is the ( un
expressed) subject. This type is common in Romance languages but not
in Germanic ones. It is no longer very productive in English, although it
seems to have been popular in earlier times: `
7
Marchand ( 1969, pp. 37-39) makes the interesting observation that such constructions
are almost invariably pejorative, which seems to be true of the examples in (35) with
the exception of drdnought , which might explain its tendency to be less well-known.
(35) cut-throat, pick-pocket, pinch-penny, do-nothing, kill-joy, know-
nothing, dreadnought, sawbones
A second [ V N] compound type is endocentric (that is, the referent is
identifed with the head noun) . The semantic relation between the verb
and the head noun is quite varied. Furthermore, it is often difcult to be
sure that the lef hand member is indeed a verb and not the homophonous
noun. New examples of this type of compound are fequently coined, but
not all attempts are plausible. Thus a punch intended to stun might be
called a stun punch, but a remark intended to ofend could not be called
an * ofend rmark:
(36) searchlight, dancehall, springboard, workbench, grindstone, play
thing, push pin, stuf bag, dive plane, trim tab, thrust plate, stab
wound, snap bean, pitchfork, stun gun, hit man, push rod, lock
waher
2. 6 Compounds with Complex Parts, and the Compound
Stress Rule
We have discussed binary nominals which are syntactically NO and which
are therefore asigned lefthand stress. We now wish to consider the stress
properties of NO constructions which have more than two leaf nodes. It
is common to fnd compounds made up of subconstituents with internal
structure. When the frst member is itself a noun compound, and the
second member is lexically simple, we gmerally fnd stress on the left:
(37) Air Force Academy, football game, money-market account, post-
ofce box, data acquisition board, fahlight battery, X-ray flm, bit
vector machine, fund-raising operations
In each of the examples in (37), a compound noun is combined with an
other noun on its right, in a second level of compounding. In the result,
the main stress is typically on the leftmost element. Sometimes, we fnd
another recursion on the same pattern, in which a doubly compound left
hand member is further compounded with a single word after it. These are
mercifully rare:
(38) water supply network repairs, windshield wiper blade replacement,
error correction code logic
It is also possible to combine a noun on the left with a compound noun
on the right, in a right-branching structure. These are perhaps slightly less
common than the comparable left-branching structures shown in (37). 8
They often occur with a compound lef hand member as well:
(39) radio direction fnder, spark-plug heat range, sink spray head, VAX
instruction set
8
For Swedish, Biaberg (1988, p. 68) claims that the right branching structure is signif
icantly less common than the left branching structure in ternary nominals.
In such constructions, the primary stress is typically found on the penulti
mate member. In general, for a noun compound [Nl N2J , if N2 is not itself
a compound, then it will give up its claim to main stress in favor of Nl.
However, if N2 is a compound, then it will retain main stress within itself.
This rule was called the COMPOUND STRESS RULE (CSR) in Chomsky and
Halle 1968; we shall also use the traditional term NUCLEAR STRESS RULE
(NSR) to refer to "phrasal" or righthand stress.
Liberman and Prince ( 1977) suggested that this rule was the same one
that assigns main stress in nouns in general. As they observed, the right
hand element of a noun is stress-dominant if and only if it branches, where
the nodes of the relevant tree are words in the case of compound nouns,
and syllables or feet in the cae of simple nouns. Hayes ( 1980) suggested
that this regularity should be recat as a simple fnal-stress rule, with the
rightmost unit ignored (extrametrical) in the case of nouns. Again, one
(simple) word is ignored i the cae of compounds, and one syllable in the
case of simple nouns. Under any formulation, the rule may be applied recur
sively to generate patterns of stress. (The efects of the so-called RHYTHM
RULE, which shifts some non-fnal stress relations in order to create a more
regularly alternating pattern, must also be taken into account. )
Across many types of noun sequences, the predictions of the CSR are
fairly well verifed. In (40) we give a set of typical examples containing
four or more nouns, with the contextually implied grouping indicated by
parentheses. The reader is invited to consider the predictions of the CSR,
and compare them to his or her own intuitions about how the examples
should be read:
(40) a. [[starter [drive gear]] clearance]
b. [ [[gear selector] [control rod]] adjustment]
c. [ [[power generating] station] [[control room] complex]]
d. [government [ [tobacco [price support]] program]]
We note in passing that the CSR's determination of main stress position
underdetermines the phrase structure in such cases-for instance, the pairs
of structures shown below yield the same main stress location (marked in
boldface) in each case:
(41) [A [[B C) Dll [[A [B ell D)
[A [B [C D)]] [[A B) [C Dll
Due to the semantically protean nature of the compound construction,
it is also not always clear what the structure of complex compounds is.
Thus tobacco price supports might be price supports for tobacco, or (almost
equivalently) supports for tobacco prices. Putting the main accent on price
(which seems to be most people's preference) settles the matter in favor of
the frst option. The example government tobacco price support program,
however, might be either [government [[ tobacco [price support] progrm
)
]]
or [[government [ tobacco [price support]] program
]
] , without any efect on
the main stress location. Of course, the situation is usually much clearer.
In the context of use from which we took it, the parsing of starter drive
gear clearance is unambiguous.
Having now considered some of the types of NO [N N] constructions in
English, and further elaborating on the stress facts for such constructions
which have more than two leaf nodes, we return to the examination of
binary modifer-noun constructions.
2. 7 Adjective-Noun Expressions
The canonical sequence [A N] difers fom [N N] in that its parent is N1
rather than NO . Usually an [A N] sequence takes main stress on the right,
in contrast to the left-stressed pattern of the [N N] sequences in (6):
(42) clear advantage, miserable bastard, comfortable chair, small dif
culty, bitter end, special feature, ample girth, heavy heart, complete
idiot, last j udgment, fuorescent lamp, slender margin, arrant non
sense, smooth operation, wonderful person, stupid question, hasty
retreat, great skill, worthless trash, enriched uranium, macroscopic
Volvox, useful widgets, rare yak, melodious zither
2.8 An Explanation for the Basic Pattern
The High Road of Phrasal Stress is: CSR in [N N] , NSR in [A N] . Syntactic
category is a very good predictor of stress patterns for [A N] , and a fairly
good predictor for [N N] . Given this observation, the simplest account
would appear to be one which derives stress placement from some syntactic
property of the construction. The account we have been assuming wherein
CSR applies in NO constructions and NSR in N1 constructions, is just
such an account. To be sure, there are complications with this story. One
complication is that we must assume that exceptions to the general pattern,
to which we turn in the next section, are exceptional from a syntactic rather
than purely fom a prosodic point of view. However, as we shall see, the
alternatives to this assumption seem even less attractive.
3 Exceptions to the "Standard Pattern"
We will present a catalog of exceptions to the High Road to Phrasal Stress,
with some commentary on its meaning.
3. 1 Lef-dominant [ A N] Sequences
Lexicalized [A N] sequences. Some of the left-stressed [A N] cases are
apparently lexically re-analyzed a nouns-i.e. , as NO. Some of these are
written as a single word, or separated by a hyphen rather than a space; as
with [N N] compounds, the typographical conventions are variable. Some
of these cases are metonymic epithets
9
hardhat, red neck-but many are
not:
g
Ofen known U bahuvrihi compound, afer the traditional Sanskrit terminolog.
(43) square knot, sof spot, sticky bun, wild man, tonic water, pink slip,
red line, black belt, blue book, golden boy, cold chisel, blind spot,
sweet spot, dry dock, straight edge, whitewah, clean room
Examples like those in (44) below, which take the usual NSR pattern (unless
contrast or whatever intervenes) , show that the CSR does not automatically
apply to fxed expressions of the form [A N) , even if their meanings are
partly or entirely idiosyncratic:
(44) white lead, red oak, green onions, brown sugar, Irish stew, hard
liquor, red herring, brown betty, blue moon, black death, musical
chairs, crazy eights, heavy metal, old maid
In order to attribute the stress pattern diference to the parent category,
we must accept that the expressions in (43) are complex words of category
NO, while those in (44) are fxed expressions or semantically idiosyncratic
phaes of category Nl . The alternative, however, is to add some feature
like [CSR) to the lexical entries of phrases.
For a few phrases of this type-indeed, for nominals of almost any
type-some speakers have lefthand stress while others have righthand stress
or can say the phrase either way. This demonstrates that any descrip
tion must make it possible to efectively specify the stress pattern of such
phrases without changing their meaning very much. However, there are a
number of subregularities suggesting that the stress pattern in these exam
ples should normally be determined fom the parent category, so that the
left-dominant expressions will be of category NO, while the right-dominant
ones will (mostly) be of category Nl . In particular, constructions for which
it is reasonable to assume an NO analysis in the general case also tend to
be pronounced with the stress predicted by the CSR. For example, [A N)
sequences that are adopted as epithets (vocative or otherwise) tend to ac
quire compound-type stress, even if the same phrase can easily also be used
a a normal phrase with righthand stress:
(45) Epithet context
Hey, big mouth!
Hey, fat head!
No more Mr. Nice Guy!
Non Epithet context
You've got a big mouth.
You've got a fat head.
You're really a nice guy
A reasonable analysis of such cases, it seems to us, is that epithet-formation
involves turning a phrase into a word-an NO-and that it is this fact alone
which accounts for the typical accentual pattern of such phrases.
Brand names, place names and plant or animal names with [A N] struc
ture also have quite a strong tendency to develop left hand stress. This
tendency is especially pronounced for metonymic names, but many others
show it as well:
(46) a. Long Beach, Hungry Horse, Red Wing, Old Town
b. Big Boy, Redbook, Green Stamps, Pure Aire, Smooth Pore, Kleen
Flex, F Flo, Final Air, Dri-Pak, Hi-Flo, E Z Kleen, Tiny C, Soft
Cote, Common Lisp, Rite Aid
c. yellowtail, bluefn, bluebird, blackbird, redwood, yellowjacket,
bluebell, bluefsh, redfsh
Again, it is reasonable to asume that in the general case, such terms have
become lexicalized and re-analyzed a NO, and that this fact accounts for
their accentual patterns.
Right-dominance is normal, however, in plant and animal names that
are semi-compositional; that is, where the right hand element is a class
name such as oak and the lefthand element is a descriptive adjective such
as red. By our analysis, such cases would have to be analyzed a N1 . This
pattern generally requires a head of some specifcity, such as leek or ea
glemore general heads such a fsh or bird are more likely to be found in
the class (46):
(47) red oak, curly dock, wild leek, bushy aster, common fa, bald eagle,
American coot, spotted rail
Some place names seem also to be semi-compositional, in this sense:
(48) Big Bend, American Falls
Finally, although most [A N] brand names are lef-dominant, a exempli
fed in (46), a few are right-dominant. Many of these have polysyllabic
right hand members; a few may be considered semi-compositional:
(49) Orange Julius, Best Western, Dutch Masters, Big Mac, Leading
Edge, Silver Cloud
We suggest that the nominals that we are calling "semi-compositional"
retain their N1 structure, despite becoming a fxed expression with a special
meaning.
What is lexicalization? Note that not all lexicalized two-word nominals
wind up with lef-word main stress. In connection with the issue under
discussion, there are in principle three types of lexicalization that a phrae
can undergo.
First, a phrase may acquire a special meaning or asociation that re
quires it to be listed in a phraal lexicon of some sort (see Di Sciullo
and Williams 1987) . Thus the compositional meaning of some topical
phrase, say mutually assured destruction or album-oriented rck, is not
enough to pin down the extra associations or technical meanings that it
has acquired. Under the present analysis, however, the fact that these
phrases are normally given phrasal stress means that they must be (mini
mally) Nl .
Second, a phrase may be treated as dominated by a lexical category,
although presumably retaining its internal structure, or at least its divisions
into words.
Finally, a former phrase may loose some or all of its internal structure
as well as becoming a lexical category. There is no question that this long
ago happened to Worcester on its way to being pronounced as if it were
written <wuster>. A more current example is high school. One of the
authors speaks a dialect that raises and fronts the diphthong [ail when it
precedes a voiceless consonant within the same word, so that the vowel in
ffe is quite diferent fom the vowel in fve. This raising applies obligatorily
to the frst vowel in high school, but not typically to the frst vowels of,
say, fy paper or pie plate, suggesting that the common word high school
ha lost its internal word boundary.
There is an interesting twist to this account which we merely mention
here. As we have noted, the normal rules of English word stress, for nouns
without internal word-level structure, say roughly to put the main word
stress on the last non-fnal heavy syllable. Thus for an (originally) two-word
phrase whose second word has at least one post-tonic syllable, the stress
pattern predicted for the third stage of lexicalization, in which the internal
word boundary is lost, is in efect, main stress on the second word. So if,
for example, red herring were a noun without internal word boundaries, its
main stress location would be the same as if it were a noun phrase. The
preponderance of monosyllabic or fnal-stressed head words in (46), and of
trochaic head words in (49), suggests that these lexicalized [A N] sequences
may generally lack internal word-level structure, at least for purposes of
stress asignment. That is, one can readily explain the distribution of stress
patterns most simply by considering such caes to belong to the third stage
of lexicalization outlined above.
Adjectives as substantives in [A N] "compounds." In other cases,
left-dominant [A N] sequences seem perhaps to involve an adjective used
as a substantive; see Levi 1978 for further discussion:
(50) athletic facilities, circulatory system, regulatory agency, electrical
tape, erogenous zone, legal work, medical department, primary
school, professional organization, solar system, tidal zone
All these are plausibly instances of productive patterns, in which a variety
of other adjectives and nouns can occur in the lefthand position, with sim
ilar meaning. In some cases-e.g. , prfessional-the adjectival form can
also be used a a substantive, but often there is no common independently
usable substantive-e. g. legal , or the independent substantive has a dif
ferent meaning-e.g. , primary. Thus we cannot claim that the lefthand
member of the examples in (50) is a true, independent substantive-a noun
that happens to look like an adjective. However, the semantic connection
between the two words in these examples seems similar to the relations
that apply in analogous [N N] examples with left-dominant stress. So, the
phrase medical building is essentially parallel to chemistry building or of
fce building, except that 'medical stuf' has to be understood for medical.
Presumably the word medicine is avoided because of the confusion with
the kind of medicine that comes in bottles.
The inadequacy of "implicit contrast": further arguments against
FCA-only theories. We might try to explain left-stressed examples like
those in (50) in terms of an "implicit contrast" between the modifer and its
alternatives. This idea has no predictive value: there is no greater amount
of "implicit contrast" in legal work than there is in lef lane. To use the
phrase lef lane almost surely implies the alternatives right and center,
and yet the main stress still falls on the second word. Here are some other
examples of normally right-dominant [A N] sequences where the adjective
is in "implicit contrast" with a small set of known alternatives:
( 51) silver medal, heavy crude, young man, big business, large family,
fesh bread
Without a defnition of "implicit contrast" that can distinguish between
the examples in (50) and those in ( 51) , such explanations must be rejected
as vacuous.
In some of the examples in (50) the head noun is a relatively empty
one, so that one might appeal to some sort of redundancy-triggered deac
centuation. However, other [A N] expressions, with head nouns that seem
equally empty, work more normally (i. e. , by the NSR) . In (52) we have
tried to use heads that are both general and common, yet in all of these
examples the primary stress falls on the right hand member:
(52) an enormous area, a nice person, a special place, a funny thing, a
strange man, a loathsome object, the thirteenth time
Again, without a defnition of "redundant head" that can distinguish (50)
fom (52), this account must also be considered to have no predictive value.
Conclusions about the left-dominant [A N] cases. So, we are left
with two fairly clear patterns and some unanswered questions about them.
There is clearly a pattern of adjectives-masquerading-as-substantives form
ing [N N]-like compounds with nominal heads, as in (50) . There also seems
to be a pattern of [A N] sequences becoming single words, and acquiring
the stress patterns thus predicted, as shown by the examples in (46) and
(45) , and presumably the examples in (43).
What determines the set of adjectives that can be used as quasi
substantives? It is surely not a random choice. Why is the substantival
usage often (not always) restricted to attributive position? Why do the
examples with monosyllabic heads in (44), such as Big Ben, blue moon,
hard times, fail to "collapse" into the single noun stress pattern along with
Big Foot, blue bird, hard ball, and all the others? In order to make progress
with these questions we probably need a substantial fragment of the rele
vant "phrasal lexicon," giving us a large enough list of examples to be able
to see patterns and test predictions.
We have motivated a distinction between "semantic" lexicalization (that
is, a fxed expression with a non-compositional meaning) and "syntactic"
lexicalization (that is, a fxed expression that is stressed a if its parent
category was NO rather than N1) . It appears that the second implies the
frst, but not vice versa. Finally, we have distinguished the category of
"morphophonemic lexicalization," in which a lexicalized compound word
loses its internal word structure, and becomes subject to phonological and
phonetic processes that would ordinarily only apply word-internally. We
assume that this can only happen to expressions that are already syntac
tically lexicalized. We have also suggested that many of the lexicalized
[A N] sequences, discounting the cases discussed above where the adjective
is functioning as a noun, may fall into this third category, though it may
be difcult in particular cases to distinguish between this possibility, and
the possibility that the cae in question really belongs to the semantic or
syntactic lexicalization clas.
3. 2 Some Other Lef-dominant Phrasal Units
Besides the atypical [A N] units just discussed, there are a few other cases
of non- [N N] noun phrase structures that are left-dominant. We survey
them below.
Compound-like cases with possessives as lef member. Sometimes
a phrae of the form [N's N] or [[N N's N] regularly prefers stress on the
frst member, although such phrases normally are stressed by the NSR:
(53) cashier's check, bachelor's degree, servant's entrance, pig's feet, ram's
horn, chef's knife, lubber's line, user's manual, goat's milk, bosun's
chair, machinists' union, sheep's wool, auto workers' union, meat
cutters' union, bull's eye, crow's nest
A familiar set of accounts are available here: lexicalization; implicit contrast
of the left constituent; low information content for the right constituent.
As in the case of lef-dominant [A N]] expressions, all of these accounts
have predictive problems.
Thus, although many of the expressions in (53) are semantically lex
icalized, there are other fxed or idiomatic [N's N] expressions that are
nevertheless right dominant:
(54) pope's nose, cat's pajama, fool's errand, busman's holiday, blind
man's buf, Zipf's Law
So, as outlined for the [A N] examples discussed above, lexicalization in
and of itself does not predict NO status.
Still, there is some evidence that lexicalization has a part in such ex
amples: again [N's N] epithets are ofen left-dominant (though one of the
authors has righthand stress in horse 's ass) as are brand names, idiomatic
plant names, etc. :
(55) a. momma's boy, cat's paw, dog's body
b. Boar's Head, Bird's Eye, Woman's Day
c. Bachelor's-button, Baby's-breath, Lady's-slipper, Solomon's-seal,
Shepherd's-purse, Dutchman's-pipe, Prince's-feather, Tinker's
weed, Lamb's-quarters, Goat's-rue, Lion's-foot, Virgin's-bower
It is interesting that metaphorically descriptive names, such a those
in (55c) , are inevitably lef dominant, while a right-dominant treatment is
given to species names in which the discoverer's name possessively partic
ularizes the name of the larger class. These cases are identical in spirit to
the "semi-compositional" [A N] names surveyed in (47) :
(56) Swainson's thrush, Lincoln's sparrow, Kirtland's warbler, Catesby's
Tillium
The parent node of lef-dominant possessives. The possessives in
left-dominant [N's N] expressions have the distributional properties of posi
tion 1 modifers: certainly, at least, they can be preceded by articles, quan
tifers, adjectives, and nominal modifers. This is in striking contrast to the
evidence supporting the view that possessive phrases normally occupy the
same peripheral determiner position as articles and demonstratives-what
we called position 4 in ( 1 ) :
(57) a. the fesh mare's milk
b. 16 Un user's manuals
c. a physics master's degree
d.*the red Sam's car
e. * 16 well-thumbed Knuth's books.
Since Unix manual and physics degree are ordinarily left-dominant, exam
ples like (57b) and (57c) suggest position 1 as opposed to position 2 status
for Unix and physics, respectively, and fom this we conclude that user's
and master's must also occupy position l.
The possessive phrases i n (56) pattern more like position 2 elements
than like normal possessives, occurring inside position 3 and other position
4 elements:
(58) a. the three Catesby's Tilliums
b. yesterday's innumerable Swainson's thrushes
We have not found any convincing examples to suggest that these right
dominant possessives, as opposed to the left-dominant ones, can occur in
side position 1 modifers. Some of the other right-dominant lexicalized
[N's N] expression are similarly distributed:
(59) the coveted pope's nose, an obvious fool's errand
These striking distributional diferences make it reasonable to treat lexi
calized possessive phrases as having a diferent parent category from normal
possessive phrases. The most plausible candidates seem to be NO for the
left-stressed cases and Nl for the right-stressed cases.
Implicit contrast in lef-dominant possessives? In many of the ex
pressions in ( 53), the possessive is in some sense implicitly contrative;
thus lamb 's wool may be distinguished fom the wool of sheep, goats, or
whatever. However, the same thing may be said of nearly any [N's N] con
struction. In the flm title The President 's Analyst, the expression might
be said to distinguish the President's analyst from the analyst associated
with anyone else, but the expression is stressed on the fnal word. The
same can be said for any of the right-dominant possessive phrases in (54)
or (56). Thus the "implicit contrast" treatment of these expressions has no
predictive value, nor any apparent descriptive appropriateness, and should
be rejected a a general explanation.
Lef-dominant [N PP] expressions. There are a fair number of com
pound-like N PP words that take main stress on the lefthand element:
(60) jack-in-the-box, brother-in-law, mother-in-law, free-for-all, hole-in-
the-wall, good-for-nothing, stick-in-the-mud, snake-in-the-grass
These caes are probably related to the general tendency of NP-Iocative
sequences to deaccent the locative phrases, regardless of whether or not
the phrae is idiomatic, and of whether the locative is a modifer of the
preceding NP or an adjunct of an earlier verb:
(61) a. Here's mud i n your eye.
b. He's got a bee in his bonnet.
c. He's got a chip on his shoulder.
d. I've got a spot on my shirt.
e. There's a multi-legged creature on your shoulder.
However, many fxed or idiomatic expressions of the form N PP (or
perhaps NP PP) tae righthand stress:
(62) man of the world, bolt fom the blue, chip of the old block, drop in
the bucket, grant-in-aid, jack of all trades, king of the hill, man in
the moon, man 0' war, manna fom heaven, pie in the sky, piece of
cake, tug of war, word of mouth
Again, epithets seem disposed to be lef-dominant:
(63) Little Miss Nose-in-the-air, Mr. Full-of-fun, light-of-my-life
Other cases. Occaionally a verb phrae or similar collocation takes on
an idiomatic nominal meaning, along with main stress on its leftmost word.
These examples are probably related to the cases in (35) :
(64) forget-me-not, thank-you-ma'am, know-it-all
However, most verb phrae nominalizations are right-dominant. Game
names are a common source:
(65) spin the bottle, pin the tail on the donkey, squeak-piggy-squeak, go
fsh, beggar my neighbor, follow the leader, catch-meif-you-can
Likewise most nominal uses of other phrasal constituents are right dom
inant:
(66) I spy, Simon says, odd man out
The kitchen sink: a potpourri of phrasal lef hand members. Here
is as appropriate a place as any to note that in some informal styles, various
phrasal categories can be feely used as prenominal modifers, with an
appropriately generic meaning. Verb and adjective phrases are particularly
common. This usage permits fee inclusion of pronouns, articles and other
things that are usually forbidden in modifers. Many such phrestop-of
the-line, hole-in-the-wall, tur-of-the-century-are fxed expressions, but
nonce formations do occur. Examples are extremely common in certain
journalistic styles, fom which the following examples are all taken. IO Both
left-dominant and right-dominant stress patterns occur in this group:
(67) an old-style white-shoe do-it-on-the-golf-course banker, the usual
wait-until-next-year attitude, a wait-until-after-the-elections scenar
io, a kind of get-to-know-what's-going-on meeting place, the like-it
or-lump-it theory of public art, state-of-the-union address, a 24-hour
a-day job, a I-percent-of-GNP guideline, a run-of-the-mill meeting,
a sweep-it-under-the-rug amendment, a middle-of-the-road format,
the state teacher-of-the-year title, a take-it-or-Ieave-it choice, the
yet-to-be-written 1987 bill, a certain chip-on-the-shoulder attitude,
make-it-fom-scratch traditionalists, Speak-Mandarin-N ot-Dialects
Month, a rob-Peter-to-pay-Paul system, the nothing-Ieft-to-chance
approach, get-out-the-vote drives, the don't-trust-anybody-over-30
crowd, national clear-your-desk day
Such examples are quite consistent with our approach to predicting stress
from structure in complex nominals, suggesting that both positions 1 and 2
may be occupied by elements projected fom (nearly) any lexical category.
3. 3 Right-dominant [ N N] Expressions
In the texts mentioned earlier in this paper, about 25% of the [N N] phrases
had main stress on the right, and essentially none of these had a plausible
peA account. We argue that these are instances of NI . l l In this section
we outline some of the semantic relations which one fnds in these casesI2
Noun modifers expressing composition. Some right-dominant
[Nl N2] sequences are of a familiar type that means something like 'N2
made of Nl' , 'N2 made with Nl', or 'N2 with Nl as a featured part' :
(68) pattern
Nl
THING-MADE-OUT-OF-
Nl
:
rubber boots, steel plate, duck soup, gold medal, corduroy suit, brass
l
OLieber (1988, 1992) also discusses examples of this kind, arguing that one needs to
allow phraal entities inside compounds in the general cae.
II
Again this is a traditional position: see Bauer 1983, pp. 104-108 for a discussion.
12
See Fudge 1984 for a discussion of some of these clases.
bed, diamond ring, nylon rope, plaster cast, wax fgure, asbestos
tile, meat pie, chicken gumbo, wood foor, mushroom omelet, carrot
halvah, rice pudding, apricot jam, corn tortillas, squash pie, beef
burrito
In some caes of common words, 'made with NI' [Nl N2] sequences
revert to lefthand stress, presumably due to lexicalization of the same type
that is involved in (43) ; one may refer back to (25) for some examples, or
consider:
(69) banana bread, rye bread, peanut butter, corn chips, potato chips,
butter cookies, oatmeal cookies, orange juice, coconut milk, aspirin
tablet, chocolate bar, chocolate milk, cornstarch, silverware
As expected if lexicalization is at issue, there are many individual difer
ences in this area; some people say peanut bt tter in place of pea! nut butter,
while others have potato sa'lad in place of pota' to salad, chocolate mi'lk in
place of cho' colate milk, etc.
In (69), note the prevalence of left-dominance in the cookie and bread
head cases, even for new coinages like zucchini bread or cinnamon cook
ies (the latter for one of the authors only), and compare this pattern
to the results of combining the same modifers with the head pie. The
left-dominance of bread tenns even extends to some forms with adjective
modifers, such as sticky bun and hard roll. These cases, in which a par
ticular head word has a more-or-Iess strong afnity for a particular stress
pattern, are similar to the familiar case of Street versus Avenue, discussed
below.
Time, place, class, etc. In other cases, NI is a time or place, loosely
defned:
(70) a. pattern PLACE- WHERE-
N2
-IS-FOUND
N2
:
garage door, basement walls, attic roof, kitchen sink, cell window,
pantry shelf, library curtains, bedroom furniture, cell membrane,
mountain pass, valley foor, college president, city employee
b. pattern TIME- WHEN-
N2
-ocCURS
N2
:
summer palace, fall weather, winter carnival, summer sausage, au
tumn leaves, spring fowers, winter cold, spring cleaning, Septem
ber morning, January thaw, Christmas dinner
Proper name modifers. Closely related are the cases in which a proper
name is used as a modifer, to establish location, type, period, source,
ownership, brand or whatever:
( 71) pattern PROPER-NAME THING
Staten Island ferry, Connecticut Yankee, US ambassador, West
Texas barbecue, Busch beer, Napoleon brandy, Pennsylvania crude,
AT&T headquarters, India ink, Tifany lamp, GB stronghold
Of course, full proper names go by the NSR, as do compound names for lo
cations, other than those that end in Street , or Land such as those in ( 72c):
(72) a. George Washington, Barnacle Bill
b. Newark Airport, Chesapeake Bay, Bryce Canyon, Park Drive,
Baxter Estates, Niagara Falls, Strawberry Fields, Union Gap,
Yosemite Lodge, Tuolumne Meadows, Adygei Autonomous Ob
last, Abiathar Peak, Cat Mousam Road, Times Square, Sturbridge
Village, Blue Jay Way
c. Bank Street, Arnhem Land
Generally, names for institutions, organizations, and so forth, also go by
the NSR, though there are exceptions ( 73b):
( 73) a. Harvard College, Locksley Hall, Coolidge High, Bell Laborato
ries, Harvard Law, Widener Library, MIT Press, Bethlehem Steel,
Stevens Tech, Yale University, Harvard Yard
b. Dudley House
Measures, methods, media. Another case of right-dominant [N N]
phrases includes various sorts of measure-phrases in the Nl position:
( 74) pattern MEASURE THING-MEASURED:
mile run, pound note, gallon jug, pint jar, two-alarm chili, three
dog night, four-man front, two-minute warning, eight-hour day, six
fgure salary, fxed-length record, 12-gauge shotgun, three-day pass,
three-minute egg
One apparent exception to this pattern is minute steak. However, the
diference between the bulk of the examples in (74) and minute steak is
that that the former are genuinely compositional instances of meaure
phrae modifcation. That is, the left hand material is measuring the head
on an appropriate (if sometimes idiosyncratic) scale. So, a mile run is a
run which is really a mile long, a three-minute egg takes three minutes
to cook, two-alarm chili is placing the chili in question at a reasonably
high point on a scale of potential gastric disturbance, and a three dog night
gives a measure of the ambient temperature of the night in question using
the number of dogs required to keep oneself warm as a metric. The term
minute steak, however, does not mean that the steak in question takes one
minute to cook, but merely that it cooks quickly.
Also, we have right-dominant [Nl N2] constructions with characteristic
methods, media or energy-sources in the Nl position, although the behavior
of this clas is erratic; see lef-dominant examples in (26) and (27) :
( 75) pattern METHOD-OR-MEDIA-FOR-
N2 N2
:
bathtub gin, gas chromatograph, electron microscope, pinhole cam
era, gasoline engine, kerosene heater, propane torch, shotgun wed
ding, gunboat diplomacy, jet engine, steam radiator, microwave
popcorn
Residual examples. Other examples of right-dominant [N N] phrases
belong to patterns whose clasifcation is not clear to us. We reproduce a
list for illustrative purposes:
(76) fy ball, discount bookstore, dictionary definition, capital gains,
weather helm, industry leader, color monitor, trial run, combina
tion lock, sacrifce single, round-trip ticket, gospel truth, battleship
grey, blood relative, tramp steamer, return ticket, precision tool,
race suicide, eggshell china, fossil man, sex education, rogue ele
phant, touch football, ball bearing, fullback draw, championship se
ries, party line, cah customer, child labor, chain reaction, touch
typist, home run, string quartet, pony express, parcel post, police
custody, polka-dot dress, box score, world bank, industry leader,
crah course, crah landing, crack regiment, smash hit, snap judg
ment, bum rap, sponge rubber, college degree, shoestring tackle, bit
part, barrier reef, toy gun
Some of these examples may follow the pattern N2 IS-A-Nl , which typically
follow the NSR a discussed for (9b) : rogue elephant , fossil man, sponge
rubber, toy gun. Notice, incidentally, that color monitor is stressed one way
if you mean ' monitor that shows color'-that is, the usual Byte magazine
reading, and the other way if you mean 'device or circuit for monitoring
colors'-say, a new improvement in TV set technology, also known a the
spectrum inspector. In the second cae, we have a clear argument-predicate
compound.
Lef-headed sequences. All of the examples of modifed nominals (with
the exception of bahuvrihi constructions) which we have discussed so far
have had the nominal head on the right and the modifer(s) on the left,
the normal ordering for English. However, there are some cases where it
appears to be necessary to asume that the head of the construction is
on the left and the modifier is on the right. We will assume that this
is simply a marked option, though it is certainly true that adjectives in
English may quite feely appear on the right of the noun they modif
under appropriate conditions: a man tall, dark and handsome (see Bolinger
1 967). In many of the examples we will see, the particular head in question
is apparently marked to occur on the lef. All of the leftheaded examples
are stressed by the NSR, hence are N1 under our analysis. For example,
(77) below contains left-headed construction where the righthand member
is an identifying name or number:
( 77) vitamin C, route 1, brand X, exit 14, peach Melba, steak diane, Cafe
Beethoven, Club Med
So, while a garage door is a door, and a company president is a presi
dent, vitamin C is a vitamin, not a letter or a programming language,
and steak diane is meat. The final examples in ( 77) are either borrowed
fom Fench, or at least modeled on French syntactic patterns, which
are more single-mindedly left-headed than English ones. Note that food
terms in particular are commonly (and productively) left-headed, and Ro
mance infuence is particularly likely in this domain: beef Wellington,
chicken cacciatori , lamb vindaloo, eggs benedict , bananas Foster. Some
of these may have questionable headedness: is lamb vindaloo a kind of
vindaloo made with lamb, or is it a kind of lamb cooked with a vindaloo
sauce?
As noted above, a number of these cases involve heads which seem sim
ply to be marked to occur on the left: so if a new vitamin were discovered,
and assigned the letter ' X' , the vitamin would have to be called vitamin
X, not * X vitamin. Similarly, exit 14 could not be called * 14 exit.
Further examples of constructions of that lean towards lef-headedness,
many of which instantiate productive patterns, are given below:
(78) pattern CLASSIFIER NAME:
Bayou Goula, Cape May, Chancellor Adenauer, Citizen Kane, Com
rade Andropov, Dr. Smith, Fort Eustis, Key Largo, King George,
Lake Erie, Lake Huron, Loch Ness, Mount Hood, planet Earth,
widow Brown
The CLASSIFIER NAME pattern can be augmented as a class by the introduc
tion of new titles borrowed from foreign languages, such as Generlissimo
Franco, but it cannot be feely used for all classifers, even for some that
seem like minor variants of cases that work fine:
(79) *Pond Walden, * Hill Blue, *bride Smith
3.4 A Possible Reason for N1 [N N]s: Is the nominal
modifer an adjective?
Leaving the left-headed cases now, we now return to a possible explanation
for the right-dominance of certain right-headed [N N] compounds, namely
the traditional notion that in such cases the lef hand noun is functioning
as an adjective. For example, for the Nl THING-MADE-OUT-OF-Nl cases
given in (68), we might postulate a null-afx derivational rule that makes
adjectives out of nouns, since the usage usually generalizes to predicative
patterns:
(80) The boots are rubber.
The plate is steel.
The medal is gold.
The suit is corduroy.
The bed is brass.
When the meaning is more like ' made with X as a relevant part' , or what
ever, the attributive noun cannot so felicitouly be used in predicate posi
tion with its head as subject:
( 81) ?The soup is duck.
And i n the majority of the right-dominant [ N N] sequences we have sur
veyed, the predicative counterpart is totally implausible:
(82) *The ball is fy.
*The jail is county.
* The door is garage.
* The membrane is cell.
All in all, the adjectival analysis of the left hand noun in N1 [N N] con
structions runs aground on the problem of limiting the distribution of the
resulting adjectives. A more straightforward idea is simply to generate the
nominal modifer in the same phrasal slot that an adjectival one would oc
cupy; that is, in position 2 of ( 1) , which we have identifed as dominated by
N
l
. The stress and the meaning of the nominally modified phrase ought
then to follow without further ado, if the rules of stress assignment and
semantic interpretation are correctly arranged.
3. 5 The Case of [ [A N] N]
As is well known among afcionados of complex nominals, although occa
sionally ignored by other morphologists, one can combine an [A N] unit
with a nominal head in a semantic relation typical of a simple [N N] com
pound with the same head. In such constructions, the main stress would
typically fall on lefthand member, CSR style, as for most [N N] compounds.
Since the left hand member itself is of the form [A N] , its internal stress
pattern normally goes by the NSR, and the result is main stress on the
middle word:
(83) toxic waste cleanup, collective bargaining agreement, balanced bud
get amendment, civil rights bill, used car business, fnancial planning
consultant, due process clause, floating underfow trap, systolic array
machine, fesh fsh shop
Both the existence of such examples and their normal stress pattern follow
straightforwardly fom our proposals.
4 The Structure of Premodifed Nominals: A Summing
Up of the Theoretical Issues
We have examined a wide range of modifed noun phrases. In this section,
we will summarize the structural analyses we have suggested for the various
types of nominals that we have discussed. In addition, we will have a few
things to say about some other nominal constructions in English and other
languages.
We will then extend our arguments against what we have called FCA
only theories of nominal compound stress, by examining the approach
of Ladd 1984. We will also discuss the distinction between morphology
and syntax: should some of constructions we have been discussing be
viewed as the product of a component of the grammar-the morpholog-
ical component-that is crucially diferent fom the component responsible
for phrasal composition? Many writers have argued that this distinction
should be made, Di Sciullo and Williams 1987 being perhaps the most vocal
supporters of this view. We will evaluate some of their arguments in terms
of the phenomena discussed in this paper, and suggest that the evidence
for such a split is not strong.
4. 1 A Summary of the Structure of the Modifying Phrase
We have seen examples of English premodifed nominals with various sorts
of supra-lexical constituents as their left hand member. At least the patterns
[A N] , and
[N PP] are fairly productive. This suggests that the lef hand
constituent of premodifed nominals is (maximally) an X
l
, where X is
commonly (though not necessarily) N
l
:
(84) [[
N
1
A N] N] equal rights amendment
[ [
N
1
N PP] N] balance of payments problems
Again, there are certainly restrictions. For example, N-plus-comple
ment or A-plus-complement modifers are somewhat restricted. There are
stylistic issues, and common or lexicalized modifers are preferred in these
cases. Examples like these seem infelicitous:
(85) *Where's my [tree in the garden] clipper.
* John is a [proud of his son] man.
Examples like those in (85) are generally taken to be ruled out by a Head
Final Filter (Fabb 1984, Di Sciullo and Williams 1987), which states that
the head of a modifer must be adjacent to the modifed word. However, the
fact that more common phrases seem unproblematic as modifers ( balance
of payments prblems) suggests that the Head-Final Filter may actually
not be a grammatical constraint, but rather a parsing constraint: as long
as the left hand component is recognizable as a unit in the context of use, its
exact construction is not a limited as constraints such as the Head-Final
Filter would suggest.
As exemplifed i n (11), the lef member of a noun compound can ordi
narily not have its own independent article. The restriction against articles
is pretty strong even for things that are clearly lexical items-thus we say
the Bronx Borough President, not * the the Bronx Borough President; see
Fabb 1984 on these. Also excluded are noun phrases specifed by a a wh
word, and any form of pronoun acting as the modifer on its own. Quan
tifers, which are often taken to be specifers of N
2
(see Abney 1987) are
normally unacceptable. So, the following examples are all impossible:
(86) *an [every kitchen] table;the [which dog] owner;the [it] hater
However, note that modifers containing numerals are not excluded:
(87) the four color theorem, the eight queens problem, three-mile mark,
a $3 million ofer
So, on the one hand we have clear evidence that in many cases the modifer
must be phrasal, and we have equally clear evidence that the modifer
cannot be a fully specifed noun phrase. We can capture such restrictions
quite neatly if we assume that prenominal nominal modifers are maximally
Nl . This will rule out modifers such as The Bronx, every kitchen, which
dog. To rule out pronouns such as it as modifers, one could appeal to
a suggestion of Paul Kiparsky (p. c. ) that closed class items (at least in
English) do not typically allow further morphological derivation. On this
account, it hater would be bad for the same reason as a derived form such
as it-ishness i bad. To be sure, other explanations are available: another
line of thought says that projections of functional categories may not serve
as modifers. This rules out pronouns and also projections of D, such as The
Bronx, as modifers. Indeed, a similar set of restrictions holds of adjectival
phrases: degree specifers such as so, that, how are obligatorily absent from
prenominal adjectival modifers:
(88) a. The houses are that large.
*the that large houses
b. How large are those houses?
*Those how large houses are over there?
c. The houses are so large.
* the so large houses13
Abney (who discusses these facts) suggests, on analogy with his analysis of
noun phrases, that full adjective phrases are really Degree Phrases ( DegP)
where the degree word is the head, just a a determiner is the head of DP,
with the AP as a complement of the degree word; see Abney 1987, p. 321:
(89)
j,j
j,
so

_
j
_
nice
}}}
One can then rule out the examples in (88) by asuming that only APs
possibly, in fact, maximally A
I
, can appear in prenominal position. This
is essentially what Abney argues, although, his analysis makes the rather
novel asumption that the adjective is the head in a construction like big
dog, that this construction is therefore an AP, and that the NP dog is
a complement of big. We have of course taken the more traditional view
that canonical prenominal modifers, namely the prenominal adjectives, are
adjoined to Nl in English.
So, it seems that for the canonical prenominal modifers, namely projec
tions of N and projections of A, there is some evidence that such modifers
are typically maximally single bar projections-Xl . If we look again at the
examples of highly-complex modifers in (67), we see that they are actu-
l3
0ne can say things like how large a house, so large a house. Such constructions are
limited, in the sense that only an indefnite singular NP can follow the degree-marked
adjective:
*
how large that house, how large houses. Abney argues that thee are
adjectival phraes (DegPs) where the head adjective takes a fully specifed noun phrae
a a complement.
ally consistent with this idea-they may be projected fom non-traditional
categories such as V, but may plausibly be analyzed as being of level X
l
or below.
4.2 More on the Distinction Between NO and NI Modifers
Having argued that the modifer of a noun is maximally X I , we now
overview the arguments for our assumptions about the structural relation
ship between the modifer and the head. It is worthwhile to start out by
placing this assumption about modifer-head relations in a broader context
of current ideas about phrase structure.
We have maintained that modifers are adjoined to either NO or NI .
The diagram in (90) illustrates the structure we have assumed for an NI
modifier such as the adjective red:
(90) [
DP
[
D
O the] [
N
2
[
N
1 [
A
red] [
N
1 [
N
O bookll]]]
We have assumed that most [N N] compounds are instances of modifer
head constructions, in this case involving adjunction to NO:
(91) [
N
O [
N
O dog] [
N
O house]]
However, we have argued at length that not all instances of [N N] modifi
cation are at the NO level; similarly, not all instances of [A N] modification
are NI . Thus we have posited structures like these: 1
4
(92) a. [
N
1 [
N
O kitchen] [
N
1 [
N
O sink]]]
b. [
N
O [
A
blind] [
N
O spot
}}
Our primary arguments involved default stress patterns and some infor
mal consideration of semantic relations, but there is also evidence from
sequence constraints among modifers. Note that one does not find [N N]
constructions which we would argue on stress grounds to be of the type
NI , occurring as the righthand member of a construction whose parent
node ought to be an NO. So, one can say cofee jar ( NO) , and plastic jar
(NI ), but one cannot get * cofee plastic jar (cf. plastic cofee jar) since
the structure would have to be *
[
N
o NO NI] , which is ruled out by X-bar
considerations given that the rightmost member of a compound in English
is generally the head (see ( 105) below).
On the other hand, complex expressions which look on the surface to be
phrasal constructions, but which we have argued on stress grounds to be
NO, can occur in the righthand position of compound nouns. The complex
14It is worth noting that the structure in (92a) is vaguely related to the proposal in
Selkirk 1984 (pp. 43-50) that right stressed compounds involve adjuncts, where the
modifer is in some sense more external than modifers in a left stressed compound,
which are taken to be arguments. Selkirk, however, s seems to suggest that at least
some compound-like constructions, such a slcl warehouse (i.e., warehouse made of
steel) might be viewed as phraal collocations (p. 247) a opposed to compounds of the
form ADJUNCT-HEAD. It is not clear from her discussion how one is supposed to
distinguish these caes.
examples in the third column are all right-branching NOs having primary
stress on the penultimate member a predicted by the eSR:
(93) user's manual Unix manual [Unix [user's manual]]
dry dock trawler dock [trawler [dry dock]]
Needless to say, prenominal noun modifers that on our analysis are
dominated by Nl , are generally reasonable candidates to stack outside
adjectives that occur at the same level. Thus the right-dominant character
of plastic cup is consistent with the possibility of plastic disposable cup,
though the alternative ordering may be slightly preferred.
By giving a structural explanation for the unavailability of examples
like * cofee plastic jar, we are claiming that such examples are ill-formed
for the same kind of structural reason that * difcult the problems is ill
formed. It is worthwhile contrasting such rigid ordering restrictions with
the much softer ordering restrictions on pre nominal adjectival modifers,
which are all introduced at the same (Nl) level, but which nonetheless
display well-known ordering preferences:
(94) large red ball ?red large ball
oversize square table ?square oversize table
One could, of course, ask whether these ordering restrictions should also be
handled by assuming intermediate structural positions or levels? For exam
ple, meaure adjectives like large could be placed in a slot "outside" color
adjectives like red, since the former prefer to precede the latter. However,
it has been observed in the literature on prenominal adjective ordering (see,
most recently, Sproat and Shih 1990b) that these ordering restrictions are
only preferences, operating in the absence of pragmatic reasons to choose a
diferent order. It is perfectly possible to choose an alternative order, given
appropriate context:
(95) I want the red large ball, not the blue one.
So it seems best to agree with the traditional view that these ordering
preferences refect some sort of natural continuum of psychological afnity,
even though it seems difcult to defne such a scale in a non-circular way.
In contrast, ordering restrictions on what we have argued to be modifers
at the NO and Nl levels seem inviolable-like the ordering restrictions be
tween specifers and modifers, suggesting that the structural analysis pro
posed for these modifers is on the right track. Something seems defnitely
degraded about (96) as compared with (95):
(96) ??I need the cofee plastic jar, not the tea plastic jar.
These considerations, as well as the stress and semantic considerations
discussed throughout this paper have therefore led us to the conclusion that
modifer-noun constructions can be either N1 or NO, no matter whether the
modifer is adjectival or nominal. We now summarize the various classes
of cases we have examined in this paper giving examples to fll out the
structural possibilities. Non-Iexicalized [A N] combinations-instances of
normal prenominal adjectival modifcation-are mostly phrasal. At least
some of the productive cases of apparent NO- level [A N] combinations may
really involve [N N] modifcation at some level, as we argued in reference
to (50) , and as suggested by the discussion in Levi 1978. Lexicalized [A N]
forms-those forms that must be listed since they have more or less id
iosyncratic or idiomatic meanings-may be either Nl or NO:
(97) Lexicalized Non-Lexicalized
N yellow pages dental operation
Nl red herring red paint
With [N N] modifcation we clearly get all possible combinations of
Productive/Lexicalized and bar level. There is no sense that [N NJ NO are
more lexicalized than [N N] phrasal combinations:
(98) Lexicalized
eyeball
rice pudding
N on-Lexicalized
Capsicum leaf
dingo stew
Finally, we have noted compound-like examples where the modifer is a
possessive. These may be either Nl or NO. In both cases the expression
tends to be lexicalized, though there are some productive subsystems, such
a those left-stressed examples headed by milk:
l
5
(99) N
l
Swainson's thrush
NO: dog's body
4. 3 Some Other Compound-like Constructions in English
and Elsewhere
We have discussed apparently left-headed nominals in English. On the bais
of their stress behavior all such examples appear to be N
l
. Many of these
constructions have a questionable status in the minds of many speakers.
Thus the prescribed plural attoreys general, which is what is expected
given a lef-headed structure, has already lost to the alternative attorey
generals, suggesting that this form has been largely reanalyzed. Other
examples, possibly because of their productivity, seem more solidly lef
headed: Mounts Whitney and McKinley. One can link these cases of post
nominal modifcation with the slightly marked, though still very productive
clas of post-nominal modifcations with adjectives: a man proud of his son,
a river broad and long, a dragon fearsome to behold.
Other compounds and compound-like words which do not ft so neatly
into the discussion of most of this paper include exocentric bahuvrihi com
pounds. Bahuvrihi compounds have been analyzed (cf. Kiparsky 1982) as
1
5
Note that some left stressed examples function as proper names and disallow preceding
modifers or specifiers in most cases, e.g., Kayne 's generalization. There is nothing
particularly surprising in this, and it seems reasonable to analyze such expressions as
themselves being proper names, which typically lack articles in English.
STRESS AND STRUCTURE OF MODIF1ED NOUN PHRASES / 167
modifers to an empty headed noun redneck T) . Such an approach seems
to solve the embarrassing problem that these constructions have no seman
tic head (a redneck is not in any sense a neck) ; however, we must assume
in addition that these cases are all NO, presumably a consequence of their
having undergone at least the second stage of lexicalization outlined in
Section 3. 1
Examples like those of (35) are presumably nominalizations of phrasal
fragments reminiscent of the Romance examples in ( 100) :
( 100) a. essuie-glace (wipe glass) 'windshield wiper'
b. tocadiscos (play records) 'record player'
c. guardaboschi (guard woods) 'forester'
Such examples are syntactic in appearance but have been transformed into
nouns, suggesting a structure along the lines of that argued for in Di Sciullo
and Williams 1987:
( 101)
toca]
discos]] ]
4.4 More on FCA-Only Theories: Ladd's Deaccenting
Argument
We have taken the traditional generative view that syntax determines stress
in the "normal" cae-abstracting away from FCA efects. In adopting this
view we are rejecting, among other approaches, Ladd's 1984 proposal that
compound stress is a result of deaccenting the head. His idea is that the
diference in representation between steel warehouse ( 'warehouse made of
steel') and steel warehouse ('warehouse for storing steel') is purely metrical,
the latter having the normal iambic phrasal structure w s and the latter
the trochaic, deaccented structure s w. This deaccenting in compounds is
explicitly claimed to be part of the more general phenomenon of phrasal
deaccenting in examples like (102) (Ladd 1984, p. 255) , where books is
deaccented for pragmatic reasons:
( 102) Ha John read Slaughterhouse Five?
No, John doesn't read books.
Ladd's baic claim is that in right-stressed modifer-head nominal construc
tions, the attribute (lefthand member) does not serve to subcategorize1
6
the
head. In the deaccented (CSR) cases the attribute does subcategorize the
head, hence the head contributes only part of what is necessary to identif
the new category. An obvious contrast is green house versus green house.
In the former the attribute green does not produce a new type of entity, but
merely serves to further specif the general category house; the accented
head therefore provides the category. A green house, on the other hand,
is a new category of entity, so house provides only part of the necessary
information, which results in its being deaccented relative to green, which
16
Ladd is not using the technical linguistic sense of subteorze, but intends the sense
of "making a subcategory out or' the head.
provides crucial additional information. The approach is argued to work
not only in cases where semantic categorization is involved, but also in cases
where this is much less clear, such as in the domain of place names. So, the
contrast between First A venue and First Street is argued to derive fom
the fact that Street, within the category of names for roads, is the more
common or expected term, and also gives the least semantic information
about what kind of road is involved (terms like Avenue, Boulevard, Alley,
carry with them connotations which Street does not have) . Thus Street is
deaccented.
While there are surely areas in which notions of subcategorization
conditioned deaccenting plays a role, the idea is problematic as a general ex
planation for the stress patterns we have seen. For example, one pragmatic
test proposed by Ladd to distinguish "favors" fom real "(sub)categories"
in food terms falls far short of correlating with stress. Ladd's idea is that
cheese sandwich is a sandwich favored with cheese and cheese merely
serves to further specify sandwich, not form a separate category: thus the
stress is phrasal. In contrast, for banana bread, one clearly needs the in
formation provided by the left hand member to determine the category of
food involved since banana bread is not really a kind of bread favored with
banana, but a new category of food similar to bread in some ways. Now,
Ladd notes that while one can felicitously ask do you want a sandwich?
and subsequently ofer a cheese sandwich, one cannot felicitously ask do
you want some bread? and ofer banana bread. It is claimed that one can
only felicitously ofer XY if XY is merely a Y favored with X, and not
a diferent category of food. This test is then supposed to correlate with
stress, right hand stress if the ofer is felicitous, left hand if it is not, for the
reasons given above. However, this test makes the wrong predictions in
whole categories of examples. For instance, most words headed with bread
are left-stressed even when their referent is a canonical instance: one can
felicitously ofer a person bread and then give them any one of wheat bread,
rye bread, white bread and a number of others. On the other hand, while
steak and kidney pie follows the normal pattern of righthand stress for
words headed by pie, it would come as quite a surprise to most Americans
to get some after assenting to an ofer of unmodifed pie. 17.
Another problem with Ladd's reduction of compound stress to general
mechanisms of phrasal deaccenting is that phrasal deaccenting can easily
produce examples of right-branching structures where all the right hand
material is deaccented:
( 103) I don't [like [buckwheat
griddle cakes]]]
Ladd's theory therefore predicts that left dominance in right-branching
compounds is possible, if the subcategories fall out right. This does not
1 7This point can be made even more strikingly with shepher's ge, a concoction wherein
ground lamb is topped with mahed potatoes and then baked.
seem to be true, even where one might expect such stressing given the
behavior of binary compounds with identical heads and lefthand members:
( 104) a. buckwheat cakes
b. griddle cakes
c. buckwheat griddle cakes
d.*buckwheat griddle cakes
Of course, (104d) is possible in FeA contexts, but it is not the default
case for such compounds, contrary to what we might expect given Ladd's
theory.
1O
4.5 Are NO Compounds Formed "in the Lexicon"?
In assuming, as we have been, that NO compounds are instances of mod
ifcation constructions on a par with N
I
modifcation constructions, we
have been at least implicitly assuming that such compounds are syntactic
objects. Doing so, of course, runs counter to most theories of morphology,
which assume that compounds are dealt with in the morphological compo
nent and that if compounds seem to share some properties with syntactic
constructions this should nonetheless not be taken as evidence that they
are syntactic (Selkirk 1982, Hoeksema 1984, Mohanan 1986, Di Sciullo and
Williams 1987, among many others) .
We would like to consider one recent attempt to justif this separa
tion of compound formation fom syntax, namely that of Di Sciullo and
Williams 1987 (henceforth D&W) , and consider whether the evidence they
adduce for this separation is reasonable. D&W suggest an approach to mor
phology whereby words are constructed in a component of the grammar,
the morphological component, which is in many ways similar to syntax,
but is both (i) diferent in detail fom syntax and (ii) separate fom syntax.
We discuss each of these points in turn. 19
Diferences of detail. A crucial diference of detail is that compounds
are right headed in English, whereas phrases are left headed, at least at a
very macroscopic level of analysis.
There are actually two issues relevant to comparing the headed ness of
phrases and compounds. One, D&W claim, is that one can always pick out
the head in a phrase because it is "the item with one less bar level than the
phrase" (p. 23) and which is generally of the same category a the phrase:
1
8
Ladd's own examples involving the head noun warehouse might seem to be a contra
diction to what we have just said since warehouse is, historically at leat, a compound.
However, it seems well within reason to asume that the analysis of warehouse as a
compound is only of historical interest and that the word today lacks an internal word
boundary. Certainly productively formed constructions with unequivocally complex
heads do not seem to stress a Ladd ought to predict: a department store for electronics
would be a electrnics deartment store, except of course in clear FCA-contexts. Yet
surely an electronics department store must be considered a subcategory of department
store, at leat if a steel warehouse is considered to be a subcategory of warehouse.
19See also Baker 1988 for a review of problems with D&W.
( 105) xn - . . . YP . . . xn-l . . . ZP . . .
They contrast this situation with the situation in a canonical compound
where the two daughters of the NO are themselves NO. They suggest that
compounds follow a diferent, non-syntactic, morphological component rule
for determining headedness, namely a modifed version of the Righthand
Head rule of Williams 1981, which states roughly that the head of a word
is its rightmost morpheme. In compounds, it seems, the syntactic notion
of head would not work.
Still, it is important to bear in mind that the ease with which one can
apply the syntactic notion of head in phrases, and the seeming difculty of
using that notion in compounds derives fom theoretical prejudice. To see
this point, consider the structure of noun phrases in Welsh such a:
( 106) meibion athrawon (sons teachers) 'teachers' sons'
This noun phrase consists of two plural nouns with no other markings, and
yet there is no question that it is left-headed, as are all noun phraes in
Welsh. In order to apply the syntactic notion of head, one would clearly
want to argue that the right hand noun athrawon 'teachers' is really the
sole member of a DP-hence could not count as the head under ( 105)
although there would be no way to ascertain the presence of the DP fom
the existence of athrawon alone. Returning to English compounds, there
is no absolute reason why one cannot assert that all prenominal modifers
within NO are phrasal; we have already discussed examples which show
that at least in some cases the prenominal modifer in the NO level must
be an Xl . So, one might assume that all pre nominal modifers are Xl just
as we tend to assume that the noun athrawon in ( 106) is really a DP, de
spite the lack of direct evidence in the particular construction in ( 106) . If
we were to take that tack, we could clearly modif the syntactic notion of
headedness to allow that the head of a construction must match the cate
gory of the parent and have a bar level not greater than that of the parent;
something like this condition must be right more generally anyway if we
are to defne headedness correctly on adjunction structures under normal
notions of adjunction. Then, the head of dog house would be clear: the
prenominal modifer dog would be an N
I
, whereas house would be an NO,
thus satisfying the conditions for headedness under the revised defnition.
Still, we don't have to adopt this strategy. The other issue, raised by
D&W's claim that English phraes are left headed is the question: left
headed at what level? Clearly, modifers, whether phrasal or not, precede
their heads in the typical case in English, and since we have assumed that
NO compounds are merely a type of modifcation, we would expect that
compounds are right-headed. Thus in order to determine the head, the
syntax would merely need to take note of the kind of relationship-i. e. ,
modifcation-holding between the members of the compound, and fx the
head on the right. D&W, in claiming that English phrases are left headed,
have missed the locally more relevant point that in modifer-head construc
tions in English the head is on the right.
Of course, in deriving compound headedness in this way, we fail to link
the right-headedness of compounds with the right-headedness of words in
English: for example, many people (e.g. , see Williams 1981, and also Lieber
1980) have argued that the head of, say, grmmaticality is the sufx -ity
since that determines the category. Williams' Righthand Head Rule explic
itly links compounds with afx right-headedness. We think, in contrat,
that the two should not be linked, at least synchronically. The fact that it
is typically sufxes and rarely prefxes that function a heads of words is
almost surely related to the cross-linguistic prevalence of sufxation, which
in turn may have psycholinguistic bases (see Cutler et al. 1985) . Thus it
seems plausible to view the right-headedness of English compounds and the
right-headedness of English afxation a a coincidence, in the sense that
they are not diverse expressions of a single fact about English grammar.
20
Indeed, we would expect languages with left-headed modifcation struc
tures by and large to also show left-headed compounds if they have com
pounding, no matter what the headedness of afxation is. This appears
to be correct, a examples fom French, discussed in Selkirk 1982, or from
Spanish, Breton "loose compounding" (Stump 1989) or Zhuang (Ramsey
1987) .
2
1
For frther discussion of this and related points see Lieber 1988,
1992, Baker 1988, and especially Clements 1989:
( 107) a. timbres-poste (stamps post) ' postage stamps'
b. los coches Chevrolet
c. Breton: kaoc' h kezeg (dung horse) ' horse dung'
(cf. paotr brav (boy fne) 'a fne boy' )
2
0
0f course, this is not to say that all other patterns would have been equally likely, or
that the two fats about English are not connected historically. As Cutler et al. 1985
show (see, e.g., pp. 727f), there is a positive cross-linguistic correlation between sufxing
and phrasal right-headedness. In particular, right-headed languages-Object-Verb and
Postpositional languages-show an overwhelming preference to be exclusively sufxing.
Left-headed languages mostly have both prefxes and sufxes, and prefx-only languages
are very rare and are apparently always phraally left-headed. English descends from
languages which were phraally right-headed, a did Latin, from which half of English
afxational morphology comes. English prehead modifcation is likely to be a living fossil
of the former right-headed syntax. So right-headed compounding may be historically
related to right-headed afxation in English, by a history which makes crucial excursions
into the syntax.
2
1We should note that Welsh, which ha left-headed modifcation structures, has a
number of frozen right-headed compounds: ysgolfeistr (school-mater) 'school mater';
rheilfordd orsaf (rail road station) 'railway station'. Some examples, such a the second
one, are clearly translation borrowings from English, and in any event this kind of
compounding is not productive in Welsh. Breton "strict compounds" (Stump 1989) are
also right-headed. Note that Welsh (and also Irish, Breton, French and Spanish) does
retain a few prenominal adjectival modifiers and so it may well be the cae that Celtic and
Romance languages are only somewhat further along the transition to left-headedness
than is English.
172 I MARK LIBERMAN AND RICHARD SPROAT
d. Zhuang: kaiS-pou4 (chicken male) 'rooster'
(cf., sail moS rau2 (book new our) 'our new book')
Concerning the Fench examples, D&W claim (p. 83) that "Selkirk misanal
ysed fxed phrases such as timbres-poste as 'lef-headed compounds, ' thus
concluding that Fench morphology is a mixture of left- and right-headed
structures (since afxation in Fench is clearly right-headed) . This is a
clear example of a generalization compromised by the failure to properly
separate syntax and morphology. "
This attack can easily be turned around, however: by insisting that mor
phological constructions, including compounds, must obey the Righthand
Head Rule, D&W make it impossible both to recognize as compounds many
constructions which any frst-order description would classif as such, and
to state the very clear generalization that compounds in a language gen
erally follow the ordering of modifer-head constructions in that language.
In the other direction, relying on the Righthand Head Rule as a correlate
of "morphological" constructions would prevent us fom noting properties
that left-headed examples in other languages may share with English com
pound nouns. Thus in caes like carne de burro (meat of donkey) ' donkey
meat' in Spanish, it is essential that the complement phrase de burro, like
the modifers in English compounds, not be a full DP. With an article,
the phrase care del burro ' meat of the donkey' would not serve to refer
to the generic substance donkey meat. Care de burro, while having clear
phrasal properties, equally clearly has properties commonly associated with
compounds.
Separation from syntax. The other half of D&W's claim about the sta
tus of compounds is based upon the apparent atomicity of words fom
the point of view of syntax. In saying that compounds are atomic, D&W
are buying into the familiar claims of lexical integrity (see, e. g. , Mohanan
1986, pp. 24-2S) which say that the syntax (or subsequent components) can
have no access to the internals of words. To substantiate this claim, D&W
present evidence that appears to show that various syntactic phenomena
fail to make use of information internal to compounds. The strength of the
evidence, however, fades rapidly on serious examination.
Consider, for example, the observation (D&W, p. 49) that rules of syn
tax cannot "discriminate among compounds whose nonhead members are
plural or singular, " so that the parks commissioner is treated by all rele
vant syntactic rules exactly the same way as the park commissioner. This
seems simply to be an observation about headedness: in exactly the same
way, relevant syntactic rules will not distinguish the commissioner of meats
from the commissioner of meat. Since plurality is not being registered on
the head of the noun phrase in either case, any syntactically relevant rule
which cares about number (such as verb agreement) will register the gram
matical number of the head and not be afected by the plurality of modifers
or complements. In this respect there is absolutely nothing special about
the compound case: one does not need to appeal to the existence of a
separate morphological component to derive the relevant properties of the
parks commissioner.
A second set of examples concerns cases like * it robber (see D& W, p. 50)
or * who killer, which are argued to show that referential expressions, which
are relevant to the syntax, may not occur within words. But, as we have
already suggested (Section 4. 1 ) , there seems to be a constraint ruling out
any kind of derivation fom closed-class items; alternatively, one could claim
that no projection of functional categories can serve as modifers. In any
case, we have argued that modifers-whether at the NO or the N
1
level
are maximally X
l
, thus ruling out a whole clas of DP or DegP modifers.
These considerations subsume D&W's facts. Referential expressions are
not in principle ruled out from occurring within words, but only those
expressions which have the above mentioned forms. Proper names can
certainly occur within compounds although D&W suggest that: (i) only
famous names really allow this, since * Bill admirer is supposedly unac
ceptable; and that (ii) even in such cases a Nixon admirer, Nixon doesn't
really refer to Nixon. We briefy examine these claims.
With respect to the second claim, it has been argued extensively in
Sproat and Ward 1987 and Ward et al. 1991, that names which are mor
phologically contained within compounds can indeed be used to refer to
particular individuals, and this can be seen by the fact that such names
can function as antecedents to pronouns:
22
( 1 08) a. There's a Thurberi-story about hisi maid . . .
b. I was reading this Peggy Noonalli book on heri years at the White
House . . .
c. We went up to Constablei country; we stayed in the village hei
was born in.
D&W's own evidence for their claim concerning the referentiality of Nixon
in Nixon admirer is the assertion that in ( 109a) , admiring Nixon is not an
essential property of a Nixon admirer, since this example is not a contra
diction, unlike ( 109b) :
( 109) a. John is a Nixon admirer i n every sense except that he does not
admire Nixon.
b. John admires Nixon in every sense except that he does not admire
Nixon.
Ward et al. ( 1991 , footnote 16) have argued that ( 109a) is only non
contradictory under the interpretation of Nixon admirer as a person with
a reliable set of traits (such as being clean-shaven, wearing a three-piece
suit and always carrying an attache case) which are independent of the
22
Contrary to the predictions of the socalled Anaphoric Island constrint (Postal 1969).
property of admiring Nixon: in other words, Nixon admirer would have to
have accreted additional meanings via institutionalization, along the lines
of Maoist. In that case a person might be said to have those properties and
thus qualif as being a Nixon admirer, even if he or she does not actually
admire Nixon; hence ( 109a) would not be a contradiction. Needless to say,
Nixon admirer has not achieved this institutionalized status in the minds
of most speakers, and the contrast in ( 109) is therefore very hard to get.
More important for D&W's argument, however, is the observation that
the putative accretion of such institutionalized meanings is largely orthog
onal to the issue of whether Nixon in Nixon admirer can be used to refer
to Nixon. Hence, even making the counterfactual assumption that the con
trast in ( 109) were a clear one, that would not serve to support D&W's
claim concerning the referential properties of compound-internal elements.
Returning to point (i), the obvious question which arises when con
fonted with an example like Bill admirer is: who is Bill? One might
suppose that Bill admirer ought to be able to mean something like 'one
who admires people with the name Bill' , but there is really no reason to
believe that it should have this reading: in its typical phraal usage, a
proper name like Bill is invariably used to refer to a specifc individual.
Therefore one would expect Bill in Bill admirer also to be used to refer
to some specifc individual, and if we do not know who that individual is,
the example seems odd. The oddity is reinforced by the fact that, as is
well-known, "agentive" nominals tend to have a "habitual" reading, and
are typically used to refer to characteristic properties. To ascribe to some
one the characteristic property of admiring some unspecifed person named
Bill seems odd indeed. Naturally, famous names do not have this problem,
since there is no question about the intended referents of the names. And
Bill admirer is, it seems to us, fne in a context where the interlocutors can
assign a defnite referent to Bill.
Still, there is no question that frst names generally seem somewhat
degraded over last names in such contexts: compare ? Noam admirer with
Chomsky admirer. At least in part this seems to have to do with the fact
admirer has an institutionalized connotation of 'admires in the professional
sense' , and this in turn requires that the admired individual be referred
to using their normal professional designation, typically the lat name. 23
But familiarity or conventional designation are clearly extragrammatical
considerations, and there is no reason to suppose that there should be a
statement in the grammar of English nominal compounds restricting the
occurrence of proper names in that context.
23
A similar constraint applies to cases of deferred reference (thanks to Mats Rooth for
the second example):
i. You'll fi nd Chomsky on the top shelf.
ii. ?#You'll fnd Noam on the top shelf.
In summary, we do not fnd any of the reasons adduced by D&W for sep
arating compounding from phrasal constructions to be compelling. There
fore it seems possible to accept the assumption that English NO compounds
are word-level objects that are formed in the syntax.
5 Two Approaches to Parsing and Stress Assignment
We turn now fom the descriptive and theoretical issues which have occu
pied us for the bulk of this paper to the issue to which we would ultimately
like to provide a solution, or at least a good approximation thereto, namely
the correct parsing of and stress assignment to modifed nominals as they
occur in unrestricted text. We outline here some of the methods we have
applied to solving this problem.
It is of practical concern that the correct stress pattern for a sequence
[N N] or [A N] cannot be generally predicted fom the preterminal sequence
of categories alone. That is, if we know that we have a sequence of two
nouns, we cannot be certain that we have an NO as opposed to an N1 for
a parent category, though if the preterminal sequence were all we could
compute, guessing NO would more often lead to a correct stress assignment
than guessing N
1
. In general, some other type of analysis is necessary. In
addition to the problem of deciding upon the appropriate stress assignment
to a binary modifed noun, there is the at leat equally difcult problem
of parsing more complex cases. As we shall see, the two problems, stress
assignment and parsing, are related in the methodologies which one can
apply to them. We shall frst turn to the problem of stress assignment, and
then consider parsing.
5. 1 Stress Assignment in Binary Nominals
The phrase safety board has main stress on the left. As we have argued
for other cases in the text, we might decide this because the phrase is an
instance of a semantic pattern, informally stated as Nl GROUP-WHOSE
CONCERN-IS-Nl. In Sproat and Liberman 1987 and Sproat 1990, we show
that a rather crude implementation of such a method can improve consid
erably on the performance of an approach based purely on the syntactic
categorization of the words in the text stream. However, as we noted in
our discussion of the semantic patterns of argument-argument compounds,
it is not at all clear on what basis a single, consistent, coherent set of such
schemata can be defned.
Alternatively, we might depend on the fact that the words in the phrase
safety board tend to yield a left-dominant stress pattern in other cases. So
one might expect that the probability that the nominal will be stressed on
the left given that the frst word is safety and the second board is greater
than the probability of righthand stress under those conditions, or in other
words:
p(s
l
l w
l
=safety) p(S
l
l w2
=board) ` p(s
2
l wl =safety) p(s
2
I
w
2
=board)
A survey of examples fom a large corpus of naturally-occurring pre
modifed nominals shows that this is indeed the case-safety as a noun
modifer wa always stress-dominant, while board when pre-modifed by a
noun was never stress-dominant. A representative sample:
( 110) a. safety belt, safety binding, safety council, safety devices, safety
equipment, safety feature, safety glasses, safety harness, safety
margin, safety meaure, safety net, safety nut, safety precautions,
safety problem, safety procedure, safety razor, safety record, safety
regulations, safety requirement, safety rule, safety school, safety
shoes, safety standards
b. Federal Reserve board, advisory board, bulletin board, chalk
board, chess board, circuit board, cutting board, dart board, draf
board, drawing board, governing board, ironing board, memory
board, particle board, planning board, promotion board, school
board, score board, skate board, supervisory board, surf board,
wall board, zoning board
Although most examples are not as clear-cut as this one, we believe that
the method can work fairly well if properly trained. Its main drawback is
that many words do not occur often enough in the needed constructions to
generate useful statistics-it seems appropriate, in such cases, to depend
on the observed behavior of "similar" words. The crux of the matter is
then the similarity metric to be used.
5. 2 Parsing Premodifed Nominals
Again, we have explored two methods. One approach would use phrasal
schemata of the kind used in assigning stress to decide upon a possible
analysis for a modifer-head string; one might frther extend such a method
along the lines of Finin 1980 by adding scores for the various schemata, and
pick the parse whose cumulative score is highest. This approach ha the
previously noted difculties of creating a reasonable set of schemata, and,
if scores are used, also the problem of juggling the interaction of possibly
ad hoc scores, endemic to diagnostic systems of this type. Nevertheless,
if the schemata are chosen so as to refect the patterns found in a certain
genre of text, the technique can work reasonably well.
A system of this kind currently under development in the context of
Bell Labs ongoing text-to-speech efort is reported in Sproat 1990,2
4
and is
an extension of earlier work reported in Sproat and Liberman 1987. The
program depends on semantic patterns of the kind described in this paper
a well as large lists of common binary nominals to attempt to compute
24
The scoring method alluded to above for picking among various possible semantic
interpretations is not implemented in the system reported in Sproat 1990, though various
a hoc heuristics are used to pick one of several possible analyses.
a reasonable structure of multiply modifed nominal phrases. As a simple
example consider a semantic grammar which contains information that ta
ble, chair and some other items are HOUSEWAR&j that kitchen, bathroom
and living room are ROOMs (and incidentally also NO, this latter informa
tion being crucial for the binary cases) j and that there is a pattern, which
says that a ROOM word and a HOUSEWARE word may be combined into
an Nl , which we will call a ROOM&THING. this rule can be thought of
a an instance of the pattern PLACE-WHERE-N2-IS-FOUND N2 discussed in
( 70). In addition to semantic grammar rules and the rather large lexicon
of binary cases, there are more traditional syntactic rules which handle
syntactic aspects of noun phrase structure, such as the treatment of deter
miners. The system in its current form can asign appropriate structure
and stress to noun phraes such as the following:
( 1 1 1 )

John's]
large
(t',ROOM&THING)
(t,ROOM)
living
room]
(t ,
HOUSEWARE)
table]]]]
In this example, living room is correctly analyzed as a modifer of N
l
.
Another module of the program responsible for assigning stress can then
take this assigned structure and determine that the main stress should be
placed on the head noun in this case.
The second approach to parsing again relies on the statistical behavior
of individual words, identifing the afnity of two adjacent words with the
extent to which they occur together more often than one would predict
baed on how often they occur individually. One reasonable measure for
such afnity might be MUTUAL INFORMATION, where the mutual informa
tion I(a, b) between events A and B with probabilities p(a) and p(b) is
defned a
( 1 12) I(a, b) log
2
J
Mutual information meaures have been used recently by Sproat and Shih
( 1990a)25 in a domain rather similar to the current one, namely the prob
lem of locating word boundaries in Chinese text. Chinese orthography, of
course, traditionally does not indicate the location of word boundaries, but
it turns out that considerations of the strengths of association between ad
jacent characters in a Chinese text, as measured by mutual information can
achieve about 95% retrieval and precision for two-character words (which
constitute the bulk of multicharacter words in Chinese text) .
However, there is a problem with defning the association measure
strictly in terms of mutual information as defned above. Instances of the
sequence ABC in fact give us no information about the relative afnity of
B for A a opposed to C, so if a signifcant faction of (say) AB instances
occur in ABC sequences, while BC is quite a bit more common outside
2
5
8ee also Magerman ad Marcus 1990 for more extensive use of mutual information in
parsing.
1 78 / MARK LIBERMAN AND RICHARD SPROAT
A B C AB BC ABC
F F F F I IX F I I X F
i
nteragency task forcers)
35 756 6909 4 10.83 00 287 10.07 IUU 4
environmental imp
act statement(s)
801 1110 4626 25 8.40 7.U 1 6 5.23 3.25 12
p
rivate arm
s
d
ealer(s)
3296 3822 1059 21 4.32 4.02 57 7.40 7.U 4
electoral college system(s)
405 1949 4691 93 10.47 9.29 62 9. 23 2.04 59
Table 1
these sequences, then a high value for I(A, B) will be misleading. It is bet
ter to compute a mutual-information-like afnity measure that excludes
the useless ABC instances fom the counts. If we use -T to refer to the
complement of the cases in which T is found, our afnity estimate for words
A and B within the ternary unit T becomes:
( 1 13)
Now, compare the right-branching ternary nominal interagency task jorce
and private arms dealer with the left-branching ternary nominals environ
mental impact statement and electoral college system. In a 12-million-word
sample of the Associated Press newswire, things work out as in Table 1 .
I n this example, the IX measure
2
6 gives the correct analyses, while the
I measure fails for the case of interagency task jorce, which ha the noted
property in that the sequence interagency task never occurs other than
in the larger fame. In general, the IX measure gives a stronger indica
tion of afnity in the correct direction, even where the I measure is also
correct.
Methods of this general type have a great deal of promise as aids to
parsing in the all-too-common cases where structural indications are weak
or lacking. However, it would be desirable to consider, in estimating the
afnity between two words, the distribution of "similar" words. 27 Also,
the generalization of the simple ternary cae to more elaborate structures
can be done in a variety of ways, and it is not trivial to fnd methods
that are both tractable and correct. We believe that the analyses given in
this paper will help to guide such explorations in sensible and productive
directions.
26
Note that we are using JJ to refer to the cae where a pair does not occur other than
in the triple.
2
7
Indeed, one of the problems encountered in using this statistical meaure a an a
ditional option in the system reported in Sproat 1990 is that there is often not enough
data about particular pairs of words in a gven triple to estimate IX, even if one uses a
moderately large corpus of about 10 million words.
References
Abney, Steven. 1987. The English Noun Phrase in its Sentential Aspect. Doctoral
dissertation, Massachusetts Institute of Technology.
Baker, Mark. 1988. Morphological and Syntactic Objects: A review of Di Sciullo
and Williams' LL ILC 1CDL1I10L 0!W01O. Morhology Yearbook 1 :259-283.
Bauer, Lauri. 1983. English Word-Formation. New York: Cambridge University
Press.
Bliberg, O. 1988. A Study of Swedish Compounds. Doctoral dissertation, Umea
University.
Bolinger, Dwight. 1967. Adjectives in English: Attribution and Predication.
Lingua 18: 1-34.
Bolinger, Dwight. 1972. Accent is Predictable (if you're a mind-reader) . Language
48:633-644.
Chomsky, Noam. 1970. Remarks on Nominalization. In Readings in English
Transformational Grammar, ed. Roderick A. Jacobs and Peter S. Rosenbaum,
184-221. Waltham, Massachusetts: Ginn and Company.
Chomky, Noam, and Morris Halle. 1968. The Sound Patter of English. New
York: Harper and Row.
Clements, J. 1989. Lexical Category Hierarchy and Syntactic Headedness in
Compounds. In Proceedings of the 6th Annual Easter States Conference on
Linguistics, 46-57.
Cutler, Anne, John Hawkins, and Gary Gilligan. 1985. The Sufxing Preference:
A Processing Explanation. Linguistics 23: 723-758.
Di Sciullo, Anne-Marie, and Edwin Williams. 1987. On the Defnition of Word.
Cambridge, Mass. : MIT Press.
Downing, Pamela A. 1977. On the Creation and Use of English Compound
Nouns. Language 53:810-842.
Fabb, Nigel. 1984. Syntactic Afxation. Doctoral dissertation, Massachusetts
Institute of Technology.
Finin, Timothy W. 1980. The Semantic Interpretation of Compound Nominals.
Doctoral dissertation, University of Illinois at Urbana-Champaign.
Fudge, Erik. 1984. English Word-Stress. London: George Allen and Unwin.
Hayes, Bruce. 1980. A Metrical Theory of Stress Rules. Doctoral dissertation,
Masachusetts Institute of Technology (distributed by IDLe) .
Hirschberg, Julia. 1990. Using Discourse Context to Guide Pitch Accent Deci
sions in Synthetic Speech: The given/new distinction and deaccentability. In
Proceedings of AAAI-90, 952-957.
Hoeksema, Jacob. 1985. Categorial Morphology. New York: Garland Publishers.
Jackendof, Ray. 1977. X-bar Syntax: a Study of Phrase Structure. Cambridge,
Mass. : MIT Press.
Kiparsky, Paul. 1982. Lexical Morphology and Phonology. In Linguistics in the
Moring Calm, ed. I.-S. Yang, 3-91. Seoul: Hanshin.
Ladd, Robert. 1984. English Compound Stress. In Intonation, Accent and
Rhythm, ed. Dafydd Gibbon and Helmut Richter, 253-266. Berlin: Mou
ton de Gruyter.
Lees, Robert B. 1960. The Grammar of English Nominalizations. Interational
Joural of American Linguistics, Publication 12.
Levi, Judith N. 1978. The Syntax and Semantics of Complex Nominals. New
York: Academic Press.
Levin, Beth, and MaIka Rappaport. 1992. -Er Nominals: Implications for a
Theory of Argument-structure. In Syntax and Semantics 26: Syntax and the
Lexicon, ed. Eric Wehrli and Tim Stowell. New York: Academic Press.
Liberman, Mark, and Alan Prince. 1977. On Stress and Linguistic Rhythm.
Lieber, Rochelle. 1981. The Organization of the Lexicon. Doctoral dissertation,
Masachusetts Institute of Technology.
Lieber, Rochelle. 1983. Argument Linking and Compounds in English. Linguistic
Inquiry 14:251-286.
Lieber, Rochelle. 1988. Phrasal Compounds in English and the Morphology
Syntax Interface. In Papers from the 24th Regional Meeting of the Chicago
Linguistic Society, 398-405. University of Chicago.
Lieber, Rochelle. 1992. Deconstructing Morhology: Word Formation in Syntac
tic Theory. Chicago: University of Chicago Press.
Magerman, David, and Mitchell Marcus. 1990. Parsing a Natural Language Using
Mutual Information Statistics. In Proceedings of AAAI-90, 984-989.
Maidment, J. A. 1989. On the Accentuation of Compounds in English. In Speech,
Hearing and Language: Work in Progress, 181-188. University College, Lon
don.
Marantz, Alec. 1989. Projection vs. Percolation in the Syntax of Synthetic Com
pounds. Manuscript, University of North Carolina.
Marchand, Hans. 1969. The Categories and Types of Present-Day English Word
Formation. Miinchen: C. H. Beck'sche Verlagsbuchhandlung.
Mohanan, K. P. 1986. The Theory of Lexical Phonology. Dordrecht: Reidel.
Postal, Paul. 1969. Anaphoric Islands. In Papers from the 5th Regional Meeting
of the Chicago Linguistic Society, 205-239. University of Chicago.
Ramsey, S. Robert. 1987. The Languages of China. Princeton: Princeton Uni
versity Press.
Roeper, Thomas, and Mufy Siegel. 1978. A Lexical Tansformation for Verbal
Compounds. Linguistic Inquiry 9: 199-260.
Safr, Ken. 1987. The Syntactic Projection of Lexical Thematic Structure. Natural
Language and Linguistic Theory 5:561-602.
Selkirk, Elizabeth. 1984. Phonology and Syntax. Cambridge, Mas. : MIT Press.
Selkirk, Elizabeth O. 1982. The Syntax of Words. Cambridge, Mas. : MIT Press.
Sproat, Richard. 1985. On Deriving the Lexicon. Doctoral dissertation, Mas-
Sproat, Richard. 1990. Stress Assignment in Complex Nominals for English Text
to-Speech. In Proceedings of ESCA Workshop on Speech Synthesis, 129-132.
Sproat, Richard, and Mark Liberman. 1987. Toward Teating English Nominals
Correctly. In Proceedings of the 25th Annual Meeting of the Association for
Computational Linguistics, 140-146. Association for Computational Linguis
tics.
Sproat, Richard, and Chi-lin Shih. 1990a. A Statistical Method for Finding Word
Boundaries in Chinese Text. Computer Processing of Chinese and Oriental
Languages 4:336-351 .
Sproat, Richard, and Chi-lin Shih. 1990b. The Cross-linguistic Distribution of
Adjective Ordering Restrictions. In Interdisciplinary Approaches to Language,
ed. Carol Georgopoulos and Roberta Ishihara, 563-593. Dordrecht: Kluwer
Academic Publishers.
Sproat, Richard, and Gregory Ward. 1987. Pragmatic Considerations in
Anaphoric Island Phenomena. In Papers from the 23rd Regional Meeting
of the Chicago Linguistic Society, 321-335. University of Chicago.
Stump, Greg. 1989. A Note on Breton Pluralization and the Elsewhere Condition.
Natural Language and Linguistic Theory 7: 261-273.
Thomaon, Richard. 1985. Some Issues Concerning the Interpretation of Derived
and Gerundive 'Nominals' . Linguistics and Philosophy 8: 73-80.
Ward, Gregory, Richard Sproat, and Gail McKoon. 1991 . A Pragmatic Analysis
of So-called Anaphoric Islands. Language 67(3) :439-474.
Williams, Edwin. 1981. On the Notions 'Lexically Related' and ' Head of a Word' .
Hungarian Derivational Morphology,

Semantic Complexity, and Semantic
Markedness
FERENC KIEFER
1 Introduction
It i a well-known observation that the meaning of some words are more
complex than others. In the simplest case the more complex meanings in
clude simpler meanings plus other components. 1 For example, the meaning
of the verb sell is semantically more complex than that of the verb give
since it contains the semantic components of give plus some other semantic
material. In general, the more semantic material the meaning of a word
contains, the more complex it is semantically. The notion of semantic com
plexity is indispensable for any study of the organization of the lexicon and
it can also be ued to predict the order in which individual lexical items
are acquired by children.
It has also been noticed that people divide things into categories, each of
which is given a name. These categories are often hierarchically organized
with several levels of abstraction. For example, plant-tree-pine-Ponderosa
pine-norther Ponderosa pine. It has been argued that the third level the
generic level) is the most basic "basic-level terms" ) . A basic-level term
such as pine, oak, elm is more useful than any term on a higher or lower
level of abstraction. Therefore, basic-level terms are acquired by children
earlier than non-basic ones. Thus, in certain cases at least, the complexity
I am grateful to Wolfgang Dressler, Edith Moravcsik, and Anna Szabolcsi for detailed
comments on earlier versions of this paper. I would also like to thank the audience of
the Conference on Lexical Semantics at Stanford ( 1987) for many insightful comments
and questions.
1 In what follows I am going to draw heavily on Clark and Clark 1977, esp. pp. 337-342
and 487-501.
Lczicul Muttcrs. Ivan A. Sag and Anna Szabolcsi, eds.
184 / FERENC KIEFER
hypothesis and the basic-level hypothesis make diferent predictions a to
order of acquisition.
It has been suggested that it is possible that semantic complexity ap
plies to verbs, adjectives and relational terms and basic-level predictions
to category names. However, this proposal does not do justice to the facts
since similar phenomena as with basic-level terms can also be encountered
among verbs, which are clearly relational. This is one of the points which
will be made in the present paper. As a consequence, a theory is called for
which is able to reconcile the two hypotheses. The main aim of this paper
is to sketch such a theory.
The meaning of a word can be more baic than the meaning of a related
word in several ways, "basicness" is not restricted to levels of utility as
stipulated for basic-level terms. In the present paper we will use the term
'semantic markedness' to describe phenomena which involve "basic ness" or
"naturalness," i. e. , we will say that of two words u and u_ , which belong
to the same semantic feld, u is semantically more marked than u_ if
u is less basic or natual than u_ . Semantic markedness is governed by
various principles such as transparency of meaning, lexicalization patterns,
restrictions imposed on the use of the words, etc. In this paper we will
concentrate on the problem concerning restrictions on use.
We will see that whenever there is a confict between the complexity and
the markedness hypothesis, the latter takes precedence over the former.
Semantic complexity contrasts with formal complexity. Formal com
plexity may have to do with morphology or with syntax. In languages such
a English morphological complexity is often explained in terms of irreg
ularity. For example, the past tense in English is formally more complex
than the expression of possession. The past tense is usually expressed
by the sufx - ed, but there are numerous exceptions. The expression
of possession, on the other hand, is quite regular. In the case of lan
guages such as Hungarian, which is an agglutinating language, morpho
logical complexity can be defned in terms of the number of morphemes
which occur in a word. Consider, for example, the verb ir 'write' and
the forms ir+at (' +
'
will be used to denote morpheme boundary through
out this paper) 'make write' , ir+at+hat ' may make write' , ir+at +hat+na
' could make write', ir+at+hat+nal 'you could make write' with increas
ing morphological complexity. In general, forms which contain more mor
phemes require a larger amount of processing time than forms with fewer
morphemes.
Syntactic complexity may be defned in terms of predicate-argument
structure, i. e. , by the number of arguments required by a predicate and the
various restrictions imposed on these arguments.
There are phenomena both in morphology and in syntax which cannot
be explained by means of the notions of morphological or syntactic com
plexity. Some of these will be touched upon briefy in the present paper.
HUNGARIAN DERIVATIONAL MORPHOLOGY / 185
Though it will be proposed to use the terms 'morphological markedness'
and 'syntactic markedness' , respectively, for these cases, nothing substan
tial will be said about "formal" markedness in this paper.
There is a strong correlation between semantic complexity and syntac
tic complexity, on the one hand, and between semantic complexity and
morphological complexity on the other. These correlations follow fom the
"coding principle" well-known in psycholinguistics which says that "com
plexity in thought tends to be refected in complexity of expression." Such
correlations do not seem to exist between the various types of markedness.
The observations to be presented in this paper are based on Hungar
ian. The semantic felds to be investigated are defned by the rules of
derivational morphology. This means that we will leave out of considera
tion semantic felds whose members are not relatable to each other by such
rules. Semantic felds of the former type make up a signifcant portion of
lexical structure in Hungarian.
The reason why I think that it is worthwhile to restrict the examina
tion of the interrelationship between complexity and markedness to seman
tic felds defned by derivational morphology is that in a language such as
Hungarian the formal structure of words refects complexity of content.
Consequently, the feasibility and usefulness of the decomposition hypoth
esis cannot be questioned. It is hoped, however, that the basic ideas put
forward here can be extended to the analysis of other types of lexical items
as well.
The principles proposed are claimed to hold for languages which ae
typologically similar to Hungarian (roughly, for agglutinating languages),
though most of them may turn out to be universal.
The main claim of the present paper can thus be summarized:
(i) Semantic complexity and semantic markedness are two diferent no
tions. Semantic complexity is based on the amount of the semantic mate
rial contained in the lexical items, semantic markedness, on the other hand,
has to do with restrictions on use, but it is also related to such notions a
productivity and semantic transparency.
(ii) Alongside of the distinction between semantic complexity and se
mantic markedness, a similar distinction between formal (morphological,
syntactic) complexity and formal (morphological, syntactic) markedness
has to be drawn.
(iii) Semantic complexity determines semantic markedness to a large
extent, however, whenever there is a confict between semantic complexity
and semantic markedness, the latter takes precedence over the former.
(iv) It is markedness and not complexity that defnes acquisition order.
(v) Markedness which is based on the notion of "level of utility" is not
restricted to category names as suggested by earlier studies but has to be
extended to relational terms a well. In this way the confict between the
two theories about acquisition order disappears.
186 / FERENC KIEFER
2 Morphological, Syntactic, and Semantic Complexity
Let's frst defne the notions of morphological, syntactic and semantic com
plexity.
2.1 Morphological Complexity
For the description of Hungarian, Hungarian being an agglutinating lan
guage, morphological complexity can be defned easily. We may simply
identify morphological complexity with the number of morphemes con
tained in a word.
2
More precisely,
( 1 ) A word Wi is morphologically more complex than another word W
j
if W
i
contains more morphemes than W
j
.
According to this defnition, (2a) is morphologically less complex than (2b) ,
and ( 2b) is morphologically less complex than (2c). Furthermore, (2b,d,e)
are all of equal complexity.
(2) a. olvas 'read'
b. olvas+tat 'make read' (causative)
c. el+olvas+tat ' make read' (causative-perfective)
d. olvas+gat 'read little by little' (frequentative)
e. el+ olvas 'read' (perfective)
Let's introduce the symbol ' <m
'
for the morphological complexity relation.
We may thus represent the complexity relations at hand as in (3).
(3) olvas <m olvas+tat <m el+olvas+tat
It goes without saying that the notion of morphological complexity has to
be defned diferently for each language type.
We know that complexity is not the only principle that determines
the acquisition of morphology. Semantic transparency and productivity
may even be more important principles than the complexity of form. In
general, of two forms of equal morphological complexity the one which is
more productive or which is more transparent in meaning is more basic or
more natura1.
3
Consider, for example,
(4) a. gereblye+z 'rake' , zongoni+z{ik) 'play the piano'
b. hal+asz{ik) 'fsh', vad+asz{ik) ' hunt'
2Clearly, defnition (I) is the simplest possible defnition of morphological complexity
which holds only for "pure" agglutinating languages of which Hungarian is not an exam
ple. In Hungarian, in addition to the number of morphemes, there are further aspects
of morphological structure that may contribute to morphological complexity, e.g., stem
alternations, "fused" morphemes, asimilation rules, etc. In the present paper we will
neglect these additional complications, however.
3
Naturalness phenomena constitute the main motivation for the introduction of the
notion of markedness in morpholog a well a in syntax and semantics. (See further
below. ) In most works complexity and markedness are not kept apart. See, however,
the recent discusions of morpholog (Dressler, Mayerthaler, Panagl and Wurzel 1988).
That transparency of meaning and productivity play an important role in the acqui
sition of derivational morpholog ha also been pointed out by Clark ( 1988).
Notice that both (4a) and (4b) are bimorphemic but whereas the sufx -z is
fully productive and quite fequent, the sufx - asz is no longer productive
in that it occurs in a small number of lexicalized items only. Consequently,
the items in (4b) may be claimed to be less basic or less natural than the
ones in (4a) .
However, afx productivity is only one factor that determines basicness.
In some cases a productive sufx can also be added to a phantom root, i. e. ,
to a bound morpheme with no clear independent meaning. Consider, for
example, the Hungarian sufx -int, which is used to derive instantaneous
verbs: emel+int 'give it a lift' , csavar+int 'give it a twist', sodor+int
'give it a twirl' , etc., where the roots are full verbs with independent mean
ings emel 'lift' , csavar 'twist' , sodor 'twine, twirl' ) . However, we may
also encounter verbs such as koh+int 'give a little cough' or b6l+int 'nod',
where the bound morphemes koh- and b6l- are semantically not transpar
ent. (That they are nevertheless identifable as morphemes is testifed by
the fact that they also occur with other derivational sufxes, e.g., koh+og
'cough' and b6l +ogat 'keep nodding' . ) In view of the diference in semantic
transparency one may argue that koh+int and b6l+int are less baic than
emel+int, csavar+int and sodor+int.
To mention yet another case where ease of perception may enter into
play consider again the sufx -z (example (4a) above) together with its
variant - zik:
(5) a. bor+oz 'sit over one's wine' , szem+ez 'ogle'
b. tanya+zik 'take shelter' , zongoni+zik 'play the piano'
Notice that in the case of (5b) the morpheme boundary between the root
and the sufx coincides with a syllable boundary, that is, syllabifcation
does not afect morphemic structure. In the case of (5a) , on the other
hand, syllabifcation yields (6) .
(6) bo&roz, sze&mez
This means that the morpheme boundary between the root and the sufx
does not coincide with a syllable boundary. In other words, the forms in
(5b) are morphologically more transparent than those in (5a) . Degrees of
morpho tactic transparency correlate with degrees of naturalness, hence the
forms in (5b) are more natural or more basic than those in (5a) . 4
In sum, then, semantic transparency, morpho tactic transparency and
productivity are principles which may run counter morphological complex
ity. The efects of these principles may be subsumed under the term ' mor
phological markedness' . This means that, in addition to a theory of mor
phological complexity, a theory of morphological markedness is called for
to explain morphological structure. In what follows I won't have anything
more to say about morphological markedness.
4For the notion of morphotactic transparency see Schaner-Wolles and Dressler 1 985.
188 I FERENC KIEFER
2.2 Syntactic Complexity
For the purposes of the present discussion we will be content with a rather
restricted notion of syntactic complexity. Consider
( 7) a. Ulok.
'I'm sitting'
b. imk ( egy levelet) .
'I'm writing ( a letter) '
c. Megirok * ( egy levelet).
Megirok egy levelet.
' I'm going to write up a letter'
The verb il 'sit' requires just a subject, the verb ir ' write' a subject and
optionally an object, the perfective verb megir 'write (up/down) ' a subject
and an obligatory object. In other words, il is a one-place predicate, ir
can be used either a a one-place predicate or as a two-place predicate,
and megir is a two-place predicate. The verb megir will be said to be
syntactically more complex than the verb ir which, in turn, will be said to
be syntactically more complex than the verb il.
To take another example, the verb beszel 'talk' and beszel-get 'converse'
are alike except that the latter requires either the plurality of agent or a
second agent represented by a with-phrase:
(8) a. Peter i Janos beszelgetnek.
'Peter and John are conversing'
b. Peter beszelget Janossal.
'Peter is conversing with John'
The verb beszelget has thus a more complex predicate-argument structure
than the verb beszel, it is syntactically more complex.
On the bais of the above observations the following defnition of syn
tactic complexity may be stipulated.
5
(9) A predicate P
i
is syntactically more complex than another predicate
Pj if P
i
contains more obligatory arguments than Pj and/or if the
arguments required by P
i
are syntactically more restricted than those
required by Pj
Again, there are good reasons to distinguish syntactic complexity from
what may be termed syntactic markedness. In Hungarian a transitive verb
may, in general, appear in sentence initial position, afer the subject but
before the object, and in sentence fnal position. For example,
5
A note of caveat is in order here. Syntatic complexity should by no means be reduced
to a simple count of the arguments. The defnition proposed under (9) is, no doubt, over
simplifed. The notion of syntactic complexity should be baed on full-Hedged syntactic
representations.
( 10) a. Janos eszi az almat.
John eats the apple
'John is eating the apple'
b. Janos az almat eszi .
'It is the apple that John is eating'
c. Eszi Janos az almat.
'John does eat apples'
It should be noted that the English translations are not the only possible
interpretations of the corresponding Hungarian sentences. Now, any native
speaker of Hungarian would readily admit t hat-quite independently of
the meaning diferences-only (lOa) sounds quite neutral, ( lOb) and ( lOc)
are somehow "marked. " The details are, of course, anything but clear.
But, again, we don't have to bother about this question in the present
context.
2. 3 Semantic Complexity
Semantic complexity can be defned on the basis of the amount of the
semantic material contained in semantic representations. For example,
causatives are semantically more complex than non-causatives since the
former, but not the latter, contain the (elementary) predicate CAUSEx, y)
in their semantic representations. In this sense, then, the causative irat
' make write' is semantically more complex than the non-causative ir 'write'.
Similarly, the causative olvastat ' make read' is semantically more com
plex than the corresponding non-causative olvas 'read'. Or, to take an
other example, the perfective forms of telic predicates are semantically
more complex than the corresponding imperfectives since they contain the
(omponents BECOMEy, z) and TERMINATEDz) in their semantic repre
sentations which are obviously lacking in the semantic representations of
the corresponding imperfectives. Thus, the form megir 'fnish writing' is
semantically more complex than ir 'write' and the form elolvas ' fnish read
ing' is semantically more complex than olvas 'read' . Yet another case can
be illustrated by the fequentative. Frequentatives are semantically more
complex than the corresponding non-frequentatives since the former, but
not the latter, contain the semantic feature REPEATEDLY PREDx, y) ) .
Thus the fequentative irogat 'write often' is semantically more complex
than the corresponding non-fequentative ir 'write' and the fequentative
olvasgat 'read little by little' is semantically more complex than the non
fequentative olvas 'read'.
So far we have been considering categories which had to do with aspect
or Aktionsart. There are of course quite a few other things which contribute
to semantic complexity. Remaining within the confnes of derivational mor
phology, notice that t here are pairs of verbs in Hungarian which difer from
each other in presuppositional properties only. For example, tagad ' deny' is
190 / FERENC KIEFER
non-factive but letagad 'deny' with the prefx le- is factive. The verbs titkol
'hide, conceal' and eltitkol 'keep to oneself, withhold' represent a similar
pair, where the latter verb is presupposition-inducing whereas the former is
not. In general, the presupposition-inducing forms are semantically more
complex than the corresponding non-inducers. Notice, frthermore, that
verbal afxes do not only change the aspectual character or the Aktionsart
of the verb to which they are attached but very often introduce further
(in general, unpredictable) semantic material. This is particularly true
of verbal prefxes except for the perfectivizing prefx meg- . Consequently,
whereas forms such a megir 'write' with the perfective prefx have got
a transparent meaning in the majority of cases, forms such as leir 'write
down, register, copy, . . . ' or felir ' write down, prescribe, charge, . . . ' show
unpredictable meanings. Both felir and leir are perfective but they con
tain further semantic material which is not present in the corresponding
imperfective forms. Hence felir may be said to be semantically more com
plex than megir which, in turn, as we already saw above, can be said to be
semantically more complex than ir.
By restricting ourselves to the semantic feld defned by derivational
morphology we need not take a stand on how the meaning of root words
should be described and we don't have to bother about the question of
how far down compositional meanings have to be looked for. By taking
the meaning of the root word for granted we will simply ak the question
of what the contribution of each derivational morpheme to that meaning
is. If the semantic contribution of a derivational morpheme cannot be
accounted for by general rule, the meaning of the derived word will be
noncompositional or idiomatic.
The above examples may sufce to illustrate semantic complexity.6
To be sure, in order to be able to exactly defne the notion of semantic
complexity one would frst have to develop a theory of semantic representa
tions. Diferent theories may lead to diferent semantic representations of
the very same item which entails that it may be characterized by diferent
6
An interesting and in many ways illuminating way of approahing this can be found
in Miller and Johnson-Laird 1976. Starting out with a set of core concepts (such a
TRAVEL, POSSESS, SEE, UTTER, SAY) and a set of baic semantic operators (such
a HAPPEN, DO, ACT, POSSIBLE, CAUSE), it is claimed that a core concept can
be refned, modulated by means of these semantic operators in several ways. Each re
fnement or elaboration of a core concept makes it semantically (conceptually) more
complex. A semantic field is defined by means of various ramifications brought about
by these semantic operations. Thus, for example, the core concept TRAVEL(x) can
be refned by introducing a presuppositional (LEAVE(x, or by specifing the man
ner ( LURCH(x)), or by introducing a causal component ( CHASE(x, y)), etc. (op. cit.
pp. 666-67). Our discussion of Hungarian derivational morpholog will be similar to
the approach advocated by Miller and Johnson-Laird in one important apet. Each
derivational extension may be taken to correspond semantically to the application of
a semantic operator (apectual, directional or other) and may thus count a a further
elaboration of the given meaning by rendering it more complex.
degrees of semantic complexity, depending on the semantic representation
one wants to adopt. For the time being we have to be content with the
following defnition.
( 1 1 ) A word u is semantically more complex than another word u_ if
the semantic representation of u contains more semantic material
than the semantic representation of u_ in terms of elementary se
mantic predicates, the types of these predicates and the operations
performed on them.
It is worthwhile noticing that the defnitions of morphological, syntac
tic and semantic complexity involve increasing levels of abstraction. The
defnition of morphological complexity is based on a straightforward count
which depends merely on the defnition of the notion of morpheme. The
defnition of syntactic complexity, on the other hand, presupposes a theory
of syntactic representations. In general, however, the syntactic properties of
predicates can be translated into these representations in a straightforward
way. Finally, the defnition of semantic complexity presupposes a theory of
semantic representations. Such a theory involves questions such as (i) what
are the primitives of the theory, (ii) how are meanings constructed out of
these primitives, (iii) what kinds of aspects of meaning have to be refected
in semantic representations. In addition, defnition ( 1 1 ) is committed to
the view that semantic representations do consist of elementary predicates.
If one is open-minded enough one has to admit, however, that the vast
majority of lexical items are not amenable to a treatment presupposed by
defnition ( 1 1 ) . For a large portion of the vocabulary one has to be content
with a description a suggested by prototype semantics. But if one wants
to make use of prototype semantics in the description of lexical items, the
question immediately arises whether it is possible to correlate degrees of
typicality or degrees of centrality and semantic complexity. I think this
question should be answered in the negative. Why should there be any
diference as to semantic complexity between, say, the word sparrow and
the word duck or between the word duck and the word penguin? Conse
quently, it seems to make sense to talk about semantic complexity only
in cases in which it is meaningful to postulate (at least partial) semantic
representations based on a set of elementary predicates.
We will see further below that-in whatever way we defne semantic
complexity-a notion of semantic markedness is called for in order to be
able to describe certain phenomena which involve baicness or naturalness.
But before embarking on the discussion of semantic markedness we will
have a closer look at semantic complexity and the relationship between
morphological complexity and semantic complexity, on the one hand, and
between syntactic complexity and semantic complexity, on the other. Once
again, in discussing these questions we will restrict ourselves to semantic
felds defned by the rules of derivational morphology.
192 / FERENC KIEFER
3 Semantic Complexity and its Relationship to Syntactic
Complexity and Morphological Complexity
To begin with, let's take a common activity verb, say, zr 'write' and its
derivational expansions. For simplicity's sake we will consider the aspec
tual prefxes a participating in the derivational processes, i. e. , as being
derivational prefxes.7 Furthermore, we will take it for granted that suf
fxation precedes prefxation. It should be made clear, however, that the
order of the two derivational processes is not quite obvious. In fact, in
principle, there are three possibilities: (i) sufxation precedes prefxation,
(ii) prefxation precedes sufxation and (iii) there is no general ordering
relation between sufxation and prefxation. However, if we were to adopt
(iii ) , we would have to give up any systematic account of derivational mor
phology. This alternative can thus be discarded. The decision between (i)
and (ii) depends on various syntactic and semantic considerations but since
this decision has no consequences for semantic complexity we will abstain
fom discussing these probleIl in more detail here.
Figure 1 shows a partial derivational expansion of the verb zr 'write' .
As can be gathered fom Figure 1, the verb zr 'write' can be expanded
by means of the perfectivizing sufx meg- to the perfective megzr 'write
down, fnish writing' or by means of one of the prefxes jel- ' up' , ki- 'out' ,
le- 'down', etc. , to the verbs jeUr 'write down, charge, . . . ' kiir ' copy out,
write in full, . . .' and leir 'write in, register, . . .' which are perfective, on
the one hand, and have often unpredictable meanings, on the other. An
other expansion of zr produces the fequentative zrogat (by means of the
derivational sufx -gat) 'keep writing, write fequently' which in turn can
be expanded by means of one of the prefxes jel-, ki-, le-, which yields
jelzrogat, kizrogat, leirogat, which have the same meanings a jelzr, kiir,
leir (see above) except that they are fequentative rather than perfective.
Then, ir can also be expanded by mean of the impersonal passive (re
fexive) sufx - 6dik, which yields ir6dik 'be/get written', which in turn
can be perfectivized resulting in megir6dik or in jeUr6dik, kiir6dik, which
are the (impersonal) passive counterparts of megzr and jeUr, kiir, leir,
respectively. The passive sufx is -atik, which can be used to produce
the pasive form iratik 'be/get written' , which again, can be perfectivized
yielding megzratik or jeliratik, kiiratik, leiratik. The causative sufx - at
7
Traditionally, prefixation is not considered to belong to derivational morpholog in
Hungarian. First of all, prefxes, in contrat to sufxes, don't harmonize with the stem
and are thus in this respect more like anterior constituents of compounds. On the
other hand, they can be separated from the stem (ust as, for example, the separable
prefxes in German) which makes them dissimilar to anterior constituents of compounds.
Semantically, they often resemble averbials, esp. prefxes such as ki- 'out', be- 'in',
k6ri/- 'around', vissza- 'bak'. The rather frequent prefx mQ- is an aspectual operator
in most cases. In other cae, prefxes behave like verbal sufxes: they are used to derive
verbs from verbs. Prefxation seems thus to occupy an intermediate position between
compounding and derivation.
CAUS
.

ASS
{
t::}
{
t}
.
Figure 1
produces the causative irat ' make write', which can further be expanded
by adding to it the perfectivizing prefx meg- resulting in megirat or one of
the prefxes Jel-, ki-, le- yielding Jelirat , kiirat, ldrat .
It should also be noted that each form i n Figure 1 can further be ex
panded by the possibility sufx - hat meaning 'may, can', e. g. , irhat ' may
be writing', megirathat 'may make write down', kiir6dhat ' may get writ
ten/copied out' , etc.
Notice that all the derivatives except for those with the prefxes Jel-,
ki-, le-, have got a transparent semantics, their meanings are fully pre
dictable on the bais of the meanings of their constituent morphemes. In
other words, the meanings of these derivatives are compositionally deriv
able. In contrast, the derivatives with the prefxes Jel-, ki-, le- are, in
general, semantically not transparent, their meanings cannot be predicted
on the basis of the meaning of their constituent morphemes. That is, these
meanings are not compositionally derivable.
It should be noted that the verbs with the prefxes Jel- 'up', ki- 'out',
le- 'down' , be- 'in' , el- 'away' , etc. , are semantically opaque just in case
they don't denote mere direction. As long as they are used in their original
meanings, they are completely transparent, e. g. , ki +megy 'go out' , be+nez
'look in' , Jel+huz 'draw up, hoist' .
194 / FERENC KIEFER
Since we have been assuming that the more semantic material a form
contains, the more complex it is semantically, in order to establish the se
mantic complexity relationship between two derivationally related forms it
sufces to determine the relative amount of the semantic material contained
in the two forms. As we saw above, perfective forms contain more semantic
material than the corresponding imperfectives and some perfective forms
(i. e. , those with a prefx whose function is more than just perfectivization)
are semantically richer than simple perfective forms, where the perfective
prefx carries perfective meaning only. In what follows we will reserve the
label 'perfective' to denote such simple perfective forms, the semantically
more complex perfectives will be termed 'qualifed perfective' . The above
observations yield the following complexity relationships.s
( 12) a. imperfective <sem
perfective
b. perfective <sem
qualifed perfective
It is also evident that causatives are semantically more complex than non
cauatives. Furthermore, it would seem that passives are, in general, se
mantically more complex than the corresponding actives. That is, we get
( 13a)-( 13b). 9
( 13) a. non-causative <sem
causative
b. active <sem
passive
Frequentatives contain more semantic material than the corresponding du
ratives, as pointed out above, hence
(14)
durative <sem
fequentative
The impersonal passive and the passive do not seem to difer fom each
other with respect to semantic complexity, that is, the relationship (13b)
is valid for both passives.
In sum, the operations denoted by PERF, CAUS, PASS, IMP.PASS,
FREQU introduce additional semantic material which motivates the rela
tionships ( 12a), (13a)-( 13b) and ( 14) . In Hungarian these operations are
refected in morphology as shown by Figure 1.
If we also take into account the possibility sufx -hat/-het, we may posit
( 15) non-modal <sem
modal
where ' modal' refers to a form expressing possibility and 'non-modal' to
the corresponding "factual."
8
We have already introduced the symbol ' <' for the relation of morphological com
plexity (cf. example (3) ) . For syntatic complexity we may use the symbol ' <_` , for
semantic complexity the symbol ' <_,` and for pragmatic complexity the symbol ' <g` ,
respectively.
9
Complexity relationships such a ( 12a) and ( 13a)-(13b) have been treated in earlier
literature under the heading of markedness. See, for example, Lapointe 1986, or for
structuralist treatments of the problem, Greenberg 1966 and Hjelmslev 1953. It ha,
however, not always been made clear whether these relationships refer to the "content
plane" or to the "expression plane."
Informally, the active, the passive and the causative may be character
ized by ( 16a)-(16c), respectively.
( 16) a. T does g
b. g is done by T
c. T causes to do g
It is therefore not at all surprising that on any analysis suggested so far the
causative is considered to be semantically more complex than the passive
and the passive semantically more complex than the active. lO That is,
we get
( 17) active <sem
passive <sem
causative
For the lexical feld of ir 'write' in Figure 1, the complexity relations for
mulated thus far yield, among other things, the following complexity or
derings.
(18) a.
b.
c.
d.
e.
ir
'write'
<sem irogat
'write little
by little'
<s
em
felirogat
<se
r felirogathat
'write up little ' may write up
by little' little by little'
ir <sem
megir <sem
megirhat
'write' ' write down'
ir <sem
irat
'write' ' make write'
' may write down'
<se
r megirat <s
em
'make write
down'
ir <se
r ir6dik <sem {rat
'write' ' be/get written' ' make write'
megirathat
' may make
write down'
felir
<sem
felir6dik <sem
felirat
'write down,
prescribe'
'befget written 'make write down,
down, be/get make prescribe'
prescribed'
In the cases discussed thus far there is an interesting correlation be
tween semantic complexity and morphological complexity. There are no
two forms such that the semantically more complex form be morphologi
cally less complex. In the case of ( 18d) , on the other hand, the semanti
cally more complex form does not always show an increase in morphological
complexity. In particular, the forms ir6dik 'be/get written' and irat ' make
write' are both bimorphemic. In sum, it would seem that the following rule
holds. 1 1
( 19) By increasing semantic complexity, morphological complexity does
not decrease.
l
OSee Lapointe 1986 (p. 224).
l l
This holds within a semantic feld defned by the rules of derivational morphology,
of course. In view of the language-specifc defnition of morphological complexity, this
correlation, too, ha to be considered to be language-specifc.
196 / FERENC KIEFER
As to the syntactic properties, the situation is as follows: ir 'write' and
irogat 'write little by little, write fequently' do not require an overt object
in contrat to irat ' make write' , megirat ' make write down' , felirt ' make
write down, prescribe', felirogat 'write down repeatedly' , megir 'write
down' , felir 'write down' which cannot occur without an overt object (ex
cept elliptically) . The impersonal passive (refexive) forms are used with
out an agent: ir6dik 'be /get written', megir6dik 'be/get written down',
felir6dik 'be/get prescribed'. The pasive form irtik, megiratik, felirtik
with the same meaning may be used with an agent, the agent is expressed
by a by-phrae. That is, we have to do here with four degrees of syntactic
complexity: the forms with an optional object are syntactically less com
plex than the forms that require an overt object which in turn are less
complex than the passive forms. Furthermore, the impersonal pasive is
syntactically less complex than the agentive passive. For the forms of ir,
then, it would seem that a similar correlation holds between semantic and
syntactic complexity as between semantic and morphological complexity.
The following rule seems to be valid:
(20) By increasing semantic complexity syntactic complexity does not
decrease.
1
2
The following example shows that the derivational expansions of a word
may all be opaque. Figure 2 depicts the derivational expansions of the verb
beszel 'talk' .
DUR PERF
'talk' 'chat, converse' 'chat, converse'
~
beszel beszelget ---. elbeszelget
DUR 1
elbeszel
,
,
t '
megbeszel
narra e
' d'
,
I5CH55
rabeszel
'persuade'
lebeszel
'dissuade'
Figure 2
What may strike us at frst glance is that, in general, the various
derivational expansions of beszel 'talk' are not semantically transparent:
beszelget 'chat, converse' is not simply the durative counterpart of beszel,
similarly megbeszel 'discuss' is not the perfective variant of beszel. Nev-
12
Since the defnition of syntactic complexity given above, however restricted it may be,
is not language-specifc, (20) could, in principle, be a good candidate for being universal.
ertheless, the forms beszelget and elbeszelget, 'chat, converse' are dura
tive and the prefxed variants of beszel are in a way perfective. Conse
quently, megbeszel ' discuss' , in contrast to beszel, must contain a resul
tative component in its semantic representation. Furthermore, beszelget
'chat, converse' presupposes at least two participants, which beszel does
not. Consequently, it seems to be quite plausible to assume that beszelget
'chat, converse' is semantically more complex than beszel and elbeszelget
(at least if used transitively with the meaning 'pass time in conversation',
see below) semantically more complex than beszelget. Similarly, megbeszel
should be considered to be semantically more complex than beszel since
it demands that the topic be explicitly specifed. The verb elbeszel ' nar
rate' is even more complex than megbeszel ' discuss' since (i) the topic must
be a story and (ii) the report must be rather detailed. Furthermore, it
is easy to see that the verbs nibeszel 'persuade' and lebeszel 'dissuade' ,
too, contain more semantic material than megbeszel ' discuss' . That is, the
complexity relations are not afected by the fact that all derivatives are
semantically opaque. This leads us to the formulation of the following
principle.
(21) The semantic complexity relations among derivationally related lex
ical items which are semantically opaque are the same as if these
lexical items were semantically transparent.
Principle (21) should be considered a a hypothesis about semantic com
plexity which must be tested on further material. 13
As to the relationship between semantic and morphological complexity,
the derivational expansions, a can easily be seen, obey rule ( 19).
Syntactically, beszelget 'chat, converse' requires either a plural argu
ment or two arguments but it is still intransitive. The verb elbeszelget has
two uses. On one of its uses it is synonymous with beszelget (i. e. , it is nei
ther transitive nor perfective). However, it may also be used transitively
with the object 'time' : 'pass time in conversation' . Megbeszel 'discuss'
and elbeszel ' narrate' always require an overt object, rabeszel 'persuade'
and lebeszel 'dissuade' have three arguments. All this seems to indicate
tat, in the cae of the derivational expansions of the verb beszel, too,
the relationship between semantic and syntactic complexity is governed by
rule (20).
The next example which we will discuss i s the basic verb of movement
mozog ' move' with the root moz- , whose derivational expansions are shown
in Figure 3.
The basic verb mozog is unspecifed as to direction and it is intransitive
and durative. From the root moz- one can derive the punctual mozdul
1
3
Principle (21) claims that the perfective operator makes the form more complex inde
pendently of whether the meaning arrived at is transparent or opaque. The same holds
true for the durative operator.
198 FERENC KIEFER
ERF
..
PUNCTUAL "
. .
..
. ..
TRANS-/ \ TRANS-
PERF-DIR {!:: }
FREQU PUNCTUAL
..
. .
. .sh
'
PEERF-DIR PERFR
.. .
..
{
. }
.
.
sh
I
f
.
..
Figure 3
' make a move' which can receive the perfective prefx meg- yielding meg
mozdul. Mozdul can also be prefxed by a directional prefx, e. g. , ki- 'out'
or el- 'away' resulting in kimozdul ' move out' or elmozdul ' move away' .
The intransitive mozog ,more precisely, the root moz- ) can be transitivized
in two diferent ways. The fequentative sufx -gat yields mozgat ' move
repeatedly' which is transitive and frequentative. This verb can be perfec
tivized by means of the perfective prefx meg- ( megmozgat ) or by means
of the perfective-directional prefx el- ( elmozgat ' move away' , . The other
transitive verb derived fom moz- is the punctual mozdi to which either
the perfective prefx meg- or one of the perfective-directional prefxes el
'away' , ki- 'out' can be added, which results in megmozdi ' move, shift' or
in elmozdi ' move away' , kimozdi ' move out' .
Notice that all the derivatives i n Figure 3 are semantically transparent.
Moreover, almost all of them can be rendered in English by the verb ' move'
which is a clear indication of how diferent Hungarian and English can be
with respect to lexical structure.
The complexity relations among the forms in Figure 3 are partly de
termined by directionality. In particular, the following complexity relation
holds.
(22) unspecifed <ser direction
where ' unspecifed' should be read "unspecifed with respect to direction."
(22) does not only account for the complexity relation between, say, meg
mozdi 'shift' and kimozdit 'get out of place' ,where both verbs are perfec
tive, but also covers the complexity relationship between such verb pairs
as nez 'look' and benez 'look in' ,where the frst verb is imperfective and
the second one has a perfective and an imperfective reading, .
Punctual verbs are more complex than durative verbs ,other things
being equal) , since only the latter, but not the former, are compati
ble with both durative and nondurative adverbials. Frthermore, punc
tual verbs can be used to express perfectivity even if they occur with
out a perfective prefx. The two sentences in (23a)-(23b) have identical
meanings.
(23) a. A ko mozdult egyet .
the stone move-Past one-Acc
'The stone made a move'
b. A ko megmozdult.
'The stone made a move'
Consequently, we may posit
(24) durative <
se
r punctual
However, mozdul and megmozdul are not synonymous. Consider
(25) a. Mar mozdul a ko.
already move the stone
'The stone starts moving'
b.
*
Mar megmozdul a ko.
As testifed by (25a) mozdul 'start moving' has an inchoative meaning
which is absent in megmozdul. In general, verbs with a perfective prefx
cannot refer to on-going processes or activities, they don't have any present
tense meaning. The verb megmozdul is semantically more complex than
the verb mozdul.
4 Pragmatic Complexity
As we have seen above, each derivational morpheme contributes to the
meaning of the "input. " There are no derivational afxes (in Hungarian)
without any content. Their contribution need not be semantic, however.
Let's explore this question in some more detail.
The diminutive sufx -i is used to form nouns fom nouns in two
ways.
(i) -i is added to the truncated form of the initial noun, e. g. , csok+i
(from csokotadi 'chocolate' ) , dir+i (fom direktor 'director' ), ov+i
(from ovoda 'kindergarten' ) , fagy+i (fom fagylalt ' ice cream' ).
(ii) The sufx - i i s added to the stem of the noun, e. g. , tab+i (fom tab
' foot' ) , comb+i (from comb ' thigh' ) , has+i (fom has 'belly' ), husi
(from hus 'meat' ) .
Words i n group ( i ) belong to school slang. They presuppose a cer
tain degree of intimacy and have ofen a jocular touch. The form in (ii),
on the other hand, belong to nursery talk: they are typically used in a
discourse in which at leat one small child participates. In neither cae,
however, does the sufx -i denote smallness. In fact, pairs such a fagy+i
200 / FERENC KIEFER
and fagylalt ' ice cream' or has+i and has 'belly' are truth-conditionally
identical. What is then the contribution of the sufx -i to the meaning
of the stem noun? It can determine the social setting of the speech sit
uation (in the case of (i) ) . The second case can also be characterized by
the label ' endearments'. Though these aspects of meaning are clearly con
ventional (they are not implicated "conversationally" ) , they are not truth
functional. Since, in principle, any conventional meaning can be accounted
for in semantics, the sufx i does not necessarily call for a pragmatic
explanation. 14
There are, however, cases which show that a derivational afx may
introduce non-conventional meaning as well. The example we are going
to consider comes from adjective morphology. 15 In addition to the posi
tive, the comparative and the superlative, Hungarian has got an excessive.
Consider
(26) a. nagy
'big'
b. okos
nagy+obb leg+nagy+obb legesleg+nagy+obb
'bigger' 'biggest' ' very biggest'
okos+abb leg+okos+abb legesleg+okos+abb
' clever' 'cleverer' 'cleverest' ' very cleverest'
The question immediately arises what the excessive means. The superla
tive already defnes the highest degree of the property at hand, Le. ,
T is
the biggest' means that fom among the things considered T possesses the
highest degree of the property at hand, that is, of being big. The excessive
cannot change this state of afairs; what it does is to introduce a presuppo
sition (and special discourse functions) . Recall that what is biggest need
not be big and who is cleverest need not be clever. In contrast, the thing
that is the very biggest must be big and the person who is the very clever
est must be clever. That is, the excessive turns a relative property into an
absolute one. The introduction of a (semantic) presupposition, as we saw
above, leads to a higher degree of semantic complexity.
As to its discourse function, notice that the excessive can be considered
as the emphatic variant of the superlative in a way. Emphasis is not neutral
contextually. For example, in general, one cannot start a conversation by
using the excessive. The excessive requires a context in which there is
a discussion about the highest degree of a property (or at least such a
discussion must be implicit in the context) . The person who uses the
excessive wants to emphasize his belief about the question: 'believe me,
this is the very best solution' or the like.
14This problem and similar phenomena were treated under the heading of morphoprag
matics in Dressler and Kiefer 1990.
1
5
Comparison of adjectives is traditionally treated in infectional morphology. In quite
a few respects, however, it is more like derivation. First of , the superlative prefx lQ-
can in no way be considered to be an infectional afx since Hungarian infectional afxes
are sufxes. Second, the semantic properties of the superlative are not comparable to
those of infectional paradigms.
The excessive does not always introduce a presupposition, however. I6
Take, for example,
(27) a. csunya csunYl+bb leg+csunyt+bb legesleg+csunyt+bb
' ugly' ' uglier' ' ugliest' 'very ugliest'
b. buta butt+bb leg+butt+bb legesleg+butt+bb
'stupid' ' more stupid' ' most stupid' lit. 'very stupidest'
Adjectives which express negative evaluation are, in general, used in an
absolute sense, i. e. , ' uglier' induces the presupposition that the things com
pared are both ugly and 'more stupid' induces the presupposition that the
things compared are both stupid. Similarly, if
T is ugliest' then T must be

ugly, and if
T is most stupid' then T must be stupid. In such cases the ex

cessive cannot induce a presupposition since this presupposition is already
there. The contribution of the excessive to the meaning of the superla
tive is thus restricted to the special discourse functions. These discourse
functions do not make the excessive semantically more complex than the
superlative but they certainly make it pragmatically more complex in a
way. This means that a defnition of pragmatic complexity is called forP
If we use 'discourse function' a a cover term for any pragmatic restrictions
or functions, then (28) holds.
(28) Of two fonns T and g, which are semantically of equal complexity, g
is pragmatically more complex than T if it difers fom the latter by
having specifc discourse functions.
Since there are good reasons to assume that the comparative is seman
tically more complex than the positive and the superlative semantically
more complex than the comparative, we get for the excessive one of the
two complexity relation (29a) or (29b) .
18
(29) a. positive <Ber comparative <Ber superlative <Bem
excessive
b. positive <Ber comparative <ser superlative
<p
excessive
where <Ber stands for "semantic complexity relation" and
<p
for "prag
matic complexity relation."
We are now in the position to reformulate principle ( 19) in the following
fahion.
(30) Within a semantic feld defned by the rules of derivational morphol
ogy, by increasing semantic or pragmatic complexity, morphological
complexity does not decreae.
16
1 discussed the presuppositional properties of adjectives at considerable length in
Kiefer 1977.
17It goes without saying that pragmatic complexity is also needed in order to account for
the diference between other types of synonyms. I!. invite versus bid/rome or roast/grill
U equivalents of interrogate. Cf. Verschueren 1985.
1
8
For a detailed account of the semantics of dimensional adjectives and the semantic
representations of the positive, the comparative, and the superlative, see Bierwisch 1987
and Lang 1987.
202 / FERENC KIEFER
5 Semantic Markedness
Semantic complexity may explain a lot but it does not explain everything.
In particular, there are certain phenomena which seem to run counter to
the predictions of the complexity theory as developed thus far. In what
follows we will discuss some such phenomena.
(i) Our frst example is taken from the semantic feld of verbs of move
ment. Recall that both megmozgat 'move' and elmozgat 'move away' were
regular derivatives of the verb mozgat 'move' (see Figure 3). Though these
verbs are not only morphologically regular but also semantically tran
parent, they have a rather restricted use. Megmozgat is not so much
used in its literal sense, but it occurs quite fequently in idiomatic expres
sions, for example, megmozgatja a kepzeletet ' quicken one's fantasy' , min
den kovet megmozgat ' leave no stone unturned' , megmozgatja a tomegeket
'rouse the masses'. Elmozgat 'move out/away' is a technical term ( 'decen
ter') and it is not part of our everyday vocabulary. Most often megmoz
gat is replaced by the transitive-punctual (and equally perfective) meg
mozdi, and in place of elmozgat the transitive-punctual-directional el
mozdi is used. Notice furthermore that megmozgat need not imply du
rativity and iterativity, that is, megmozgat may be synonymous with meg
mozdi. Consequently they exhibit the same degree of semantic complex
ity. Similarly, elmozgat and elmozdi may also be semantically identi
cal. Yet megmozdi is more useful than megmozgat because it can be
employed in a larger number of situations. Its use is not collocation
ally restricted. In this sense megmozdit is more normal or more natural
than megmozgat. Similar considerations hold for the pair elmozgat and
elmozdi. Since elmozgat is only used as a technical term (in the sense
of 'decenter' ) , it is not necessary to show that it has a more restricted
use than elmozdit. Again, elmozdit is more normal or more natural than
elmozgat.
(ii) The second example concerns the passive in Hungarian. Notice
that the passive suf xes - atik/-etik, -tatik/-tetik are morphologically fully
productive: ad+atik 'be given' , mond+atik 'be said' , nez+etik 'be looked
at' , ker+etik 'be asked' , tanit+tatik 'be taught' , enged+tetik 'be allowed' ,
etc. However, the passive is hardly used i n present-day Hungarian: i t occurs
almost exclusively in certain idiomatic expressions. In all other cases it is
replaced by the corresponding active form. Compare
(31) a. Nem adatott meg nekem.
'It was not granted to me'
b. Engedtessek meg, hogy . . .
lit. ' let .it be allowed that . . .
(31a) is rare and (31b) is stylistically marked. Normally, the passive (32a)
would always be replaced by the ( impersonal) active (32b) .
(32) a. Megtiltatott.
'It wa forbidden'
b. Megtiltottak.
'They have forbidden it'
( meg- : perfective prefx, tilt: verb stem ' forbid' , - at: passive sufx,
- ott : Past Tense, - ak: 3rd Person Plural)
Consequently, in spite of the morphological productivity of the passive,
passive constructions are infrequent, unusual or even odd.
The impersonal passive - 6dik is, no doubt, more often used than the
agentive passive. Consider
(33) a. Az ajt6 j6l csuk6dik.
'The door closes properly'
b. A titok kitud6dott.
'The secret is out'
( csuk: verb stem 'close' , - 6d: impersonal passive, - ik: 3rd Person
Singular, tud: verb stem ' know' , ki- : prefx 'out' , - 6d: impersonal
passive, - ott: Past Tense) .
In general, the impersonal passive is more useful than the agentive passive
in Hungarian. I argued above that the agentive passive is semantically
more complex than the impersonal passive. One might thus conjecture
that the degrees of complexity correlate with the levels of utility: the more
complex a form is, the less useful it is. Although this may very often be
the case, this correlation is not always valid. Recall that the causative was
considered to be semantically more complex than the agentive passive. The
causative, too, is morphologically fully productive in Hungarian. Consider
(34) a. Peter tanittatja a gyerekeit.
lit. Peter have taught the children-his
'Peter provides for the education of his children'
b. Anna ruhdt csinaltat.
'Anna has a (new) dress made'
Sentences such as (34a)-(34b) are also widely used in Hungarian. In view
of these facts the causative has to be considered more normal or more nat
ural than the passive which seems to run counter the complexity hierarchy
repeated here under (35) .
(35) active <ser passive <s
er causative
(iii) In some cases the perfective prefx meg- can be used to express
single events. Notice that some verbs seem to express habitual or iterative
meaning in virtue of their lexical potential. For example, cs6kol 'kiss',
ldtogat 'visit' , dicser ' praise' , k6stol 'taste' , etc. As soon as these verbs are
turned perfective, however, the habitual or iterative meaning disappears
and what is left is the meaning referred to as 'single event' . Compare
204 / FERENC KIEFER
36) a. Peter cs6kolja Katit.
' Peter is kissing Kati' iterative)
b. Peter megcs6kolja Katil .
'Peter is going to kiss Kati' single event)
37) a. Peter l6togatja Katit .
' Peter visits Kati' iterative)
b. Peter megl6togatja Katil.
' Peter is going to visit Kati' single event)
Even if the forms megcs6kol and meglatogat are not perfective-resultative,
i. e. , they are not in the same way perfective as, for example, the verbs megir
'write down' fom ir 'write' )or megepa 'build' fom epa ' build' )implying
a result a letter or a house) , they share two fundamental properties with
them. First, their internal temporal structue ha to be considered a
an indivisible whole, hence perfectivity excludes iterativity. The ' single
event' interpretation is a consequence of this property. Second, the verbs
megcs6kol 'kiss once) ' and megl6togat 'visit once) ' , in general, cannot
refer to the present, their potential time reference is non-present. There
are thus good reaons to asume that the prefxed forms are semantically
more complex than the corresponding forms without a prefx.
For some reason, however, sentences such a 36b) and 37b) express
the preferred way of looking at the world. Of course, this ha nothing to
do with iterativity, sentences such a 38a)-38b) sound quite normal.
38) a. Peter tobbszor megcs6kolta Katit .
' Peter kissed Kati several times'
b. Peter tobbszor meglatogatta Katit.
' Peter visited Kati several times'
Apparently, our preference goes for 'perfectivity' , that is, for completed
action, or terminated event in cases such as cs6kol 'kiss' , latogat 'visit' and
k6stol 'taste'. No wonder, then, that the perfective forms of these verbs
are more fequent than the corresponding unprefxed verbs, as they sound
more normal or more natural. And they do this in spite of the fact that
they are semantically more complex.
iv) A similar argument can be made for some other types of prefxed
verbs. Thus, for example, verbs of movement are more ofen ued with
directional prefxes than without them. In general, we go in somewhere
or come out fom somewhere, we come down fom somewhere or we go up
somewhere. Direction is implicit in the case of verbs such as come and go,
it has to be inferred on the basis of the respective positions of the speaker
and the hearer and some other aspects of the speech situation. In certain
contexts the utterances in 39a)-39b) or the ones in 40a)-40b) may be
considered equivalent.
(39) a. Pisti megy mar.
' Steve is already going'
b. Pisti mar elmegy.
' Steve is already leaving'
(40) a. Pisti jojjon mar!
' Pisti should come'
b. Pisti jojjon mar ide!
' Pisti should come here'
In (39b) and (40b) the direction of the movement is made explicit: el
' away' and ide 'here' . It is arguable that elmegy 'go away' is semantically
more complex than megy 'go', but at the same time the former is more
explicit about the direction of the movement, hence less marked, than the
latter. The same holds true for the pair idejon 'come here' and jon 'come' .
Quite often a directional prefx of the verb also appears as a case sufx
on the noun which denotes the goal of the movement. For example, bemegy
0 terembe 'go into the hall', ralep 0 ldbdra 'step on someone's foot' . The
verbal prefx can be omitted under certain circumstances. One such case
is if the directional noun is moved into focus position, i. e., in front of the
verb: a terembe megy, a ldbdra lep. This is, no doubt, a syntactically more
complex construction than the previous one. The situation is similar in
the other cases in which the verbal prefx is omissible. This state-of-afairs
leads to a higher fequency of the prefxed verbs in comparison with the
unprefxed ones. That is, bemegy 'go in' is less marked than megy 'go' and
ralep 'step on' can be claimed to be less marked than lep 'step' . Once again,
higher semantic complexity means lesser degree of semantic markedness.
In the previous sections we have been discussing four cases which showed
that two forms of equal semantic complexity need not be equally natural
or normal or that semantically more complex forms may be more natural
or normal than their semantically less complex equivalents. In these cases
naturalness or normality has to do with restrictions on the use of the given
form. That is, the more restricted its use is, the less natural or normal
it will be. As a quasi-technical term ' natural' or 'normal' can be equated
with ' unmarked', and ' unnatural' or 'unusual' with 'marked' . Naturalness
is of course a matter of degree hence the markedness principle which covers
the cases discussed above can be formulated as follows.
(41) Of two derivationally related forms T and y, y is semantically more
marked than T if there are more restrictions imposed on the use of
y than on that of T.
If there are no restrictions imposed on T and y, principle ( 41) does not
apply.
Except in cases where markedness relations are derivable fom (41 ) , or
from some markedness principle, obtain, markedness can be taken to be
identical with complexity. This principle is formulated in (42) .
206 / FERENC KIEFER
(42) In a semantic feld defned by derivational morphology, the marked
ness relations are identical to the complexity relations, except in
cases in which special markedness relations obtain.
Principle (42) holds true for semantic markedness and semantic complex
ity only. It is doubtful whether similar principles can be formulated for
morphological and syntactic markedness.
It would seem that it is markedness and not complexity (and this holds
for semantic, syntactic and morphological markedness alike) which should
be relatable to processing time and acquisition order. 19 Though no syn
tactic study has been undertaken so far with respect to the acquisition of
the derivational morphology of Hungarian, the preliminary data which we
have at our disposal seem to support this claim.
20
6 Summay
To summarize, by examining some examples fom Hungarian derivational
morphology defnitions for morphological complexity, syntactic complexity
and semantic complexity were proposed. The defnition of morphological
complexity was claimed to be language-specifc (or language-type specifc)
whereas the other two defnitions should be considered to be universally
valid. It wa pointed out, however, that the proposed defnition of syntactic
complexity is too narrow as a general defnition of syntactic complexity
since it is based solely on predicate-argument structure.
It was further claimed that markedness relations are diferent fom
complexity relations on all levels considered. The principles which govern
markedness relations are thought to be universal but (some of) the partic
ular markedness relations may be language-specifc. In general, marked
ness has to do with naturalness phenomena. Some examples were given
for morphological markedness and syntactic markedness was discussed in
more detail. In general, there seems to be no direct relationship between
morphological markedness and morphological complexity, on the one hand,
and between syntactic markedness and syntactic complexity, on the other.
In contrast, semantic markedness seems to be identical with semantic com
plexity except where special markedness relations obtain.
Also, a few words were said about pragmatic complexity which can be
used to distinguish two forms of equal semantic complexity in terms of
discourse function.
19 As for the hypothesis that semantic complexity determines acquisition order, see Clark
and Clark 1977 (pp. 498-500) and the references cited there. The hypothesis about
"baic level terms" is due to E. Rosch and her colleagues. (See, for example, Brown
1973 and Clark and Clark 1977 (pp. 500, 527-530) and the references cited there.)
20See, for example, Fazeka 1989. Some of the data on which the present study is baed
are taken from a as yet unpublished material collected by Zita Reger from the Reearch
Institute for Linguistics of the Hungarian Academy of Sciences and by Csaba Pleh from
the Institute of Psychology at Budapest University and their collaborators.
Acquisition order is claimed to be determined by markedness rather
than by complexity. If no special markedness relations obtain, however,
semantic markedness is thought to be identical with semantic complexity.
A strong correlation was found between semantic and syntactic com
plexity: if semantic complexity increases, then in general, syntactic com
plexity cannot decrease. There is also a relationship between semantic
complexity and morphological complexity. More precisely, it was shown
that morphological complexity cannot decrease if semantic or pragmatic
complexity is increased. Cross-linguistic studies should reveal to what ex
tent these relationships are universal.
References
..1987. .Studia Grammatica 26-
27:91-286.
.1973. A First Language: The Early Stages. ..
.
1988. ....
...
..
.. 1977. Psychology and Language. .
.
. 1990. ..
.Contemporary Morphology, ..
69-77. .
.
. ....
.1988. Leitmotifs i n Natural Morhology. ..
.
...1986.
Markedness. ..
A. 1989. .
.
.1966. Language Universals with Special Reference to Fea
ture Hierarchies. ..
..1953. Prolegomena to a Theory of Language. ..
1977. Theoretical Linguistics

5: 135-173.
.1987. ..Studia Grammatica
26-27:287-458.
1986. ..
Markedness,
.. .219-239.
..
208 / FERENC KIEFER
. 1976. Language and Perception.
.. .
...1976.
Cognitive Psychology 8:382-439.
. .1985.
...
....Acta
Linguistica Academiae Scientiarum Hungaricae 35: 153-149.
1985. What People Say They Do with Words. .
.
Focus-Based Inferences in Sentence

Comprehension
GYORGY GERGELY
1 Introduction
The aim of this paper is to investigate the nature of discourse inferential
processes in comprehending complex sentences. It will be argued that the
interpretational structure that the listener constructs for a sentence in dis
course is a joint function of its abstract semantic content and its discourse
context (see e. g. , Bierwisch 1981 , Gergely and Bever 1986). The inferen
tial processes involved in computing such a representation are constrained
by diferent sources of discourse information, such (i) topic-focus (TF)
structure, (ii) interclausal semantic relations, and (iii) pragmatic knowledge
structures, like scripts. The present study will demonstrate that such dis
course cues, when available, are employed directly and interactively in the
construction of a discourse interpretational representation for the sentence,
even before its full syntactic and semantic processing is complete.
The experimental studies to be discussed were carried out in Hungar
ian, a non-lndo-European, agglutinative, "free" word order language, where
the discourse functions (topic versus focus) of constituents are systemat
ically related to surface structural positions in the sentence. Therefore,
Hungarian seems ideally suited for the investigation of discourse inferential
The experimental work reported in this paper forms part of my Ph. D. dissertation
(Columbia University, 1986). I am grateful to Tom Bever, Dave Townsend, Csaba
Pleh, Marylene Cloitre, Lou Ann Gerken, and Steve Dopkins for their intellectual and
personal support during the research. I thank Herb Clark, Dan Morrow, Richard Meier,
Anna Szabolcsi, and Laszlo K:Jman for their detailed comments on an earlier draft
of this paper. The experimental work wa carried out at the Department of General
Psycholog of the Eotvos Lorand University in Budapest; I wish to thank Sandor Illyes,
Ilona Bark6czi , and Csaba Pleh for making the research facilities of the University
available to me, Andra Varga for statistical, Karoly Halmai, Laszlo Bernath, Bela
Marian, and Janos Geier for technical help. During the writing of the fnal version of
this paper I wa supported by a research grant from the MTA-Soros Foundation.
Lexical Matter
s
. Ivan A. Sag and Anna Szabo1csi, eds.
210 / GYORGY GERGELY
processes during sentence comprehension, since, apart from intonational
cues, it also provides locally unambiguous structural cues for the on-line
identifcation of TF structure.
The paper will be organized as follows. Section 2 will discuss some of the
limitations of earlier approaches to discourse inferential processing, which
used only of-line, post-sentence measures. In Section 3, I will argue that
the sentence focus is directly accessed during on-line speech comprehension
and plays a central role in the discourse interpretation of sentences. In Sec
tions 4 and 5, I will develop some hypotheses about the role of focus-based
inferences in the comprehension of adversative subordinate constructions.
Section 6, will describe how the structural positions in the Hungarian sen
tence can act a cues for the on-line identifcation of the discourse roles
of surface constituents. Then, Section 7 will summarize the results of an
earlier study ( Gergely 1986, 1991a, 1991b) which examined the role of
focus-baed inferences in establishing discourse antecedent relations during
the processing of subordinate constructions. In Section 8, I will report a
new study which investigates the role of focus-based predictive inferences
in the processing of adversative subordinate constructions. Finally, in Sec
tion 9, I shall discuss the implications of the results for current models of
speech comprehension.
2 The Role of Discourse Information in Sentence
Comprehension
Early research on sentence comprehension (see Fodor, Bever, and Garrett
1974) was preoccupied with investigating how the listener employs gram
matical cues in the surface sequence to assign underlying syntactic rela
tions that are necessary to identify the semantic structure of the sentence.
Typically, it was assumed that " . . . understanding a sentence requires the
recovery of its semantic representation" (Fodor, Fodor, and Garrett 1975) ,
which then serves a the input to the inferential processes that fx the in
terpretation of the sentence within its discourse context. The implication
of this approach was that the processes integrating the sentence into the
discourse structure take place only after its underlying linguistic structure
has been fully assigned, i. e. , not before the end of the sentence.
More recent approaches, however, (e. g., Marslen-Wilson, Tyler, and Sei
denberg 1978, Marslen-Wilson and Tyler 1980, Riesbeck and Schank 1978,
Johnson-Laird 1984, Crain and Steedman 1985) have tended to abandon
this strictly modular view, and assume that as the listener proceeds in the
sentence, he attempts to construct a discourse interpretational structure
directly. These more interactive models pay increaing attention to the
top-down infuence of higher-order contextual variables on speech process
ing. As a result, researchers started to investigate the role of discourse
informational cues in the inferential processes that are involved in com-
Focus-BASED INFERENCES IN SENTENCE COMPREHENSION / 211
puting the contextual interpretation of sentences. However, until recently
the diferent types of discourse information (such as pragmatic knowledge,
TF tructure, or inter-propositional semantic relations) were typically stud
ied in isolation, and their efects on comprehension were demonstrated
only indirectly, using of-line measures. As a result, the on-line nature
of the utilization of discourse cues in comprehension remained largely
unspecifed.
For example, the construct of scripts (e.g. , Schank 1975; Schank and
Abelson 1977; Bower, Black, and Turner 1979) provides a useful famework
for conceptualizing how pragmatic knowledge might bear on the interpreta
tion of sentences in discourse. Scripts are characterized as complex memory
structures that represent our shared cultural knowledge about the typical
structure of events and situations in the world in terms of structured sets
of propositions. In discourse, speakers often make reference to information
that has not been explicitly introduced, but that they assume is available in
the listener's memory. Therefore, listeners often have to make linking infer
ences baed on their pragmatic knowledge to establish coherent referential,
causal, and temporal relations between successive propositions. Scripts
contrain such inferential processes by providing the interpretational do
main over which such relations are established (e. g. , Sanford and Garrod
1981) .
The relevance of scripts i n sentence processing has been demonstrated
in a number of studies (e.g. , Bower et aL 1979, Dufy 1983, Gibbs and
Tenney 1980, Townsend 1983). The typical fnding is that after a script
has been activated by the previous text, a subsequent sentence is compre
hended faster when its content is predictable fom the script. These results
suggest that the processing of a sentence can be facilitated by script-based
expectations about its content. In fact, it has been proposed that sen
tence comprehension in normal discourse relies heavily on generating ex
pectations through directly accessing contextual and pragmatic knowledge
(Riesbeck and Schank 1978) .
The studies demonstrating such script-based efects, however, typically
examined overall reading times for whole sentences. As a result, they
cannot tell us whether the script-based expectations exert their infuence
only afer the syntactic and semantic processing of the sentence ha been
completed, or whether they already afect the more immediate levels of
processing. In fact, it was not until relatively recently (see e. g. , Sharkey
and Mitchell 1985, Gergely 1991a) that the on-line efects of higher-order
contextual contraints on such local aspects of speech processing as word
recognition have received systematic attention.
The situation was similar in studies (see Clark and Haviland 1977)
investigating another major source of discourse information that constrains
the construction of interpretational representations: what I shall call here
the topic-focus (TF) structure of sentences. Sentences provide cues that
allow the listener to identif two distinct parts of the expressed proposition:
(i) what is variably called "topic," "old," or "given" information, that
the sentence is "about, " and that has typically been already established
(or is readily inferable) in the listener's memory; and (ii) what is called
"comment," or "new" information, that the speaker asserts about the topic,
and that he believes to be not yet represented in the listener's memory
(see Halliday 1967; Chafe 1970, 1976; Chomsky 1971 ; Jackendof 1972;
Reinhart 1981) . The most prominent constituent, both semantically and
phonologically, of the latter part of the proposition is the sentence "focus"
which receives the primarys stress of the sentence (e. g. , Chomsky 1971 ;
Sgall, Hajicova, and Benesova 1973, Szabolcsi 1981) .
In a series of experiments, Clark and Haviland (Clark and Haviland
1974, 1977; Haviland and Clark 1974) demonstrated that during compre
hension the listener relies on TF structure to relate the new information
in a propositional sequence to some previously established antecedent rep
resentation in the discourse. In particular, they hypothesized that the
identifcation of the sentence topic leads to a backward search for a match
ing antecedent in the discourse structure. If no direct match is found,
the listener has to infer an antecedent on the basis of his pragmatic
knowledge.
But Clark and Haviland's data consisted of post-sentence meaures of
overall comprehenion times. Clearly, while such studies do imply that
listeners rely on the discourse functions (topic versus focus) of constituents
for establishing discourse antecedent relations, they, nevertheless, fail to
show how and when this process takes place during on-line processing. For
example, is the backward search for a discourse antecedent initiated as soon
as the topic of the sentence is identifed? Or is the information provided by
TF structure employed only after the full syntactic and semantic processing
of the sentence is complete?
3 The On-Line Use of TF Structure in the Discourse
Integrational Processing of Sentences
To answer such fner-grained questions about the processing of discourse
informational cues, one needs to employ on-line measures, as exemplifed
by the studies of Cutler ( 1976) and Cutler and Fodor ( 1979). These authors
demonstrated that listeners actively search for the focus while they process
a sentence, as shown by shorter phoneme-monitoring latencies when the
target word is focused than when it is not. Their results indicate that
the focus is diferentially processed already before the end of the sentence
is reached (see also Crain and Steedman 1985). This fnding, then, is
consistent with the hypothesis that the on-line identifcation of TF structure
directly initiates a backward search for a discourse antecedent.
In this paper I shall argue that TF structure, apart fom triggering a
backward search for antecedents, is also involved in generating predictive in
ferential processes that serve the discourse integration of complex sentences.
In particular, I shall examine how during the on-line processing of subor
dinate constructions TF structure intercts in setting up interpretational
representations with other major sources of discourse information, namely,
interclausal semantic relations encoded by conjunctions (see Townsend and
Bever 1978), and pragmatic knowledge stored in memory structures such
as scripts (e. g. , Schank and Abelson 1977, Bower et al. 1979) and lexical
stereotypes ( Gergely and Bever 1986) .
4 The Interpretational Structure of Adversative
Subordinate Constructions
As an example, let us frst examine the structure of the inferential rela
tions that the listener has to establish when comprehending an adversative
subordinate construction. This analysis will be central for the empirical
studies that follow, as the latter were designed to test diferent aspects of
the processes through which listeners construct the interpretational struc
ture for such constructions. Consider the initial ' though' clause in ( 1) and
the two possible continuations (2a) and (2b):
(1) Though yesterday John cleaned the bathroom,
(2) a. Mrs. Jones is still convinced that it is Mary who does all the
cleaning in the house.
b. the tiles still looked a bit dirty.
The general interpretational schema for adversative constructions (see
Figure 1) can be summarized as follows (see also Dakin 1970, Townsend
and Bever 1978, Bever and Townsend 1979). The initial 'though' in ( 1)
informs the listener that some expectable consequence of the frst clause
will be denied in the main clause. The particular consequence, however, is
not explicitly specifed: it has to be inferred. For example, the listener has
to reconstruct diferent inferential paths leading fom ( 1) to (2a) and (2b) .
For (2a) the implicit consequence of ( 1) is the expectation that because it
was John who cleaned the bathroom, Mrs. Jones would change her belief
about the unequal division of domestic labor between John and Mary. On
the other hand, for (2b) the implicit efect is quite diferent: given that the
bathroom has been cleaned, it is expected that the tiles would not look
dirty.
How does the listener identif the inference that the speaker intends to
deny, when there are always a number of typical consequences that follow
from a given proposition? From the initial 'though' clause, he can generate
a set of candidate consequences based on his knowledge of typical cause
efect relations (see Figure 1) . However, the abstract propositional content
of the clause does not indicate which of those is intended by the speaker.
..
Causal In!"ence / \ ve<sative Relation
Implicit Proposition:
IP
l
, . . . , IPn constitute the set of expectable consequences
baed on the initial 'though' clause proposition.
Figure 1 Schematic Representation of the Interpretational Structure
of Adversative Subordinate Constructions
Does the listener have to wait until the main clause is processed, whose
content allows him to identify the denied consequence retrospectively?
Townsend and Bever ( 1978) reported evidence supporting this possi
bility (see also Bever and Townsend 1979, Townsend 1983) . They used
on-line tests to assess the accessibility of (a) the meaning, and (b) the su
perfcial aspects (such as the serial order of words) of the representation
of incomplete clauses. They found that the semantic content of an initial
'though' clause was less available than that of a corresponding main or
'if' clause, while the superficial aspects of the clausal representation were
more accessible in the initial 'though' claue. Townend and Bever (1978)
hypothesized that the full semantic processing of an initial 'though' clause
is postponed and the clause is kept in a relatively superfcial representa
tional format, precisely because the listener " . . . cannot determine which
cause-efect relation the speaker is denying" before processing the second
clause.
5 The Role of TF Structure in Directing the Inferential
Processing of Adversative Subordinate Clauses
In contrat, it can be argued (see Gergely 1986, 1991a, 1991b) that the TF
structure of an initial 'though' clause does provide information that can be
used by the listener on-line to restrict the set of expectable consequences
of the proposition to the one that is most likely to be intended by the
speaker. This is so because speakers, following a Gricean convention of
language use, tend to mark the inferential basis of the denied consequence
as the informational focus of the 'though' clause. Consider the sentence
pairs (3) and (4) below:
(3) a. Though Daddy praised his daughter for the excellent dinner, she
was still not happy.
b. ?Though Daddy praised his daughter for the excellent dinner, she
was still not happy.
(4) a. ?Though Daddy praised his daughter for the excellent dinner, it
was, in fact, his son who prepared it.
b. Though Daddy praised his daughter for the excellent dinner, it
wa, in fact, his son who prepared it.
The initial subordinate clauses are identical in the four sentences in all
respects save for one: in the placement of contrastive stress (indicated by
italics) , marking the focus of the clause. In (3a) and (4a) the focused ele
ment is the verb 'praised', while in (3b) and (4b) the object noun 'daughter'
receives contrastive focus. In the case of (3a) and (3b) the fnal main clause
contains the anaphoric pronoun 'she' whose antecedent is in both cases the
object NP 'his daughter' from the initial clause. Notice, however, that this
fnal clause is a perfectly natural continuation of the initial clause (3a) , but
sounds rather awkward following the frst clause in (3b).
We can account for this diference in the following way. It can be hy
pothesized that when the listener identifes the focus of the initial 'though'
clause, he will compute a focus- based inference, based on his pragmatic
knowledge of typical cause-efect relations, about the implicit consequence
that the speaker intends to deny. Furthermore, I shall assume (see Section
7) that this focus-baed consequence will be "foregrounded" (i. e. , become
diferentially activated) at the clause-boundary, because it contains the ex
pectable antecedent argument for the next clause.
Thus, in (3a) the speaker focuses the fact that the daughter was praised.
Since the typical efect of being praised is to feel happy, by hypothesis, the
listener will compute the implicit consequence proposition 'the daughter is
happy' . This proposition contains the topic NP of the initial subordinate
clause, which, in (3a) , serves as the antecedent for the anaphoric pronoun
'she' in the second clause. Therefore, on the assumption that the back
ward search for a coreferent during the processing of the second clause will
frst access the foregrounded (and, therefore, most available) consequence
proposition, the focus-baed inference in (3a) can be argued to facilitate
the smooth integrational processing of the second clause.
Not so in (3b), however, where the antecedent NP 'his daughter' is
contrastively focused. By contratively focusing the object NP, the speaker
makes implicit reference to a contrastive set of entities other than the object
that could have been praised, but were not (see Jackendof 1972, Chafe
1976). Therefore, the inferred focus-based proposition will contain the
entities of the contrative set, but not the focused object 'his daughter'
which is, however, the antecedent for the pronoun 'she' of the second clause
in (3b). Thus, in (3b) the speaker misleads the listener when, violating the
Gricean cooperative convention, he denies in the main clause a consequence
that is not based on the focus of the initial subordinate clause. As a result,
the backward search for antecedent for the main clause pronoun 'she' will,
at frst, have to be aborted, as the foregrounded focus-based inference does
not contain the right antecedent NP. For this reason, (3b) sounds awkward
when compared to (3a) .
In (4a) and (4b), however, the situation is reversed. Here the antecedent
for the subject NP 'his son' of the second clause of (4b) is contained in the
contrastive set of the inferred focus-based proposition, and so the back
ward search for antecedent is facilitated. In contrast, the foregrounded
focus-based consequence in (4a) contains the topic NP 'the daughter' of
the subordinate clause. Therefore, the backward search for a matching an
tecedent for ' his son' in the second clause will, at frst, have to be aborted,
since, by hypothesis, it will initially access the foregrounded consequence
proposition. As a result, (4a) sounds awkward in contrast to (4b) .
Therefore, if the above hypothesis is correct, we would expect a good
deal of inferential processing to take place in an initial 'though' clause. This
prediction, however, seems to contradict Townsend and Bever ( 1978) who
found that the full semantic processing of an initial adversative clause is
postponed until the second clause is reached. To resolve this contradiction,
in an earlier study I replicated Townsend and Bever 1978 using Hungarian
sentences (see Gergely 1986, 1991a). I argued that Townsend and Bever's
subjects processed initial 'though' clauses only superfcially because they
could not identify the focus of the clause fagments used in the experiment.
In English, information about TF structure is typically provided by (a) con
textual cues, and (b) intonational cues such as focal stress, which normally
falls on the last word of a simple clause (see e. g. , Jackendof 1972). The
clause fagments used by Townsend and Bever, however, appeared in isola
tion, and were recorded with "normal intonation" with the last word of the
clause cut out. As a result, none of the constituents was clearly marked as
the focus of the clause. It is possible, therefore, that at the point of testing,
subjects have not yet assigned focus role to any of the clausal constituents.
Consequently, they could not carry out the hypothesized focus-based infer
ential processing of the initial ' though' clause.
In Hungarian, however, there are clear structural cues (see Section 6 be
low) that can help the listener identif the clausal focus long before the end
of the clause. As a result, listeners should be able to assign discourse func
tions to constituents even when hearing isolated and incomplete clauses. In
fact, the replication of Townsend and Bever ( 1978) in Hungarian showed
no diferences in either kind of on-line measure between initial 'though'
clauses, on the one hand, and corresponding main or 'if' clauses, on the
other (see Gergely 1986, 1991a) . This indicates that the Hungarian initial
'though' clauses were as fully processed as the other clause types.
In sum, it seems that when cues for TF structure are available in an
initial ' though' clause, listeners will compute focus-based inferences to an
ticipate the content of the main clause. The studies to be discussed below
examine this hypothesis in more detail. Since the experimental work was
carried out in Hungarian, in the next section I shall briefy describe how
the surface marking of the TF structure of Hungarian sentences difers
from English.
6 Structural Cues for the Identifcation of Discourse
Functions in Hungarian
In English TF structure is ofen marked only by intonational cues. The
focused element receives the primary or "focal" stress, which in simple
declarative sentences falls on the fnal word (see Halliday 1967; Chafe 1970,
1976; Akmajian 1973; Chomsky 1971; Jackendof 1972). Other elements
can also be marked as the focus, if they receive contrastive stress (e. g. ,
Jackendof 1972). Thus, though there are certain syntactic constructions,
such as clef sentences, which mark the discourse functions of constituents
explicitly, in general, syntactic structural positions do not provide cues for
the identifcation of TF structure during processing.
In contrat, Hungarian-a non-Indo-European, agglutinative language
with a rich inectional system and so-called "fee" word order-marks
grammatical relations and discourse functions by clearly distinct struc
tural cues (see e.g. , . Kiss 1981, 1987; Horvath 1981 ; Pleh 1982; Szabolcsi
1981) . Grammatical relations, such a subject, object, indirect object, etc. ,
are encoded by local cues in the form of case-marker sufxes attached to
the nouns. This allows the suface order of the major constituents to be
relatively fee: practically any permutation of subject, verb, object, and
adverbial yields a grammatical sequence.
I
The resulting versions of a sen
tence, however, are not interchangeable, they receive diferent discourse
interpretations. In fact, in Hungarian, discourse functions are related to
particular surface structure positions in a systematic manner. Therefore,
the structural positions of the Hungarian sentence can function as confg
urational cues of high cue validity for the identifcation of the discourse
roles (topic versus focus) of constituents.
The basic structural positions of the Hungarian sentence can be schem
atic-ally depicted as follows (see . Kiss 1981, 1987):
[
T
I ' T
2 , . . .
J [FJ
V
[
X
l
, x2, . .
However, word order within a major constituent (e. g. , Det Adj N) is fxed, and some
further restrictions, such a obligatory OV order for indefnite objects, apply. In fact,
constituent order is really free only with respect to grammatical functions and caes, the
position of question words, negated complements, or quantifed phrases cannot be freely
varied (see e.g. , . Kiss 1987).
2
The syntactic characterization of Hungarian sentence structure is controversial at
present. Working within the framework of generative grammar, some linguists (e. g. ,
Dezs6 1965, Kiefer 1967, Horvath 1981) propose a fully confgurational base structure
for the Hungarian sentence of the form [
s
NP[vp V NP] ] . The diferent permutations of
the verb and its complements are derived by various reordering rules such a subject
post posing and focus movement (see Horvath 1981). In contrat, . Kiss (1981, 1987)
developed within the GB framework ( Chomsky 1981) a partially non-confgurational
approach in which the underlying structure of the Hungarian sentence ha the V in initial
The sentence focus always appears in the immediately preverbal posi
tion (Le. , the "focus slot" : [FJ) and carries the focal stress of the sentence
(when there is one). 3 This position can be occupied only by a single ele
ment. The element(s) (if any) in the position preceding the focus (i. e. , the
"topic slot" : [Tl ' T2 , . . . J ) is/are the topic(s) of the sentence. The verb can
optionally be followed by frther complements ( [Xl , X2 , . . . J) which, when
conveying new information, receive secondary stress, or, if they are known,
remain unstressed.
The verb in its stem form can itself become the focus in which cae it
carries the primary stress of the sentence. However, the experiments to
be reported below rely heavily on a special feature of Hungarian grammar:
namely, that verbs often form complex predicates when combined with a
class of aspect-marking adverbial particles, called verbal modifers or con
verbs (e. g. , be /=' in' /, or el /='away' /) (see . Kiss 1981, 1987; Ackerman
and Koml6sy 1983; Szabolcsi 1986) which indicate the perfectivity of the
action.4 Complex verbs of this type can appear in two forms. On the one
hand, the verbal modifer can occupy the F slot, in which case it is prefxed
to the verb (e.g. , ' el mosta' in (5a) below) . Alternatively, the focus slot
may be flled by some other constituent. In that case, the modifer ha to
appear in a postverbal position as a separate element (e.g. , ' mosta el
'
in
(5b) below) .
position followed by its complements (including the subject) generated in an arbitrary
order as sisters to the V. The diferent surface permutations are generated by movement
rules (such a topicalization and focusing) which move the post verbal complements into
the syntactic Topic and Focus positions in front of the verb. For a critical discussion of
these positions, see Abraham and de Mey 1986, the papers by Kalman, Proszeky, and
Varga in Kenesei 1987.
3
As pointed out recently by Kalman (1985) (se also the papers by Proszeky and Varga
in Keneei 1987), there is a clas of Hungarian sentences, the socalled "level-prosody"
sentences, in which no element receives focal stress. Such neutral sentences have several
main stresses one of which falls on the element appearing in the syntactic focus position.
In such ''fat'' sentences the syntactic focus slot is not interpreted a the communicative
focus of the sentence. What is important for our present purposes, however, is that when
the sentence does have a communicative focus (as in socalled "corrective" sentences,
se Kalman 1985) it will always occupy the syntactic focus position. Therefore, this
structural position acts a a confgurational cue of high cue validity for the listener who
attempts to identify the focus of the sentence. For further details on the treatment of
the syntactic focus position in generative grammar, see Selkirk 1984, Schmerling 1980,
Farka 1986, Abraham and de Mey 1986, and Rochemont 1988.
4The verbal modifer is, in fat, only one kind of constituent which can be incorporated
into the Hungarian verb to form a complex predicate. Other incorporated constituents
include a bare N of object, subject, adverbial, or predicate function, or an Adj of predi
cate function. Incorporated constituents are not referring expressions, they add semantic
features to the verb and together they form a single semantic unit. However, they are
afected by syntatic operations: e.g., if another complement is moved into F position,
they surface postverbally. Their correct grammatical (syntactic versus lexical) treat
ment is presently subject of a controversy (see . Kiss 1987, Horvath 1981, Ackerman
and Komlosy 1983, Kenesei 1983, Szabolcsi 1986, Farkas 1986, Abraham and de Mey
1986).
TOPIC FOCUS
(5) a. Nagymama elmosta a tanyerokat a tegnapi hazibuli utan, . . .
I[T: Grandmother (=nom)] [F: away ( =V-mod)] + washed the
plates +t (=acc) yesterday +i (=adj ) party afer, . . . j
(' Grandmother has washed the plates after yesterday's party' )
TOPIC FOCUS
b. Nagymama a tanyerokat most a el a tegnapi hazibuli utan, . . .
I [T: Grandmother (=nom)] [F: the plates + t ( =acc) 1 washed
away (=V-mod) yesterday +i (=adj) party after, . . . 1
(' It was the plates that Grandmother has washed after yesterday's
party')
These examples illustrate two important points about the diference be
tween the encoding of TF structure in Hungarian and English. First, in
Hungarian the placement of the verbal modifer around the verb provides
a clear structural cue for the asignment of discourse functions to con
stituents, even when contextual and suprasegmental cues are not present.
Secondly, this structural cue can appear long before the end of the sentence
or clause is reached. With these points in mind let us return to the role of
TF structure in sentence comprehension.
7 The Role of Focus-Based Inferences in Discourse
Antecedent Matching
The hypothesized focus-based inferential processes (see Section 5 above)
involved in establishing across-clause antecedent relations in subordinate
constructions, such as (3) and (4), were tested in a previous study (Gergely
1986, 1991a, 1991b) whose main results I shall now summarize. The analy
sis of (3) and (4) above suggests that an object NP of an initial subordinate
clause can be expected to resurface as an argument of the following main
clause with higher likelihood when it is topic in the initial clause than
when it is contrastively focused. By hypothesis, it can be predicted that
an object NP, that is topic, will be foregrounded fom an initial subordi
nate clause at the clause-boundary, while a corresponding object NP, that
receives contrastive focus, will not.
To test this hypothesis the experiment examined the on-line accessibility
of initial clause object nouns in complex Hungarian sentences using an on
line word recognition task. Word probes were presented either before the
end of the initial clause (see (6a) and (6b) ) , or following the frst constituent
of the fnal clause (see ( 7a) and (7b) ) . The initial clauses tested were
either subordinate ( 'if' and 'though') or main clauses. In the critical cases,
the tested object noun (indicated by italics) always appeared in a serially
identical, preverbal position in the clause, however, it wa either the topic
of the clause (as in (6a) and ( 7a) ) , or it was contrastively focused (as in
(6b) and (7b)) .

(6) a. Ugyan 1 az estelyenl a herceget l bemutattal a baro a . . . 1 ******
1
HERCEGETI
IThough 1
at the party 1 the prince+t (=acc)] 1
in
( =V-mod 'be-')] showed \=introduced\ 1 the baron ( =nom) \to\
the . . . 1 I
( ' Though at the party the prince was introduced by the baron to
the . . . ' )
TOPIC
b. Ugyan 1 az estelyenl a herceget l mutatta be 1 a baro a . . . 1 *******
1
HERCEGETI
IThough 1
at the party] j the prince +t (=acc)] 1 showed

in I=introducedl 1 the baron (=nom) Itol the
.
|
I
(' Though at the party it was the prince whom the baron intro
duced to the . . . ' )

(7) a. Ugyan 1 a gerillakl a herceget l elraboltak, 1 az ornagy a . . . 1 ******
1
HERCEGETI
IThough 1 the guerrillas (=nom) 1 the prince +t (=acc)]
1
away- ( =V-mod 'el-' )] robbed \=kidnapped\, 1 the major

( =nom) the . j I
( ' Though the guerrillas have kidnapped the prince, the major the
. . )

b. Ugyan a gerillakl a herceget l raboltak el, 1 az ornagy a . . .

*********1 HERCEGETI
IThough 1 jthe guerrillas (=nom)]
|
the prince +t ( =acc)]

|
robbed away, (=V-mod 'el') \=kidnapped
\ 1 the major (=nom)
the . . . 1 I
( ' Though it was the prince that the guerrillas have kidnapped, the
major the . . . ' )
The sentences were presented to the subjects visually constituent-by
constituent (indicated by vertical lines in (6)-(7)) in a subject-paced read
ing paradigm. The last constituent was followed by a warning signal )
shown for . 5 sec. Then the word probe was presented and the subjects had
to decide by pushing a YES or a ' NO' button whether the probe was from
the sentence just read or not. (For further details, see Gergely 1986, 1991a,
1991b) . The main results of the experiment are summarized in Figures 2
and 3 below:
RT
920
(osee)
900
880
860
840
820
800
780
760
740
720
700
Before the
clause-boundary

MAIN
`
`
`

`

g THOUGH /=' UGYAW/

IF /=' HA' /
After the
clause-boundary
Figure 2 The Accessibility of Initial Clause Topic Objects
as a Function of Clause Type and Clause Completeness:
Mean Probe Recognition Times
RT
920
(osee)
900
880
860
840
820
800
780
760
740
720
700
Before the
clause-boundary
MAIN
g THOUGH /='UGYAN'/
-
After the
clause-boundary
Figure 3 The Accessibility of Initial Clause Focus Objects
as a Function of Clause Type and Clause Completeness:
Mean Probe Recognition Times
Figure 2 shows that, as predicted, the topic objects from initial subor
dinate clauses (both 'if' and 'though' ) were more accessible when tested
during the second clause than before the end of the initial clause (F( l, 92)
12. 62, p < . 01) . In contrast, when the corresponding object noun were
contrastively focused, no increase in target accessibility was observed a
a function of the clause-boundary (see Figure 3). These fndings support
the hypothesis that topic objects fom initial subordinate clauses are fore
grounded at the clause-boundary as likely candidates for second clause an
tecedent matching.
The results also indicate that the selective foregrounding of an initial
clause constituent is a joint function of (i) the discourse role (topic versus
focus) of the constituent, and (ii) the sentence-initial conjunctions (' if' and
'though') marking the particular semantic relation between the two clausal
propositions. Thus, when no information is available about the interclausal
semantic relation, as in initial main clauses, the listener ha no sufcient
basis to compute predictive inferences about the expectable content of the
next clause. As a result, neither topic nor focus targets get foregrounded
at the clause-boundary (see Figures 2 and 3). In fact, irrespective of their
discourse role in the initial main clause, constituents became less accessible
after the claue-boundary than before it ( F( 1, 92) 14.04, p < . 01). This
can be interpreted a being due to a process of recoding at the end of
the clause, which discards the clausal representation from working memory
when no predictive hypothesis about its likely involvement in the processing
of the next clause can be generated.
To sum up: the results of the experiment support the hypothesis that
the asignment of discourse functions to constituents, apart fom initiating
a backward search for antecedent, also starts a complementary predictive
process. As a result, the initial clause constituent, that can be expected to
serve a the discourse antecedent for the next clause, is fore grounded (Le. ,
diferentially activated) at the clause-boundary. The discourse integra
tional function of the hypothesized process is that of feeding the backward
search for an antecedent in the upcoming clause by making the most likely
candidate fom the initial clause more accessible for efcient antecedent
matching.
8 Focus-Based Predictive Inferences in Clausal Processing
8. 1 Introduction
In this section I shall report an experiment to test the hypothesis that
during the comprehension of an initial ' though' clause (see Figure 1 ) the
listener computes focus-based inferences to predict the propositional con
tent of the upcoming clause. The hypothesized function of computing such
a predictive inferential structure is to facilitate the processing of the main
clause by anticipating its content.
I assume (see also Crain and Steedman 1985, Johnson-Laird 1984,
Marslen-Wilson et al. 1978) that as the listener proceeds in the sentence,
he attempts to directly construct an interpretational model of the speaker's
intended meaning. In doing so, he accesses his general pragmatic knowl
edge represented in memory structures such as scripts. The activated script
structure (together with the overall discourse structure) serves as the in
terpretational domain for the constructed representation. I assume that
TF structure guides the construction of such a representation by fore
grounding the focus-related parts of the interpretational domain. I hy
pothesize that the foregrounded parts receive a higher level of temporary
activation than the remaining ( backgrounded) aspects of the constructed
representation.
Consider the contribution of TF structure to the interpretation of initial
main clauses such as (5a) and (5b) above repeated here as (8a) and (8b):
TOPIC FOCUS
(8) a. Nagymama elmosta a tanyerokat a tegnapi Mzibuli utan, . , .
/[T: Grandmother ( =nom)] [F: away ( =V-mod)] + washed the
plates +t (=acc) yesterday +i (=adj ) party after, . . . /
( ' Grandmother ha washed the plates after yesterday's party')
TOPIC FOCUS
b. Nagymama a trnyerokat most a el a tegnapi Mzibuli utan, . . .
/[T: Grandmother (=nom)] [F: the plates +t (=acc)] wahed
away (=V-mod) yesterday +i (=adj ) party after, . . . /
( 'It was the plates that Grandmother has washed after yesterday's
party')
When hearing these clauses, the listener accesses the corresponding 'dish
washing' script in his memory. The script specifes that the typical input
state of the object (DIRTY) is transformed into its characteristic output
state ( CLEAN) as a result of washing. In (8a) the focus position is flled by
the verbal modifer el ( elmosta = 'has washed') which expresses the perfec
tivity of the wahing event. By hypothesis, then, the representation of the
object ('the plates' ) will be foregrounded with its output state property,
i. e. , CLEAN.
The situation i s diferent, however, i n (8b). Here the fact that Grand
mother has washed something is not in focus, and so it will be in the
backgrounded part of the representation. The contrastive focus on the ob
ject 'the plates' implies a set of other things which also could have become
washed, but were not. Therefore, it is hypothesized that in the constructed
representation for (8b) this implicit contrastive set will be foregrounded.
Since this set contains potential objects of washing, it is predicted that they
will be represented with the foregrounded property of their characteristic
input state, i. e. , DIRTY.
The present study tested these predictions i n an on-line probe nam
ing task. Subjects were presented with a word probe describing either the
foregrounded, or the backgrounded part of the constructed representation.
I assume that if foregrounding results in higher level of activation, then
a related word probe will be relatively more accessible (see e. g. , Morton
1969, Marslen-Wilson and Welsh 1978) due to spreading activation (e. g. ,
Collins and Lofus 1975) fom the foregrounded aspects of the interpreta
tional structure. Thus, the word probe ' dirty' , which describes the typical
input state of the potential objects of a washing-event, should be primed
in (8b), where DIRTY is hypothesized to be a foregrounded property of
the representation. It should be less accessible, however, in (8a) , where the
foregrounded object property is CLEAN.
Similar predictions can be developed for corresponding initial 'though'
(=' ugyan' in Hungarian) clauses, as in (9a) and (9b) below:
(9) a. Ugyan Nagymama elmosta a tanyerokat a tegnapi hazibuli utan,
b. Ugyan Nagymama a tanyerokat mosta el a tegnapi hazibuli utan,
In Section 4 I argued that the sentence-initial 'though' cues the listener to
generate predictive inferences about the content of the upcoming clause.
Let us consider (9a) . For the corresponding main clause in (8a) it wa
predicted that focusing the perfective verbal modifer 'el' would foreground
the representation of the object of washing with its typical output state
property, CLEAN. The initial 'though' , however, indicates that the next
clause will contain the denial of the focus-based consequence of the initial
clause (see Figure 1) . As a result, in (9a) the construction of the predictive
structure involves a further, adversative inferential step that transforms
the foregrounded representation of the output state of the object into its
opposite, i. e. , DIRTY. Therefore, in contrat to (8a) , the word probe 'dirty'
should become relatively more accessible in (9a) .
For (9b) , however, no corresponding facilitation of 'dirty' is predicted.
Here the predicate is not focused, and so the consequence of wahing
( CLEAN) will be in the backgrounded part of the representation. The
foregrounded focus-based consequence, that is expected to be denied in the
next clause, has to do with the contrstive choice as a result of which ' the
plates' became the object of washing rather than the other potential objects
in the implied contrative set. Therefore, the listener's hypothesis about
the content of the speaker's denial will not involve the opposite ( DIRTY)
of the typical consequence of washing ( CLEAN). By hypothesis, then, in
(9b) the word probe " dirty" should not be primed.
For initial ' though' clauses the hypothesized predictive inferences are
jointly determined by the TF structure of the clause and by the semantic
value of 'though' . This suggests that, in general, the content of the con
structed representation for an initial subordinate clause varies as a function
of the kind of interclausal semantic relation. To test this hypothesis, the
Focus- BASED INFERENCES IN SENTENCE COMPREHENSION / 225
experiment also examines 'while' mialatt in Hungarian) clauses such as
( lOa) and ( lOb) below:
( 10) a. Mialatt Nagymama elmosta a tanyerokat a tegnapi Mzibuli utan,
b. Mialatt Nagymama a tanyerokat mosta el a tegnapi Mzibuli utan,
The initial ' while' in ( 10) instructs the listener that the event in the up
coming main clause takes place simultaneously with the event expressed in
the initial clause. This predictive cue does not imply adversative content for
the main clause. For 'while' clauses, therefore, the projected expectations
about the simultaneous event should not contain the meaning of 'dirty' as
a foregrounded element in either the verb-focus or the object-focus version
of ( 10). Thus, no facilitation of the word probe 'dirty' is predicted in either
( lOa) or ( lOb).
Finally, the experiment also examined the temporal characteristics of
the hypothesized priming efects. Are the inferential processes initiated as
soon as their basis in the clause (i. e. , the element in the focus position) is
identifed, or are they delayed until the full clause has been processed? To
fnd out, the word probes were presented in two serial positions: either (i)
before the end of the clause (but following the focus position) , or (ii) after
the clause-boundary.
Another question concerns the diference in the processing functions of
the hypothesized inferences in initial subordinate clauses, on the one hand,
and in corresponding main clauses, on the other. I argued that in initial
'though' clauses the focus-based inferences are predictive: t heir function is
to facilitate the processing of the main clause by anticipating its content.
Therefore, the priming efect in (9a) should be present at (or carry over to)
the beginning of the main clause.
In initial main clauses, however, the hypothesized focus-based infer
ences have no predictive function. The results of the previous study
(Gergely 1986, 1991a, 1991b), discussed in Section 7 above, indicate that
an initial clause representation is discarded fom working memory due to
recoding at the clause-boundary, unless the presence of an initial sub
ordinate conjunction overrides this process. In (8b), therefore, the hy
pothesized priming efect is predicted to disappear following the clause
boundary.
8. 2 Method
The experiment examined naming times for visually presented word probes
at diferent points in the sentence. Subjects heard three kinds of Hungarian
initial clause ( main, 'though' , and 'while' ), each in two focus versions: (i)
with a perfective verbal modifer in focus position prefxed to the verb, and
(ii) with a contrastively focused object noun in focus position. The verbs
always expressed causative events (e. g. , ' wash' ), which imply pragmatically
(but not necessarily) a typical change of state in their object (e. g. , from
DIRTY to CLEAN) . Probes were presented either (i) before the end of the
clause, or (ii) after it. The critical probes always expressed the state (e.g. ,
DIRTY) that is the opposite of the typical output state of the object (e. g. ,
CLEAN) implied by the clausal predicate (e.g. , ' wash' ) .
Subjects
Forty eight subjects (28 females and 20 males, with a mean age of 24
years) participated in the experiment. They were all students at the Eotvos
L6nind University in Budapest; their mother tongue was Hungarian.
Materials
Twelve lists were created, each containing 12 experimental and 16 filler
sentence fagments. A given experimental fagment appeared in a list in
one of two Discourse Type versions: (i) the focus position being flled by
the verbal modifer prefxed to the verb (the "verbfocus" version) as in
( 1 1a) , or (ii) with the object in focus ( "object-focus" version) as in ( 11b)
below:
TOPIC FOCUS
( 1 1 ) a. Ugyan Nagymama elmosta a tanyerokt a tegnapi hazibuli * utan,
**
/Though [ T: Grandmother (=nom)] [ F: away- (=V-mod 'el- ' )] +
washed the plates + t (=acc) yesterday + i ( =adj ) party * after,
. . . **/
( 'Though Grandmother has washed the plates after yesterday's
party' )
TOPIC FOCUS
b. Ugyan Nagymama a tanyerokat mosta el a tegnapi hazibuli * utan,
**
/Though [T: Grandmother ( =nom)] [F: the plates +t ( =acc)]
washed away (=V-mod 'el' ) yesterday +i ( =adj ) party * after, . . .
**/
( ' Though i t was the plates that Grandmother has washed after
yesterday's party')
The verb-focus fagments always had an ' S V-mod + V L Adv Phrase'
constituent order. The object-focus versions were always ' S L V V-mod
Adv Phrase' . That is, the object-focus fagments were identical with the
verb-focus versions, except for the change in word order: the object NP ap
pearing in the preverbal focus position, and the verbal modifer surfacing in
a post verbal position. All experimental fagments contained equal number
of words. Across subjects each fagment appeared equally often in the three
Clause Types: either as an initial main, 'though', or 'while' clause. Thus,
an experimental fagment was tested in six Clause Type/Discourse Type
versions. The sixteen fllers varied in length, and were either initial main
or subordinate clauses introduced by conjunctions other than 'though' or
'while'.
Each fragment was paired with two kinds of probes: "related" or "unre
lated. " The "related" probes were always state-descriptive terms express
ing the opposite of the typical output state of the object implied by the
verb. Thus, the "related" probe for ( l la) and (Ub) was 'dirty'. The "un
related" probes, however, described a state that was not related to the
state of the object (either before or after the causative event has occurred) .
For example, the "unrelated" probe for ( l la) and ( l lb) was 'torn'. Each
"relevant" probe of a fagment also appeared as an "unrelated" probe for
another fagment, and vice versa. The two kinds of probes appeared with
equal frequency throughout the lists in a random order. The fllers were fol
lowed by probes that either could or could not be part of a typical, "good"
second clause continuation of the fragment.
Across subjects, an experimental fragment appeared in each of the 12
combinations of the three Clause Types, two Discourse Types, and two
Probe Types. Accordingly, 12 lists were created and a fagment appeared
once in each list as a token of one of the 12 combinations. The 12 fagment
types appeared in an identical randomized order in each list, randomly
intermixed with the 16 fllers. The order of the 12 experimental types was
randomized so that the three clause types appeared with equal frequency
over the list. Across lists the 12 experimental fagments were rotated so
that a fagment appeared once in each of the 12 experimental positions.
Each list was presented to two subjects.
The lists were recorded in two completeness versions with normal in
tonation. In the Incomplete Clause version the experimental (but not the
fller) fagments were recorded so that the last word of the clause wa not
heard by the subjects. Instead, they heard a 50 msec 1000 Hz warning tone
in both ears at the point where the experimental fragment wa interrupted
(as indicated by the single asterisk , *
,
in ( l la) and ( Ub) ) . The fller fag
ments were flly complete even in the Incomplete Clause condition so that
subjects would not develop strategies to cope with truncated, incomplete
fagments only.
In the Complete Clause condition, both the experimental and the fller
fagments were complete initial clauses. Just after the end of the clause
( indicated by two asterisks ' **
' in ( Ua) and ( Ub) ) subjects heard a 50
msec 1000 Hz warning tone in both ears. In normal speech there is a short
pause between the end of the frst clause and the beginning of the second
clause. To allow for clause-boundary bound processes to take efect before
the on-line test, the probe word in this condition followed the warning tone
by a 250 msec delay. The lists were preceded by 8 practice trials that varied
in length and structural type similarly to the fllers.
Procedure
Subjects sat in an armchair in a dimly lit room facing a screen placed
approximately 12 feet in front of them. A microphone was placed at the
level of the subject's head slightly to his right in such a way that it would
not block his view. The lists were presented through headphones to both
ears. Subjects heard a 50 msec 1000 Hz warning tone in both ears two sec
before the presentation of a fagment.
The subjects were instructed to listen to the clause fagments trying
to fully comprehend their meaning. They were told that the sentences
would be interrupted at some ( unspecifed) point by a warning tone fol
lowed by the presentation of the probe on the screen. In the Incomplete
Clause condition the warning tone immediately started the timer, stopped
the tape-recorder, and opened the shutter of the slide projector. In the
Complete Clause condition, there was a 250 msec delay between the warn
ing tone and the appearance of the probe. There were 24 subjects in both
completeness conditions.
Subjects had to perform two tasks. First, they had to name the probe
word as fast as they could. Naming times fom the appearance of the probe
until the subject's verbal response were recorded automatically by a timer
(BP 2045982) located in the other room. To insure that subjects fully
understood the fagments, a second task was introduced. After they had
named the probe, subjects had to judge whether or not it "would ft into
a good, typical, well-sounding continuation of the interrupted sentence. "
Their response was recorded by the experimenter. Following the completion
of the second task, the experimenter restarted the tape-recorder to present
the next sentence. Each list was preceded by 8 practice trials. Overall, the
experiment lasted for approximately 45 minutes.
8. 3 Results
Scores that were more than two standard deviations above each subject's
mean on a given condition were excluded fom the data analysis. Such
scores constituted less than 3% of the responses. The results were ana
lyzed by a three-way analysis of variance with one grouping factor and two
within-subject variables. The grouping factor had two levels: the Incom
plete Clause and the Complete Clause conditions. The two within-subject
variables were: (i) Clause Type with three levels (main, 'though' , and
'while' ), and (ii) Discourse Type with two levels ( "verb-focus" and "object
focus" ) . Since both sentence fragments and subjects were fully crossed with
experimental types, separate subject-based and item-based ANOVAs were
performed.
Overall, Clause Completeness did not show a main efect (Subjects:
F( 1, 46) . 63, ] ` .6; Items: F( 1, 22) 2.07, ] < . 2). There was no main
efect of Clause Type in the item-based analysis (F(2, 44) 1 .57, ] < .25),
however, the subject-based analysis showed a slight efect (F(2, 92) 2. 81,
] < . 07) . The subject analysis also showed a marginal interaction between
Clause Type and Clause Completeness (F(2, 92) 2.42, ] < . 1) , but no
such interaction was apparent in the item-based analysis (F(2, 44) 1 .07,
] < . 4).
There was a signifcant main efect of Discourse Type in both the subject
(F( I , 46) 5. 15, ] < .03) and the item ( F( I, 22) 4.48, ] < .04) anal
ysis. Furthermore, Discourse Type and Clause Type showed a signifcant
interaction (Subjects: F(2, 92) 6.09, ] < .005; Items: F(2, 44) 3. 71,
] < . 04). Clause Completeness, however, did not interact signifcantly with
either Discourse Type or Clause Type.
Figures 4 and 5 illustrate the pattern of "related" probe naming times5
for the 'incomplete clause' , and for the Complete Clause conditions, re
spectively.
RT
960
(osec)
940
920
900
880
860
840
820
800
780
760
740

Verb-Focus

>
..
. THOUGH /;' UGYAN' /

WHILE /;' MIALATT'/

MAIN
Object-Focus
Figure 4 Mean "Related" Probe Recognition Times
in the Incomplete Clause Condition
Overall, the results supported the tested hypotheses. As predicted, Dis
course Type had a clear efect on probe recognition in both initial main and
initial 'though' clauses in the Incomplete Clause Condition (see Figure 4).
Furthermore, for 'though' clauses this efect carried over to the Complete
Clause Condition as well ( Figure 5).
The signifcant Clause Type Discourse Type interaction is largely
due to the opposite efect of Discourse Type on main versus 'though' in the
Incomplete Clause condition. Figure 4 shows that the verb-focus version of
initial 'though' clauses resulted in signifcantly fater probe naming times
than the corresponding object-focus version (Subjects: F( I, 92) = 8.90,
5
The relatively long mean probe naming times are due to the fact that the probes were
often rather long (ranging from 2 to 5 syllables, with a mean syllable length of 2.95),
morphologically complex word-formations (e.g., ' megsavanyodott' 1= [it] turned sour/,
'kiszaradtak' 1= [they] dried out/)
H
_ 960
(m-ec)
940
920
900
880
860
840
820
800
780
760
740
..
.
.
.
o
Verb-Focus
.
.
.
.
,
.
.
.
.
.
.
.
.
q wutLE)=' tL11)
'
.
utN
.
q 1uCUcu )=ucvN')
.
.
.
.
Lb ject-Focus
Figure 5 Mean "Related" Probe Recognition Times
in the Complete Clause Condition
p < . 001; Items: F( 1, 44) ~ 1 1.94, p < .001) . The opposite prediction for
initial main clauses was also borne out by the data (Subjects: F( 1, 92) =
6. 18, p < . 01; Items: F( 1, 44) = 7.37, p < 0. 01) (see Figure 4).
Figure 5 illustrates that for the object-focus version of initial main
clauses, the probe facilitation, present in the "incomplete clause" condi
tion, disappeared afer the clause-boundary. In contrast, for initial 'though'
clauses the probe facilitation efect observed in the verb-focus version of the
Incomplete Clause condition was present in the Complete Clause condition
a well (Subjects: F( 1, 92) = 5.32, p < 0. 01; Items: F( 1, 44) ~ 6.34,
p < 0. 01) . Finally, as expected, the diferent discourse versions of initial
'while' clauses showed no priming efect in either completeness condition.
8.4 Discussion
The results of the experiment make it possible to evaluate three alterna
tive hypotheses about how the listener identifes .he implicit consequence
of an initial 'though' clause that is denied by the speaker in the fnal main
clause. First, Townsend and Bever (1978) proposed that the listener can
not anticipate fom an initial ' though' clause which cause-efect relation
the speaker is denying. As a result, he computes no inferences based on
the adversative clause until the main clause is reached. In this view, the
identifcation of the particular consequence denied by the sentence takes
place only retrospectively, during the processing of the fnal main clause.
Alternatively, the listener, when processing the initial 'though' clause,
could immediately compute fom his pragmatic knowledge the exhaustive
set of typical consequences based on all parts of the clausal proposition. He
would select the intended one only later, on the bais of the second clause.
Finally, according to the present hypothesis, the listener could rely on
the TF structure of the initial 'though' clause to restrict the set of ex-
pectable consequences to only those that follow from the focused part of
the clause. This way he could anticipate the adversative content of the
fnal clause even before the clause-boundary.
The results support the fnal hypothesis. First, naming times for a word
probe, whose meaning is in adversative relation to the focus-based conse
quence of the initial 'though' clause, were signifcantly decreased both when
tested before the end of the clause (Figure 4) and following the clause
boundary ( Figure 5). This indicates that by the time the main clause is
reached, subjects have generated a predictive hypothesis about its adver
sative content, resulting in the observed probe facilitation. This rules out
the frst hypothesis that listeners postpone the inferential processing of an
initial 'though' clause.
The result demonstrates that when the TF structure of an initial ad
versative clause is clearly marked (a in the cae of the tested Hungarian
sentences) , the listener accesses his knowledge of typical cause-efect re
lations to compute focus-based predictive inferences about the content of
the fnal clause. This is also in line with the results of the replication
of Townsend and Bever ( 1978) using Hungarian sentences ( Gergely 1986,
1991a) , where, in contrast to the English results, no evidence was found
for a relatively shallow processing of initial 'though' clauses (see Section
5 above). Furthermore, the fact that probe facilitation was present al
ready before the end of the clause ( Figure 4) indicates that the focus-based
inference is generated even before the full clause has been processed.
This fnding, however, is not sufcient to rule out the second hypothesis
that the whole set of typical consequences, based on all parts of the initial
clause, would be computed from pragmatic knowledge. But the fact that
the word probes were primed only in the verb-focus version of the initial
' though' clause, rules out the latter alternative as well. The selective na
ture of the facilitation efect demonstrates that subjects inferred only those
consequences that were based on the focused part of the clause.
The results indicate that the sentence-initial ' though' modifes the pro
cessing of the frst clause by invoking a focus-baed adversative inference
whose function is to predict the content of the upcoming clause. This is
also supported by the fnding that when no initial conjunction is present,
as in the corresponding main clauses (see Figures 4 and 5) , the verb-focus
version does not result in probe facilitation.
Furthermore, the priming efect present in the initial 'though' clauses,
was absent in the corresponding 'while' clauses in both completeness con
ditions. This can be attributed to the diferent semantic value of 'while'
whose dominant interpretation is temporal rather than adversative. As
such, it is likely to cue the listener to develop predictive hypotheses about
a simultaneous event in the upcoming main clause. Since there was no spe
cifc contextual information to suggest that this next clause event would
have an adversative content, the predictive structure computed did not
contain the opposite of the output state of the object in either the verb
focus or the object-focus version of the initial 'while' clause. As a result,
no probe facilitation was observed in either of these conditions (see Figures
4 and 5).
It seems, however, that some predictive inference of a diferent con
tent was, nevertheless, computed during the initial 'while' clause as well.
This is suggested by the fact that the 'while' clause data do not show
the object-focus probe facilitation that was present in corresponding ini
tial main clauses. It seems, therefore, that the sentence-initial 'while' also
modifes the processing of the initial clause, but in a manner diferent from
'though' .
These findings are in line with the general hypothesis that the focus
based inferential structures computed for initial subordinate clauses serve a
diferent processing function than those generated for corresponding main
clauses. I argued that sentence-initial conjunctions, cueing the listener
about the interclausal semantic organization of the sentence, invoke focus
based inferences to predict the content of the second clause. In initial main
clauses, however, no such predictive cues are available, and so the observed
focus-based inferences have no interclausal predictive relevance.
This interpretation is also supported by the temporal characteristics of
the priming efects. On the assumption that the focus-based inferences in
the initial 'though' clause serve to facilitate the processing of the upcom
ing clause, it was predicted that the resulting probe facilitation would be
present at the beginning of the fnal clause. The fact that the priming ef
fect in the verb-focus version was present afer the clause-boundary in the
case of initial 'though' clauses (Figure 5) clearly supports this hypothesis.
In contrast, since the focus-based inferential structure constructed for an
initial main clause serves no predictive function, it becomes discarded from
working memory at the end of the clause. This is shown by the fact that
the priming efect, present in the object-focus version before the end of the
main clause (Figure 4), disappears in the after-clause condition (Figure 5) .
This fnding is also in line with the results of my earlier study ( Gergely
1986, 1991a, 1991b) discussed in Section 7, where targets fom an initial
main clause were found to be less accessible after the clause-boundary than
before it.
The relatively early presence of the priming efects in both the main
and the adversative clauses indicates that the focus-based inferences tested
are computed more or less directly, certainly before the full syntactic and
semantic processing of the clause is completed. This supports the interac
tive models of speech comprehension (e. g. , Marslen-Wilson and Tyler 1980,
Johnson-Laird 1984, Crain and Steedman 1985), which hold that the difer
ent sources of discourse-relevant information are employed as they become
available to constrain the direct construction of a discourse representation
of the sentence.
The results also support the assumption that the focus-based inferences
directly access pragmatic knowledge of the typical structure of events. The
critical verbs in the experiment (e. g. , 'wash') always expressed a typical
(but not necessary) change of state in their object (e. g. , fom DIRTY to
CLEAN) .
6
It was hypothesized that when the object is contrastively fo
cused, the listener constructs a representation of the implied set of poten
tial objects. In doing so, he accesses his pragmatic knowledge in which the
typical input state of the predicate event is specifed, and so the objects
in the contrastive set will be represented with the input state property
foregrounded. The fnding that the object-focus version of the main clause
primed the adjective probe that described this input state, dearly supports
the hypothesis. Thus, it can be concluded that the on-line construction of
an interpretational representation during sentence comprehension has di
rect access to pragmatic knowledge structures such as (e. g. , Schank and
Abelson 1977, Bower et al. 1979, Sharkey and Mitchell 1985) and lexical
stereotypes ( Gergely and Bever 1986).
Finally, let us compare the focus-based inferential priming efects dem
onstrated in the present study to other cases of contextual facilit0ry efects
on word recognition such as lexical priming (e.g. , Becker 1979; Fischler
and Goodman 1978; Meyer and Schvaneveldt 1971; 1975) or script-based
priming (see Sharkey and Mitchell 1985). In lexical priming the presence
of a prior context word (e. g., DOCTOR) that is semantically or associa
tively related to a target word (e.g. , NURSE) results in faster recognition
of the latter. Current theories of word recognition (e. g. , Morton 1969,
Marslen-Wilson and Welsh 1978) assume that such priming efects are due
to the temporal lowering of the recognition threshold of the target word as
a result of spreading activation coming through the semantic/associative
links fom the context word that has just been activated ( see Collins and
Lofus 1975) .
At frst, one might think that the probe facilitation in the present exper
iment is a further case of lexical priming. For example, spreading activation
propagating through the associative link fom the focused predicate ' wash'
could prime the word probe 'dirty' . However, there are two aspects of the
results suggesting that the priming efect brought about by focus-based in
ferences is a qualitatively diferent kind of contextual facilitory efect than
associative lexical priming.
The frst concerns the selective nature of the observed priming efects.
Focus-based priming of e.g. , the probe 'dirty' (with the perfective verbal
modifer 'el' of the complex predicate 'elmosta' (=' has washed') in focus
position) occurred only when the clause was introduced by 'though' . Had
6
Note that 'wah' is only a "quasi-causative" verb in so far a it is possible to wash
clean things, a well a to wah something without succeeding in making it clean. All
the critical verbs were quai-causatives of this kind.
this efect been due to automatic activation through associative lexical
connections, one would have expected comparable probe facilitation in the
corresponding verb-focus versions of initial 'while' and main clauses as well.
However, no such efects were observed.
Secondly, focus-based inferential priming seems to have a diferent time
course than lexical associative priming. Previous studies indicate that
lexical priming is both very short lived (e. g. , Warren 1972, Neely 1977,
Loftus 1973) and can be easily disrupted by even one single unrelated
intervening item (e. g. , Meyer and Schvaneveldt 1971; Gough, Alford, and
Holley-Wilcox 1981; Foss 1982). However, in this study there were always
fve or six intervening words between the focused item and the word probe
(see ( l la) and ( l lb) ) . In addition, the priming efect in the 'though' clause
condition survived even an intervening clause-boundary.
These considerations indicate that the observed facilitory efect is not
a case of simple lexical priming, as the latter is more localized and subject
to rapid automatic decay. Focus-based inferential priming seems to be a
more robust, higher-order contextual facilitation efect. It disappears only
when the representational structure underlying the efect is discarded fom
working memory. In main clauses, this takes place at the clause-boundary
as shown by the lack of priming in the afer-clause condition. In initial
'though' clauses, however, where the function of the focus-based inferential
structure is predictive, the probe facilitation is present even after the clause
boundary.
Inferential priming seems more similar, therefore, to the robust higher
order contextual facilitation of script-associated words demonstrated by
Sharkey and Mitchell ( 1985) . These authors found that the facilitation of
words associated with a script that was activated by the previous text, is
not subject to automatic decay, but is deactivated by discourse-relevant
control cues indicating script-switch in the text.
However, it seems that the priming efects in the present study cannot
be accounted for solely in terms of script activation either. Rather, the
observed probe facilitation is the result of an interactive inferential process
that is a joint function of (i) script-based knowledge, (ii) TF structure, and
(iii) interclausal semantic relation. The respective contribution of these fac
tors to the construction of the inferential structure underlying the priming
efect can be characterized a follows:
The predicate expression of the initial subordinate clause (e.g., 'washing
the plates') activates the corresponding script structure in memory (e. g. ,
the "dishwashing" script) which acts as the interpretational domain over
which the inferential structure is computed. The semantic value of the
particular subordinate conjunction restricts the size of the active interpre
tational domain by identifying those parts of the script that are relevant for
the computed inference. For example, the initial 'though' indicates that the
next clause will relate to an implicit consequence of the adversative clause.
This information restricts the active interpretational domain to that part of
the script which specifes typical cause-efect relations. Then TF structure
further restricts the active script domain to only those cause-efect rela
tions that are based on the focus of the clause. 7 Finally, the conjunction
'though' , which cues the listener that the next clause will deny the implicit
consequence, specifes a further adversative inferential step in constructing
the predictive structure, which transforms the focus-based consequence in
the active script domain into its opposite.
9 Conclusions
To sum up: the studies discussed in this paper examined the role of three
sources of discourse information in the on-line processing of complex sen
tences: (i) TF structure, (ii) interclausal semantic relations, and (iii) prag
matic knowledge. The results indicate that listeners employ these cues
interctively, as they become available in the clause, to constrain the con
struction of an interpretational model of the speaker's intended meaning.
Overall, the reported experiments support recent interactive models of
speech processing (e.g. , Marslen-Wilson et al. 1978, Marslen-Wilson and
Tyler 1980, Johnson-Laird 1984, Crain and Steedman 1985) , which argue
that the listener accesses the available contextual information directly, and
starts to construct a discourse interpretational representation for a sentence
even before its full syntactic and semantic structure has been identifed.
The present studies demonstrate that in this process a central role is played
by focus-based inferences that have direct access to pragmatic knowledge
represented in script-like memory structues. Focus-based inferences serve
several diferent processing functions during clausal processing: they are
involved in establishing implicit discourse antecedent relations as well as
in setting up implicit inferential structures linking clausal propositions in
discourse. 8
7
Discourse context typically provides even more information helping the listener to
further constrain the set of expectable consequences that the speaker's choice of focus
implies. Consider the sentence 'Though it was the plates that John ha wahed, Mother
was still not satisfied' in a context where it is known that (i) Mother had asked John
to wah the dishes, and (ii) the bulk of the dirty dishes consisted of plates. Clearly, the
listener could anticipate from the initial 'though' clause that the main clause will deny
John's expectation that his choice to do the plates would be sufcient to satisfy mother.
B
It should be pointed out that the discourse inferential processes demonstrated in the
present experiments seem to ft well the more recent discourse-oriented approaches to
lexical semantics, such as the one proposed by Kalman and Szab6 (1990). Kalman
and Szab6 distinguish a part of the discourse representation, called an "enclosure,"
that contains implicit pragmatic information asociated with predicates, which, however,
doesn't belong to the propositional content to which the speaker has committed himself
truth-conditionally, called the "rigid" part of the discourse representation. Under certain
conditions (e.g., for recovering implicit antecedents), enclosures can be "unpacked,"
i.e., their contents can be raised into the rigid part of the discourse representation.
Within this framework, the demonstrated focus-baed inferences can be interpreted as
The observed priming efects indicate that focus-based inferences result
in the foregrounding of certain parts of the constructed representation mak
ing them more accessible for on-line processes of discourse integration. The
temporal pattern of this selective activation process is related to the partic
ular processing function that the focus-based inference serves. For example,
in an initial main clause, where the inference which contributes to the con
struction of the clausal representation is nonpredictive, the foregrounded
inferential structure becomes deactivated at the clause-boundary due to
recoding at the end of the clause. On the other hand, when sentence-initial
conjunctions provide predictive cues about the likely content of the next
clause, as in initial subordinate clauses, focus-based inferences generate in
terclausal predictive hypotheses to facilitate the integrational processing of
the upcoming clause. In such cases, the presence of the sentence initial
subordinate conjunction acts as a cue to block the process of recoding at
the end of the clause, and so the predictive inferential structure does not
get deactivated at the clause-boundary.
References
.. 1986. Topic, Focus, and Confgurationality.
1983.
.
1973.
A Festshrif for Moms Halle,
...
1979. .
.Journal of Experimental Psychology: Human Perception
and Performance 5:252-259.
.1979. .
.Sentence Processing:
Psycholinguistic Studies Presented to Merrill Garrett, .
. ..
..1981. ..
The Child's Construction of Language, .
.. 1979. .
Cognitive Psychology 1 1 : 177-220.
.1970. Meaning and the Structure of Language.
.1976.
.Subject and Topic, 26-55. .
one kind of processing operation which accesses the encyclopedic knowledge represented
in enclosures during discourse processing.
ocusASED INFERENCES IN SENTENCE COMPREHENSION /
...
Semantics: An Interdisciplinary Reader in Philosophy, Linguis
tics, and Psychology, .
.
.. Lectures on Goverment and Binding.
.. ..
Exlaining Linguistic Phenomena,
. .
.. ..
.Discourse Production and Comprehension,
.
.
.Psychological Review
..
Natural Language
Processing, ...
.
...
Perception and Psychophysics
..
Cognition
Joural of Linguistics
..
Computational Linguistics
.
.Natural
Language and Linguistic Theory
.
..Joural of Experimental Psychology: Human Perception and
Performance
.
.Linguistic Inquiry
.. The Psychology
of Language. ....
..Cognitive Psychology

Discourse Integrational Processes in Sentence Compre
hension: A Study in Comparative Psycho linguistics.
.
Free Word Order and Discourse Interpretation: Ex
perimental Studies on Hungarian Sentence Processing. .

..
Language and Cognitive Processes.
/ GYORGY GERGELY
..
.Cognition

. .
Joural of Psycholinguistic Research
... .
Perception of Print: Reading Research in Experimental
Psychology, ...
.. .
Journal of Linguistics
......
..Joural of Verbal Learning and
Verbal Behavior
. Aspects of Hungarian Syntax and the Theory of Grammar.
Semantic Interpretation in Generative Grmmar. .

.. .
Mental Model. .. .
. .Approaches to
Hungarian, Vol. 1: Data and Descriptions,
. .Proceed
ings of the 7h Amsterdam Colloquium, .
..
.
.
.
Approaches to Hungarian, Vol. 2: Theories and Analyses.
On Emphasis and Word Order in Hungarian. .

.
E. ..
Linguistic Inquiry
E. Confgurationality i n Hungarian. .
.
...American Joural
of Psychology
.... .
Cognition
.... .
Studies in the Perception of
Language, .. .
.
.... .
.Cognitive
Psychology
FocusBASED INFERENCES IN SENTENCE COMPREHENSION / 239
Meyer, David E. , and Roger W. Schvaneveldt. 1971. Facilitation in Recognizing
Pairs of Words: Evidence of a Dependence Between Retrieval Operations.
Joural of Exerimental Psychology 10: 227-234.
Meyer, David E., Roger W. Schvaneveldt, and Margaret G. Ruddy. 1975. Loci of
Contextual Efects on Word Recognition. In Attention and Performance V,
ed. P. M. A. Rabbitt and S. Dornic. New York: Academic Press.
Morton, J. 1969. The Interaction of Information in Word Recognition. Psycho
logical Review 76:165-178.
Neely, James H. 1977. Semantic Priming and Retrieval from Lexical Memory:
Roles of Inhibitionless Spreading Activation and Limited-capacity Attention.
Journal of Experimental Psychology: General 106:226-254.
PIeh, Csaba. 1982. Subject or Topic in Hungarian. In Hungarian Linguistics, ed.
Ferenc Kiefer. Amsterdam: John Benj amins.
Reinhart, Tanya. 1981. Pragmatics and Linguistics: An Analysis of Sentence
Topics. Philosophica 27( 1) : 53-94.
Riesbeck, Christopher K., and Roger C. Schank. 1978. Comprehension by Com
puter: Expectation-based Analysis of Sentences in Context. In Studies in
the Perception of Language, ed. Willem J. M. Levelt and Giovanni B. Flores
d'Arcais. Chichester: Wiley.
Rochemont, Michael. 1988. FOCUS in Generative Grammar. Amsterdam: John
Benjamins.
Sanford, Anthony J. , and Simon C. Garrod. 1981. Understanding Written Lan
guage: Explortion of Comprehension Beyond the Sentence. New York: Wi
ley.
Schank, Roger C. 1975. A Story Understander. Technical Report 3. Yale Uni
versity. AI Project Research Report.
Schank, Roger C. , and Robert P. Abelson. 1977. Scripts, Plans, Goals, and
Understanding. Hillsdale: Erlbaum.
Schmeriing, Susan. 1980. The Proper Treatment of the Relationship between
Syntax and Phonology. Paper presented at the 55th annual meeting of the
LSA, December 1980, San Antonio, TX.
Selkirk, Elizabeth. 1984. Phonology and Syntax. Cambridge, Mass. : MIT Press.
Sgall, Petr, Eva Hajicova, and Eva Benesova. 1973. Topic/Focus and Generative
Semantics. Kronberg Taunus: Scriptor Verlag.
Sharkey, Noel E. , and D. C. Mitchell. 1985. Word Recognition in a Functional
Context: The Use of Scripts in Reading. Journal of Memory and Language
24: 253-270.
Szabolcsi, Anna. 1981. The Semantics of Topic-focus Articulation. In Formal
Methods in the Study of Language, ed. Jeroen Groenendijk, Theo Janssen,
and Martin Stokhof. Amsterdam: Mathematisch Centrum.
Szabolcsi, Anna. 1986. From the Definiteness Efect to Lexical Integrity. In
Topic, Focus, and Confgurtionality, ed. Werner Abraham and Sjaak de Meij.
Amsterdam: John Benjamins.
Townsend, David J. 1983. Thematic Processing in Sentences and Texts. Cognition
13: 223-61.
Townsend, David J. , and Thomas G. Bever. 1978. Interclause Relations and
Clausal Processing. Joural of Verbal Learing and Verbal Behavior 17:509-
21.
Warren, R. E. 1972. Stimulus Encoding and Memory. Joural of Experimental
Psychology 94:90-100.
Combinatory Grammar and

Projection from the Lexicon
ANNA SZABOLCSI
Objective
This paper addresses the relation between the lexicon and syntax, and sug
gests that projection fom the lexicon should be the constitutive principle
of grammar. The argument will consist of a general part in Section 1 and
a specifc part in Section 2.
In Section 1, I will ask to what extent the assumptions concerning the
nature of lexical items determine the form of syntactic representations in
various theories. As an illustration, I will briefy examine the treatments of
unbounded dependencies in Government and Binding theory, Head-driven
Phrase Structure Grammar, and Combinatory Categorial Grammar. I will
argue that the concept of lexical items as functions underlies all three of
them. However, only CCG recognizes this within its theoretical fames, and
therefore only CCG can use it as an explanatory principle.
In Section 2, I present a case study on anaphora as a specifc exam
ple of the connection between the lexicon and syntax within this theory.
I will show that a simple lexical distinction between duplicators and fee
variables interacts with the independently motivated machinery of CCG to
derive what is commonly stipulated in Binding Theory in syntax. Section
2. 1 will be concerned with refexives, and Section 2. 2 with pronouns. After
showing that the combinatory approach inescapably leads to a theory a la
Reinhart, in Section 2. 3 I will extend the proposal to VP-ellipsis, and elim-
The research reported in this paper wa supported by a fellowship from IREX at the
1987 Linguistic Institute at Stanford University, and one from the British Council in
the Centre for Cognitive Science in Edinburgh in 1988. I wish to thank the members of
those communities, a well a of MIT, for insightful objections and suggestions. I am fur
thermore indebted to Jef Goldberg, Lazl6 Kalman, Katalin E. Kiss, Edith Moravcsik,
Stanley Peters, and Mark Steedman for recent discussions.
Lexical Matter
s
. Ivan A. Sag and Anna Szabo1csi, eds.
242 / ANNA SZABOLCSI
inate an apparent problem caused by strict readings with non-referential
antecedents.
1 The Projection Principle: One Filter or the
Whole Story?
In current theories of grammar the relation of the lexicon to syntax is
regulated by some form of the Projection Principle:
( 1) Syntactic representations are projected from the lexicon i n that they
observe the pertinent properties of lexical items (where "pertinent"
and "observe" are defned theory-internally)
This principle is used to impose a necessary, but not sufcient, condition on
well-formed syntactic representations; they must not contradict the prop
erties of the lexical items they contain (subcategorization properties in the
frst place) . That granted, however, they may be subject to any kind of fur
ther conditions (pertaining to empty elements and binding, for instance) .
Given that nothing constrains the nature of such conditions with reference
to the nature of lexical items, the extent to which the lexicon eventually
determines the form of syntactic representations need not be overwhelming.
This conclusion, while perfectly compatible with the technical status of
the Projection Principle, is in confict with the constraining power many
linguists attribute to it. Intuitively, what we want may be stated as make
as much of the lexicon as possible, or, more strongly, take lexical items to
be axioms and derive sentences as theorems. This intuition would require
that projection from the lexicon be the constitutive principle of grammar.
One obvious part of this program is to develop a sophisticated empiri
cal theory of the lexicon that explains, for instance, why certain argument
structures are attested and others are not. All "interesting" issues about
grammar should belong here, and I will not be concerned with it in this
paper. The other part of the program, on which I will focus, is to show
how "uninteresting" our beloved syntax actually is. This involves making
specifc asumptions about the general nature of lexical items that, so to
say, "predict syntax." I believe that the crucial reason why syntactic repre
sentations tend to be underdetermined by the lexicon is that most theories
are extremely neutral as to what kind of objects lexical items are. Given
this neutrality, the choice of syntactic tools is extremely free, so the "laws"
of syntax turn out to be rather independent of "lexical substance" .
1. 1 Government-Binding Theory and Head-Driven Phrase
Structure Grammar
To illustrate this independence, let us consider two examples, one fom
GB as in Kayne 19831 and another fom HPSG as in Pollard 1984, 1988
l
The insights of Kayne 1983 are s featured in versions of GB, based on Chomsky
1986a. Se Koster 1987 a well.
COMBINATORY GRAMMAR AND PROJECTION FROM THE LEXICON / 243
and Pollard and Sag 1987. GB is a theory in which there is no prevalent
specifc mathematical commitment a to what kind of entities linguistic
objects, lexical items among them, are. HPSG is a theory that makes a very
general but mathematically precise commitment: it takes linguistic objects,
including lexical items, to be feature structures in a strict technical sense.
In this section I will frst briefy review their treatments of "extraction"
and "parasitic gaps" . I will then point out that in neither theory does the
nature of the generalizations concerning the well-formedness of syntactic
structures follow naturally from the view of lexical items that the theory
adopts.
Consider the following examples:
(2) (a gangster) who Mary said relatives of _ thought Kim eliminated
(3) (a gangster) who Mary said the police thought Kim eliminated
(4)
*
(a gangster) who Mary said relatives of _ thought Kim died.
For (2) to be well-formed in Kayne's theory, it must be possible to assign a
structure to it that conforms to the X-bar theory of phrase structure and
the theta theory of subcategorization in the frst place. I assume that the
reader can imagine how this would work. Secondly, given that gaps are
flled by an empty category, (2) must satisf the Empty Category Principle
(ECP) that Kayne defnes as follows:
(5) Empty Category Principle:
An empty category e must have an antecedent a such that
a. a governs e, or
a c-commands e and there exists a lexical category X such
that X governs e, and a is contained in some g( overnment)
projection of X, and
b. the union of the g-projection sets of all the empty categories
e } , . . . t@ bound by the same antecedent a must constitute a
subtree.
(6) Y is a g-prjection of X if
a. Y is an X-bar projection of X or of a g-projection of X, or
b. X is a structural governor, and Y immediately dominates W
and Z, where Z is a governed (Longobardi's amendment) max
imal projection of a g-projection of X, and W and Z are in a
canonical government confguration. (English being a VO lan
guage, the canonical government conguration is WZ. )
( 7) The g-projection set of e contains
a. every ] that is a g-projection of X, the governor of e, and
b. e itself.
Now consider ( 2) . Note that Kayne takes S and Sf to be projections
of V. In diagram (8) below only branching nodes are given. Nodes are
labelled with the numbers of the relevant clauses of the defnition.
(8)
:::
(6a,5a)

COMP Y"(6a)
I
who NP
I
Mary
Y'(6b)

Y Y"(6a,5b)
I .
said NP(6b) Y'(6b)

N PP(6a) Y Y"(6a)
I I
relatives P NP thought NP Y' (6a)
I I
of Kim Y NP
I
eliminated
e
l
Let us frst see whether e} is legitimate in view of the ECP ( 5). e} 's
antecedent who c-commands e} but it is too far up in the tree to govern
it. On the other hand, there exists a lexical category, eliminate, that
governs el . So the question is whether who is contained in a g-projection of
eliminate. We check this against defnition (6) , proceeding fom bottom to
top. Vi and V" are X-bar projections, and hence g-projections, of eliminate.
V" is in a canonical government confguration with its sister thought, hence
the next Vi is a g-projection of eliminate. The V" on top of it is an X
bar projection of Vi and hence a g-projection of eliminate. With similar
steps we get far the topmost VIII S-bar). It indeed contains the
c-commanding binder who. The relation between who and the "real gap"
el is therefore legitimate, is also attested by the grammaticality of (3).
Now take the relation between who and e2. Of governs e2. PP is an
X-bar projection, hence a g-projection, of oj. PP is in a canonical govern
ment confguration with relatives; hence NP is a g-projection of oj. V" is
not an X-bar projection of this NP, nor is NP in a canonical government
confguration with its sister. Hence V" is not a g-projection of of. There
fore, if e2 were the only gap, it could not be connected to who, as is also
attested by the ungrammaticality of (4). However, the g-projection of of
that reaches NP forms a subtree with that of eliminate. Hence, in view of
the second clause of ECP, e2 is saved; it can be "parasitic" on e } .
Consider now the treatment of the same example in Pollard 1984 and
1988. For (2) to be well-formed in HPSG, it must be obtainable by uni
fing the feature structures corresponding to its lexical items. Eliminate
ha, for instance, alongside with features specifying its part of speech and
morphological properties, a feature SUBCAT, whose value is (NP, NP) , and
an empty category e has a feature SLASH with value (NP) . It is an impor
tant property of the grammar that it contains, in addition to the general
principles of unifcation, principles specifcally pertaining to such particular
features. The informal summary of the principles that govern the behavior
of SUBCAT and SLASH is modeled after that of Pollard ( 1988):
( 9) Remove symbols from the font of the SUBCAT list one by one,
observing the Subcategorization Principle (SCP) :
Match the symbols with a complement.
( 10) Binding Inheritance Principle ( BlP) :
a. For each binding feature B, the value of B on the mother is ob
tained by starting with the value on the head daughter and ap
pending (host dependency) or merging (parasitic dependency)
into it the values on the other daughters, in the order more
oblique to less-oblique.
b. Binding features are not pased up if they are specifcally dis
charged in the construction.
( 11) Head Feature Principle ( HFP) .
For other featues, the value on mother is obtained from the
head daughter.
( 12) SLASH and REL are binding featues.
Caution: only the relevant portion of feature structures is given in ( 13).
We start fom el again. Unifcation of eliminated with the trace el
yields the feature structure V[SUBCAT (NP) , SLASH (NP) ] . Unifcation of
this with Kim yields V with [SUBCAT ( )] by the SCP and [SLASH (NP) )] by
the BlP. [SLASH (NP) ] will propagate up by the BlP. At the node dominat
ing relatives of e2 thought Kim eliminated el we have merger since another
SLASH value turns up, coming from e2. Note, incidentally, that nothing
in ( 10) would prevent this latter SLASH fom propagating even in the ab
sence of e l , in which case (4) needs to be excluded by an extra constraint
on the BlP (Pollard 1984, p. 175) . At the top we fnd who with another
binding feature, REL, of value NP, and Mary said relatives of thought Kim
eliminated with [SUBCAT ( ) , SLASH (NP) ] . Finally, a topicalization rule
allows these to unif, emptying SLASH and retaining NP in the REL-stack.
Let us return to Kayne' s proposal now. The frst thing to notice is
that the presence of who needs to be guaranteed by the ECP, for the prin
ciples pertaining to subcategorization are already satisfed by the empty
categories e l and e2. Second, the g-projection mechanism seems entirely
stipulative in the sense that there is no independent reason in the grammar
to make H expect that "pathbuilding" between the governor of the gap and
its antecedent will play a role in the grammaticality of sentences. Viewed
( 13) V[REL (NP)]
.
NP[REL (NP)]
|
who
NP[SLASH (NP)]
.

V[SLASH (NP)]
V[SUBCAT (NP), SLASH (NP)]
.
N[SUBCAT (PP)]
'
PP[SLASH (NP)] V[SUBCAT (NP, S)] V[SLASH (NP)]
relatives

thoght
P[SUBCAT (NP)]
'
of
NP[SLASH (NP)]
'
e
2
V[SUBCAT (NP), SLASH (NP)]

NP
'
Kim
V[SUBCAT (NP, NP)]
'
NP[SLASH (NP)]
|
eliminated e
l
from another angle, Kayne's proposal is not arbitrary, however. Suppose
that heads are not merely said to "subcategorize for" specifers and com
plements but are admitted as functions of these. Then, using the notation
f : a b, viz. , function: domain co-domain, (6a,b) can be said to be
about the following constellation (assume association to the right) : X is
a function from e (and possibly other arguments) to X ma , and W is a
function fom xma (and possibly other arguments) to Wma:
( 14) a. X: . . . e . . . xmax
b. W: . . . " xma " . . . wma
To say now that a path is defnable between e and Wma is to notice
that X and W as given above are composible functions. The path
forming operation can be schematized as a general version of function
composition:
( 15) Ae[W . . . (X . . . (e) . . . ) . . . J
Frthermore, the subtree condition suggests an extension of this oper
ation which, in addition to composing W and X, identifes one argument
of W with an argument of X. Let us say that W and X are connected in
this cae:
( 16) Ae[W . . . (e) . . . (X . . . (e) . . . ) . . . J
Now we can say that Kayne's defnition of g-projection is nothing other
than an algorithm that specifes how the items intervening between the
empty categori(es) and the antecedent can be assembled into one big func
tion by composition and connection (plus application to pick up other ar
guments on the way). This is an interesting result because we may now
feel we understand the rationale behind this rather complex mechanism.
First, we understand what a g-projection path is: the result of application,
composition and connection. Second, we understand why g-projection, as
opposed to any arbitrary graph-theoretic notion, may play a distinguished
role in grammar: because heads are function-like and it is in the nature of
functions to apply, compose and connect. Third, we even understand why
g-projection is specifcally relevant for structures with gaps: because com
position and connection allow a function to extend itself and thus combine
with a distant argument. The presence of who on top will now be directly
related to the subcategorization of the governor(s) at the root(s) of the
path. Note, though, that while these considerations provide glorious j us
tifcation for Kayne's ECP, they remain external to his theory since GB
does not explicitly identif lexical items with subcategorization frames as
functions.
The situation is even more striking i the cae of HPSG. Pollard ( 1988)
demonstrates, for example, with exemplary clarity, that the Subcatego
rization Principle could be paraphraed a "Combine signs as if heads were
functions applying to complements a arguments" , and the use of SLASH
plus Binding Inheritance can be paraphraed as "Propagate information as
if you were composing (connecting) functions" . It remains a fact, though,
that HPSG's lexical items are not even uncommitted entities but are enti
ties committed specifcally not to be functions. SUBCAT and SLASH are
just features on a par with, say, suppletion features. Therefore every bit of
their functional behavior needs to be imposed by brute force (by means of
the SCP and the BIP) . That is to say, the recognition of their fnctional
behavior is not merely external but rather orthogonal to the theory they
are part of.
1. 2 Combinatory Categorial Grammar
We have seen that the "function metaphor" appears insightful in connection
with important properties of syntactic representations. It may thus be
interesting to consider a grammar that elevates this metaphor to the status
of a theoretical claim and thereby lets the kind of operations found in
syntax be a natural consequence of the kind of entities it takes lexical
items to be.
Categorial grammar is known for adopting the view that lexical items
with subcategorization fames are to be conceived of a functions/functors.
One respect in which versions of categorial grammar difer quite signif
cantly ha to do with what abilities of functions they make use of in syn
tax. For extensive discussion, see the contributions in Oehrle et al. 1988.
The version that I will explore below began to take shape in Ades and
Steedman 1982 and Szabolcsi 1983, and has come to bear the name "com
binatory categorial grammar" (CCG) , for reaons that will become clear
shortly. Its main strategy can be summarized as follows.
( 17) a. No appeal is made to a phonetically empty placeholder or some
equivalent instance of hidden application to satisf the subcat
egorization requirements of the head when it does not combine
immediately with its intended argument.
b. The observations concerning functional behavior that are built
into GB/HPSG in the form of constraints are turned into oper
ations that construct the well-formed representations, and only
those, directly.
For an expression to be well-formed in this theory, it must be possible to
asemble its lexical items in a strictly monotonic fashion by rules that are
applicable whenever they meet categories they are defned for, and whose
output is not subjected to further constraints. Rules ( 18)-(21) constitute
a representative portion of the grammar. Their input ought to be more
restricted, but I will adhere to these overly general formulations to let the
reader see what can be achieved without fxes.
The rules in ( 18)-(21) are given in the same format as derivations will
be. For example, ( 18a) can be spelled out as follows: Let a be an expression
of category Y / X, the category of functors that look for an X on their right
and return a Y, and let b be an expression of category X. They can be
concatenated in this order to yield ab of category Y. In ( 18b) we have
a of Y\X, a leftward-looking functor fom X to Y, hence b of X must
precede a. (Notice that it is the orientation of the slash that indicates the
directionality of the functor; the domain category is always to the right
of the co-domain category. ) In both cases the result of concatenation is
interpreted by applying a' , the fnction interpreting a, to b' , the argument
interpreting b. Underlines bear an index that has no statu in the theory
but merely indicates, for the sake of transparency, what semantic operation
is performed, and which version of concatenation is used.2
The output of each semantic operation is spelled out in lambda terms
for convenience. Notice that we use exactly application, composition and
connection-the operations noted in the preceding section. (The fourth
rule, lifting, is a kind of auxiliary device that is necessary but is of no
particular theoretical interest in the present context. )
( 18) Concatenation interpreted as application: a' (b')
. a b - b a
Y/X X X Y\X
A
Y Y
( 19) Concatenation interpreted as composition: Ax[a' (b' (x) )]
(20)
i. a b ii. b a
Z/Y Y/X Y/X Z\Y
Z/X
Concatenation interpreted as connection:3
I. a b ii.
(Z/Y)/X Y/X
Z/X

Z/X
Ax[a' (x) (b' (x))]
b a
Y/X (Z\Y)/X
Z/X
(21) Category lifing interpreted as: Af[J(a')]
i. a
Z

T
Y/(Y\Z)
.
The derivation of (2) will proceed a follows:
2For the choice of index letters, see (23) below.
a
Z
Y\(Y/Z)
3The connection operation, introduced in Szabolcsi 1 983, is renamed U substitution in
Steedman 1987, 1988.
(22) who relatives of thought Kim eliminated
S/( S/NP) (S/(S\NP)) /NP (S\NP)/S S/( S\NP) (S\NP)/NP
Bi
S/NP
Bi
(S\NP)/NP
S
i
S/NP
Ai
S
The (S\NP) /NP category of eliminated specifes that it is a functor with
two NP arguments, of which the frst (object) is expected from the right
and the second (subject) is expected from the left. In this sentence there
is nothing in the position where eliminated expects its object to be, and it
is a remarkable feature of the grammar that it does not insert a phonet
ically empty element, either. So eliminated combines directly with Kim.
This cannot be by application, though, since the subject NP is not the frst
argument of the transitive verb (and there is nothing in the grammar to
change argument order) . The subject and the transitive verb can combine
via composition, which is possible ifthe subject also bears an appropriate
functor category, cf. the input conditions of ( 19i) . (Whether a noun phrae
in the nominative comes with the functor category S/(S\NP) fom the lex
icon, or ( 21i) lifts NP to S/(S\NP) , is immaterial for our present concerns. )
The composed functor Kim eliminated i s exactly like eliminated i n that it
expects an NP fom its right. The same holds for thought Kim eliminated
that we compose in the next step.
Now comes relatives of. Its internal structure can be ignored; the im
portant point is that, once more, it has nothing in the position of the
prepositional object and is therefore a functor. Relatives of and thought
Kim eliminated match the input conditions in (20i) , so they can be con
nected into a big "forked" fnctor that is looking for one NP ( who) which,
when found, will satisfy both prongs. (This grammar, just like that of Pol
lard ( 1988) , would require an extra restriction to rule (4) out. As things
stand now, S\NP can be lifted to S\(S/( S\NP) ) , whence thought that Kim
died can compose with relatives of of ( S/(S\NP))/NP. ) The trivial com
position steps that take care of said and Mary are omitted to save space.
Finally, accusative who must bring its S/(S/NP) category fom the lexicon.
The reason is partly semantic ( who does not denote an entity, hence can
not be just NP), and partly syntactic (lifting as in (21) preserves the order
of combination and hence cannot be responsible for the "left-extracting"
property of who) . Wo combines with (Mary said) relatives of thought
Kim eliminated by application.
Now, do the lack of traces and the use of complex operations, as in
( 17ab), represent two arbitrarily juxtaposed grammatical strategies, or are
they logically coherent? It seems they are, since as Steedman ( 1987, 1988)
observes, there exists a system of logic that has essentially these features as
its defning properties. This is combinatory logic, initiated by SchOnfnkel
( 1924), developed, among others, by Curry and Feys (1958), and expanded
in a most enjoyable fashion by Smullyan ( 1985) . Combinatory logic is a sys
tem with the same potential expressive power as the lambda calculus. But
while the lambda calculus uses abstraction and bound variables, combina
tory logic appeals solely to functional operations (combinators), to achieve
the same results. That is, it difers from the lambda calculus in precisely
the same way as the grammar just sketched difers fom GB and HPSG.
For a fast-and-easy introduction to combinators, note the following. In
( 15)-(16) and ( 19)-(21) we were dealing with "composed functions" , "con
nected functions" , and "lifted objects" . The combinators "compositor"
(B), "connector" (S) and "lifter" (T) are the operations themselves that
we get by abstracting fom those functions/objects. In (23) I defne them
both in usual lambda terms and in standard combinatory notation. The
latter should be read as follows: B is that operation which, when applied to
f, g, and x in this order, returns f(gx) , etc. Left-associativity is asumed,
i. e. , (ao)c aoc.
(23) a. Compositor: B Af AgAx[f(gx)]
Bfgx f(gx)
b. Connector: S AhAgAx[h(x) (gx)]
Shgx hx(gx)
c. Lifer: T AxAg[g(x)]
Txg gx
Application is not a combinator; it is the interpretation of the concate
nation of any two combinatory terms. However, it is possible to defne a
combinator
A
that does exactly this, AgAx[g(x)] .
The fact that combinatory logic has the same potential expressive power
a the lambda calculus entails that variable binding can be dispensed with.
To illustrate what sameness of expressive power means in more complex
caes, consider two examples in (24), which will also be useful below:
(24) a. ST expresses the same as AfAx(f(x) (x)] because
STfx Tx(fx) fxx
b. BB(ST) expresses the same as VAgAx[f(gx) (gx)] because
BB(ST)fgx B(STf)gx STf(gx) T(gx)(f(gx
f
gx) (gx)
It may be interesting to point out that the operations we found useful
in devising a grammar of English are not merely defnable in combinatory
terms but actually all correspond to rather fundamental combinators. B,
S and T represent those independent operation types which, were they
supplemented with an identifcator or a cancellator and allowed to apply
recursively, would actually yield the power of the full lambda calculus.
Since that power is neither necessary nor desirable in our syntax, I will
continue to restrict my attention to combinators with specifc linguistic
motivation. Another important assumption that I make is that combinators
are typed (cf. Hindley et al. 1972 and Morrill and Carpenter 1987. For
instance, in (23a,b,c), f is to have a type of the form P * R, g of Q P,
1 of Q, and h of Q " (P R) . Typing will be assumed but not notated
throughout the discussion.
In sum, the pursuit of the "function metaphor" appears to have led us
to the "combinator metaphor" . Again, the question is whether we should
leave it as a metaphor, or we should asign it the status of a theoretical
claim. The latter possibility would mean that we try to devise a restrictive
theory of grammar by letting the spirit and techniques of combinatory logic
delimit our analytic options.
In what follows I will make one move in this direction. I will present a
case study of anaphors and pronouns, and argue that, in view of both its
positive and negative results, combinatory logic can serve a a guideline for
their study.
Caveat: In the rest of the paper I will continue to use lambda terms
alongside with, and sometimes even instead of, combinatory terms. I wish
to point out that this is only because combinators are unfamiliar to many
readers. Lambda terms will merely serve the purpose of exposition; they
have the same status as a paraphrase in ordinary English would.
2 Anaphora-Lexical Semantics and Syntax
The suggestion that natural language syntax implements combinators,
rather than explicit variable binding, wa motivated above by a specifc
kind of example in which the variable would be phonetically empty. In
those caes it i easy to argue that we only need to account for the place
ment and interpretation of overt words, and thus the use of variables is
a mere artifct of some theories. Consider the phenomenon of anaphora,
however. The items himself and him below can be looked upon a bound
variables, but both they and their binders are in every respect normal ar
guments of the verb, so they cannot be dispensed with:
(25) a. Everyone loves himself.
b. Everyone thinks that Mary loves him.
Modulo technical details, standard theories of anaphors and pronouns
have two components. The lexical component is rather meager; these items
are assigned a so-to-say minimal interpretation, namely, that of a fee vari
able. The syntactic component is rather rich: it consists of a mechanism
for binding those variables plus a set of constraints on what they can, or
must, be bound to. See Chomsky 1981, 1986b; Chierchia 1988; Pollard
1984; among others.
The crucial binding component of these theories is not reproducible in
our grammar. The reason is that combinatory logic does not merely allow
us to handle specifcally extraction and parasitic gap structures without
bound variables: it has no variable binding at all. 4
Alongside with the above theories of binding, a number of proposals
have been put forth to the efect that (sentences containing) a refexive
can be interpreted as (containing) a kind of argument reducer, in lambda
terms, A/AX(fXX] . Applied to a two-place function, say, AZAy[SEE(z) ( y)] it
returns a one-place function obtained by identifing SEE's arguments, viz. ,
Ax[SEE(x)(x)]. See, for example, Quine 1 960, Geach 1972, von Stechow
1979, Keenan 1987, and Kanski 1987.
Given that combinatory logic has the same expressive power as the
lambda calculus, this proposal does have a straightforward equivalent in
our grammar. The requisite combinator is known as W, the duplicator in
Curry and Feys 1958.
(26) Duplicator: W VAX[/XX]
W/x /xx
where / is of P (P Q), x of P
Notice that W is the combinator that we defned a ST in (24a) .
The argument reducer proposals in literature have been essentially se
mantic in nature. They were not intended to account for the syntactic
constraints that binding theories are preoccupied with. Keenan 1987 even
argues that semantics is all that there should be to it: matters of interpre
tation are independent of matters of form like constituency.
The question that is interesting to H is this: What happens if we
incorporate the duplicator account of anaphors into the grammar outlined
in 1 .2? In the rest of the paper I will focus on three aspects of this question.
In 2. 1, I will examine how the duplicator account of refexives interacts with
the syntax of CCG as in ( 18)-(21 ) to make predictions concerning "Principle
A" properties. In 2. 2, I will examine what the theory entails for pronouns,
free and bound, and point out that it practically derives Reinhart's ( 1983)
results. In 2. 3, I will extend the proposal to VP-ellipsis to resolve a problem
facing any Reinhart-like theory of anaphora.
2. 1 Refexives
Standard binding theories attribute three primitive properties to anaphors:
the necessity for there to be a binder, the prominence condition on the
binder, and the locality condition on the binder. I suggest that the simplest
possible account of the frst of these is to interpret refexives as duplicators
in the lexicon. Given that refexives are lexical items in need of some
4That is, we have combinatory terms with the same meaning B f(x) and combinatory
terms with the same meaning U AX[J(X)
'
but the latter is not obtained from the former.
meaning in any cae, any other treatment would involve that we assign the
"wrong" lexical meaning to them, and go on to put it "right" in syntax.
Note a problem now: W is an operation over functions, whereas himself
is an argument of the verb. But notice that turning a two-place fnction
into a one-place function is essentially the same as providing one of its
arguments. The conceptual gap is bridged by function-argument structure
reversal, viz. , lifing. This tells H what kind of a noun phrase a refexive
is: nothing but a lifed kind.
Lifting has ample syntactic motivation in our grammar (see especially
Steedman 1986 and Dowty 1988), so refexives are by no means excep
tional in having lifted kind of categories. The only peculiarity is that while
an entity-denoting item like Ma may come fom the lexicon with the
category NP and get lifed in syntax, the same categories are assigned to
refexives directly in the lexicon, matching their meaning. In this regard
the treatment of refexives is like that of quantifers. See (27) , in which
items in the frst row are annotated with a category and an interpretation
that suit the subject position, and those in the second, the object position.
I continue to use lambda terms for the reader's convenience:
(27) everyone Mary * sheself
S/(S\NP) S/(S\NP) S/(S\NP)
V
Vy[fy] -t[ f
m
]
everyone Mary herself
(S\NP) \ ( (S\NP) /NP) (S\NP)\ ( (S\NP) /NP) (S\NP) \ ( (S\NP) /NP)
-g-yVx[gxy] -g-y[g
m
y] -g-x[gxx]
A simple example is then derived as below. Herself is a leftward-looking
functor with the duplicator interpretation AgAX[gXXJ . It applies to the
transitive verb sees to yield a verb phrae S\NP with the interpretation
Ax[SEE(x) (x)J . The subject quantifer everyone is a rightward-looking
functor with interpretation VVy[fyJ . It applies to the verb phrase and
yields a sentence with the desired interpretation Vy[SEE(y)(y)J .
(28) Everyone sees herself
S/(S\NP) (S\NP) /NP (S\NP) \ ( (S\NP) /NP)
A
ii
S\NP
Ai
S
Now consider how this proposal accounts for the ungrammaticality of
(29a,b) : 5
5
(29a,b) cannot be ruled out with reference to a morphological gap. They are equally
ungrammatical in Hungarian, a language that ha nominative anaphors in the subject
position of NP. (On the clause-like structure of the Hungarian noun phrae, see Szabo1csi
1984.)
(29) a.
*
Sheself lef.
b.
*
Sheself sees everyone.
As for (29a), the non-existence of sheself of category S/(S\NP) is due to the
fact that the duplicator is by defnition a two-place function, which cannot
be the interpretation of a one-place functor. No such simple semantic
explanation can be given for ( 29b) , however. To wit, shes elf might also
be expected to bear the lexical category (S/NP) /( (S\NP) /NP
)
, which may
well be interpreted a a duplicator and bind the subject to the object.
This is essentially what Keenan points out when he invokes his Nominative
Reference Condition to eliminate (29b).
The situation i n our grammar i s slightly diferent, however. Notice that
the dangerous category above cannot be obtained by lifting NP but only
by applying the compositor B to S/(S\NP) unarily. To make this more
transparent, in (30) let u use the ad hoc labels subj and obj, and ignore
directionality:
(30) a. (S/NP) /((S\NP) /NP) is ( Sl obj) I ((Sl subj) l obj)
S/(S\NP) is SI (Sl subj)
b. T of subj may be (Sl obj) I ((SI obj) l subj)
T of obj may be (Sl subj) ! ((S! subj) l obj)
c. B of S! ( S! subj) may be ( S! obj) ! ( ( S! subj) ! obj)
Now, i n distinction to Lambek calculi, for instance, the syntax i n ( 18)
(21) does not include unary B (division) . We only used B to interpret
concatenation. This is not accidental. There are various kinds of disaters,
some but not all mentioned in (the revised version of) Szabolcsi 1987,
that unary B may cause in syntax. Hence the problematic category is not
a standard category for noun phraes in English: even Mary will never
acquire it. So all we need to assume in order to exclude ( 29b) is that
refexives may not have lexical categories that are not available in syntax
to normal noun phrases. Which, in fact , is the null hypothesis.
It may be concluded that by lexically interpreting refexives a dupli
cators, plus assigning them to categories in conformity to independently
motivated syntactic assumptions, we predict that they "are bound by a
more prominent argument" . But no binding mechanism and c-command
condition need to be stipulated. Further details of this proposal, including
the treatment of two-complement verb cases, pied piping, and interaction
with extraction and coordination are developed in Szabolcsi 1987.
That said about the requirement of a more prominent binder, let us
turn to the locality condition on anaphors. It is to be observed that locality
cannot appear as a natural condition in a system that includes composition.
Consider, for instance:
(31) a.
*
Mary believes that John loves herself.
b. Who does Mary believe that John loves?
The extraction structure (31b) shows that believe that John loves of cate-
gory ( S\NP) /NP can be obtained by composition. If herself in (31a) applies
to such a composed transitive verb, it will "get bound" by the subject Mary.
So, in case the treatment of anaphors is to be part of this system, we must
resort to brute force to capture the locality condition. The brute force
method is, basically, to require that the duplicator only apply to functors
that are lexical in some sense. (Cf. Chomsky's 1986b minimal complete
functional complex. )
Without going into details regarding the precise defnition of lexicality,
let us ask how sad one should be about this. Is the locality condition part
and parcel of the notion anaphor, where by "anaphor" we mean an item
that must be bound by a c-commanding argument, i.e., which is interpreted
as a duplicator?
The existence of long-distance anaphors has been widely recognized
for quite some time. Most of them reside within NPs and are exclusively
subject oriented, that is, they appear to have rather peculiar restrictions.
A very interesting cae fom Modern Greek is reported by Iatridou 1986,
however. Greek ha two anaphors in our sense, of which ton eafon tau
is bound within, and ton idhio outside, its governing category. The data
Iatridou presents also indicate that this latter, long-distance anaphor has
no funny restrictions:
(32) 0 Yanisi ipe ston Costaj oti i Maria
k
aghapa ton idhio
i/j/
*k
said to that loves himself
A plethora of further relevant data can be found in Keenan 1988.
This indicates that locality can in general be divorced fom the core
notion of anaphor, contrary to what binding theories suggest. The local
ity condition (lexicality requirement) may really be a brute force device
employed by natural language to facilitate processing. Clearly, it is very
useful for the hearer if binding ambiguities are reduced by having diferent
forms for the duplicator; but there may be nothing more to it. While my
proposal (as it stands, at leat) is unrevealing with respect to what locality
conditions diferent languages may impose on their anaphors, it may be
taken to be revealing in the sense that it predicts locality to be a more or
less ad hoc matter.
2.2 Pronouns
There is another phenomenon that the present theory makes predictions
about, namely, bound versus fee pronouns.
If the local binding condition on refexives is a more or less ad hoc
matter, there is in principle no obstacle to extending the class of anaphors
i. e. , lexical duplicators-to include items which only difer from refexives in
that they are subject to no, or diferent, locality conditions. This is good
news because, given that our combinatory logic ha no variable binding
(assignment switching) mechanism, having a duplicator kind of meaning is
the only chance for an item to get interpreted as a bound variable in the
usual sense. Besides this, combinatory logic only ofers free variables. But
those are like any name: they start out free and remain fee. They may
only co-refer with other referential expressions on independent grounds.
These considerations imply that pronouns must be multiply ambiguous.
(33) a. He[bound] is a member of the class of anaphors. As a frst approx
imation, it may be assigned the same lexical meaning(s) as refex
ives, whence it is subject to the same prominence/constituency
requirements, and to possibly diferent locality conditions.
b. He[free] is baically deictic. It represents arbitrarily many difer
ent fee variables, each having its value fxed once and for all.
Consider now the anti-locality condition on pronouns ( "free in its gov
erning category" ) . The main point to note is this. The relation between
he [bound] and its antecedent is recognizable within the combinatory the
ory, but the relation between he [fee] and the item it happens to co-refer
with is not. Hence we may hope to be able to impose an anti-locality
condition only on the former, but not on the latter.
Anti-locality for he (bound] can be captured by something like the op
posite of the lexicality requirement for refexives or, more interestingly, by
utilizing a combinator that is independently necessary for standard cases of
pied piping. The combinatory equivalent of having a feature inherited by
mother fom daughter is discussed in Szaholcsi 1987, following Steedman
(p.C. ) :6
(34) If a noun phrase is interpreted as a, its pied piper version is inter
preted as C(B(Ba)B)or B(Ba)B,and has a matching category. For
example: X\( X/NP)-its pied piper version: (X\( X/Y \( Y/NP) .
For instance, if a is Wand Y is PP, the use of the pied piper category will
allow to himself to inherit anaphorhood fom himself. 7 But notice now
that the "feature" that "percolates" from NP to Y is necessarily "inert"
within Y. Thus, by making pronouns obligatory S-pied pipers, we let S
inherit anaphorhood fom he and him and, at the same time, we guarantee
that him has no antecedent within its minimal S. So let us assume that
he(bound] has no simple W interpretation but is at least C(B(BW)B) or
B(BW)B,viz. , AgAfAX[f(gx)(x)] .
(35) a. Everyone thinks he[bound]
S/VP VP/S (VP\(VP/S /(S\NP)
likes John
S\NP
Ai
VP\(VP IS)
AfAx[f(likes(j)(x (x)]
6
C i the permutator, B(T(BBT (BBT) or V>XAy[YXJ .
7
Pace Pollard's "Evidently, there is no principled analysis of pied piping in an ex
tended categorial framework like Steedman's without the addition of a feature-passing
mechanism for unbounded dependencies" (1988, 421).
b. Everyone thinks John likes him[bound] .
S/VP VP/S S/VP VP/NP (VP\(VP /S
) )
\(S/NP
)
+
c. Everyone thinks John likes hi m [bound]
S/VP VP/S S/Vp VP/NP (VP\(VP /S
) )
\(S/NP
)
S/NP
A
.
VP\(VP/S
)
>1 >x[J(likes(x ) (
j
) )
(x)]
He/him [bound] , the S-pied piper, can only combine with S/NP or S\NP,
whence it cannot be bound to anything within its own clause. As was
pointed out above, anti-locality for he/him[fee] cannot be accommodated
in this way and must therefore be attributed to independent mechanisms.
The pictue emerging here is not only technically but also conceptually
diferent fom that of standard binding theory, which specifes only that
pronouns are fee in their governing category. However, it squares rather
well with an alternative theory proposed in Reinhart 1983. Reinhart ar
gues that the conation of coreference and bound variable interpretation
leads to enormous complications because the two are empirically diferent.
A pronoun may be coreferential with another referential expression even
if the latter does not c-command it; moreover, their (non-)coreference is
afected by pragmatic factors. On the other hand, a pronoun can only be
bound by a quantifer that c-commands it fom a distance at s-structure:
i.e. , bound variable interpretation is as strictly syntactic a refexivization.
In conclusion, she proposes to distinguish the syntactic theory of bound
anaphora from the pragmatic theory of coreference. See also Partee 1978.
It is extremely interesting to note here that acquisition studies by
Wexler and Chien fom 1988 provide evidence in support of Reinhart's dis
tinctions. In previous experiments they had found that English-speaking
children's performance on locality properties of refexives increaes steadily
fom age 2;6 to ahnost perfect performance at 5; 6. On the other hand,
children in the age range 5; 6 to 6; 6 still do not show that they have the
knowledge that a pronoun may not have a local c-commanding antecedent.
However, in the study reported in this paper they tested the hypothesis
that, in line with Reinhart's proposal, principle B really applies to bound,
and not coreferential, pronouns. They found that children at age 6 violate
principle B approximately 50% of the time with coreference, but less than
15% of the time in the case of a bound variable pronoun. s
B
The signifcance of Chien and Wexler's earlier results ha been challenged by Grimshaw
and Rosen 1988: they argue that children know, although they may not obey, principle B.
The 1988 study involving bound variables is not subjected to criticism i n Grimshaw and
COMBINATORY GRAMMAR AND PROJECTION FROM THE LEXICON I 259
These results, if correct, are especially important for combinatory gram
mar. As was pointed out, the above treatment of refexives and bound ver
sus free pronouns is the only kind of treatment the logical foundations make
available. Therefore it makes a much stronger case than Reinhart's the
ory, which formulates empirical observations within a grammar explicitly
incorporating the lambda-calculus.
Despite its acknowledged appeal, Reinhart's theory ha been criticized
in the literature fom various angles. In this paper I will not be concerned
with the problem raised by the possibility of a pronoun being bound by
a quantifer that does not c-command it at s-structure. (See a suggestion
in the revised version of Szabolcsi 1987. ) Rather, I will turn directly to a
third crucial binding phenomenon in Reinhart's theory, viz. , VP-ellipsis.
2. 3 VP-Ellipsis
It is well-known that elliptical VPs may be ambiguous between the "strict"
and the "sloppy" readings in the following kind of context:
(36) Felix hates his neighbors and so does Max.
a. sloppy reading: "and Max hates Max's neighbors"
b. strict reading: "and Max hates Felix's neighbors"
Reinhart ( 1983, 150-51) points out that the sloppy reading is obtained if
Felix binds his in the antecedent claue, and the strict reading is obtained
if his is merely coreferential with Felix. Reinhart is primarily interested in
the sloppy case, and her pertinent claim is generally accepted. On the other
hand, her claim concerning the strict reading faces a serious problem. If the
strict reading is contingent on accidental coreference, then it is predicted to
be available only if the antecedent itself is referential. But this prediction
is refuted by quantifcational antecedents, which also support the strict
reading in a slightly diferent context, viz. , when we are not dealing with
coordination:
(37) Every man mentioned his merits before Mary did.
a. sloppy: "before Mary mentioned her merits"
b. strict: "before Mary mentioned his merits"
Gawron and Peters ( 1990) discuss this point extensively and take it to be
one important piece of evidence against the correctness of Reinhart-like
theories of pronominal anaphora. They propose a system whose innova
tions include the postulation of three, rather than two, kinds of anaphoric
relations: co-variation (cf. co-reference) , role-linking (cf. binding) and, cru
cially, co-parametrization for cases like (37b) . The reader is referred to
their book for details.
Rosen's paper, however, and it appears to me that Grimshaw and Rosen's fndings are
perfectly compatible with the results concerning the diferential acquisition of conditions
on binding and coreference.
Let us see what the combinatory theory of anaphora has to say about
example (37). At frst glance, it appears to be impossible to give an account
of (37b) . The reasoning, along the lines of Gawron and Peters, goes as
follows. We have binding in the frst clause, so its VP must be interpreted
a >x[ mentioned(x
'
s merits) (x)J . If the ellipted VP echoes this, we get the
sloppy reading; fne. In order to get the strict reading, we should start
out with an antecedent VP meaning containing a fee variable, echo this,
and then bind both occurrences of that variable. But we cannot get free
variables bound. Hence (37b) seems disatrous. In a sense, it is even more
disatrous for this theory than for Reinhart's since, a I noted above, what
is stipulated in her theory is a matter of logical necessity here.
This reaoning ha an important presupposition, namely, that strict
readings with non-referential antecedents need to be accounted for specif
ically by the machinery for pronominal anaphora and, therefore, if the
pronominal anaphora machinery ha nothing to say about it, the availabil
ity of this reading is evidence against the machinery.
In what follows I will suggest that this presupposition is false, and
hence the reasoning is wrong. First, I will observe that strict readings
with non-referential antecedents arise even in contexts without any overt
anaphora. Therefore the existence of (37b) does not directly refect on
the viability of proposals concerning pronouns. Second, I will ask whether
the treatment of the natural clas of phenomena to which (37b) belongs
really requires that we go beyond the tools of combinatory grammar and
introduce a novel device like co-parametrization. I will argue that it does
not. I will outline an account in CCG, and point out that it requires only a
minimal extension of what we already have, because good old composition
plus duplication will do "co-parametrization" for us. In this paper I will
not investigate how other theories that do not have these would derive the
critical examples.
The frst step is entirely theory-independent. The consideration of two
examples will sufce to show that the problem and, consequently, the treat
ment of (37b) is independent of the specifc abilities of pronouns. One
is (38):
(38) Every man mentioned himself before Mary did.
a. sloppy: "before Mary mentioned herself"
b. strict: "before Mary mentioned him"
Given that (38) contains a refexive, it shows that we must be able to
produce the strict reading with a non-referential antecedent even under
the canonically syntactic conditions refexive binding is subject to. Second,
consider (39):
(39) Who did you mention before Mary did?
"Which T, you mentioned x before Mary mentioned x"
As the interpretation makes clear, 39)presents precisely the same problem
as 37b) and 38b) . But it is a simple extraction structure, so the prob
lem must be independent of the treatment of overt anaphoric elements on
the whole.
With these observations in mind we may now proceed and see how the
phenomenon can be handled in CCG. It is to be emphasized that in this
paper I will not develop a detailed account of VP-ellipsis but concentrate
only on the points relevant for the argument as sketched above. The frst
task is to accommodate the simplest case of VP-ellipsis, with which we start:
40) John lef before Mary did.
Observe that the interpretation of 40) involves duplication, namely: be
fore ( lef( Mary ( lef) . One way to capture this would be to say that did
is a duplicator, i.e. , it has the same meaning as items like himself, but
its typing is diferent: the function whose arguments it identifes has VP,
rather than NP, arguments. This proposal encounters two kinds of dif
culty. One is the fact that many of the pertinent examples work also in
the absence of did ( John lef before Mary, etc. ) . The second is that this
proposal would extend easily to 39) but, for rather technical reasons, not
to 38b) . Therefore I will largely ignore the contribution of did and use
the same Wa the interpretation of a syntactic operation. Technically, W
can be implemented as a new interpretation for concatenation, see 41a) ,
or as a unary type-change) rule that applies to the material in the second
clause, as in 41 b) . Given that here the use of W is followed by applica
tion, the choice makes no diference. I sketch both options because binary
W may be more intuitive, but subsequent steps will have to utilize the
unary one.
41) a. lef before Mary did
S\NP ( (S\NP) \ (S\NP) )/S S/(S\NP)
B
.
( (S\NP) \ (S\NP /(S\NP)
W
..
S\NP
b. lef before Mary did
S\NP ( (S\NP) \(S\NP /S S/(S\NP)
B,
( (S\NP) \ (S\NP / (S\NP)
W
(S\NP)\(S\NP)
A
..
S\NP
For the reader's convenience, in what follows I will spell out the interpre
tation of each line in lambda terms. Intermediate steps will be omitted if
given in a previous example:
(41' ) a. B(before) (Mary did)
AaAbAc[a( bc)] (ApAgAX [before (
p
) (g ) (x )] ) (AJ[J( m)] )
AbAC[A
p
AgAX
[before (
p
)(g ) (x
) ] ( bC)](AJ[J(m)])
= AbAcAgAx[b
efore(bc) (g) (X)] ( AJ[J(
m)] )
=
AcAgAx[bef
ore(
V
[J(m)] (
c)) (g) (x)]
ACAgAx[before(cm) (g) (x)]
W(B(before) ( Mary-did)) (left)
= AaAb[abb] (AcAgAx[before(cm) (g )(x )] )( left)
Ab[AcAgAx[before( cm ) (g ) (x ) ] ( b )(b )] (left)
AbAx[before(bm) (b) (x)] (left)
Ax[before(left( m) ) (left ) (x)]
b. B(before) (Mary-did) = ACAgAx[before(cm) (g) (x)]
W(B(before) ( Mary-did))
= AaAb[ abb] ( ACAg AX [before( cm ) (g) (x )] )
AbAx[before(bm)(b) (x)]
W( B(before) (Mary-did)) (left)
AbAX [before(bm) (b) (x)] (left)
= Ax[before(lef(m) ) (left ) (x)]
This accommodates the simplest cae of VP-ellipsis. Now consider ( 39):
(39) Who did you mention before Mary did?
Sentence (39) presents a problem on its own right. W echoes a VP-meaning
but, the direct object being extracted, there is no VP in the antecedent
clause. Recall that our grammar ha no traces.
The problem of (39) is in fact eay to solve with the tools our grammar
ha had al along. In (41b) the segment W(B(before Mary)) is combined
with lef by application. But it can equally well combine with mention by
composition-and that is all we need. A straightforward execution of this
idea is given in (42a). Once more, (42b) presents an alternative that makes
no diference here but will underlie the next derivation. In (42b) a unary
combinator BBW is used, followed by application. BBW is obtained by
composing the above motivated steps W and B, and serves no other end
than lumping these two together in the said order. For its expression in
lambda terms, recall (24b) .
(42) a. mention before
Mary did
(S\NP) /NP ( (S\NP) \(S\NP))/S S/(S\NP)
Bi
S\NP) \
( S\NP) ) / (S\NP)
--W
(S\NP) \ (S\NP)
B
ii
(S\NP)/NP
b. mention before Mary did
(S\NP)/NP ( (S\NP)\(S\NP) )/S S/(S\NP)
B
.
( (S\NP) \ (S\NP / (S\NP)
BBW
( (S\NP) /NP) \ ( (S\NP) /NP)
^
(S\NP)/NP
Note the replacement of the two occurrences of b in (42'a) with cd: this
has the efect of "co-parametrization" .
(42' ) a. W(B(before) (Mary-did)) = AbAx[before(bm) (b)(x)]
B ((W (before) (Mary-did ) (mention)
= AaAcAd[a( cd)] (AbAx[before(bm) (b ) (x )] ) (mention)
= Ac'd[ 'b'x[before( bm) ( b) (x)] (cd)] (mention)
= Ac,d,x [before ( cdm) c)(x)] ( mention)
= Ad'x[before( mention( d) (m (mention( d (x)]
b. B(before) (Mary-did) = 'cAg'x[before(cm)(g) (x)]
BBW = ,f'h,y[f(h
y
) ( h
y
)]
BBW(B(before) (Mary-did
= AI 'hAy[f(hy) ( hy
)] (AcAg'x[before(cm) (
g
) (x)] )
= AhA
y
'x[before( hym)(h
y
) (x)]
BBW(B(before)(Mary-did (mention)
= AYAx[before(mention(
y
) (m( mention(
y
(x)]
In view of the informal suggestion made at the outset, with these we
must have everything ready to cater to the ambiguity with anaphors. To
avoid the treatment of possessive pronoun, I will only consider (38) here.
The sloppy reading is derivable by mimicking (41b). For compactness,
S\NP will be abbreviated as VP whenever this does not afect intelligibility.
(43) mention himself before Mary did
VP/NP VP\(VP/NP) B
i
A
..
W
VP (S\NP)\ VP
^tt
S\NP
(43' ) W(B(before) (Mary-did = 'bAx[before(bm) (b) (x)]
himself = W = AI Ay[fyy]
W(mention) = AI,y[fyy] (mention) = ,y
[mention(
y
) (
y
)]
W(B(before ) ( Mary-did) ) (W(mention) )
= AbAx[before( bm) ( b) (x)] (AY[ mention(
y
) (
y
)] )
= Ax[before( mention( m)( m (mention( x )( x ]
Consider now the strict reading. The correct semantic result obtains if we
apply himself, the duplicator to the interpretation of mention before Mary
did derived in (42) :
(44) W(BBW(B(before) (Mary-did) ) (mention) )
>f >z[f zz] ( >
y
>x[before(mention(y) (m) ) (mention(y)) (x)] )
>z[>y
>x[before(mention(
y
) (m) ) (mention(y
) ) (x )] (z ) (
z
) ]
>z[before( mention( z) ( m) ) ( mention( z) ) (z)]
We could indeed do this and thus derive the strict reading by simply mim
icking (42) if we had a wrap operation in syntax:
( 45) mention himself before Mary did
hiInself mention before Mary did

As a matter of fact, the grammar I am working with does not have wrap.
This is not the only case for which wrap would be needed, however. An
account of VP-internal anaphora presupposes that the neutral linear order
of complements in English is the reverse of their semantic order. Therefore
in Szabolcsi 1987, a simulation of wrap is developed, using forward mixing
composition (i.e. , disharmonic composition with the direct object acting a
principal functor) . Since that is an operation we otherwise do not want to
set fee in syntax, it is pushed back into the lexicon in the form of unary
B with idiosyncratic slahing. Using this lexically derived category, wrap
can be broken into two concatenation steps.9
With these technicalities in mind, consider (46). Notice that it utilizes,
and thus provides the ultimate motivation for, the composite combinator
BBW developed in (42b) :
(46) B
lex
(himself)
(VP\(VP /NP) ) /( (VP /NP)\(VP /NP) )
everyone
S/(S\NP)
mentioned
VP/NP
VP
before Mary-did
B
BBW
(VP /NP)\(VP /NP)
^
VP\(VP/NP)
^
S
9This is not the only, or even the best, imaginable simulation of wrap in CCG. It is
possible that a nicer solution will also allow us to dispense with UU and, hence,
U the interpretation of a syntactic rule in general.
(46') BBW(B(before) (Mary-did)) ~ AhAqAx[before(hym) ( hy) (x)|
Blex( himselj) ~ BW ~ AJAgAV[J(gV)] (AkAZ[kzz] ) = AgAVAZ[gVZZ]
BW(BBW(B(before ) (Mary-did) ) )
~ AgAvAz[gvzz](AhAYAx[before(hym)(hy ) (x)] )
= AAz
AhAqAx
before( hym) ( hy) (x)] ( v) ( z) (z) ]

~ AVAz[before(vzm) ( vz) (z)]
BW(BBW(B(before) (Mary-did) ) ) (mention)
= Az[before( mention( z) ( m) ) (mention( z)) (z)]
ArtX(f x] (Az[before(mention( z) (m) ) (mention( z) )(z)] )
~ Vx[before(mention(x) ( m) ) (mention(x) ) (x)]
Finally, note that this treatment does not carry over to VP-ellipsis in
coordination or in discourse. But that makes just the right prediction. In
contradistinction to (36) with referential Felix, (47a) and (47b) have no
strict readings: 1o
( 4 7) a. Every man corrected himself, and so did Mary.
b. Every man corrected himself. So did Mary.
Let me summarize what we have done as follows. We need some new
tool to account for the simplest case of VP-ellipsis, viz. , John lef before
Mary did. We choose this tool to be the rule W. Once we add that, the
tools we have always had in the grammar, composition (B) in particular,
will allow us to derive Wo did you mention before Mary did? This lat
ter, in conjunction with the apparatu necessary for pure refexivization,
will derive the sloppy as well as the strict readings of Everyone mentioned
himselJ before Mary did. In this way, the existence of the strict reading of
VP-ellipsis with non-referential antecedents ceases to constitute an argu
ment against Reinhart-like theories of pronominal anaphora in general and
against the combinatory theory in particular.
3 Conclusion
In Section 1 a specifc problem wa raised concerning the implementation
of the Projection Principle. It was pointed out that in various theories
of grammar the nature of syntactic representations is underdetermined
by the nature of lexical items. An examination of accounts of extrac
tion and parasitic gaps suggested that taking lexical items to be functions
allows us to establish a signifcant link between the two. The "fnction
1
1 tentatively asume that in coordination and in discourse do ha the interpretation of

a free variable and corefers with the antecedent VP. The fact that VP-variables cannot
be deictic need not rule out this analysis; it may be due to an independent requirement
that the referents ofVP-variables be linguistically individuated. I am grateful to Frederic
Evans, Jorge Hankamer and Ivan Sag for discussion of this point.
metaphor" was then shown to lead to the "combinator metaphor" in a
natural way.
In Section 2 the combinatory proposal was confonted with a new phe
nomenon, namely, anaphora. A selection of specifc problems concerning
refexives, pronouns, and VP-ellipsis was examined. At the technical level it
was argued that combinatory logic ofers a simple lexical-based treatment
of anaphora which, in conjunction with the grammar independently moti
vated in Section 1 , factors out and derives a set of coherent properties of
anaphora without any specifc stipulation. At the intuitive level I hope to
have shown that the use of combinators in grammar has its own heuristic
value and ofers genuine insights.
References
Ades, Anthony, and Mark J. Steedman. 1982. On the Order of Words. Linguistics
and Philosophy 4(4): 517-558.
Chierchia, Gennaro. 1988. Aspects of a Categorial Theory of Binding. In Catego
rial Grammars and Natural Language Structures, ed. Richard Oehrle, Emmon
Bach, and Deirdre Wheeler. Dordrecht: Reidel.
Chomsky, Noam. 1981. Lectures on Goverment and Binding. Dordrecht: Foris.
Chomsky, Noam. 1986a. Barers. Cambridge, Mas.: MIT Press.
Chomsky, Noam. 1986b. Knowledge of Language. New York: Praeger.
Curry, Haskell B. , and Robert Feys. 1958. Combinatory Logic I. Amsterdam:
North-Holland.
Dowty, David R. 1988. Type Raising, Functional Composition, and Non
Constituent Conjunction. In Categorial Grammars and Natural Language
Structures, ed. Richard Oehrle, Emmon Bach, and Deirdre Wheeler. Dor
drecht: Reidel.
Gawron, Jean Mark, and Stanley Peters. 1990. Anaphora and Quantifcation in
Situation Semantics. CSLI Lecture Notes No. 19. Stanford: CSLI Publica
tions.
Geach, Peter T. 1972. A Program for Syntax. In Semantics of Natural Language,
ed. Gilbert Harman and Donald Davidson. Dordrecht: Reidel.
Grimshaw, Jane, and Sarah T. Rosen. 1 990. The Developmental Status of the
Binding Theory or Knowledge and Obedience. Linguistic Inquiry 21(2): 187-
222.
Groenendijk, Jeroen, Disk de Jongh, and Martin J. B. Stokhof (eds. ) . 1987. Pro
ceedings of the 6th Amsterdam Colloquium. Amsterdam University: Institute
for Language, Logic, and Information.
Hankamer, Jorge, and Ivan A. Sag. 1976. Deep and Surface Anaphora. Linguistic
Inquiry 7:391-426.
Hindley, J. Roger, B. Lercher, and Jonathan P. Seldin. 1972. Introduction to
Combinatory Logic. New York: Cambridge University Press.
Iatridou, Sabina. 1986. An Anaphor Not Bound in its Governing Category.
Linguistic Inquiry 17(4):766-772.
Kaliski, Zbigniew. 1987. Logical Symmetry and Natural Language Reciprocals.
In Proceedings of the 1987 Debrecen Symposium on Logic and Language, ed.
Imre Ruzsa and Anna Szabolcsi. Budapest: Akademiai Kiad6.
Kayne, Richard. 1983. Connectedness. Linguistic Inquiry 14(2): 223-249.
Keenan, Edward L. 1987. Semantic Case Theory. In Proceedings of the 6th
Amsterdam Colloquium, ed. Jeroen Groenendijk, Disk de Jongh, and Martin
J. B. Stokhof, 109-133. Amsterdam University: Institute for Language, Logic,
and Information.
Keenan, Edward L. 1988. Complex Anaphors and Bind L- In Papers from the 24th
Regional Meeting of the Chicago Linguistic Society. University of Chicago.
Koster, Jan. 1987. Domains and Dynasties. Dordrecht: Foris.
Lambek, Joachim. 1958. The Mathematics of Sentence Structure. American
Mathematical Monthly 65: 154-170.
Morrill, Glyn, and Bob Carpenter. 1987. Compositionality, Implicational Logics,
and Theories of Grammar. Manuscript, Edinburgh University.
Oehrle, Richard, Emmon Bach, and Deirdre Wheeler (eds. ). 1988. Categorial
Grammars and Natural Language Structures. Dordrecht: Reidel.
Partee, Barbara. 1978. Bound Variables and Other Anaphors. In Proceedings of
TINLAP 2, ed. David L. Waltz. University of Illinois, Urbana.
Pollard, Carl. 1984. Generlized Phrse Structure Grmmars, Head Grammars
and Natural Languages. Doctoral dissertation, Stanford Universtiy.
Pollard, Carl. 1988. Categorial Grammar and Phrase Structure Grammar: An
Excursion on the Syntax-Semantics Fontier. In Categorial Grammars and
Natural Language Structures, ed. Richard Oehrle, Emmon Bach, and Deirdre
Wheeler. Dordrecht: Reidel.
tions.
Quine, Willard V. O. 1960. Variables Explained Away. Proceedings of the
American Philosophical Society 104(3) :343-347.
Reinhart, Tanya. 1983. Anaphora and Semantic Interretation. London: Croom
Helm.
Sch6nfnkel, M. 1924. Uber die Bausteine der mathematischen Logik. Mathema
tische Annalen 92:305-316.
Smullyan, Raymond. 1985. To Mock a Mockingbird. New York: Knopf.
Steedman, Mark J. 1985. Dependency and Coordination in the Grammar of
Dutch and English. Language 61: 523-568.
Steedman, Mark J. 1987. Combinatory Grammars and Parasitic Gaps. Natural
Language and Linguistic Theory 5(3):403-349.
Steedman, Mark J. 1988. Combinators and Grammars. In Categorial Grammars
and Natural Language Structures, ed. Richard Oehrle, Emmon Bach, and
Deirdre Wheeler. Dordrecht: Reidel.
Szabolcsi, Anna. 1983. ECP in Categorial Grammar. Manuscript, Max Planck
Institute for Psycho linguistics, Nijmegen.
Szabolcsi, Anna. 1984. The Possessor That Ran Away Fom Home. The Lin
guistic Review 3:89-102.
Szabo1csi, Anna. 1987. Bound Variables in Syntax (Are There Any?) . In Pro
ceedings of the 6th Amsterdam Colloquium, ed. Jeroen Groenendijk, Disk
de Jongh, and Martin J. B. Stokhof. Amsterdam University: Institute for
Language, Logic, and Information. Revised version in Semantics and Con
textual Expression, ed. Renate Bartsch et al. ( Dordrecht: Foris, 1989).
Szabo1csi, Anna. 1990. Across-the-Board Binding Meets Verb Second. In Gram
mar in Progress: GLOW Essays for Henk van Riemsdijk, ed. Joan Mascaro
and Marina Nespor, 409-418. Dordrecht: Foris.
Szabo1csi, Anna. 1991. Antecedent-Contained VP-Deletion: A Misnomer.
Manuscript, University of California at Los Angeles.
von Stechow, Arnim. 1979. Deutsche Worstellung und Montague Grammatik. In
Linear Order and Generative Theory, ed. Jirgen Meisel and Martin D. Pam,
319-488. Amsterdam: John Benjamins.
Wexler, Ken, and Yu-Chin Chien. 1988. The Acquisition of Locality Principles
in Binding Theory. GLOW Newsletter 20.
I
The Lexical Entailment Theory of
Control and the Tough-Construction
.
For well over a decade, research on a variety of fonts has indicated that the
role of the lexicon in the statement of various grammatical phenomena had
been very much underestimated in earlier work within Generative Gram
mar. A serious look at lexical items-their subcategorization possibilities,
their meaning, and the relationship among items-has shown that phenom
ena once thought to necessitate complex syntactic apparatus can instead
be accounted for by independently motivated properties of lexical items.
This is not to say that relegating some process to the lexicon necessarily
results in an overall simplifcation of the grammar-it does so, of course,
only when the process can be handled by independently motivated appara
tus. But one of the most successful applications of the lexical program has
been what I will refer to as the lexical entailment theory of control, as put
forth in, e.g. , Chierchia 1984, Dowty 1985 and others. The central claim
here is that control is nothing more than a fact about lexical meaning.
The bulk of research on the lexical entailment theory of control has
dealt with cases of the "understood subject" relation. Yet there are also
cases in which a higher is the "understood object" of some embedded
constituent; the paradigmatic case of this being the tough construction, as
exemplifed in ( 1) :
1
*This paper owes a considerable debt to David Dowty, both for his work on control and
for numerous discussions of much of the material herein. Moreover, part of this paper
is baed on our joint work in Dowty and Jacobson 1988. I would also like to thank
Bob Levine for much useful discussion. This research was supported by NSF grant
#BNS9014676.
I will not deal here with the closely related construction exemplifed in (i):
i. The rock is too heavy for me to move.
For the most part, the analysis proposed here extends straightforwardly to this construc
tion, although there are some diferences between this and the tough construction. For
Leztcc/ Mcttero. Ivan A. Sag and Anna Szabolcsi, eds.
270 / PAULINE JACOBSON
( 1 ) That rock is hard for me to move.
The purpose of this paper is to show that the lexical entailment theory
of control extends straightforwardly to this construction. In particular, I
will argue that there is no syntactic relationship between the subject of the
tough adjective and the "gap," and this then raises the question of how it
is that the subject is understood a flling the gap position. I will suggest
that this can be handled purely by lexical meaning, exactly as in the cae of
control of a "missing" subject. A thorough treatment of either the syntax
or the semantics of this construction is far beyond the scope of this paper,
but it is hoped that the analysis here can be fne-tuned in a more complete
account.
The tough construction exhibits a seemingly bafing array of proper
ties. As discussed in Chomsky 1977, the "gap" following move behaves
in many respects like a gap in a wh-construction; most notably, it can be
embedded indefnitely far down fom the tough adjective (although these
do decay more quickly than the corresponding wh-movement gaps; I have
no explanation for this):
(2) a. That rock is hard for me to try to move.
b. That house is pretty hard to imagine Bill buying.
c. That house is pretty hard to imagine Bill thinking Mary would
buy.
And, a originally pointed out in unpublished work by Maling and Zaenen,
paraitic gaps are possible here:
(3) ? John is hard to convince fiends of to invite.
But unlike run-of-the-mill wh-constructions, the apparent "binder" for
the gap is in an "A-position" rather than an "A-position," and in this
respect, the tough construction more closely resembles a control or rais
ing construction. And of these two, tough sentences are again somewhat
schizophrenic. As is typical in raising but not in run-of-the-mill control
cases, the tough adjective imposes no selectional restrictions on the sub
ject position; this property is ofen taken as a diagnostic for raising rather
than control. Along these lines, some of the adjectives in the tough clas
at least arguably assign no B-role to the subject. Indeed, there ha been
some controversy in the literature over just this point (see, e.g. , Lanik
and Fiengo 1974 and Hukari and Levine 1990), and the notion "assigning
a B-role" is notoriously difcult to pin down. Nonetheless, we can at least
note that sentences like (4a) and (4b) have the same truth conditions:
example, in the too-Adj construction it is much more difcult to get a deeply embedded
gap. Thus compare (ii) and (iii) :
ii. ? John is hard for me to imagine Mary wanting to meet.
iii. ?* John is too obnoxious for me to imagine Mary wanting to met.
I return to this very briefly in Section 4. 1.
CONTROL AND THE TOUGH-CONSTRUCTION / 271
(4) a. That rock is impossible for me to move.
b. It's impossible for me to move that rock.
Moreover (pace Lasnik and Fiengo 1974) at least certain idiom chunks can
appear here (Berman 1973), and even a case like (6) is surprisingly good
on the idiomatic reading:
(5) Careful attention was very hard to pay to that boring lecture.
(6) ?The cat would be quite easy to let out of the bag.
It is presumably these kinds of facts which frst led Postal ( 1971) to an
alyze this construction in a fashion parallel to raising, and a more recent
version of the raising analysis is developed in Bayer 1990. But in other re
spects tough adjectives behave more like run-of-the-mill control verbs and
adjectives than like raising adjectives. I postpone the demonstration of this
point until Section 2, as one of the central claims of this paper is that the
lexical entailment theory of control extends to this construction.
Of course the above discussion presupposes a sharp distinction between
raising and control; a distinction which has been maintained in most work
since Rosenbaum 1967 and which is further argued for in Jacobson ( 1990,
to appear). However, Dowty ( 1985) proposes that a raising verb or adjec
tive is nothing more than a control verb or adjective which simply fails to
assign selectional restrictions or a B-role to the controller argument posi
tion. Although I have argued in the works cited above that this analysis is
incorrect for the usual raising caes and that these instead involve syntactic
function composition, I will be claiming here that Dowty's basic analysis
is correct for the tough construction.
Thus my analysis of the tough construction combines the syntactic anal
ysis defended in detail in Fodor 1983 and Jacobson 1984 with a semantic
analysis baed on the lexical entailment theory of control. A related anal
ysis is developed in Pollard and Sag 1992. The gist of the analysis of this.
Tough adjectives subcategorize for a VP which contains a gap of the same
(or at leat similar) sort found in wh-constructions. This wa originally
proposed in Gazdar 1980, and Gazdar's analysis itself is quite similar in
spirit to the analysis of Chomsky 1977. This general kind of analysis is
also adopted in Gazdar, Klein, Pullum, and Sag 1985, but in their version
an elaborate set of conventions are developed to ensure feature matching
between the subject of the tough adjective and the gap. In Jacobson 1984
and Dowty and Jacobson 1988 it was argued that there is no such matching,
and arguments for this will be reviewed here. In other words, the subject
of the tough construction in no way syntactically "links" to the position of
the gap; the linkage is purely one of lexical entailments.
The organization of this paper, then, is a follows. Section 1 reviews the
basic premises of the lexical entailment theory of control, both to fesh out
this theory and to lay out some background assumptions used in the analy
sis here. Section 2 turns to several issues regarding the tough construction,
and in particular argues against any analysis which posits syntactic connec
tivity between the subject and the gap. Section 3 turns to the semantics,
and extends the lexical entailment theory of control to this construction.
Section 4 outlines some of the remaining problems and peculiarities with
the tough construction, and suggests some directions towards solutions.
The baic remarks will be developed using the apparatus of Categorial
Grammar, but most of the material contained here does not crucially rely
on the premises of Categorial Grammar. For this reason, I will keep the
categorial apparatus to a minimum.
1 The Lexical Entailment Theory of Control
Let us frst turn to an ordinary control case such as (7) :
(7) John wants to win.
The lexical entailment theory of control, as discussed in, e. g. , Chierchia
1984 and Dowty 1985, embodies two main premises regarding (7). The
frst is that syntactically the complement here is a bare VP with no sub
ject. The second premise is parallel: semantically this complement denotes
not a proposition, but a property (that is, it denotes the same kind of
object that VPs in general do).
2 Thus want denotes not-as conventional
wisdom has it-a relation between individuals and propositions but rather
a relation between individuals and properties. Of course more needs to
be said; some explanation must be given for the fact that John in ( 7) is
the "understood subject" of to win. Under this view, the explanation for
this resides completely in the meaning of want-the "control" relation is
merely one of entailments asociated with a lexical item. A subject control
verb, for example, is one which entails something about the individual de
noted by its subject having the property denoted by its VP complement;
call this a control entailment. Of course this is not to say that the verb
or adjective entails that the individual denoted by the subject actually has
the property denoted by the VP complement, and the particular control
entailment will vary according to the verb in question. In the cae of want,
for example, we might say roughly the following: want denotes a relation
between individuals and properties, such that for any individual 2 which
stands in the want' -relation to a property P then in the "want-world" of
2 2 has the property P.
A variety of evidence for this view is adduced in Chierchia 1984, Dowty
1985, Chierchia and Jacobson 1985 and others; we will briefy review just
some of this here. Perhaps the most compelling argument for this view is
2
These two premises are logically independent, and so all four possible combinations of
these have been proposed at various times. For example, Klein and Sag (1985) propose
an analysis of control constructions whereby the complement is syntatically a VP but
denotes a proposition; the semantics thus supplies the subject argument. And Chierchia
(n.d. ) explores an analysis whereby the complement is a syntactic S with a PRO subject
which denotes a property. We will not consider either of these analyses here.
one discussed in detail in Dowty 1985, which is that this is the null hy
pothesis. By this I mean that this view of control requires no additional
apparatus in grammatical theory which is not independently motivated.
As a consequence, extra principles would in fact be needed to rule out the
existence of a verb whose meaning contained a control entailment. Thus on
the syntactic side, nothing extra is needed to allow for a verb or adjective
which takes a VP complement; indeed, something extra is needed to rule
out such a lexical item.
3
To be sure, many theories do provide such extra
principles: in many theories it is asumed either implicitly or explicitly that
all VPs must contain overt syntactic subjects, and hence a bare VP could
not serve as complement to some verb. But this requires an extra stipu
lation, and does not follow fom any independently motivated principle of
grammar. The reasoning on the semantic side is parallel: there is no reason
not to expect a verb or adjective which denotes a relation between individ
uals and properties. Moreover, in any theory, the meaning of a lexical item
will have certain entailments associated with it, and so there is no reason
not to expect lexical meanings to have associated control entailments.
To pursue this point further, consider the standard alternative theory of
the control relation, according to which want takes a sentential complement
and denotes a relation between individuals and propositions. Such a theory
requires a co-indexing mechanism, but in addition it too must contain a
statement quite similar to the one above regarding the meaning of want.
Here want denotes some relation between individuals and propositions such
that an individual 2 stands in the want' -relation to a proposition just in
cae is true in the "want world" of 2. While the meaning postulate given
above under the lexical entailment theory is, perhaps, slightly more com
plex than this meaning postulate, the important point is that the lexical en
tailment theory makes use only of apparatus which is required in any theory.
Note moreover that under the lexical entailment view of control, we
would expect to fnd a verb or adjective taking a property-denoting com
plement, but where the meaning of that verb or adjective did not have an
asociated control entailment. And, in fact, there are such cases; the so
called "pro-arb" cases as in (8) can be viewed in j ust this way (see Chierchia
and Jacobson 1985) :
( 8) Polluting the atmosphere is harmful to the children.
Nor is there any reason not to expect some verb which takes a property
denoting complement and two individual-denoting arguments and which
entails something about its two arguments jointly having the property de
noted by the VP Complement. While such cases of split controllers are
rare, they do seem to exist, and have been discussed by Postal ( 1970),
Jackendof ( 1972), and more recently Dowty and Jacobson ( 1988):
3
Although i t is true that nothing subcategorizes for a tensed VP complement, and so
this remains to be explained under the view here.
(9) ? John argued with Mary about buying themselves a new house.
Control entailments are, of course, not the only sorts of entailments
associated with lexical items. I will assume that "selectional restrictions,"
when not purely pragmatic in nature, are also facts about lexical meaning
the meaning of fghten, for example, is such that if 2 stands in the
frighten' relation to @, then @ is a sentient being. Ladusaw and Dowty
1988 and Dowty 1988 analyze {-roles in a similar way; to say that a verb
or adjective assigns a particular {-role to one of its arguments is to say
that part of its meaning includes a particular entailment or a clustering
of entailments of relevance to that argument. Under this view of {-roles
and selectional restrictions, there is no reason to expect that every verb
or adjective will have a {-role or a selectional restriction entailment associ
ated with each of its argument positions (the (-Criterion notwithstanding).
This point is discussed in detail in Dowty 1985; I will be suggesting that
some tough adjectives may be just such a case.
Before leaving this, it should be noted that there are, of course, well
known arguments to the efect that a controlled VP contains a syntactic
subject. The familiar arguments generally center on agreement facts of the
type shown in (10):
(10) The woman persuaded the man to shave himself/
*
herself.
If the embedded VP contains a subject, then agreement between this sub
ject and the refexive reduces to normal agreement; without a subject some
other conventions are needed to ensure that the refexive agrees in features
with its "understood subject." However, it has often been proposed that
this kind of agreement is also non-syntactic and should be accounted for in
the semantics. While open questions remain concerning how best to treat
agreement in languages with "syntactic gender," the important point is
that agreement is found in a variety of cases where it would be difcult, if
not impossible, to provide a syntactic account. Space precludes a discussion
of this issue here, but see Pollard and Sag 1988 and Dowty and Jacobson
1988 for two recent discussions of this point and for two diferent although
closely related non-syntactic accounts of agreement, including agreement
in languages with syntactic gender.
2 Synta of the Tough-Construction
2. 1 Preliminaries
Before turning to the syntactic analysis, a few preliminary remarks are in
order. The frst concerns the status of the material for me in ( 11) :
( 11) That rock is hard for me to move.
Following Bresnan 1971, Lasnik and Fiengo 1974, and others, I will assume
here that this is a PP within the matrix VP. While the fundamental points
of this paper do not crucially hinge on this, it is nonetheless worth consid-
ering some evidence for it in view of recent work which claims instead that
for is a complementizer where me is in subject position (see, e. g. , Gazdar,
Klein, Pullum, and Sag 1985; Hukari and Levine 1990) . Thus frst, as dis
cussed in Chomsky 1973, the PP here can both font and extract, and it
can appear in other PP positions:
( 12) a. For whom is the rock easy to move?
b. For Bill, the rock is eay to move.
c. The rock is easy to move for Bill.
As is well-known, the sequence Comp-Subj normally cannot occur in these
positions:
(13) a. It's unlikely for John to win.
b. *For whom is it unlikely to win?
c.*For Bill, it's unlikely to win.
d.*It's unlikely to win for Bill.
Nonetheless, Hukari and Levine (1990) challenge this argument, and claim
that the PPs in ( 12) are not "displaced" PPs. Their evidence for this is
based on an observation of Samuel Bayer's (cited in Hukari and Levine
1990) which centers on the behavior of the adjective worth. Like a typical
tough adjective, worth selects a VP complement with a gap, and it can also
have a PP in the positions shown in ( 12):
(14) a. For whom is the book worth reading?
b. For Bill, the book is worth reading.
c. The book is worth reading for Bill.
However, here the PP cannot precede the complement; in view of this
Hukari and Levine conclude that there is no evidence that the PPs in ( 12)
are "displaced" PPs:
( 1 5) *The book is worth for Bill reading.
But while the behavior of worth is quite interesting, it seems to have lit
tle relevance to the question at hand. For the hypothesis that me in ( 1 1 ) is
in subject position explains neither the existence of sentences like ( 12) nor
the peculiar behavior of worth. On the other hand, the hypothesis that for
NP is in all caes a matrix PP-combined with one principle concerning lin
ear order-can account for the full range of facts. Thus assume that worth
is like other tough adjectives in taking both a PP argument and a VP with
a gap; the diference being that here the VP must be a gerund. Moreover,
there appears to simply be some principle regarding linear order to the
efect that the gerundive VP cannot be separated from the adjective while
infnitive VPs show no such restriction. Some independent confrmation for
such a principle comes fom the following contrast:
(16) a. That book was easier than pie to read.
b.*That book is worth more than you think reading.
Additional evidence for the PP hypothesis is the fact discussed in Bres
nan 1971 that dummies do not occur in this position; under all accounts of
the distribution of dummies that I am aware of this fact follows immediately
under the PP analysis but not under the subject analysis:
( 17) a. ?
*
This park would be easy for there to be a riot in.
b.* John would be eay for it to be obvious that Bill likes.4
Related to this is the fact that the tough adjective imposes selectional re
strictions on the constituent following for; roughly, it must be something
with certain abilities. This is not surprising, a the meaning of the con
struction is such that some claim is being made about the ability of this NP.
( 18) , then, is decidedly strange, and we also fnd contrats such a ( 19) :
(18) ?
*
John would be eay for that picture to frighten.
(19) a. John is easy for Mary to show that picture to.
b.* John is easy for that picture to be shown to by Mary.
Again, Hukari and Levine (1990) argue that these facts do not give evidence
regarding the syntactic structure; they show only that the NP following for
is in some sense a semantic argument of the tough adjective. They propose,
therefore, that there is a mismatch here between the semantic and the syn
tactic structure, and that there i a meaning postulate which ensures that
the lower subject i a semantic argument of the adjective. However, I know
of no way to state such a meaning postulate. Leaving aside for the moment
the status of John in the above examples, the claim that for Mary is in
subject position in ( 19a) means that easy' takes a one argument the de
notation of for Mary to show that picture to. But note that the constituent
that picture to be shown to by Mary in (19b) has exactly this same denota
tion. A meaning postulate could impose certain constraints regarding the
set of possible denotations for this argument position, but there is no way
that it could access the syntactic structure of this argument. 5
4Tough gaps are generally somewhat bad in tensed Ss, a in (i):
i. ?? John is hard for me to believe that Bill likes.
Nonetheless, ( 17b) is robustly worse than (i).
5
However, Hukari and Levine (1990) construct an ingenious argument which seems to
show that the material following for can inded be in subject position. This centers on
(i), where a paraitic gap occurs within this constituent:
i. John would be hard for friends of to admire.
If for fends of were a matrix PP, then the appearance of this gap would be quite
surprising. The reaon is that under the sort of analysis here, this PP is not within what
we might call the gap domain. If, on the other hand, a tough adjective subcategorizes
for a subject with a gap and fends of is in subject position, then it is not surprising
that the gap can occur here.
I have no explanation for the appearance of the paraitic gap in (i). Nonetheless,
the fact that the tough adjective imposes selectional restrictions on the NP following
for seems almost impossible to explain if this NP can be in subject position. And note
thateven when this NP contains a paraitic gap the same selectional restrictions hold;
The second preliminary remark concerns the nature of the gap in the
complement VP. Following the general line frst put forth in Chomsky 1977
I assume that this is much like the gap in a wh-construction. For conve
nience, I will here follow the basic line on extraction developed in Steedman
1987. Under this view, an extraction gap is nothing more than the failure
of some argument to be introduced at the expected time. In other words,
to use the usual notation of Categorial Grammar, the constituent to move
in ( 1 1 ) is a VP /NP. I will ignore here the question of whether or not slashes
should indicate the direction in which they want their argument, and thus
a category of the form A/B should be interpreted as an item which wants
a B to give an A and does not specif whether it wants the B to the left
or to the right. It should, however, be noted that a number of researchers
have suggested that extraction "gaps" are not the same as arguments which
have not been introduced in the normal way and have therefore proposed
that there is a distinction between an ordinary VP /NP and a VP with an
extraction gap (see, e. g. , Jacobson 1987b, 1989; Oehrle 1990) . Such propos
als have suggested that an extraction gap be treated more along the lines
developed within GPSG, and have thus notated a VP with an extraction
gap a VP[SLASH NPj. Should this turn out to be correct it will not afect
the basic premises of this paper; we need only say that to move in ( 11) is
a VP[SLASH NPj rather than VP /NP. I will briefy return to the status of
the gap in Section 4. 1.
The third preliminary concerns the status of Adjective Phrases and
verbs like be. Consider, for example, an ordinary sentence with an adjective
like (20) :
(20) The ball is red.
It is clear that the ball is an argument of the adjective here, even though
it is not the actual subject of the adjective but rather of the verb be.
Following the analysis laid out in Jacobson 1990, we will therefore assume
that an adjective like red is of category A/NP; in general the term "AP" can
be understood a an abbreviation for an item of category A/X (for X any
category) . What is peculiar about the category A, however, is that it rarely
surfaces; adjectives rarely directly take subjects.6 Rather, APs generally
occur as arguments of verbs like be, and any syntactic restrictions imposed
by the adjective on the subject are inherited by be. In Categorial Grammar
terms, this can easily be stated by assuming that be is of the category
(S/a) /(A/a) ; a here is a variable over categories which is contained within
the lexical specifcation of be; this means that be is marked so as to inherit
thus (ii) is quite bad:
ii. * John would be hard for pictures of to frighten.
6
Although adjectives do directly take subjects in a case like (i), discussed in Akmajian
1984:
i. With Barbara angry, you
'
ll never settle out of court.
the subject selection properties of the adjective. We will call such an item
a lexcal inheritor. I will asume further that the meaning of be is the
identity fnction on properties. 7
2. 2 The Syntactic Analysis
The syntactic analysis of ( 1 ) is now straightforward. Hard is of category
( (A/NP)/(VP/NP) ) /Pp
8
and so it combines frst with the PP for me and
then with the VP /NP to move to give the A/NP hard for me to move. This
in turn is the argument of be, and the result is the S/NP is hard for me
to move. Note that although NP occurs twice in the category of hard this
is an accident; I am claiming that hard is not a lexical inheritor, and thus
its category i not ( (A/a)/(VP/a) ) /PP. (In fact, the above category for
hard will be revised slightly below so that the subject position need not be
an NP. )
Consider now a case like (2b) :
(2) b. That house is pretty hard to imagine Bill buying.
The analysis of extraction in Steedman 1987 allows for a constituent like
to imagine Bill buying to compose to form a complex transitive verb; that
is, it too ha the category VP /NP. Consequently, this too can serve a
argument of hard. This, then, is much like the analyses in Chomsky 1977,
Gazdar 1980, Fodor 1983, Jacobson 1984, Pollard and Sag 1992 and others
whereby a tough adjective subcategorizes for a constituent with a wh-type
of gap. The fact that the gap can be embedded indefnitely far down and
the fact that parasitic gaps are possible here follows immediately in all of
these analyses.
Of course there are some well-known diferences between gaps in the
tough construction and gaps in wh-constructions; a discussion of many of
these can be found in Hukari and Levine 1987; 1991) . I do not have a
solution to all of these diferences, but I will briefy return to this one of
these in Section 4. 1.
2. 3 The Lack of Syntactic Connection Between the
Subject and the Gap
Although the baic analysis is quite familiar, there are two alternatives in
the literature. The frst analysis is similar, but proposes that there must
'Alternatively, following the general line of Chierchia ( 1984), an AP might denote
what Chierchia refers to U a nominalized property, where the meaning of be maps a
nominalized property (the individual correlate of some function) into its corresponding
function.
s
Or perhaps its category is ( (A/NP)/(VP /NP[ACC] ))/PP. This i an adaptation of the
proposal in Gadar, Klein, Pullum, and Sag 1985; they point out that restricting the
gap to accusative positions acounts for the fact (discussed originally in Schachter 1981)
that tough gaps are impossible in subject position:
i. * John is hard for me to believe likes Mary.
CONTROL AND THE TOUGH-CONSTRUCTION I 279
be category matching between the subject position and the gap position
(Gazdar, Klein, Pullum, and Sag 1985) . The motivation for this stems
from agreement facts of the type shown in ( 21)
(21 ) That man i s hard for me to persuade to shave himself/
*herself.
Suppose that agreement is indeed a matter of syntactic feature matching.
Then there must be some set of conventions to ensure that the features
on the "missing" NP following persuade match those of the refexive, and
there must thus be additional conventions to ensure that the features on
the subject match those on the missing position. But as discussed earlier,
there are proposals to treat agreement semantically, and if it is correct that
agreement is non-syntactic then no such feature-matching conventions are
needed for (21) . (For additional discussion of this point, see Jacobson
1987b and Dowty and Jacobson 1988. ) Put somewhat diferently, I will
below review some evidence which suggests that the tough construction
does not in fact impose any syntactic matching between the subject and
the gap. Given this, this construction provides further support that the
kind of "matching" in ( 21) is a semantic matter.
The second proposal derives originally fom Postal 1971 (see also Postal
and Ross 1971) , and claims that the tough construction is a species of rais
ing. A more recent version of this, based on my analysis of raising in
Jacobson 1990 is developed in Bayer 1990. In the raising analysis of Ja
cobson 1990, a raising verb like seem is (roughly) of category SiS and
concomitantly denotes a function fom propositions to propositions. This,
then, is much like the conventional analysis of raising verbs, which derives
originally fom Rosenbaum ( 1967). There is, however, no movement of the
"raised" NP. Rather, such a verb must function compose with its comple
ment; we will thus notate the category of seem as 8rS. Thus in a sentence
like John seems to be tall, seems does not take its sentential argument
in the normal way. Instead, it function composes with the S/NP to be
tall to give a new S/NP seems to be tall, and this takes John as subject.
By the defnition of function composition, any category requirements im
posed by the embedded VP will be inherited by the composed expression.
Thus in general seem will compose with a complement of category S/ X
to give an expression of category S/ X. The semantics is similar; seem'
function composes with be-tall' and the resulting VP ha the meaning
AP[seem'(be-tall' (P))J. Thus in this case the fact that John is the un
derstood subject of be-tall is not a matter of lexical entailments; it follows
directly fom the compositional semantics.
Bayer ( 1990) suggests that the tough construction should be analyzed
similarly. In his analysis, hard is (to use the present notation) of category
(ArVP)/PP. In ( 1) , it frst combines with the argument for me. At the
next step, however, it does not take an ordinary VP argument, but rather
function composes with a VP /NP to give an A/NP, and this will then occur
as argument of be.
9
Thus if this or some other version of the Raising analy
sis were correct, then the tough construction would provide no evidence for
the lexical entailment theory of control. Under any kind of Raising analysis
(by which I mean either movement or function composition) , the fact that
the rock is the "understood object" of to move in ( 1 ) would follow directly
from the compositional semantics. Note further that, as in the analysis of
Gazdar, Klein, Pullum, and Sag 1985, any kind of Raising analysis predicts
that there will be syntactic matching between the category requirements of
the gap position and the subject of the tough construction. In the function
composition analysis, for example, a phrase like hard for me will fnction
compose with a VP I X to become an AI X, and so the category which the
AP wants as subject is inherited fom the position of the gap. Moreover,
since be is a lexical inheritor (see Jacobson 1990 for motivation for this
claim) , the category restriction will again be passed up to the subject po
sition. We turn then frst to some evidence specifcally against any kind
of Raising analysis (whether via function composition or via movement) ,
and then more generally to evidence against any analysis which requires
syntactic matching between the gap and the tough subject.
2. 3. 1 Tough versus Rising
Perhaps the strongest piece of evidence against any sort of Raising anal
ysis is a point which is actually noted in Bayer 1990 and centers on the
phenomenon termed Null Complement Anaphora by Hankamer and Sag
( 1976) . By way of background, Jacobson ( 1990; to appear) demonstrates
that raising and control verbs and adjectives exhibit very diferent behavior
with respect to this process and that this follows as a consequence of the
fact that these have very diferent argument structures. Thus many control
verbs allow their VP complement to be missing:
(22) John tried.
(23) John hopes.
A diagnostic for the phenomena under discussion here is that the "miss
ing" complement is not understood as an existentially quantifed property;
rather, it is understood anaphorically and refers to some property which
is salient in the discourse context. Note, then, that all of the following are
possible, and the "missing" complement is most naturally understood as
the meaning of to climb a mountain:
l
O
(24) John tried to climb the mountain, and Bill also tried.
9This is not quite Bayer's analysis; in his analysis be is of category SjONP, and so
function composes with the A/NP. For evidence that be is not a function composer but
takes APs a arguments, see Jaobson 1990.
laThe missing complement can also be understood a any other property salient in the
discourse context; this is an instance of what Hankamer and Sag ( 1976) called deep
anaphora.
{perSUaded}
(25) I ordered John to climb the mountain, but I haven't
taught
{perSUaded}
yet ordered
Bill.
taught
I asume, then, that there is a lexical argument drop process. In the present
term, this lexical rule can be stated as in (26) (this will be generalized
later ) -
(26) A lexical item 0
1
of category X /(S/NP) maps into a homophonous
item 0
2
of category X (for X a variable over categories)
0_ = o (P)
P here is a fee variable over properties, and so its value is contextually
supplied. This is not an entirely productive lexical process; while most or
all object control verbs allow it, some subject control verbs do not:
(27) * John
{-:Pted}
to climb the mountain, and Bill
{
wants }
also
attempted
.
That this process exhibits a certain amount of lexical idiosyncrasy is not
surprising under the analysis here; like many lexical rules, the rule in (26)
is only semi-productive.
Control adjectives also generally allow this process, although one caveat
is perhaps in order. Some such adjectives appear to have pure intransitive
homophonous variants, where there is no anaphorically understood prop
erty. Take, for example, a case like (28) :
(28) Sam is happy.
Surely the most natural interpretation of this does not involve an under
stood property as complement of happy; whether happy ever allows the
Null Complement Anaphora reading is unclear to me. Nonetheless, many
control adjectives do have a reading in which a property salient in the
discourse context is understood as the argument of the adjective. This
becomes clear in a case like (29):
(29) a. John is very eager to leave, but Bill is much less eager.
b. John is willing to serve on committees but Bill is not particularly
willing.
But now let us turn to the situation with raising verbs and adjectives.
As discussed in detail in Jacobson (1990; to appear), these do not work in
this way. Thus note that none of the following are possible on the relevant
reading:
(30) * John seems to be tall, and Sam also seems.
(31) *1 expect Mary to win, but don't expect Bill.
(32) * John is certain to win, but Mary is not certain.
As mentioned above, the lexical rule in (26) does not apply with all control
verbs, and so one might think that (30)-(32) are simply accidental gaps,
just like (27) . However, the contrat between raising verbs/adjectives and
control verbs/adjectives is far too systematic to be a case of lexical idiosyn
crasy; see Jacobson ( 1990, to appear) for a detailed demonstration of this
point.
Moreover, the fact that raising verbs and adjectives do not allow the
VP complement to be missing follows immediately under either the clasical
account of raising or on the function composition analysis. To illustrate
with the latter approach: 1 claimed that (as conventional wisdom has it)
a Raising verb or adjective syntactically selects for an S rather than a VP
complement. As discussed earlier, then, seem is of category S/oS, expect is
of category (S/NP) ;oS, and cerain is of category A;oS. The raising efect
is the result of syntactic function composition of such an item with an
S/NP (i. e. , a VP) . This means that the VP "complement" is not a lexical
argument of the raising verb or adjective, and so it obviously could not
be missing due to a lexical process which removes an argument position.
A lexical argument drop process could, of course, remove the S argument
position (as in I expect) , but then there could be no "raised" NP. The
reason is that the slot for this NP in the upper clause is created only
when the verb or adjective function composes with an S/NP; the "raised"
NP slot is the slot which is inherited from the complement as a result of
function composition. A movement analysis of raising makes exactly the
same prediction; if the S argument is removed by a lexical argument drop
process there will be no embedded S fom which an NP can raise.
Thus if the tough construction involved either movement fom the po
sition of the gap or function composition with a VP /NP,we would predict
that this complement cannot be missing. But (as noted in Bayer 1990) , it
indeed can be, and in this respect a tough adjective behaves like a typical
control adjective. This is demonstrated by the following, which are good
on the anaphoric reading:
(33) a. ?This rock is impossible for me to move, and that one is impossible
for John.
b. ?This rock is hard for me to move, but that one is easy for John.
(34) a. This rock is impossible for me to move, and that one is equally
impossible.
b. This rock is hard for me to move, but that one is quite easy.
For reaons that 1 do not understand, these are most natural if the PP
complement is also missing (as in (34) ) ; but this does not afect the point
at hand. The grammaticality of (33) and (34) is completely mysterious un
der a raising analysis, but is not at all surprising under the sort of analysis
proposed hereY Under the present analysis the subject, the PP, and the
VP /NP are all arguments of the tough adjective, and so the VP /NP argu
ment position may be removed by a generalized version of the lexical rule
in (26) (the generalized version is given below in (62) ) . Note, incidentally,
that the rule in (26) should be generalized in any case since S complement
positions may also be deleted (see Grimshaw 1976) . In view of the fact that
both S and VP complement positions may be deleted, it is not surprising
to fnd that VP /NP complement positions may also be.12
2.3.2 The Lack of Connectivity Efects
Frther evidence against a Raising analysis comes fom the fact discussed
in detail in Jacobson 1982, 1984 and Dowty and Jacobson 1988 that the
category of the subject is not inherited fom the category allowed in the
position of the gap. Such facts also argue againt analyses like that of Gaz
dar, Klein, Pullum, and Sag 1985 which require feature matching between
these two positions. We will thus consider here two ways in which the cat
egory requirements in the gap position need not match the category of the
subject. 13 The frst is major category mismatch. By way of background,
1 J
Jacobson (1990) discusses two other lexical processe which apply to control verbs
and/or adjectives but not to raising verbs and/or adjectives. As it turns out, tough
adjective also do not undergo these processes, and so Bayer ( 1990) uses this in support
of the raising analysis. However, Levine ( 1991) points out that an adjective like rady
in (i) also does not undergo these lexical processes:
i. The beans are ready for me to cook.
Since the raising analysis cannot plausibly be extended to (i), the failure of tough adjec
tives to undergo these processes provides no evidence in support of the raising analysis
for the tough construction. Rather, it would appear that the failure of tough adjectives
and adjectives like rady to undergo these processes is due to the fact that these take
VP /NP complements rather than full VP complements. See also Jacobson ( 1990, fn. 29)
for discussion of this point.
1
2
Another possible diference between typical raising caes and the tOugh construction
centers on scope facts. It ha sometimes been noted that in a sentence like (i) the subject
NP ha only wide scope with respect to the adjective:
i. Every book is hard for me to read.
I am, however, not convinced that the narrow scope reading is always impossible; this
sems to emerge in (ii) and in (iii), where it is clearet in (iii) with stress on cuc[.
ii. A whole book would be hard for me to read.
iii. Every book would be hard for me to read.
If the narrow scope reading is possible here, then this does not jeopardize the account of
control to be developed in Section 3; Chierchia ( 1984) and Dowty ( 1985) demonstrate
that control verbs can be asigned a meaning such a to allow for the narrow scope
reading.
Actually, the most obvious cae of this kind of mismatch is discussed in detail in
Jacobson 1984; this concerns the fact that a subject pronoun will be nominative while
the gap is in a position normally marked accusative:
i. He1Him is hard for me to pleae.
However, both Hukari and Levine 1987 and Bayer 1990 provide mechanisms to account
for this mismatch under the asumption that matching is, in general, required, and so I
will not deal with cae mismatch here.
it is well-known that there are verbs like hope which allow S objects and
not NP objects (Rosenbaum 1967):
(35) a. r hope that Bill will win.
b.*r hope that outcome.
But the reverse case also exists; Grimshaw (1982) points out that there is
a clas of verbs like express, refect, and capture which allow NP and not
S complements despite the fact that their meaning is compatible with a
propositional object:
(36) a. This principle expresses the fact that languages are learnable.
b.*This principle expresses that languages are learnable.
(37) a. This theory captures the fact that languages are learnable.
b.*This theory captures that languages are learnable.
That there is no semantic problem here is evidenced by the following pairs
of sentences:
(38) This principle expresses something. It's that languages are learnable.
(39) This theory captures something. It's that languages are learnable.
Taken together, the two sentences in (38) (under the relevant understand
ing of it) entail that this-principle' stands in the express' relation to the
proposition that-languages-are-learnable, and this is a perfectly coher
ent discourse. Thus the problem cannot be a semantic one; rather, express
subcategorizes for an NP and not for an S object.
A similar phenomenon occurs with infnitive complements. For many
speakers, the verb dislike (in contrast to like) surprisingly takes only NP
objects:
(40) a. r like to play poker.
b.*r dislike to play poker.
Again (38) shows that the problem i not a semantic one:
(41) r dislike something. It's to play poker.
Grimshaw further observes that while a verb can choose the major
category of its object, selection for subject position in English is much
more limited. Why this is so is not clear, but it does indeed appear to be
the cae that any verb or adjective which can take an S subject can also
take an NP, provided, of course, that the NP has the appropriate kind of
meaning. For example, we have pairs like (42) and (43) :
(42) a. That Bill will leave is likely.
b. That outcome is likely.
(43) a. That Bill will leave i obvious.
b. That fact is obvious.
On the other side of the coin, any verb or adjective which takes a propo
sitional NP a subject can also take an S. Counterexamples to this claim
have occaionally been adduced; thus Pollard and Sag (1992) cite examples
like the following of an adjective which appears to take only an NP subject:
(44) a. ??That images are waterproof is incoherent.
b. That idea/claim/proposition that images are waterproof is inco-
herent.
I fnd (44a) , and Pollard and Sag's other examples, only marginally deviant.
But in any case the problem here seems to be semantic, as the following
pair of sentences shows little improvement:
(45) ??Something is incoherent. It's that images are waterproof.
Similar remarks hold for VPs in subject position; if a verb allows a VP it
also allows an NP, and if it allows an NP with the appropriate meaning it
also allows a VP.
As further evidence that there is no selection for NP versus 8 in subject
position Grimshaw points out that passives corresponding to (36b) and
(37b) are fne:
(46)
( 47)
That language is learnable is expressed by that theory.
That language is learnable is captured by that theory.
This follows under the assumption that passive is a lexical process (or a
phrasal one in the sense of Bach 1979) ) . Assuming that pasive maps a
verb (or, in Bach's terminology, a TVP) which subcategorizes for an NP
object into an intransitive verb, and assuming there i no selection for NP
versus 8 in subject position, then it follows immediately that the pasive
of 6T]I6SS can take an 8 subjectY
It is unclear how best to capture the lack of category selection in subject
position. One possibility, which is discussed in Jacobson 1987a and which
I adopt here for convenience, is to assume that there are actually no lexical
items of the form S/NP, S/8, A/NP, etc. in English. Rather, assume that
English grammar contains a super-category SUB, and that most verbs and
adjectives actually have the category S/SUB. A further rule specifes that
SUB can be NP, S, or VP. (Another possibility is that there is a lexical
redundancy rule. ) This is not to say that there is never selection for subject
position. I assume that in sentences like (48a) and (48b) the VPs have the
category S/there and Slit respectively:
(48) a. There was a riot.
b. It is obvious that John came.
14 (40b) does not have a corresponding pasive, or the pasive is at leat marginal:
i. ?*To play poker is disliked by everyone.
This, however, ha nothing to do with the fact that a VP cannot occur in subject position
but is probably due to some constraint requiring the controller to be in an argument
position. Note that the pasive of (40a) is equally marginal:
ii. ?*To play poker is liked by everyone.
I assume further that there and it cannot instantiate SUB; this means
simply that they are not allowed in the default case but must be specifcally
selected.
Now consider the following tough sentences, which are discussed in
Dowty and Jacobson 1988:
(49) That language is learnable is hard for any theory to capture/express.
(50) ?To win the lottery is hard to imagine anyone disliking.
There is indeed some awkwardness in (50) , but this does not seem to be
due to any problem with category mismatching; (51) is equally awkward:
( 51) ?To lose one's money is hard to imagine anyone liking.
The interesting point is that (49) has an S subject, but the gap is in a
position in which only NPs are allowed. Similarly, (50) contains a VP
subject and an NP gap. 15
Under both the raising analysis and the analysis which posits match
ing between the tough subject and the gap, this is quite surprising. How
ever, it follows immediately under the non-matching analysis proposed here.
Since there is no selection for major category in subject position, hard is
not actually of the category shown earlier, but is, rather, of the category
( (A/SUB) /(VP /NP) ) /PP. The gap must be an NP, but the subject position
can be anything which can instantiate SUB. Moreover, any other way (that
I can see) to capture the fact that subject selection is fee will make the
same prediction. Note further that pairs like (52) do not provide evidence
for a requirement that the category of the subject match that of the gap:
(52) a. *It's hard for me to imagine anyone hoping that outcome.
b.*That outcome is hard for me to imagine anyone hoping.
(52b) is bad simply because the hard requires that its VP complement con
tain an NP gap, while hope allows only S as complement. Thus (contra a
claim made in Hukari and Levine 1991) , the analysis here does not pre
dict that an NP subject should be allowed when the gap is in a position
permitting only Ss.
In the case of (49) , we fnd that a complement type which is barred in
the gap position is nonetheless allowed in subject position, since subject
15Larry Horn (among others) has pointed out to me that category mismatch of the kind
found in (46) is also found in topicalization:
i. ?That language is learnable, no theory can capture.
I am not sure what conclusion should be drawn from this. First, it does not seem quite
a good as (46), although the judgments are delicate. Moreover, there are other ways
in which there can be mismatch in the topicaiization construction. For example, some
speakers allow sentences like (ii):
ii. Him, I think likes Bill.
(Heavy stress is required on him a is always the case with topicalized pronouns.) The
topicalization construction, then, might simply turn out to be another construction
without connectivity.
position is generally free. But we also fnd the opposite sort of situation
with regard to dummies. Thus dummies are in general not allowed in
subject position-they are allowed there only if explicitly sanctioned by the
verb or adjective in question. Moreover, while there have been attempts
to reduce all co-occurrence restrictions centering on dummies to semantic
restrictions (e. g. , Sag 1982) , it would appear that the ability of a verb
or adjective to require a dummy in subject position and to select among
dummies is a syntactic fact. Consider, for example, the analysis of dummies
in Gazdar, Klein, Pullum, and Sag 1985 and Dowty 1985. Here the meaning
of the VPs in (53) involve vacuous A-abstraction; their meanings are as
represented in (54) :
(53) a. It is obvious that John left .
b. There was snow in New Hampshire.
(54) a. obvious-that-John-Ief' ~ AP[obvious' (left' (John' ) )]
b. was-snow-in-New-Hampshire' =
AP[in'(New-Hampshire') (snow')]
But under this kind of analysis, the subject makes no contribution to the
meaning and hence nothing in the semantics would ensure that the subject
is a dummy, let alone that it is the correct dummy. In other words, if
this general kind of analysis is correct, then there is nothing semantically
deviant about sentences like the following:
(55) a.*Bill is obvious that John left.
b. *There is obvious that John left.
Presumably, then, dummy selection is at leat in part a syntactic matter;
I assume that obvious in (53a) is of category (A/it)/S, while the verb be in
(53b) is of category ((S/there)/PP)/NP.
Given that dummies are allowed only where specifcally sanctioned by
a given verb, we predict that if a dummy is sanctioned in the position of
the gap, it nonetheless cannot occur in subject position since the subject
category is not "inherited " fom the position of the gap. And indeed this
prediction is borne out:
(56) a. It's hard for me to imagine John ever being nice.
b. ?John is hard for me to imagine ever being nice.
(57) a. It's hard for me to imagine there ever being enough snow in New
England.
b.*There is hard for me to imagine ever being enough snow in New
England.
(58) a. It's hard for me to imagine it ever being obvious that there's a
level of LF.
b.*It's hard for me to imagine ever being obvious that there's a level
of LF.
As is well-known, a gap in the position following imagine in (56b) is far
fom impeccable. Nonetheless, there is a robust contrast between this and
the sentences in (57b) and ( 58b) . This is completely unexpected both under
a Raising analysis and under an analysis positing featue matching.
The tough construction is thus quite interesting, in that it appears
to provide a good diagnostic for syntactic versus semantic co-occurrence
restrictions. Whether a given co-occurrence restriction is a syntactic or a
semantic one is not an easy question to answer in any given cae, but those
co-occurrence restrictions which are preserved in the tough construction
are just those which are plausibly analyzed a semantic, while those which
are not are just those which appear to be syntactic. Thus, as noted in
the introductory section, we fnd separable idiom chunks as in (5) and (6) ;
but many researchers have concluded on quite independent grounds that
such idiom dependencies are a semantic matter (see, e. g. , Wasow, Sag,
and Nunberg 1983) . Moreover, a shown in ( 21) , there is agreement in
person, gender, and number between the subject and the gap. But this too
ha a semantic account, and there is a good deal of independent evidence
that such matching cannot be handled in the syntax (for discussion see,
among others, Pollard and Sag 1988, Dowty and Jacobson 1988) . And while
there are indeed cases where the tough subject seems to inherit a major
category requirement fom the position of the gap, such cases can be shown
independently to be cases where the matching requirement is semantic.
Consider, for example (59), which is cited by Hukari and Levine a evidence
for the claim that there is category connectivity in the tough construction:
(59) a. I demand that you be/*are there on time.
b. That you be/
*
are there on time is hard for me to demand.
But the fact that demand requires a subjunctive complement is a semantic
one; this can be demonstrated by the fact that this co-occurrence restriction
is preserved across sentences:
(60) I' m demanding something. It's that you be/
*
are there on time.
On the other side of the coin, (49) and (50) show a mismatch in major
category; but (38), (39) and (41) demonstrate independently that here the
category selection is not a semantic phenomenon. And the distribution of
dummies is also a likely candidate for a syntactic treatment, and dummy
selection is also not preserved in the tough construction. It seems quite
reasonable to conclude, then, that there is no syntactic matching required
in this construction-the only matching requirements will be those imposed
by the fact that the gap is subject is "understood" in the gap position. We
thus turn now to the semantic analysis.
3 The Control Relation
Since I am claiming that there is no sense in which the tough subject is
in the position of the gap at some syntactic level, the question arises as to
how it is interpreted as the "missing object. " Earlier versions of this kind
of syntactic analysis within the general famework assumed here made use
of a distinguished variable in order to capture the semantic linkage. For
example, consider again ( 1) That rck is hard for me to move. Jacobson
( 1984) , following the general line in Gazdar 1980, took the meaning of the
VP/NP to be move'(P
1
) where P, is a distinguished variable. Hard for
me combines with this argument, and the associated semantics both ap
plies hard'(me') to the meaning of the complement and at the same time
A-abstracts over this variable. The result, then, is that ( is) hard for me
to move translates as AP
1
[hard' (me' ) (move' ( P
1
))] , and this ultimately
applies to the subject argument. But in addition to whatever theoretical
and/or aesthetic objections that one might level against the use of distin
guished variables, this kind of solution-and, perhaps more importantly,
any similar solution involving the binding of a variable in the complement
does not appear to be viable.
Thus consider the Null Complement Anaphora case (33b) which was
discussed earlier:
(33) b. ?This rock is hard for me to move, but that one is easy for John.
(For eae of exposition, I will demonstrate the point here with respect
to (33b) where the PP is present; the same point could be made for the
slightly more natural sentences in (34) . ) The "missing" material in the sec
ond conjunct can be understood as to move, and on this interpretation that
rck is understood as the object of move. Under the distinguished vari
able analysis, however, the semantic linkage between the subject and the
gap is accomplished only when the complement combines with the tough
adjective. If the complement position were removed by a lexical argument
drop process, then there would be no complement here, no distinguished
variable, and hence no way for that rck to bind the gap position in the
understood complemet. As mentioned above, this problem is probably
not unique to the distinguished variable analysis. Any similar attempt to
establish semantic connectivity via a variable which is somehow bound by
the subject will most likely face the same difculty, as will any account of
control which relies on a co-indexing mechanism between the subject and
the gap. For the second conjunct contains no syntactic complement, and
thus there is no position to be co-indexed with the controller.
The lexical entailment theory of control, on the other hand, extends
straightforwardly to this construction. In fact, the tough construction is
especially interesting because two arguments of the tough adjective are
controllers. The subject is understood as the object of the VP /NP, while
the PP is the understood subject of this complement. (Notice, incidentally,
that this fact is especially problematic for any theory of control which must
designate a single argument as a "controller. " ) Thus a tough adjective has
three arguments: the subject, the PP, and the VP/NP. Under the view
290 , AULlNE JACOBSON
of extraction asumed here, any VP /NP-whether complex or simple
denotes a two-place relation R. (This follows both under the account of
extraction developed in, e. g. , Steedman 1987 and in the account within
GPSG. ) For simplicity, let us take this relation to be one between two
individuals (rather than a relation between two generalized quantifers) .
The claim, then, is that each tough adjective entails something about the
individual denoted by the PP standing in the R relation to the individual
denoted by the subject.
The point is most easily illustrated with respect to an adjective like
impossible, as its meaning is less complex than most other tough adjec
tives. Impossible denotes a three-place relation holding among an individ
ual 2 (the denotation of the subject argument) , a two-place relation R,
and an individual @ (the denotation of the PP argument) . Moreover, it
contains a meaning postulate to the efect that if 2, @, and R stand in the
impossible' relation, then there is no possible world out of some suitably
defned set in which @ stands in the R relation to 2. Of course some of
the most interesting questions center on just what is the relevant set of
possible worlds; presumably these are worlds in which 2 and @ have the
same properties in this world, and other things may also have to be held
constant. The meaning of adjectives like hard and easy are much more
complex and would probably require considerably more apparatus, but the
essential point with respect to the control relation remains the same. In
each case, the claim is that the control relation arises because the adjec
tive entails something about the denotation of the PP argument and the
denotation of the subject argument standing in the relation which is the
denotation of the VP /NP.
Viewed in this way, the fact that at leat some tough adjectives impose
no selectional restrictions on the subject is not particularly problematic.
As noted earlier, selectional restrictions can be seen simply as further con
straints on an argument position which arise fom the lexical meaning of
the adjective; these adjectives thus seem to have few or no other associated
entailments of relevance to the subject position. (They do, however, im
pose selectional restrictions on the PP argument, which simply means that
they do have additional entailments associated with this position.) Similar
remarks apply to O-roles. If we follow the basic line of Ladusaw and Dowty
1988 and Dowty 1991 that a O-role is just a label for one or a clustering of
entailments associated with a given argument position, then an adjective
like impossible seems to simply lack any such entailment with respect to
the subject position.
It is, incidentally, an open question as to just what if any entailments
the tough adjectives do impose on the subject position. Some of these
adjectives (and nouns) in this class appear to have a more robust rela
tionship to the subject (see especially Lasnik and Fiengo 1974 and Hukari
and Levine 1990 for discussion) . This can be seen in cases like (61a) and
particularly (61b) where, as has ofen been noted, it appears that the dif
fculty of performing the action is being ascribed to some property of the
subject NP:
(61) a. Those new bindings are really hard for me to faten.
b. Those new bindings are a bitch for me to fasten.
Impossible, on the other hand, seems to have minimal entailments as
sociated with the subject position beyond, of course, the control entail
ment. But the key point here is simply that if indeed some tough ad
jectives assign no O-role to the subject position, then this is not a par
ticularly problematic state of afairs for the lexical entailment theory of
control.
Note that the Null Complement Anaphora cae exemplifed in (33b) is
also handled easily under this analysis:
(33) b. ?This rock is hard for me to move, but that one is easy for John.
To demonstrate this, we need to frst generalize the lexical rille given ear
lier in (26) . Two modifcations are needed. First, since 5, VP, and VP/NP
complement positions all can delete, the rille needs to be generalized ac
cordingly . Second, since we have been asuming that a tough adjective frst
combines with its PP argument and then with the VP /NP, the rule needs
to be generalized so that it is not only the outermost argument position
which can delete. 16 Therefore, we will revise (26) as (62) :
(62) A lexical item 01 of category (X / Y)/ Z (for Y = 5, VP, or VP /NP)
maps into a homophonous 02 of category X / Z (for Z any or no
category)
o AZ[o ( Z) ( Y)]
where Z is a variable of the semantic type of Z, and Y is a fee vari
able of the semantic type of Y (hence, a fee variable over proposi
tions is supplied if Y S; a free variable over properties is supplied
if Y VP; and a free variable over 2-place relations is supplied if Y
:: VP/NP) .
This means that the lexical item eaSY1 of category ( (A/SUB)/(VP /NP) ) /PP
maps into the homophonous item eaSY2 of category (A/SUB) /PP. Us
ing R as a fee variable over 2-place relations, the meaning of easY2 is
Ax[easy; (x) (R)] , and this is the item which occurs in the second conjunct
in (33b) . Since R is free, its value is contextually supplied, and it can,
therefore, pick up the relation move' which is the meaning of the VP /NP
in the frst conjunct. Hence when move' is supplied as the understood
relation, the meaning of the sentence is such that the-rock' and John'
16However, the assumption that the PP is the frst argument is one which I have made
for convenience only; it could be that a tough adjective is actually of the category
( (A/SUB)/PP)/(VP/NP) where the PP complement wraps in. If this is the cae, then
the rule could perhaps be stated so a to delete only the outermost argument position.
stand in the easy-relation to the move' relation. This in turn entails
something about John' standing in the move'-relation to the-rock', and
consequently the rock is the understood object of the missing complement
in the second conjunct. Thus the fact that the subject is understood as
the object of a non-overt complement follows quite easily under this view
of control.
More complex cases of control as in (63) are equally straightforward:
(63) John is hard for Sam to persuade to run for ofce.
Following the basic line in, e. g. , Bach 1979, let us assume that persuade
frst combines with its VP argument, then with its object argument, and
fnally with its subject. This means that persuade to run for ofce is a
complex transitive verb, and so it denotes the relation which can be repre
sented as persuade' run-for-ofce' ) . The meaning of hard is such that it
entails something about this relation holding for the pair Sam' , John' ) .
However, the persuade' relation in turn is such that if this relation holds
among two individuals 1 (the subject argument) , @ and some property P,
then something else is entailed about @ having the property P. Thus by
the transitivity of entailments, (63) entails something about John' having
the to-run-for-ofce' property. Under the sort of account of agreement
discussed in Dowty and Jacobson 1988, it is also predicted that there is a
semantic mismatch in ( 21) . There we suggested that the meaning of a VP
like to shave herself is a function which is undefned for certain objects,
and in particular it is undefned for the denotation of that man. Since (21)
ultimately entails something about the denotation of that man having the
property denoted by the lowest VP ( to shave himself) there is a semantic
mismatch.
4 Remaining Questions
The above analysis is by no means defnitive; there remain a number of pe
culiarities associated with the tough construction which are not accounted
for here. To deal with all of these is far beyond the scope of this paper,
but I will conclude with a brief look at two of these.
4.1 Island Efects
The frst remaining question concerns one of the major diferences between
the tough construction and the wh-construction: wh-constructions are is
lands, while tough sentences (often) are not. Thus, as pointed out in Chom
sky 1977, wh-constructions not only preclude other wh-gaps, but a tough
gap is also disallowed in a wh-construction:
(64) a.*Which book did you decide who to give to?
b.*That book is hard for me to decide who to give to.
However, the tough construction itself is (in some cases) not an island;
both wh-gaps and other tough gaps are possible here. Thus sentences like
(65a) are well-known; note that (65b) which contains two tough gaps is
also possibleY
(65) a. Which violin are sonatas easy to play on?
b. That violin is hard to imagine any sonata being eay to play on.
Broadly speaking, there are two main approaches to island efects. On
one, the fact that wh-constructions are islands is attributed to the fact
that these contain a gap. While the details of course vary according to
the particular theory, the idea in all of these accounts is that two gaps are
not permitted in the same domain. The other locates the island efect to
the wh-construction itself: some principle simply precludes certain types
of constituents (including relative clauses and questions) from containing
gaps. An example of this approach can be found in Gazdar, Klein, Pullum,
and Sag 1985, where island efects are the result of a feature cOoccurrence
restriction according to which a slash cannot occur on a wh-node. (This of
course assumes that "slash" is a feature; but this basic kind of account could
be translated into a theory with the view of extraction asumed here. ) Each
of these approaches has certain advantages. The frst accounts for the fact
that extraction gaps are, in general, possible in wh-construction containing
resumptive pronouns rather than gaps. The second, however, generalizes
to wh-constructions without gaps. Thus whether clauses are also islands,
even though these contain no gaps. Moreover, the italicized constituent in
(66a) is also an island, but there is little reaon to assume here that the
embedded constituent who saw the woman contains any kind of "gap"
subject position (see, e. g. , Gazdar, Klein, Pullum, and Sag 1985) :
(66) a. I wonder who saw the woman.
b.*Which woman do you wonder who saw?
Since we are assuming that a tough gap is much like a wh-gap, the island
diferences shown in (64) and (65) are problematic only under an account
17Hukari and Levine ( 1991) (follOwing earlier work within GPSG) claim that island
efects are due to a principle precluding two "slahes" of the same type on a single node.
To acount for the fact that a tough gap and a wh-gap are possible in the same domain
(a in (65a)) they propose that there are two separate features here: SLASH (for wh
gaps) and GAP (for tough gaps) . But this does not account for the ungrammaticality of
(64b) (which contains one gap of each type), and it also incorrectly predicts that (65b)
should be ungrammatical. In fact they do claim that sentences of the general form in
(65b) are bad, but I fnd (65b) a good as any other tough sentence where a gap is
deeply embedded. In support of the claim that two tough gaps are impossible, Hukari
and Levine cite the following sentence from Brame 1979:
i. *This violin would be hard for these sonata to be eay to play on.
But the problem with (i) ha nothing to do with the existence of two tough gaps, a the
following is equally bad:
ii. *It would be hard for these sonata to be eay to play on this violin.
The deviance of both of these is predicted by the fact that the pp does not denote
something with abilities, and thus the selectional restrictions on this PP position are
violated.
of extraction which attributes the island efect to the presence of two gaps
within the same domain. If instead the efect is due to the fact that a
gap cannot be located in certain kinds of constituents (those which I am
calling "wh-constituents" ) then there is little reason to expect the tough
construction to be an island.
But the situation is considerably more complex than it would appear
at frst glance. Notice that in, for example, (65a) , the tough gap is im
mediately within the VP complement of the tough adjective. This, then,
raises the following question: If a tough gap is more deeply embedded,
does the complement remain amenable to a wh- (or, another tough) gap?
Interestingly, the answer is no: if the tough gap is more deeply embedded,
then the tough construction becomes an island. This is evidenced by the
contrasts in (67) :
(67) a. Which violin is it hard to imagine John being able to play this
sonata on?
b. This sonata is hard to imagine John being able to play on that
violin.
c. *Which violin is this sonata hard to imagine John being able to
play on?
(67c) is quite strikingly ungrammatical, and is just as bad a a run-of-the
mill island violation a in (68):
(68) *Which violin do you know which sonata John is able to play on?
This suggests several things, although the remarks here should be taken
a quite tentative. First, it may well be that some distinction is needed
between gaps which occur locally within the domain in which they are
sanctioned and gaps which can be embedded indefnitely far down. One
might, for example, assume that there is a distinction between the ordi
nary VP INP of Categorial Grammar and a VP with an extraction gap ;
we will notate the latter as a VP[SLASH NP] (see, e. g. , Jacobson 1987b,
1989; Oehrle 1990). If such a distinction is correct, this will not afect the
semantic analysis developed above; a VP INP and a VP[SLASH NP] can be
given the same semantic treatment, where both denote 2-place relations.
Second, this suggests that a tough adjective selects for either a VP INP or a
VP[SLASH NP] , while a wh-word takes a its complement an S[SLASH NP].
Third, categories of the form X[SLASH Y] are "islands"-they cannot con
tain additional slash features, while categories of the form X I Y can. Thus
in (65) the complement contains a local gap, and is therefore of the category
VP/NP; hence it can have an additional SLASH feature. In (67c), o
n
the
other hand, as in a wh-extraction case, the complement is a VP[SLASH NP]
and so cannot contain another SLASH. Moreover, some confrmation for a
distinction between local and non-local gaps comes fom the fact noted in
fn. 1 that (for some speakers at least) missing object constructions of the
kind in (69) do not allow embedded gaps:
(69) a. The rock is too heavy for me to move.
b.*The rock is too heavy for me to imagine Bill moving.
It may be, then, that too heavy selects only for a VP /NP and not a
VP[SLASH NPj .
4.2 Missing PPs
In the discussion above, I have dealt only with cases in which the tough
adjective contains an overt PP complement. However, this complement can
quite easily be absent, as in (70):
(70) The rock is impossible to move.
There are two issues surrounding this construction: just how is the missing
PP argument interpreted, and how is it that this understood argument
is also the understood subject of to move? The second question can be
answered straightforwardly under the lexical entailment theory of control.
As a point of departure, take the analysis proposed in Chierchia 1984.
Chierchia suggests that the lexical item impossible in (70) is derived from
the corresponding item which deletes the PP argument slot as follows: (I
am taking slight liberties here with Chierchia's actual proposal, but none
which afect the point at hand) :
(71) a
1
of category ( (A/SUB)/(VP/NP) ) /PP maps into an item a2 of
category (A/SUB) /(VP/NP)
a = A
R
[Ay[3x[
a
(x) (
R
) (y)lll
What this means is that if some individual @ (the denotation of the sub
ject argument) stands in the a; relation to the relation R denoted by the
VP /NP complement, then there is some individual 2 such that @, Z, and R
stand in the a relation. Now this is not quite right, since the missing PP
complement is not understood as existentially quantifed. The most natu
ral reading for (70) is one where the missing PP is understood generically,
or has an interpretation much like the NP one. Thus let us reformulate the
semantic part of Chierchia's proposal as in (72) :
(72) a
;
= A
R
[Ay[a

(
on
e' )(
R
) ( y)]]
The key point is that it follows that one' is also the understood subject
of the VP; this basic point is discussed in detail in Chierchia 1984. Take
the case where this rule applies to a lexical item like impossible. The
lexical item derived by the rule (71-72) is such that if a sentence like
(70) is true, then the-rock' and one' stand in what we will now call the
impossible-relation to move'. Note, though, that impossible is simply
the relation discussed earlier, and it therefore entails something about one'
and the-rock' standing in the move'-relation. It follows, then, that the
understood argument of impossible is also the understood subject of the
complement.
But this is not a fll account, for the missing argument of impossi
ble need not be understood a one'. As detailed in Kimball 1971 , it
can also be understood as the meaning of some NP in a higher clause,
a in ( 73):
(73) a. John thought that the rock would be impossible to move.
b. John thought that the movie would be hard to convince himself
to like.
Just how to account for this second reading is unclear. It is tempt
ing to posit another lexical rule here by which a free variable is sup
plied in the relevant argument position, where the value of that vari
able will be contextually supplied. However, this predicts that the value
for this variable can be any individual salient in the discourse. Yet it
seems difcult, if not impossible, to get the relevant reading in a case
like (74) :
(74) John is always procrastinating. ?
*
This last homework assignment
was particularly difcult to convince himself to do on time.
Compare (74) to ( 75) , where the PP i overt:
(75) John is always procratinating. This lat homework asignment was
particularly difcult for him to convince himself to do on time.
Thus the problem of how to establish the control relation between John
and the "missing PP" in (73) remains an unsolved problem.
There are, then, several peculiarities of the tough construction which
remain unexplained. Nonetheless, I hope to have established two main
points. First, there is no syntactic linkage between the tough subject and
the gap. Second, the semantic linkage can be handled straightforwardly
under the lexical entailment theory of control. Thus this construction pro
vides additional evidence for the hypothesis that control is only a matter
of entailments associated with lexical items.
References
Akmajian, Adrian. 1984. Sentence Types and the Form-Function Fit. Natural
Language and Linguistic Theory 2: 1-23.
Bach, Emmon. 1979. Control in Montague Grammar. Linguistic Inquiry 10: 515-
532.
Bayer, Samuel. 1990. Tough Movement Function Composition. In Proceedings
of the 9th Annual West Coast Conference on Formal Linguistics, ed. Peter
Sells. Stanford: CSLI Publications.
Berman, Arlene. 1973. Adjectives and Adjective Complement Constructions in
English. Doctoral dissertation, Harvard University.
Brame, Michael. 1979. Essays Toward Realistic Syntax. Seattle: Noit Amrofer.
Bresnan, Joan. 1971. Sentence Stress and Syntactic Tansformation. Language
47:257-281.
Chierchia, Gennaro. 1984. Topics in the Syntax and Semantics of Infnitives and
Gerunds. Doctoral dissertation, University of Massachusetts. Distributed by
the Graduate Linguistics Student Association, University of Massachusetts.
Chierchia, Gennaro. To appear. Anaphora and Attitudes De Se. In Language in
Action, ed. Renate Bartsch et al. Dordrecht: Foris.
Chierchia, Gennaro, and Pauline Jacobson. 1985. Local and Long Distance
Control. In Proceedings of NELS 16, 57-74. Amherst: Univeristy of Mas
sachusetts, GLSA.
Chomsky, Noam. 1973. Conditions on Tansformations. In A Festschrif for
Morris Halle, ed. Stephen R. Anderson and Paul Kiparsky, 232-286. New
York: Holt, Rinehart and Winston.
Chomsky, Noam. 1977. On Wh-movement. In Formal Syntax, ed. Peter Culicover,
Thoma Wasow, and Adrian Akmajian. New York: Academic Press.
Dowty, David. 1985. On Recent Analyses of the Semantics of Control. Linguistics
and Philosophy 8:291-331.
67(3):547-619.
Dowty, David, and Pauline Jacobson. 1988. Agreement as a Semantic Phe
nomenon. In Proceedings of the 5th Annual Eastern States Conference on
Linguistics, ed. Joyce Powers and Ken de Jong. Columbus: Ohio State Uni
versity.
Fodor, Janet D. 1983. Phrase Structure Parsing and the Island Constraints.
Linguistics and Philosophy 6: 163-223.
Gazdar, Gerald. 1980. English as a Context-fee Language. Manuscript, Univer
sity of Sussex.
Gazdar, Gerald, Ewan Klein, Geofrey K. Pullum, and Ivan A. Sag. 1985. Gen
eralized Phrase Structure Grammar. Cambridge, Mass.: Harvard University
Press, and Oxford: Basil Blackwell.
Grimshaw, Jane. 1976. Wh-Constructions and the Theory of Grammar. Doctoral
dissertation, University of Massachusetts.
Grimshaw, Jane. 1982. Grammatical Relations and Subcategorization. In Sub
jects and Other Subjects: Proceedings of the Harvard Conference on Gram
matical Relations, ed. Annie Zaenen, 35-55. Bloomington: Indiana University
Linguistics Club.
Hankamer, Jorge, and Ivan A. Sag. 1976. Deep and Surface Anaphora. Linguistic
Inquiry 7:391-426.
Hukari, Thomas, and Robert Levine. 1987. Rethinking Connectivity in Un
bounded Dependency Constructions. In Proceedings of the 6th Annual West
Coast Conference on Formal Linguistics, ed. Megan Crowhurst, 91-102. Stan
ford University: Stanford Linguistics Association.
Hukari, Thomas, and Robert Levine. 1990. The Complement Structure of Tough
Constructions. In Proceedings of the 7th Annual Eastern States Conference
on Linguistics. Columbus: Ohio State University.
Hukari, Thomas, and Robert Levine. 1991. On the Disunity of Unbounded
Dependency Constructions. Natural Language and Linguistic Theory 9( 1) :97-
144.
298 / PAUL1NE JACOBSON
Jackendof, Ray. 1972. Semantic Interretation in Genertive Grmmar. Cam
bridge, Mass. : MIT Press.
Jacobson, Pauline. 1982. Evidence for Gaps. In The Nature of Syntactic Repre
sentation, ed. Pauline Jacobson and Geofrey K. Pullum, 187-228. Dordrecht:
Reidel.
Jacobson, Pauline. 1984. Connectivity in Phrase Structure Grammar. Naturl
Language and Linguistic Theory 1: 535-581.
Jacobson, Pauline. 1987a. Generalized Phrase Structure Grammar (review arti
cle). Linguistics and Philosophy 10: 389-426.
Jacobson, Pauline. 1987b. Phrase Structure, Grammatical Relations and Dis
continuous Constituents. In Syntax and Semantics 20: Discontinuous Con
stituency, ed. Geofrey Huck and Almerindo Ojeda, 27-69. New York: Aca
demic Press.
Jacobson, Pauline. 1989. A(nother) Categorial Grammar Account of Extrac
tion. Paper presented at the Conference on Categorial Grammar at the LSA
Summer Linguistic Institute, Tucson.
Jacobson, Pauline. 1990. Raising as Function Composition. Linguistics and
Philosophy 13:423-475.
Jacobson, Pauline. To appear. Raising Without Movement. In Control and
Grammatical Theory, ed. Richard Larson et al. Dordrecht: Kluwer.
Kimball, John. 1971. Super-Equi-NP-Deletion as Dative Deletion. In Papers from
the 7th Regional Meeting of the Chicago Linguistic Society, ed. D. Adams et al.
University of Chicago.
Klein, Ewan, and Ivan A. Sag. 1985. Type-driven Tanslation. Linguistics and
Philosophy 8: 163-202.
Ladusaw, William, and David Dowty. 1988. Toward a Nongrammatical Account
of Thematic Roles. In Syntax and Semantics: Thematic Relations, ed. Wendy
Wilkins, 61-73. New York: Academic Press.
Lanik, Howard, and Robert Fiengo. 1974. Complement Object Deletion. Lin
guistic Inquiry 5:535-571.
Levine, Robert. 1991. On Missing Object Constructions and the Function
composition Analysis of Raising. Manuscript, Ohio State University.
Oehrle, Richard. 1990. Categorial Frameworks, Coordination, and Extraction. In
Proceedings of the 9th Annual West Coast Conference on Formal Linguistics,
ed. Peter Sells. Stanford: CSLI Publications.
tions.
Pollard, Carl, and Ivan A. Sag. 1988. An Information-based Theory of Agreement.
In Proceedings of the Parasession on Agreement, ed. Diane Brentari, Gary
Larson, and Lynn McCleod, 236-257. Chicago: Chicago Linguistic Society.
Pollard, Carl, and Ivan A. Sag. 1992. Head-Driven Phrase Structure Grmmar.
Chicago: University of Chicago Press, and Stanford: CSLI Publications.
Postal, Paul. 1970. On Coreferential Complement Subject Deletion. Linguistic
Inquiry 1 : 439-500.
Postal, Paul. 1971 . Crss- Over Phenomena. New York: Holt, Rinehart, and
Winston.
Postal, Paul, and John R. Ross. 1971 . Tough Movement Si, Tough Deletion No!
Rosenbaum, Peter. 1967. The Grammar of English Predicate Complement Con
structions. Cambridge, Mass. : MIT Press.
Sag, Ivan A. 1982. A Semantic Theory of NP-movement Dependencies. In The
Nature of Syntactic Representation, ed. Pauline Jacobson and Geofrey K.
Pullum, 427-466. Dordrecht: Reidel.
Schachter, Paul. 1981. Lovely to Look At. Linguistic Analysis 8:431-448.
Steedman, Mark J. 1987. Combinatory Grammars and Parasitic Gaps. Natural
Wasow, Thomas, Ivan A. Sag, and Geofrey Nunberg. 1983. Idioms: An Interim
Report. In Proceedings of the XIIIth Interational Congress of Linguists, ed.
Shiro Hattori and Kazuko Inoue, 102-115. Tokyo: CIPL.
II
A Lexical Analysis of Icelandic Case
IVAN SAG, LAURI KARTTUNEN, AND JEFFREY GOLDBERG
1 Introduction
The facts of performance have much to tell us about our knowledge of lan
guage. In performance-that is, in real-time language use-our knowledge
of language is deployed in wondrous harmony with knowledge of the world,
knowledge of the utterance situation, knowledge of our fellow interlocutors,
and so forth. Given the extent to which interpretation depends on such fac
tors, i. e. the extent to which linguistic knowledge alone fails to determine
interpretation, it remains quite mysterious how we humans are able to use
language to communicate determinate messages accurately and efciently.
One fndamental problem that will occupy the study of language in the
decades to come is one we will refer to here a the Resolution Problem: the
problem of how diverse kinds of knowledge-linguistic and otherwise-are
efciently integrated in real-time language processing. Fom a linguistic
perspective, taking this problem seriously means that we must ask our
linguistic descriptions to do more for us than they have in the past. It is not
enough that grammars express linguistic generalizations. If our grammars
are ultimately to help us address the Resolution Problem (and explain other
properties of performance) , then they should be designed in accordance
with certain criteria that emerge fom basic observations about the nature
of performance. Some of the relevant observations about human language
processing are these:
( 1) Partiality: In real-time, partial descriptions of linguistic structure
are accessed incrementally.
Flexibility: The partial information from these descriptions may be
We would like to thank Avery Andrews, Joan Maling, and Annie Zanen for their help.
This research wa supported in part by a gift to Stanford University from The System
Development Foundation, and in part by a grant from the National Science Foundation
(BNS-8511687).
Leztccl Mctter Ivan A. Sag and Anna Szabo1csi, eds.
302 / IVAN SAG, LAURI KARTTUNEN, AND JEFFREY GOLDBERG
fexibly integrated with non-linguistic information in real-time pro
cessing.
Order-Independence: Linguistic information is not consulted in any
fxed order; there may be task-specifc variations in the order in
which relevant information is accessed.
Reversibility: In general, linguistic knowledge is consulted equally
well by diverse kinds of processors, e. g. , those that perform compre
hension and those that perform production.
While such observations about language use could in theory be explained
in a variety of ways, and perhaps made consistent with a number of ar
chitectures for the design of grammar, it is evident that certain existing
grammatical architectures-and not others-begin to make sense of these
facts of performance.
We are in a better position to begin to explain the observations in ( 1)
i f we render linguistic knowledge declaratively, i. e. , as a static system of
constraints, as is the cae, for example, within work in the traditions of
phrase structure and categorial grammar. These systems of grammar em
ploy no destructive rules whose order of application with respect to other
rules must be crucially fxed. Hence these systems of grammar are appro
priately unbiased with respect to comprehension and production. Lexical
entries and grammar rules in phrae structure and categorial systems can
be interpretd a nothing more than constraints on structures that provide
reliable partial information about linguistic structures, rather than infor
mation that is true of a structure only at some point in its derivation. And
because the information involved is highly restricted, the data structures
to be computed in incremental processing of substrings are highly con
strained. It thus becomes possible to formulate precise models of process
ing that attempt to reflect the observed properties of partiality, flexibility,
and order-independence.
By contrast, transformational models including GB models where
' Move a' is interpreted literally) provide little help in explaining the per
formance properties listed above. Rules perform mappings fom total rep
resentations of a sentence at a particular "level of representation" to other
total representations, and these mappings have typically involved deletions,
permutations, insertions, and the like. Moreover such mappings are not in
general reversible; nor are their computational properties in general well
understood. There is no clear way to model fexible, integrative process
ing of partial transformational descriptions, though one can imagine that
a performance model based on transformational grammars would require
some mechanism for mapping transformational descriptions onto some set
of constraints that either are otiose or else closely resemble those posited
within phrase structure and categorial theories.
These considerations strongly suggest that constraint-based linguistic
A LEXICAL ANALYSIS OF ICELANDIC CASE / 303
descriptions like phrae structure grammars and categorial grammars hold
greater promise for reconciling the theory of linguistic knowledge (or "com
petence" ) with the facts of language processing (performance) . However,
the fact remains that many complex and well-studied sets of language data
constitute a serious challenge for purely constraint-based, non-derivational
linguistic theories such a these. 1 In this paper, we discuss one of these chal
lenges, namely the problem of "quirky" case in Icelandic, which has been
analyzed by Thrainsson ( 1979) , Andrews ( 1976, 1982, 1990) and Yip et al.
(1987), among others. The data we discuss appear to warrant an analysis
that employs transformational rules, constraining equations (Kaplan and
Bresnan 1982), or some other non-monotonic syntactic device. The analysis
we propose, however, functions in a purely monotonic, order-independent,
process-neutral fashion, and hence constitutes a performance-compatible
description of this troublesome set of facts. This description involves only
lexical entries and schematic phrase structure rules, which together spec
if a set of constraints that the well-formed structures of Icelandic must
satisf.
Constraint-baed analyses in general, and ours in particular, rely heav
ily on lexical information. The complexity of this information raises new
questions about how lexical information is structured and how lexical
generalizations are expressed. In light of these concerns, we also dis
cuss the implications of our analysis for the organization of the Icelandic
lexicon.
2 The Problem of "Quirky" Case
Icelandic verbs, in the default circumstance, take subjects whose cae must
be nominative:
(2) Drengurinn kyssti stWkuna bilnum
the-boy.NOM kissed the-girl.ACC in the-car.DAT
'The boy kissed the girl in the car. '
But Icelandic ha a number of verbs that assign quirky (i. e. , non
nominative) case to their subjects. In (3)-(6) we see examples with ac
cusative subjects.
2
(3) Mig langar ao fara
me.ACC longs to to-go
'I long to go to Iceland. '
til
i
slands.
to(LOC) Iceland
(4) Hana dreymdi ur haf.
she.ACC dreamed about the-sea
'She dreamed about the sea.'
For some discussion, see Kay 1986, Sag et al. 1985, and Sag 1991.
2
All of the data we will discuss here are taken from Andrews 1982.
304 / IVAN SAG, LAURl KARTTUNEN, AND JEFFREY GOLDBERG
(5) Mig velgir vic setningafroi.
me.ACC is-nauseated at syntax
' Syntax turns my stomach. '
( 6) Drengina vantar mat.
the-boys.ACC lack food.ACC
'The boys lack food. '
In (7)-( 10) we see examples with dative subjects.
(7)
(8)
Honum mrltist vel i kirkjunni.
He. DAT spoke well in the-church
' He spoke well in church.'
Henni akotnaoist bill.
her.DAT lucked-onto car. NOM
' She got possession of a car by luck. '
(9) StUlkunni svelgdist a sUpunni.
girl.DAT mis-swallowed on the-soup
'The girl swallowed the soup wrong. '
( 10) Barninu batnaoi veikin.
the-child. DAT recovered-from the-diseae.NOM
'The child recovered fom the diseae. '
And we also fnd verbs that take genitive subjects a shown i n ( 1 1 )-(12).
(11) Verkjanna grtir ekki.
the-pains.GEN is-noticeable not
'The pains are not noticeable'
( 12) Konungs var pamgao von.
the-king.GEN wa thither expectation. NOM
'The king wa expected there. '
I n all three instances, the quirky case i s the only possible one for the subject
in question.
Raising verbs are what makes Icelandic cae assignment intriguing. If
the infnitival complement's verbal head does not asign quirky cae, then
the "raised" subject of a fnite form of a subject raising verb must be
nominative in cae:
( 13) Hann viroist elska hana.
he.NOM seems to-love her.ACC
'He seems to love her. '
This i s what one might expect. But i n examples like the following, the
raised subject takes the quirky case assigned by the verb that is the head
of the infnitival complement:
(14) Hana viroist vanta peninga.
her.ACC seems to-lack money
' She seems to lack money. '
( 15) Barninu viroist hafa batnao veikin.
the-child. DAT seems to-have recovered-from the-disease
'The child seems to have recovered from the disease. '
(16) Verkjanna viroist ekki grta.
the-pains. GEN seems not to-be-noticeable
'The pains don't seem to be noticeable.'
Note that in ( 15) , the quirky case NP is separated from its case assigner
by two subject raising verbs ( viroist and hala) .
Similar facts may be observed i n the case of object raising constructions.
If the head of the infnitival complement is a quirky case assigner, then the
raised object NP must bear the appropriate quirky case:
( 17) Hann telur mig vanta peninga.
he.NOM believes me.ACC to-lack money
'He believes that I lack money. '
( 18) Hann telur barninu hafa batnao veikin.
he believes the-child.DAT to-have recovered-fom the disease
'He believes the child to have recovered from the disease. '
( 19) Hann telur verkjanna ekki grta.
he believes the-pains.GEN not to-be-noticeable
' He believes the pains to be no noticeable. '
But if the infnitival complement's verbal head is not a quirky case assigner,
then the raised NP must be accusative in cae a shown in (20).
(20) peir telja Mariu hafa skrifao ritgeroina.
they believe Mary. ACC to-have written the-thesis
' They believe Mary to have written her thesis. '
In short, Icelandic exhibits what appears to be a syntactic non-mono
tonic case assignment system. It is one that could be described in trans
formational terms by assuming ( 1 ) that quirky verbs assign case to their
subjects prior to the application of raising rules (or, if one prefers, ' Move
a' ) , (2) that fnite subject raising verbs optionally assign nominative case
to their subjects, (3) that object raising verbs optionally assign accusative
case to their objects, and that (4) all NP's must be assigned case somehow
(the "Case Filter" ) . 3
But the derivational (movement transformations) and non-monotonic
(fltering conditions) character of such an analyses is inconsistent with the
performance-based design considerations we discussed at the outset. Per
haps there is a version of this analysis that can be made to provide answers
to the puzzles posed by performance, but it is certainly not any analysis
based on the Cae Filter that has been discussed in the literature we are
3
Chomsky 1981 and elsewhere.
familiar with. All such analyses presume derivational mappings between
fully specifed syntactic representations, with the Case Filter functioning
a a non-monotonic constraint that must be checked only afer all trans
formational rules have had an opportunity to apply.
In another infuential analysis of Icelandic cae dependencies, Yip et al.
( 1987) explain an impressive array of complex facts by adapting certain
idea fom non-linear phonology (Goldsmith 1976, 1990) . The essence of
their analysis is a factoring of the case assignment patterns into the inter
action between a basic nominative-accusative tier and a restricted set of
lexical and semantically based exceptions. Their analysis also relies on a
principle fltering derivations where an NP receives no cae. Moreover, their
approach clearly employs cyclic application of syntactic rules (p. 241f.)
whose efect is to destructively modify the initial (lowest cycle) cae value
asigned to a given NP. Although their analysis expresses generalizations
about Icelandic cae not captured by previous approaches, it fails to do
so in a way that is compatible with the declarative, constraint-satisfaction
architecture for the design of grammar.
There are alternatives that one can imagine. Following the style of
Neidle's (1982) LFG analysis of Russian, one might employ constraining
equations (Kaplan and Bresnan 1982) to ensure that case is asigned to
every NP in a syntactic structure. Like the transformational analysis, this
approach would have quirky verbs obligatorily assign cae to their sub
jects, while raising verbs asigned the appropriate default case optionally.
A raised NP is identifed with the subject of the appropriate complement,
hence if nouns are lexically specifed with constraining equations requiring
that the appropriate cae must get asigned to them, then this analysis
accomplishes the same efect a the transformational one, without trans
formational rules.
But such an analysis is also potentially objectionable on performance
grounds, for the constraint that it relies on functions in a non-monotonic
fashion. Constraining equations are defned as conditions that must be
met by functional descriptions, that is, they are not just constraints on
functional structures. Rather, they are crucially defned as flters on the
process that maps constituent structures to functional structures.4 Perhaps
models of language processing could be devised to deal with limited non
monotonicity of this sort, S but this remains an open issue. 6
The question of how to reconcile limited non-monotonicity with the
facts of performance is one that need not arise. Appearances to the contrary
4 An analysis essentially equivalent to this can be formulated using Kay's ANY variable,
one which must receive a value in a complete linguistic description.
5
So, for example, Johnson's (1988) discussion of completeness and consistency con
straints in LFG.
6We will not discuss here the alternate LFG analysis of Icelandic cae proposed by
Andrews ( 1990).
notwithstanding, the generalizations of Icelandic quirky case asignment
can be elegantly expressed within an analysis that involves no destructive
operations, no non-monotonic constraints, and no transformational deriva
tions. It is to this analysis that we turn directly.
3 A Lexical Solution
In any adequate lexical theory, verbs must be provided with some mecha
nism for selecting for their various dependents, including subjects, objects,
oblique NP's or PP's, and complements of various sorts. We will asume,
following a long-standing tradition of research in Categorial Grammar,7
that the syntactic arguments of a verb are hierarchically arranged accord
ing to a more or less traditional notion of "obliqueness" . This notion ha
also played a role i Relational Grammar analyses of grammatical relations,
in Keenan and Comrie's (1977) cross-linguistic work on relative clause for
mation, and in the LFG theory of default controller assignment (Bresnan
1982) .
To implement this idea, we will asume a simplifed version of the notion
of SUBCAT list, developed independently by Pollard and Sag within Head
Driven Phrae Structure Grammar (Pollard and Sag 1987, 1992) and by
researchers developing the PATR-I I system at SRI International (Shieber
et a1. 1983) . The order of elements on the SUBCAT list corresponds not to
surface order (Le. , the linear order in which the phonological realizations
of the elements occur), but rather to the order of obliqueness, with more
oblique elements appearing later than (i. e. , to the right of) less oblique
elements. Following long-standing tradition, PP (or VP or S) complements
are treated a more oblique than direct objects when both occur, and direct
objects in turn are more oblique than subjects. Thus the SUBCAT list for
an intransitive verb contains exactly one NP, corresponding to the verb's
subject; and the SUBCAT list for a (strict) transitive verb contains exactly
two NP signs, the frst corresponding to the verb's subject and the second
to its direct object. Verbs of higher valence have appropriately specifed
longer SUBCAT lists.
Thus the lexical representation of each verb specifes partial information
about the number and kind of syntactic dependents that that verb can
combine with. Each element on the SUBCAT list is a partial specifcation
of information about a particular syntactic dependent of the verb, and
the actual complements the verb combines with must be consistent with
the partial information the verb selects for. Verbs may thus select (for
any of their dependents) : grammatical category, agreement features, cae,
valence, and so forth.
A typical fnite, transitive verb form like kyssti ('kissed' ) would be spec
ifed a follows:
7See, for example, Dowty 1982a,b.
( 21)
[
.
nom
np
1
3;d
smg
Fom this lexical specifcation, it follows that kyssti can combine only with
an accusative object and a 3rd person singular subject
And we have already introduced sufcient machinery to provide an anal
ysis of the fnite forms of quirky subject verbs in Icelandic. These forms
simply specif diferent information about their subjects' cae, a illustrated
in (22)-(24) .
( 22) langar:
(23) askotnaoist:
(24) grtir:

np ]
acc .
np ]
dat
But this simple method of imposing case restrictions on verbal dependents
is insufcient for the raising constructions discussed earlier, as we will now
show.
The most straightforward treatment of raising constructions in lexically
baed theories of grammar (e. g. , Bresnan 1982, Pollard and Sag 1987, 1992)
involves identifing the raised element with the unexpressed subject of the
complement. Under the assumptions we are making here, com
plements are phrases whose value contains exactly one element,
corresponding to the unexpressed subject of the "downstairs" verb. Rais
ing is thus accomplished by positing lists like the following for
raising verbs.
(25) Simple subject raising verb:
np| ,
. 1
Simple object raising verb:

np
[

np| , .
vp
l )
inf
, J
)
BWe use the attribute CAT here U an expository convenience. In the theory of grammar
developed in Pollard and Sag 1987, 1992, category information is analyzed partially in
terms of values for the feature HEAD. In virtue of the Head Feature Principle, category
information is projected from verbs to VPs and Ss, from nouns to 'N's and NPs, and so
forth. Nothing hinges on this in the present discussion.
A LEXICAL ANALYSIS OF ICELANDIC CASE ] 309
Here the tags (e.g., I) indicate identity (or structure-sharing) between the
VP complement's unexpressed subject (the one element on its SUBCAT
list) and the appropriate other dependent of the verb.
This lexical approach to raising phenomena would appear to give the
right result for complements headed by quirky verbs. Assuming the
SUBCAT specifcations given above hold for both fnite and non-fnite verbal
fonns, the infnitive form of an accusative subject verb (e.g., vanta) would
give rise to a VP specifed as [SUBCAT (NP[acc] ) ] . The subject raising verb
viroist, specifed a in (25), would then combine with the VP vanta peninga
to form a VP that is also specifed a [SUBCAT (NP[acc] ) ] , a shown in (26) .
(26) VP[SUBCAT ([) ]
.
V
[SUBCAT ( [, ) ]
I
VP[SUBCAT ([)]
6
viroist vanta peninga
And this VP can combine only with an accusative subject to form such
sentences a Hana.ACC viroist vanta peninga.
Similarly, the lexical analysis of raising correctly predicts that quirky
case selections will be inherited in object raising:
(27) VP[SUBCAT { ilNP [nom] ) ]
V
[SUBCAT ( il, , I) ]
|
NP[gen] !VP[SUBCAT () ]
1 6
telur verkjanna
ekki grta
The object of telur ('believe' ) must appear in whatever case the VP com
plement requires of its (unexpressed) subject.
The problem that now arises concerns non-quirky verbs. If we assume
that non-fnite fonns of non-quirky verbs select for nominative subjects
(like their fnite counterparts) , then the lexical raising analysis works fne
for simple subject raising cases, but incorrectly predicts that the object
of telur will also be nominative when the VP complement is headed by a
non-quirky verb. As we saw earlier, such raised objects must be accusative
in case. If, on the other hand, we were to assume that the non-fnite forms
of non-quirky verbs selected for accusative case subjects, then we would
have no account of the fact that in subject raising, the matrix subject
(e.g., of viroist must be nominative if the VP complement is headed by a
non-quirky verb. Nor can we allow such verbs to make disjunctive cae
selections (nominative 0I accusative) for their subjects, because this is not
sufcient to block the incorrect choice of cae in either raising construction.
The problem here is that the notion of default case assignment is not
being expressed. The default case for subjects of fnite verb forms is nom-
inative; the default case for direct objects is accusative. The non-quirky
verbs obey these defaults; the quirky verbs do not. This is the fundamental
intuition of the solution we now propose.
Let us assume that there are two distinct features that nominals bear:
CASE and DCASE, where CASE, as before, corresponds to the actual cae
that the NP in question will bear and DCASE corresponds to the default
case that a nominal bears in a given syntactic position. Nominals have no
inherent DCASE value, but verbs may subcategorize for subjects or objects
that have particular CASE values or particular DCASE values. The difer
ence between quirky and non-quirky verbs can now be analyzed a follows:
(28) Non-quirky verbs require that the DCASE and CASE values of their
subjects be identifed (structure-shared) .
Quirky verbs impose no such identity, and select a particular CASE
value for their subject.
Thus the SUBCAT lists for all forms of quirky verbs are just a before
(see (19) above). They specif CASE values for their subjects, but say
nothing about the subjects' DCASE values. The lexical entries for non
quirky verbs, on the other hand, will all contain the following information:
(29)
( [CAT
SUBCAT CASE
DCASE
Note that no particular value for either CASE or DCASE is assigned to the
subject in (29) . The only information specifed i that the two features
have the same value, however that may be determined.
We may now articulate the rest of our analysis and illutrate its conse
quences. First, it is a general fct about realized subjects in Icelandic (in
fnite and non-fnite contexts alike), that their default cae value i nomi
native. This information is presumably to be asociated with the grammar
rule that introduces subjects. The fact that a realized subject9 must bear
the specifcation [DCASE nom) has important consequences for non-quirky
verbs. Since the CASE and DCASE values of the realized subject of a non
quirky verb are identifed, it follows that the actual cae (determined by
the CASE value) of such a subject must be nominative. This is true not
only for the subjects of fnite forms of non-quirky verbs, as we saw earlier
(e.g., (2) ) , but also for examples like (30), where the subject of a non-fnite
form is realized.
(30) Mer viroist hun elska hann.
me. DAT seems she. NOM to-love him.ACC
'It seems to me that she loves him.'
9Given the range of modifers that exhibit default nominative cae (Andrews 1990),
this analysis may have interesting consequences for the treatment of the unexpressed
subjects in such modifers.
For quirky verbs, on the other hand, the constraint that realized subjects
are specifed as [DCASE nom] has no particular consequence. The actual
case of the subject (determined by the CASE value) is assigned by the
individual quirky verb, and nothing requires the CASE and DCASE values
of the subject to be the same.
Second, the SUBCAT specifcations for object raising verbs will no longer
be as in (25) above, but rather will be as shown in (31 ) .
(31) Simple object raising verbs:
(
[CAT
S
U
BCAT
[
CAT np
[
,
J
DCASE
np
FORM
[CAT
acc]
,
SUBCAT
vp
l )
inf
,J
|
Again, this will have important consequences when the VP complement
of an object raising verb is headed by a non-quirky verb. In that event,
since the complement's subject's DCASE and CASE values are identifed
and since the complement's subject is identifed with the direct object of
the object raising verb, it follows that the object must be accusative. If the
VP complement is headed by a quirky verb, again nothing of consequence
follows-the case of the raised object will be determined by the CASE value
asigned by the downstairs quirky verb.
These observations are illustrated in fgures (32)-(35) .
(32)
(33)
Finite non-quirky verb:
S
.
[
CAT np
1
[CAT
CASE
vp ]
nom
SUBCAT _,
DCASE nom
V[fn] NP[acc]
| '
Drengurinn.NOM kyssti stulkuna.ACC
'The boy' 'kissed' 'the girl'
Finite quirky verb:
S
.
[
CAT
CASE
DCASE
np
1
acc
nom
Mig.ACC
'me'
[CAT vp ]
SUBCAT _,
.
V [fn]
langar
' longs'
[CAT VP]
FORM inf
ao fara til islands

'to go to Iceland'
312 I IVAN SAG, LAURI KARTTUNEN, AND JEFFREY GOLDBERG
(34) Object raising verb with non-quirky VP complement:
V[fn)
[SUBCAT ([, m, @))
telja
'believe'
VP[SUBCAT ([))
[CAT
m CASE
DCASE
np
1
ace
ace
Marfu.ACC
'Mary'
[CAT vp ]
@
SUBCAT (m)
hafa skrifa ritgeroina

't<have written
the-thesis'
(35) Object raising verb with quirky VP complement:
V [fn]
[SUBCAT ( [, m, @)]
'
telur
'believes'
VP[SUBCAT ([))
[CAT
m CASE
DCASE
I
np
1
gen
ace
verkjanna. GEN
'the-pains'
[CAT vp ]
@
SUBCAT ( r)
ekki grta
'not be-noticeable'
In our analysis, subject raising verbs are exceptional only in not spec
ifying that their subject's CASE and DCASE values are identifed. Finite
forms of subject raising verbs, e.g. , viroist, like all fnite verb forms, select
for subjects that are [DCASE NOM] . Again, this ha consequences when
the VP complement is headed by a non-quirky verb, but not when the VP
contains a quirky head. This is illustrated in (36) and (37) .
(36) Subject raising verb with non-quirky VP complement:
S
.
[
CAT
[ CASE
DCASE
np
1
nom
nom
Hann. NOM
'he'
[CAT vp ]
SUBCAT ([)
V [fn] [CAT vp ]
[SUBCAT ( [, m)]
m
SUBCAT ( [)
'
viroist
'seems'
elska hana.ACC
't<love her'
(37) Subject raising verb with quirky VP complement:
S
.
[CAT
[ CASE
DCASE
np
1
ace
nom
Hann.ACC
'he'
[CAT vp ]
SUBCAT (m)
V [fn] [CAT vp ]
[SUBCAT ( [, m)]
m
SUBCAT ( [)

viroist
'seems'
vanta peninga
't<lack money'
The lexical entries we have proposed, together with general principles
that directly constrain syntactic information in phrase structure all pro
vide partial information that must be satisfed by the phrase structures of
Icelandic. But the constraints need not be satisfed in any particular or
der. Our analysis thus accounts for the puzzles of quirky cae in Icelandic
in a purely monotonic fahion, i.e. , in terms of constraints that may be
satisfed in any order and which provide fexible, process-neutral, partial
information about the structures of the language.
4 Expressing Lexical Generalizations
Although fully specifed attribute-value matrices for individual verbs in
our analysis of Icelandic may appear to be somewhat complex, their lexical
entries may be stated very concisely by generalizing over classes of similar
lexical items. A set of features common to a clas of words can be defned
just once so that the defnition is shared by all the entries which have
that property. In the P ATR-I I notation, such named properties are called
templates.
1O
A simple template defnition consists of path equations that
describe an attribute-value matrix. Templates can also be defned in terms
of other templates; in this way a lexicographer may construct inheritance
hierarchies of generalizations. In the PATR-I I formalism, such defnitions
are not necessarily monotonic because the information in a higher ranked
template may override some specifcation that is inherited fom a template
with a lower rank. However, our analysis of Icelandic verbs can be stated
under the simpler asumption that templates are composed by unifcation
without overriding.
In this section we show how the complex feature structures for Icelandic
verbs can be decomposed to a small number of simple lexical properties.
The only technical problem in converting from the semi-formal notation we
have used so far to explicit attribute-value structures is the representation
of a list-the value of the SUBCAT feature. In a manner reIniniscent of
CAR and CDR in Lisp, we use the attributes FIRST and REST for the head
and the tail of a list, respectively. A special atom, none, marks the end of
a list. For example, a list of two elements, (A B) , would be designated a
in (38) .
(38) (FIRST) = A
(REST FIRST) = B
(REST REST) = none
In defning our frst template, Verb, we leave the REST argument un
specifed so that this basic defnition can be shared by all types of verbs
that have at least one argument-a subject-on their SUBCAT list:
10
A very similar device is incorporated into the theory of the "hierarchical lexicon"
presented in Flickinger 1987 and Pollard and Sag 1987, Chapter 8.
(39) Verb: (CAT) verb
(SUBCAT FIRST CAT) np
In terms of the semi-formal notation we have used so far, the defnition of
Verb in (39) is equivalent to (40) , where ' . . .' means that the number of
elements on the list is unspecifed.
(40) SUBCAT ( [CAT np] . . . )
The next two defnitions are straightforward:
(41) Finite: (FORM) fn
Infnitive: (FORM) inf
And the following set of templates serve to subcategorize verbs on the bais
of the number and type of their non-subject arguments:
(42) Strict-Intransitive: (SUBCAT REST) none
(43) Transitive: (SUBCAT REST FIRST CAT) = np
(SUBCAT REST FIRST DCASE) acc
(44) Strict- Transitive: (SUBCAT REST REST) = none
The efect of (42) is to limit the SUBCAT list of the verb to a subject. Note
that (43) makes the default case of an object be accusative irrespective of
whether the verb is fnite or non-fnite.
The two types of complementizable verbs we have discussed are char
acterized by the templates (45)-(46). The subject raising verbs merge the
subject role of the complement with the subject role of the matrix verb,
the object raising verbs equate the object of the matrix with the subject
of the complement clause.
(45) Subject-Raising:
(SUBCAT REST FIRST CAT) = vp
(SUBCAT REST FIRST FORM) inf
(SUBCAT REST FIRST SUBCAT FIRST) = (SUBCAT FIRST)
(SUBCAT REST REST) none
(46) Object-Raising:
(SUBCAT REST REST FIRST CAT) = y
(SUBCAT REST REST FIRST FORM) inf
(SUBCAT REST REST FIRST SUBCAT FIRST)
(SUBCAT REST FIRST)
(SUBCAT REST REST REST) = none
The next two templates characterize non-quirky intransitive and transitive
verbs.
(47) Normal-Subject:
(SUBCAT FIRST CASE) = (SUBCAT FIRST DCASE)
(48) Normal-Object:
(SUBCAT REST FIRST CASE) = (SUBCAT REST FIRST DCASE)
Templates (49) and (50) defne two types of quirky verbs. ll
(49)
(50)
Accusative-Subject: (SUBCAT FIRST CASE) = acc
Genitive-Subject: (SUBCAT FIRST CASE) gen
Using the defnitions in (45)-(50) , we can now defne useful composite
templates:
(51) Intransitive Verb: Verb ^ Strict-Intransitive
Tansitive Verb: Verb ^ Tansitive / Strict-Transitive
Subject-Raising- Verb: Verb ^ Subject-Raising
Object-Raising- Verb: Verb ^ Tansitive ^ Object-Raising.
Because the defnitions of individual templates do not introduce any con
ficts, the interpretation of these composite defnitions is straightforward:
the defnition of an object raising verb, for example, is a simple union of
the path equations that defne the templates Verb, Transitive, and Object
Raising. It describes the attribute-value matrix in (52).
(52) Object raising verb:
CAT verb
FIRST [CAT np]
SUBCAT
REST
FIRST m
[CAT np ]
DCASE acc
REST
[
FIRST
[
M
:
lJ
SUBCAT [FIRST m]
REST none
Note that this is a notational variant of (31) above.
We now have all the general properties needed for the entries of indi
vidual verbs. For example:
(53) a. kyssti:
Finite / Trasitive- Verb ^ Normal-Subject / Normal-Object
b. langar:
Finite ^ Equi- Verb ^ Accusative-Subject
c. telja:
Infnitive / Normal-Subject / Object-Raising- Verb
d. grta:
Infnitive / Intransitive- Verb ^ Genitive-Subject
e. viroist:
Finite ^ Subject-Raising- Verb
I1
We leave open here the possibility, sugeted by Yip et ( 1987), that there are
semantic regularities underlying some of the quirky cae asignments we assume here.
316 I IVAN SAG, LAURI KARTTUNEN, AND JEFFREY GOLDBERG
As a consequence of the defnitions we have given for the various lexical
classes, the syntactic behavior and case assignment properties of these verbs
are correctly and completely specifed by the simple lexical entries in (53) .
5 Conclusion
At frst glance, The problem of "quirky" cae in Icelandic seems to involve
non-monotonic (overriding) behavior of default and exceptional cae as
signment principles. However, a we have shown in this paper, the relevant
information, reifed in terms of the attributes CASE and DCASE, can be
treated quite naturally in terms of simultaneous constraint satisfaction.
An interesting feature of the analysis we have proposed is that it cru
cially depends on partial information. For example, the subject raising
verbs asign cae to their subject by letting it be determined by the verb
in the complement clause. When a subject is overtly realized, the relevant
grammar rule also requires that subjects be specifed a [DCASE nom] . And
the lexical entry for object raising verbs require their objects to be speci
fed a [DCASE acc] . These specifcations, together with a set of identities
specifed by the constraints of the grammar interacting with those of the
lexicon, are what account for the realization of case when the complement's
verb is non-quirky.
And because the CASE and DCASE values of the subjects of quirky
verbs are distinct, it follows that when a quirky verb appears within the
complement of a raising verb, the DCASE specifcations just mentioned have
no efect on the cae of the raised element, which is determined entirely
by the quirky verb within the complement. It seems to us that the use
of partial information, together with the distinction between default case
and actual case, provides a much simpler and more satisfactory account of
Icelandic cae than earlier descriptions.
One shortcoming of the analysis presented here is that it provides no ac
count of the nominative-accusative template efect whose utility is demon
strated by Yip et al. (1987) . We believe, however, that their generalizations
can be introduced into our analysis, once rules of case association are seen
to apply not to ordered phrase structures (as Yip et al. assume) , but rather
to hierarchical syntactic argument structures of the sort provided by our
SUBCAT feature. In this way, we can incorporate the central insight of
the Yip et al. analysis into an analysis that avoids the destructive rule
application (cyclical reasignment of non-lexical case) that their analysis
employs.
Finally, our analysis embodies a lexical architecture that we believe
to be of considerable value in the context of the increasing trend within
grammatical theory to locate crucial syntactic information within more
complex lexical entries. The lexicon need not be just a list of forms and
meanings, coupled with idiosyncratic grammatical information. Rather,
i n virtue of the organization provided by templates of the sort we have
used here, the lexicon can be viewed as an elegant hierarchical (or even
heterarchical) databae, where every signifcant generalization about classes
of words may be expressed in compact terms. As complexity of syntactic
rules is continually eliminated in favor of more complex lexical entries, the
development of such methods for achieving lexical compaction is not just
a desideratum, it is a necessity.
References
Andrews, Avery. 1976. The VP Complement Analysis in Modern Icelandic.
In Proceedings of NELS 6 (Montreal Working Papers in Linguistics) , 1-22.
Amherst: University of Massachusetts, GLSA.
Andrews, Avery. 1982. The Representation of Case in Modern Icelandic. In The
Mental Representation of Grammatical Relations, ed. Joan Bresnan. Cam
bridge, Mass. : MIT Press.
Andrews, Avery. 1990. Case Structures and Control in Modern Icelandic. In
Syntax and Semantics 24: Modern Icelandic Syntax, ed. Joan Maling and
Annie Zaenen, 187-234. San Diego: Academic Press.
Bresnan, Joan. 1982. Control and Complementation. In The Mental Represen
tation of Grammatical Relations, ed. Joan Bresnan. Cambridge, Mass.: MIT
Press.
Chomsky, Noam. 1981. Lectures on Goverment and Binding. Dordrecht: Foris.
Dowty, David. 1982a. Grammatical Relations and Montague Grammar. In The
Nature of Syntactic Representation, ed. Pauline Jacobson and Geofrey Pul
lum, 79-130. Dordrecht: Reidel.
Dowty, David. 1982b. More on the Categorial Analysis of Grammatical Relations.
In Subjects and Other Subjects, ed. Annie Zaenen, 115-153. Bloomington:
Indiana University Linguistics Club.
Flickinger, Daniel. 1987. Lexical Fules in the Hierarchical Lexicon. Doctoral
dissertation, Stanford University.
Goldsmith, John. 1976. An Overview of Autosegmental Phonology. Linguistic
Analysis 2: 231-68.
Goldsmith, John. 1990. Autosegmental and Metrical Phonology. Oxford: Bail
Blackwell.
Johnson, Mark. 1988. Attribute- Value Logic and the Theory of Grammar. CSLI
Lecture Notes No. 16. Stanford: CSLI Publications.
Kaplan, Ronald, and Joan Bresnan. 1982. Lexical-Functional Grammar: A For
mal System for Grammatical Representation. In The Mental Representation
of Grammatical Relations, ed. Joan Bresnan. Cambridge, Mass. : MIT Press.
Kay, Martin. 1986. Parsing in Functional Unifcation Grammar. In Natural
Language Parsing, ed. David Dowty, Lauri Karttunen, and Arnold Zwicky.
New York: Cambridge University Press.
Keenan, Edward, and Bernard Comrie. 1977. Noun Phrase Accessibility and
Universal Grammar. Linguistic Inquiry 8:63-99.
318 / VAN 5AG, AlRI hARTTUNFN, AND JFFFRFY lOIDBFRG
Maling, Joan, and Annie Zaenen (eds. ). 1 990. Syntax and Semantics 24: Modem
Icelandic Syntax. San Diego: Academic Press.
Neidle, Carol. 1982. Cae Agreement in Russian. In The Mental Representation
of Grammatical Relations, ed. Joan Bresnan. Cambridge, Mass. : MIT Press.
PoUard, Carl, and Ivan A. Sag. 1987. Information-Based Syntax and Semantics,
Vol. I: Fundamentals. CSLI Lecture Notes No. 1 3. Stanford: CSLI Publica
tions.
Pollard, Carl, and Ivan A. Sag. 1992. Head-Driven Phrase Structure Grammar.
Chicago: University of Chicago Press, and Stanford: L5 Publications.
Sag, Ivan A. 1991 . Linguistic Theory and Natural Language Processing. In
Natural Language and Speech, ed. Ewan Klein and Frank Veltman. Berlin:
Springer-Verlag. Symposium Proceedings. Brussels, November 1 991 .
Sag, Ivan A. , Ronald Kaplan, Lauri Karttunen, Martin Kay, Carl PoUard, Stu
art Shieber, and Annie Zaenen. 1985. Unifi cation and Grammatical Theory.
In Proceedings of the 5th West Coast Conference on Formal Linguistics, ed.
Mary Dalrymple, Jefrey Goldberg, Kristin Hanson, Michael Inman, Christo
pher Pinon, and Stephen Wechsler. Stanford University: Stanford Linguistics
Association.
Shieber, Stuart, Hans Uszkoreit, Fernando Pereira, Jane Robinson, and Mabry
Tyson. 1983. The Formalism and Implementation of PATR-II. In Research
on Interactive Acquisition and User of Knowledge, ed. Barbara Grosz and
Mark Stickel. Menlo Park, CA: SRI International. (SRI Final Report 1 984) .
Thrainsson, Hoskuldur. 1979. On Complementation in Icelandic. New York:
Garland.
Yip, Moira, Joan Maling, and Ray Jackendof. 1987. Case in Tiers. Language
63: 21 7-250.
Author Index
Abelson, Robert P., 211, 213, 233
Abney, Steven, 1 33, 162, 163
Abraham, Werner, 218
Ackerman, Farrell, x, 55, 58, 75, 218
Ades, Anthony, 248
Akmajian, Adrian, 277
Alford, Jack A. , Jr., 234
Allen, Margaret R., 111
Alsina, Alex, 62
Anderson, Mona, 8
Anderson, Stephen R. , 14, 55, 1 1 1
Andrews, A very, 303, 306, 310
Aronof, Mark, 94, 111
Bach, Emmon, 31 , 47, 285, 292
Baker, Mark, 2, 169, 171
Barwise, Jon, 96
Bauer, Lauri, 156
Bayer, Samuel, 271, 275, 279, 280,
282, 283
Becker, Curtis A., 233
Belletti, Adriana, 17
Benesova, Eva, 212
Bennett, Michael, 47
Berman, Arlene, 271
Bever, Thomas G. , 209, 210, 213,
216, 230, 231, 233
Bierwisch, Manfed, 201, 209
Blaberg, L., 146
Black, J. B., 21 1
Bolinger, Dwight, 133, 135, 159
Bouchard, Denis, 85, 90, 107
Bower, Gordon H., 211, 213, 233
Brame, Michael, 293
Bresnan, Joan, 56, 58, 61, 62, 71,
76, 274, 276, 303, 306-308
Brown, Roger, 206
Burzio, Luigi, 1 1
Carlson, Greg, 44
Carpenter, Bob, 252
Chafe, William L. , 212, 215
Chien, Yu-Chin, 258
Chierchia, Gennara, xvi, 252, 269,
272, 273, 278, 283, 295
Chomsky, Noam, xi, xvi, 132, 139,
147, 212, 252, 256, 270,
271 , 275, 277, 278, 292
Clark, Eve, 111, 183, 186, 206
Clark, Herb, 111, 183, 206, 211, 212
Clements, J. , 171
Collins, Allan M. , 224, 233
Comorovski, Ileana, 95
Comrie, Bernard, 307
Crain, Steven, 210, 212, 222, 235
Culicover, Peter, 62
Curry, Haskell B. , 251, 253
Cutler, Anne, 171, 212
Dakin, Julian, 213
de Meij, Sjaak, 218
Dezso, Laszlo, 217
Di Sciullo, Anna Maria, 123, 124,
133, 150, 162, 167, 169
Downing, Pamela A., 141, 143
Dowty, David, viii, x, xvi, 5, 7, 13,
21, 22, 30, 31, 44, 57, 60,
62, 63, 142, 254, 269, 272,
320 / LXlCAL NATTR5
273, 279, 283, 286-288, 290,
292
Dressler, Wolfgang U. , 186, 187
Dufy, S. , 21 1
Esau, Helmut, 1 1 1
Fabb, Nigel, 138, 162
Farkas, Donka, xi, xv, xvii, 88, 95,
218
Fauconnier, Gilles, 99
Fazekas, A. , 206
Fellbaum, Christiane, 8
Feys, Robert, 251, 253
Fiengo, Robert, 270, 271 , 274
Filip, Hana, 49
Fillmore, Cnar!tH, bb, 63, 128
t u| u, Timothy W. , 176
Fischler, Ira, 233
Flickinger, Daniel, 313
Fodor, Janet D. , 210, 271, 278
Fodor, Jerry A. , 210, 212
Franlois, Jean, 30
Fudge, Erik, 143, 156
Garrett, Merrill F. , 210
Garrod, Simon L. , 211
Gawron, Jean Mark, 63, 259, 260
Gazdar, Gerald, 271 , 275, 278-280,
283, 287, 289, 293
Geach, Peter T. , 253
Gergely, Gyorgy, xiv, 209, 21 0, 21 1 ,
213, 214, 221 , 225, 231 ,
233
Gibbs, Ray W. , 21 1
Gleitman, Lila, 23
Goldberg, Jefrey, 23
Goldsmith, John, 306
Goodman, George \. , 233
Gough, Phillip B. , 234
Greenberg, Joseph H. , 194
Grimshaw, Jane, 1 1 , 1 13, 258, 283-
285
Gropen, Jess, 23
Gruber, Jefrey S., 7
Hajicova, Eva, 212
Hale, Kenneth, 8
Halle, Morris, 132, 147
Halliday, Michael A. K. , 212
Hankamer, Jorge, 280
Hasegawa, Nobuko, 1 13, 1 14
Haviland, Susan E. , 21 1 , 212
Hayes, Bruce, 132, 147
Heinamaki, Orvokki, 46
Hindley, J. Roger, 252
Hinrichs, Erhard, 7, 31
Hirschberg, Julia, 1 34
Hjelmslev, Louis, 194
Hoeksema, Jacob, 169
Hoepelman, Jakob, 31
Holley-Wilcox, Pamela, 234
Horn, Laurence R. , x, 94, 106, 286
Horvath, Julia, 21 7
Householder, Frederick W. , x, 123
Hualde, Jose Ignacio, 122
Hukari, Thomas, 270, 275, 276, 278,
283, 286, 288, 290, 293
Inoue, Kazuko, 1 1 3
Jackendof, Ray, 7 , 56, 1 3 1 , 132,
21 2, 21 5, 216, 273
Jacobson, Pauline, xvi, xvii, 271-
274, 277-283, 285, 286, 288,
289, 292, 294
Johnson, Mark, 306
Johnson-Laird, Phillip N. , 190, 210,
222, 235
Joseph, Brian, 95
Kageyama, Taroo, ll3, ll 6
Kalman, Laszlo, 218, 235
Kanerva, Jonni M. , 61 , 62
Kanski, Zbigniew, 253
Kaplan, Ronald, 303, 306
Kawakami, Shin, 115
Kay, Martin, 128, 303
Kayne, Richard, 242, 243, 245, 247
Keenan, Edward L., 253, 255, 256,
307
Kempchinsky, Paula, 85, 89
Kenesei, Istvan, 21 8
Keyser, S. Jay, 8
Kiefer, Ferenc, 201, 21 7
Kimballg John, 296
Kiparsky, Paul, 85, 94, 1 1 1 , 163, 167
Kiss, Katalin E., 217
Klein, Ewan, 271, 275, 278-280, 283,
287, 293
Koml6sy, Andras, 218
Koster, Jan, 242
Kriia, Manfred, viii-x, 7, 31, 36, 38
Laczk6, Tibor, 55, 75
Ladd, Robert, 133, 135, 161, 167
Ladusaw, William, 62, 274, 290
Landau, Barbara, 23
Lang, Ewald, 201
Lapointe, Steven G. , 194, 195
Lasnik, Howard, 270, 271, 274
Lees, Robert B. , 137
Leisi, Ernst, 31
Levi, Judith N. , 137, 141, 142, 151,
166
Levin, Beth, 14, 55, 56, 71, 72, 76,
137
Levine, Robert, 270, 275, 276, 278,
283, 286, 288, 290, 293
Levinson, Stephen, x
Liberman, Mark, xii, 132, 147, 175,
176
Lieber, Rochelle, 137, 156, 171
Link, Godehard, 31, 45
Loftus, Elizabeth F. , 224, 233, 234
Magerman, David, 177
Maidment, J. A. , 135
Maling, Joan, 270
Marantz, Alec, 137
Marchand, Hans, 145
Marcus, Mitchell, 177
Marslen-Wilson, William D. , 210, 222,
224, 233, 235
Mayerthaler, Willi, 186
McCawley, James, x, 94, 106, 123
Mchombo, Sam, 62
Meireles, J., 85, 87
Mester, Armin, 113
Meyer, David E. , 233, 234
Miller, George A. , 190
Mitchell, D. C. , 21 1, 233, 234
Miyagawa, Shigeru, 111, 1 13
Mohanan, t P., 169, 172
AUTHOR INDEX / 321
Moravcsik, Edith, 46
Morton, J. , 224, 233
Neely, James H. , 234
Neidle, Carol, 306
Nunberg, Geofrey, 288
O'Connor, Mary Catherine, 128
Oehrle, Richard, 248, 277
Padilla-Rivera, J., 87, 103
Panagl, Oswald, 186
Partee, Barbara, 47, 258
Paul, Hermann, 1 1 1
Perlmutter, David, 1 , 1 1
Perry, John, 96
Peters, Stanley, 259, 260
Picallo, Carme, 85, 87
Pinker, Steven, 23, 55
Platzack, Christer, 31
PlEh, Csaba, 206, 217
Pollard, Carl, 242-244, 247, 252, 271,
274, 278, 285, 288, 307,
308, 313
Poser, William J. , 113
Postal, Paul, 1 , 11, 173, 271, 273,
279
Prince, Alan, 132, 147
Pullum, Geofrey t, 271 , 275, 278-
280, 283, 287, 293
Pustejovsky, James, 7
Quine, Willard van Orman, 30, 253
Ramsey, S. Robert, 171
Raposo, Eduardo, 85, 87
Rappaport, Maika, 14, 55, 56, 71,
72, 76, 137
Reger, Zita, 206
Reinhart, Tanya, 107, 212, 258-260
Riesbeck, Christopher , 210, 21 1
Rizzi, Luigi, 17
Roberts, Ian, 8
Rochemont, Michael, 218
Roeper, Thomas, 137
Rohrer, Christian, 31
Rooth, Mats, 174
Rosch, Elinor, 206
322 / LXlCAL ATTER5
Rosen. Carol, 258
Rosenbaum, Peter, 271, 279, 284
Ros, John R. , 279
Ruddy, Margaret G. , 233
Ruwet, icola, 88, 89, 103
Safr, Ken, 139
Sag, Ivn A., 243, 272, 274, 275,
278-280, 283, 285, 288, 293,
303, 307, 308, 313
Saiki, Mariko, 1 16
Salamanca, D. , 87
Salkoff, Morris, 55
Sanford, Anthony J. , 2 1 1
Scha, Remko, 43
Schachter, Paul, 278
Schnurr- Wolles, L. , 187
Schank, Rogrr L. , 2 1, 2 1 1 , 2 1 3, 233
Srh111l`I1ing, Susan, 218
Schonfinkel, M. , 251
Schvalleveldt, Roger W. , 233, 234
Schwartz-Norman, Linda, 14
Seidenberg, Mark, 210
Selkirk, Elizabeth, 137, 145, 164,
1 69, 171, 218
Sgall, Petr, 212
Sharkey, Noel E. , 211, 233, 234
Shih, Chi-lin, 165, 177
Siegel, Mufy, 137
Simpson, Jane, 56
Sproat, Richard, xii, 1 37, 139, 165,
1 73, 1 75-178
Steedman, Mark, 210, 2 12, 222, 235,
248, 249, 251, 254, 257,
277, 278, 290
Stump, Greg, 171
Suner, Margarita, 87, 103
Szabo, Zoltan, 235
Szabolcsi, Anna, xv, xvii. 55. 92.
21 2, 21 7, 2 1 8. 248. 249.
254, 257, 259. 264
Tenney, Yvette J. , 2 1 1
Tenny, Carol, viii , 7 , 1 9
ter Meulen, Alice, 3 1
Thomason, Richard, 139
Thrainsson, Hoskuldur, 303
Toman, Jindrich, I I I
Townsend, David J. , 2 1 1 , 2 13, 216,
230, 23 1
Turer, T. J. , 2 1 1
Tyler, Lorraine K. , 2 1 0, 235
Van Val i n, Robert, 1 1 , 1 3
Vendlrr, Zeno, 30, 60
Verkuyl , Henk J . , 7, 3 1 , 46
Verschueren, Jef, 201
von Stechow, Ari m, 253
Ward, Gregory, 1 73
Warren, R. E. , 234
Was ow , Thomas, 71, 288
Welsh, A. , 224, 233
Wexler, Ken, 258
Wierzbicka, Anna, 49
Williams, Edwin, 4, 1 23, 124, 133,
1 50, 162, 167, 169-171
Wurzel, Wolfgang V. , 1 86
Yip, Moira, 303, 31 5, 316
Zaenen, Annie, x, 11, 56, 58, 60-64,
76, 270
Zaring, L. , 87, 1 03
Zec, Draga, 93
Zubizarreta, Maria-Luisa, 57
Zwanenberg, Weicher, 1 1 1
Subject Index
Accomplishment, 29, 30
acquisition, 23
order, xiii, xiv, 206
activities, 29
adjective
comparative, 120
superlative, 120
adjective-noun compounds, xii, 148,
151
adverbial
apect-marking, 218
directional, 51
durative, 30, 51
locative, 51
time-span, 30, 51
adversative, xiv
inference, 231
subordinate clauses, 213, 214
afected arguments, 9, 10
afectedness, viii, 8, 10, 23
agentive-headed compounds, 137
pseudo-object le
f
hand members,
138
restrictions on lef hand mem-
bers, 138
agglutinative language, 209, 217
agreement, 274, 288, 292
aktionsart, 29, 189
fequentative 51
anaphora, xvi, 241, 252, 253, 255-
260, 265, 266
deep, 280
null complement, xvi, 280, 281,
289, 291
anaphoric island constraint, 173
antecedent matching, 222
argument structure, viii, x, xiii, 2, 3,
21, 23
argument-argument compounds, 137,
140, 143
argument-predicate compounds, 137,
140
aspect, D, x, 2, 4, 10, 11, 13, 16-18,
20, 21, 23, 29, 31, 38, 60,
189, 192
apect-marking adverbial, 218
atelic predicate, viii, 30, 34, 35
atomic reference, D, 32
atomicity of word, 172
Backward search for discourse an-
tecedent, 212, 213, 216
basic lexical entry, 64
baic-level term, 183
Baque, xi, 122
binding, xi, xv, xvi, 251-253, 255,
256, 258-260
blocking, viii, x-xiii, 94, 1 1 1
Brazilian Portuguese, 87
Breton, 171
Cae, xv, xvii, 303, 305
flter, 305, 306
case-marker sufxes, 217
categorial grammar, xv, 241 , 248,
272, 277, 294, 302, 303,
307
causativization, xiii, 56
324 / LXlCAL NATTR5
cause-efect relation, xiv, 213, 230,
235
change, 3, 6, 8, 19-21
of location, 79
Chinese, 1 77
closed class IlCm8, 1 63
collective rC0dtug8, 44
combinator, xv, 241 , 248, 251-253,
256, 257, 259, 260, 262,
264-266
competence, 303
complex predicate, , 58, 65, 70,
21 8
complexit.y, formal, xi i i , I b4
compound
bnhuvrihi , I 4b, 1 66
derived-nominal headed, 1 39
EUglIsh
orthographic conventions, 1 36
modifers, 132
stress rule, 146
with possessive modifiers, 153
implicit contrast theories, 1 55
conjunction, 232
constraining equations, 303, 306
constraint, xvii, 302, 303, 306, 31 3,
31 6
construction grammar, 1 28
control, xi, xv, xvi, 269-272, 280,
289, 291 , 295, 296
controller choice, 88, 103
count nouns, 29
cumulative
readings, 43
reference, ix, 30, 32
Czech, 46, 49
Deep anaphora, 280
default cae, 307, 309, 310, 314, 316
deferred reference, 1 74
defniteness, 49
degree of agentivity, 88
delimitedness, 5, 7, 1 1 , 19, 24
derivational morphology, xiii, 185,
1 92
derived-nominal headed compound,
1 39
direct internal argument, 4, 6-8, 10,
1 5, 23
directional adYerbial 5 1
discourse
uncIIOu8 In !uug0tI0n, 21 7
informational cues, 210
representation, 232
disjoint reference, 85
dummy, 276, 287, 288
durative adverbial, 30, 51
Efected object8, 41
elementary predicates, 1 91
elsewhere pri nci ple, x, xi , 90, 1 07,
J
tng\8h, xii, xiv, 120
compound, 136, 1 69
noun phrases, 1 31
epithets, 149
event, 4, , 1 2
tokens, 34
types, 34
external argument, 4, 1 1 , 1 7, 18, 24
extraction, xvi, 243, 253, 255, 261,
262, 265
gap, 277
Finnish, i x, 46
"fat" sentences, 21 8
focus, 2 12, 21 4, 21 7, 2 1 8
movement, 2 1 7
focus-based inference, xv, 209, 210,
21 5, 231-236
in discourse antecedent match
ing, 21 9
predictive, 2 10, 231
clausal processing, 222
priming, 233
processes, 21 9
foregrounding, 222, 223, 236
formal complexity, 184
"free" word order, 209, 21 7
Fench, 86-89, 167, 1 71
frequentative
verb form, xiii
aktionsart , 5 1
function composition, xvi, 247-249,
251, 260, 271, 279, 280,
282
Gap, xv, xvi, 247
extraction, 277
parasitic, 243, 244, 270, 278
German, 46
gerund-headed compounds, 139
pseudo-object left hand members,
139
given information, 212
goal, 14, 15, 19
government-binding theory, xv, 4,
85, 241-243, 247, 248, 302
graduality, 42
Gricean convention, 214, 215
Head-driven phrase structure gram
mar, xv, 241-244, 247, 248,
307
head feature principle, 308
head-fnal flter, 162
headedness
of afxed words, 171
of compounds and phrases in
English, 169
hierarchical lexicon, 313
Hindi, 51
Hungarian, ix-xi, xiii, xiv, xviii, 55,
58, 64, 71, 72, 81, 90-92,
95, 96, 184, 185, 209, 217,
254
discourse functions, 217
sentence structure, 217
topic-focus structure 219
Icelandic, xvii, 301-317
idiom chunks, 288
incremental theme, viii, 7, 22, 62-63
indirect internal argument, 4, 7
inference
adversative 231
focus-based, xv, 209, 210, 215,
231-236
predictive, 213, 222, 224
information
cues, 210
SUBJECT INDEX / 325
given, 212
new, 212
mutual, 177
inheritance hierarchy, 313
interactive model, 232, 235
interclausal semantic relations, xiv,
209, 213, 232, 235
internal argument, ix, 3, 4, 10, 11,
13, 17-20
indirect, 4, 7
island, 292-294
Italian, 167
iterativity, 39, 40
Japanese, 112
Language processing, 301, 303
lattice, ix, 31, 32
left-headed modifed nominals, 166
stress pattern in, 151
lexical
apect 60
entry, vii, 3
form, 69
integrity, 172
item, xiii, 241-244, 248, 253,
265
mapping theory, x, 55-81
priming, 233
semantics, x, xiii, 1, 18, 21, 60,
252-256
versus syntactic analysis, xv,
241-244
lexicalization, xiii, 150
lexicon, vii-xv, 241, 242, 254, 264
locative
adverbial, 51
alternation, D, 14, 15, 23, 55,
56, 58, 64
inversion, 57
Manner of motion, 79
mass nouns, 29
measure constructions, 29, 51
measuring out the event, D, 2-7,
12-20, 23
middle, 10, 11
monotonicity, xvii, 57, 68, 248, 303
326 / LXlCAL ATR
morpholexical operation, x, 55, 56,
80
morphological
complexity, xiii, 184-187, 191,
192, 194-197, 201 , 206, 207
component, 133, 169
construction, 125
markedness, 187
morphosyntactic operation, 7 56,
80
morphotactic transparency, 187
motion, 3, 21
moV(' I t , 302, Jh
IIIl1tllal informat ion, 1 77
DI\mrH H modifirr8, 1 73
lI('gation, 5 1
new information, 2 1 2
nominal reference, 30
with possessive modifers, 154
non-binary premodifed nominal, 146,
161
nondelimited event, 5, 7
nonstative verb, 3
noun phrase, xii, 1 31
noun-noun compound, xii, 156, 1 60
noun-PP expressions
left-stressed caes, 155
nuclear stress rule, xii, 1 46
null complement anaphora, xvi, 280,
281, 289, 291
O-ellipsis, 1 16
obliqueness, 307
observed object, 31
obviative complement, 85
overlap, 32
overriding, 313, 316
Paraitic gap, 243, 244, 270, 278
partitive, ix, 46
pasive, xiii, 8, 11, 57, 285
path equations, 313, 315
lAH-\, 307, 313
perfective, D, x, xiii, 49
performance, 301-306
periphratic verb, xi, xii, 1 12, 122
phraal modifers, 132
plural, 29
Polish, 49
possessives, distributional propertie
of, 1 54
pragmatic complexity, 201 , 206, 207
predicate-argument structure, 188
predictive inference, 213, 222, 224
pre modifed nominal
P
9
versus P
I
caes, I b4
automatic parsing
statistical methods, 1 77
and stress assignment, I Tb
:utc strrss- patter, 1 48
except. ions, 1 48
"i mplicit contrast" , I h/
i ntcrspeakcr variation, I 49,
1 57
Ladd's "subcategorization" the-
ory, 167
deaccenting in, 168
determiners in, 162
non-binary, 146, 161
pronouns in, 162
quantifers in, 162
semantics of
classifcatory schemes, 141
'connected-with' theory, 142
structure of, 162
syntactic restrictions on left hand
member, 162
with phrasal modifer, 156, 161
prenominal adjective ordering restric-
tions, 165
presuppositional properties, 189
pro-arb, 273
process nominals in compounds, 139
processing time, xiv, 206
productivity, 186
progressive, i x, 46
proto-agent, 21, 22, 63
proto-patient, viii, 7, 21, 22, 63, 78
protorole, viii, x, 63
prototype semantics, 191
psych verb, 16
Quantifcation, 51
quantized reference, ix, 30, 32
quirky cae, xvii, 303-316
Raising, 304-316
construction, xvii, 270, 271, 279-
282, 286, 288
function composition analysis
of, 279, 280
recoding
at clause-boundary, 225
at end of clause, 222, 236
reference
atomic, ix, 32
cumulative, ix, 30, 32
deferred, 174
disjoint, 85
nominal, 30
with possessive modifers, 154
quantized, 30, 32
singular, 32
referential expressions within com-
pounds, 173
relational grammar, 307
resolution problem, 301
result nominal, 139
resultative preverb, x, 18, 59
Romance, xi, 85, 87, 96
compounds, 167
Romanian, xi, 88, 89, 95
Scenario, 45
script, xiv, 211, 233, 234
selectional restriction, 270, 271, 274,
276, 290, 293
self lefthand member of compound,
140
semantic
complexity, 7111 xiv, 183-186,
189-192, 194-197, 200-202,
206, 207
felds, 185
markedness, xiii, 184, 185, 191,
205-207
representations, 189
transparency, 186
semantics
lexical x, xiii, 1 , 18, 21, 60,
252-256
of pre modifed nominal, 141, 142
prototype, 191
Serbo-Croatian, xi , 90, 93-95
SUBJECT INDEX / 327
set terminal point, 34
singular reference, ix, 32
Slavic, D, 46
Spanish, 87, 89, 167, 171, 172
stress patterns, 131
discourse-conditioned vs. discourse
neutral, 134
of argument-argument compounds,
143
structure sharing, 309, 310
SUBCAT, 307-316
subject, 85, 308, 309
postposing, 217
summativity, 39, 43
Swedish, 146
syntactic
complexity, xiii, 184, 188, 207
markedness, 188, 206
Telic predicate, viii-x, 30, 34, 35,
61, 69, 74, 80
template, 313-317
temporal
constitution, 29, 30
order, 33
trace, 33
thematic relations, N1117 1, 2, 11,
19, 31, 34, 62, 71
theme, 14
-criterion, 274
8-role, 270, 271, 274, 291
'though' clause, 213, 214, 231
time dependency, 101
time-span adverbial, 30, 51
topic, 212, 217
topic-focus structure, xiv, 209, 211,
214, 217, 219, 223, 230,
235
topicalization, 218, 286
tough construction, xvi, 269-283, 288-
296
transformational rules, xv, xvii, 302,
303, 305
transitivity, 76
Unaccusative, 1 1 , 12, 57, 58, 73, 76
unbounded dependencies, xv, 241
unergative, 11, 12
328 / LBXlCAL NATTBR
unifcation, 313
uniqueness
of events, 39, 46
of objects, 39, 44
universal alignment hypothesis, 2
universal theta assignment hypoth
esis, 2
Variable, 241 , 251-253, 256-258, 260,
265
verba efciendi, 46
verb
compound, 145
frequentative, xiii
nonstative, 3
performance, 46
periphrastic, xi, xii, 1 1 2, 122
psych, 16
quirky, 309
verbal modifers, 21 8
VP-ellipsis, xvi, 241, 253, 259-262
Welsh, 170, 171
wh-construction, 270, 271, 277, 292,
293
wh-extraction case, xvi, 294
'while' clause, 225, 231
word
afxed, 1 71
atomicity, 1 72
order, 209, 217
Zhuang, l 71
Lb uDCaOD5
CCutC OC5
The titles in this series are distributed
by the University of Chicago Press
and may be purchaed in aademic
or university bookstores or ordered di
rectly from the distributor: Order De
partment, I Ic1c S. Langely Avenue,
Chicago, Illinois rcrz.
A Manual oj Intensional Logic. Johan
van Benthem, second edition, revised
ad expaded. Lecture Notes No. I.
I5BN 0-937073-29-6 (paper) , 0-937073-
30-X (cloth)
Emotion and Focus. Helen Fay Nis
senbaum. Lectue Notes No. 2. I5BN
0-937073-20-2 (paper)
Lecture, on Contemporary Syntactic
Theories. Peter Sells. Lecture Notes
No. 3. I5BN 0-937073-14-8 (paper) ,
0-93707313-X (cloth)
An Introduction to Unifcation-Based
Approache, to Grammar. Stuart M.
Shieber. Lecture Notes No. 4. I5BN
0-937073-00-8 (paper) , 0-937073-01-6
(cloth)
The Semantics oj Destructive Li,p. Ian
A. Maon. Lecture Notes No. 5. I5BN
0-937073-06-7 (paper) , 0-937073-05-9
(cloth)
An Euay on Facts. Ken Olson. Lec
tue Notes No. 6. I5BN 0-937073-08-3
(paper) ,
0-937073-05-9 (cloth)
Logics oj Time and Computation.
Robert Goldblatt, second edition, re
vised and expanded. Lectue Notes
No. 7. I5BN 0-937073-94-6 (paper) ,
0-937073-93-8 (cloth)
Word Order and Constituent Structure
in German. Hans Uszkoreit. Lecture
Notes No. 8. I5BN 0-937073-10-5 (pa
per) , 0-937073-0-1 (cloth)
Color and Color Perception: A Study
in Anthropocentric Realism. David
Russel Hilbert. Lecture Notes No. 9.
I5BN 0-937073-16-4 (paper) , 0-937073-
15-6 (cloth)
Prolog and Natural-Language Analysis.
Ferando C. N. Pereira and Stuart M.
Shieber. Lecture Notes No. 10. I5BN
0-937073-180 (paper) , 0-937073-17-2
(cloth)
Working Papers in Grammatical The
ory and Discourse Structure: Inter
action, oj Morphology, Syntax, and
Di,course. M. Iida, S. Wechsler, and
D. Zec (Eds.) with an Introduction by
Joan Bresnan. Lecture Notes No. 11.
I5BN 0-937073-04-0 (paper) , 0-937073-
25-3 (cloth)
Natural Language ProceS6ing in the
I 80s: A Bibliography. Gerald Gaz
dar, Alex Franz, Karen Osbore, and
Roger Evans. Lecture Notes No. 12.
I5BN 0-937073-28-8 (paper) , 0-937073-
26-1 (cloth)
InJormation-Based Syntax and Seman
tics. Carl Pollard and Ivan Sag.
Lecture Notes No. 13. I5BN 0-937073-
24-5 (paper) , 0-937073-23-7 (cloth)
Non- Well-Founded Sets. Peter Aczel.
Lecture Notes No. 14. I5BN 0-937073-
22-9 (paper) , 0-937073-21-0 (cloth)
Partiality, Truth and Persistence. Tore
Langholm. Lecture Notes No. 15.
I5BN0-937073-34-2 (paper) , 0-937073-
35-0 (cloth)
Attribute- Value Logic and the Theory oj
Grammar. Mark Johnson. Lecture
Notes No. 16. I5BN 0-937073-36-9
(paper) , 0-937073-37-7 (cloth)
The Situation in Logic. Jon Barwise.
Lecture Notes No. 1 7. I5BN 0-937073-
32-6 (paper), 0-937073-33-4 (cloth)
The Linguistics oj Punctuation. Geof
Nunberg. Lecture Notes No. 18. I5BN
0-937073-466 (paper) , 0-937073-47-4
( cloth)
Anaphora and Quantifcation in Situa
tion Semantics. Jean Mak Gawron
and Stanley Peters. Lecture Notes
No. 19. I5BN 0-937073-48-4 (paper) ,
0-937073-49-0 (cloth)
Propo,itiofl1 Attih,u: The Role of
Cont.nt in Logic. Lang"age. and
Mind. C. Anthony Anderon ad
Joe ph Owens. LectU Notes No. 20.
ISBN (937073-5(4 (paper). (937073-
51-2 (cloth)
Lite ..t. 4d Cognition. Jerry R.
Hobbs. LectU Notes No. 21. ISBN
(937073-52-0 (paper). (937073-53-9
(cloth)
Sitution Theory and It, Application,.
lol. t . Robin Cooper. Kuniaki Mukai.
and John Perry (Eda. ) . Lecture Notes
No. 22. ISBN (937073-54-7 (paper).
093773- 55-5 (cloth)
Thc Llnguge of Fir.t- Order Logic
(i"c1"ci"g the Mlcinto,h program_
7areki , World). Jon 8Mwise atd
John Etchemendy. second edition. re
vaed atd expanded. Lecture Notes
No. 23. ISBN (937073-74-1 (paper)
Lexical Maller . Iva A. Sag and Anna
Szabolcsi. editors. Lecture Notes
No. 24. ISBN (9370736&O (paper) .
(93707365-2 (cloth)
Tar.ki, World. Jon BMwise and John
Etchemendy. Lecture Notes No. 25.
ISBN (937073-67-9 (paper)
Situation Theory and It. Application"
Vol. 2. Jon Bawise. J. Mark Gawron.
Gordon Plotkn. Syun Tutiya. editors.
Lecture Notes No. 26. ISBN 0-937073-
7(9 (paper). (937073-71-7 (cloth)
Literlte Programming. Donald E.
Knuth. Lecture Notes No. 27. ISBN
(937073-8(6 (paper). (937013-81-4
(cloth)
Normalization, Cut-Elimination and
the Theory of Proof,. A. M. Ungar.
Lecture Notes No. 28. ISBN 0-931013-
82-2 (paper) . (931013-83-0 (cloth)
Lect"re, on Linear Logic. A. S. Troel
stra. Lecture Notes No. 29. ISBN
(937073-17-6 (paper). (931013-18-4
(cloth)
A Short Introd"ction to Modal Logic.
Grigori Mints. Lecture Notes No. 30.
ISBN (93701315-X (paper) . 0-931013-
16-8 (cloth)
Lther Lo11 tes
1strDuteU Dj 1L
Agree ment i n Nat ural Language: Ap
proach." Theorie., DelCr.ption, .
Michael Barlow and Chales A. Fergu
son (Eds. ) . ISBN 0-937073-02-4 (cloth)
Paper. from t he Second International
Work,hop on Japane.e Syntax.
William J. Poser ( Ed. ) . ISBN
931073-38-5 (paper) . 0-937073-39-3
(cloth)
The Proceedi"g. of the Seventh W.,t
Coa,t Conference on Formal Ling"i,.
tiel (WCCFL 7). ISBN 0-937073-4(7
(paper)
The Proceeding. of t he Eighth W .. t
Coa.t Conference on Formal Ling"i.
tic. (WCCFL
8
). ISBN 0-937073-48
(paper)
The Phonology-Syntax Connection.
ShMon Iea ad Draga Zec (Eds.)
(copublished with The University of
Chicago Press) . ISBN (226381Q5
(paper). 0-22&38101-3 (cloth)
The Proceeding. of the Ninth We.t COalt
Conference on Formal Lingui.ti"
(WCCFL 9). ISBN 0-937073-64-4
(paper)
Japane .. /Korean Lingui.ti". Hajime
Hoji (Ed. ) . ISBN 0-93107351-1 (pa
per) . 0-931013-563 (cloth)
Experiencer Subject. in South A,ian
Language.. Manindra K. Verma
and K. P. Mohanan (Eds. ) . ISBN
931073-60-1 (paper) . 0-937073-61-X
( cloth)
Grammatical Relation.: A Cro .. -
Theoretical Per'peclive. Katarzyna
Dziwirek. Patrick Farell. Errapel
Mejia Bikandi (Eds. ) . ISBN (9370
73
-
63-6 (paper) . 0-937013-62-8 (cloth)
The
Proceeding. of the Tenth Welt
COI.t
Conference on Formal Lingui.ti"
(WCCFL 10). ISBN 0-931073-19-2
(paper)
OOK5 )5tDuCU
D Lb
The Proceedinga of the Third Weat Coaat
Conference on Formal Linguiatica
(WCCFL 3). (ti e.ss)I5BN 0-937073-
45-8 (paper,
The Proceedings of the Fourth West
Coaat Conference on Formal Lin
guiatica (WCCFL 4). ( tii. ss) I5BN
0-937073-45-8(paper,
The Proceedings of the Fifth Weat Coaat
Conference on Formal Linguiatica
(WCCFL 5). ( ti e.ss)I5BN 0-937073-
45-8 (paper,
The Proceedings of the Sith West Coast
Conference on Formal Linguistics
(WCCFL 6). (ti. ss) 5BN 0-93707J-
45-8 (paper,
Hausar Yau Do Kullum: Intermediate
and Advanced Le880n8 in Hausa Lan
guage and Culture. Wilia H. Leben,
Ahmadu Bello Zaria, Shekrau B.
Mak, and Lawan Danladi Yalwa.
( tis. ss) I5BN 0-937073-b8-7(paper,
Hau8ar Yau Do Kullum Workbook.
Wilia H. Leben, Ahmadu Belo
Zaria, Shekarau B. Mak, and
Lawan Danladi Yalwa. ( t. se) I5BN
0-93703-b9-5 (paper,
LtUCtD_ C5
)5tDuCU D Lb
Titles distributed by CSLI may be
ordered directly from CSLI Publica
tions, Ventura ), Stanford Univer
sity, Stanford, California 11c-1I I or
by phone (1I)Iz1-IIIz or (1I)Iz1-
I1. Orders can also be placed by e
mail ([email protected]) or FAX
(1I)Iz1-cI.
All orders must be prepaid by
check, VISA, or MaterCard (include
card name, number, expiration date) .
For shipping and handling add tz. c
for frst book and tc.Ifor each addi
tional book; tI. I for the frst report
and tc.z for each additional report.
California residents add I7sales tax.
For oversea shipping, add t1. c
for frst book and tz. zfor each addi
tiona book; tz.zfor frst report and
tc.I for each additional report. All
payments must be made in US cur
rency.

Lexical Matters

Uploaded by

Copyright:

Available Formats

Lexical Matters

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lexical Matters

Uploaded by

Copyright:

Available Formats

LXC^L

argument must be the syntactic object of the verb. What is

3y[y x 1 ATOM(y, P)J J J

Complex Predicates and

Blocking of Phrasal Constructions by

The Stress and Structure of

Hungarian Derivational Morphology,

T is ugliest' then T must be

T is most stupid' then T must be stupid. In such cases the ex

1977. Theoretical Linguistics

Focus-Based Inferences in Sentence

at the party 1 the prince+t (=acc)] 1

at the party] j the prince +t (=acc)] 1 showed

away- ( =V-mod 'el-' )] robbed \=kidnapped\, 1 the major

the prince +t ( =acc)]

Semantic Interpretation in Generative Grmmar. .

On Emphasis and Word Order in Hungarian. .

Combinatory Grammar and

hiInself mention before Mary did

before( hym) ( hy) (x)] ( v) ( z) (z) ]

1 tentatively asume that in coordination and in discourse do ha the interpretation of

Simple object raising verb:

ao fara til islands

hafa skrifa ritgeroina

You might also like