Gennaro Chierchia - Semantics

PROPERTIES, TYPES AND MEANING
VOLUME TWO
STUDIES IN LINGUISTICS AND PHILOSOPHY
formerly Synthese Language Library
Managing Editors:
GENNARO CHIERCHIA, Cornell University
PAULINE JACOBSON, Brown University
FRANCIS J. PELLETIER, University ofAlberta
Editorial Board:
EMMON BACH, University of Massachusetts atAmherst
JON BAR WISE, CSLI, Stanford
JOHAN VAN BENTHEM, UniversityofAmsterdam
DA VID DOWTY, Ohio State University, Columbus
GERALD GAZDAR, UniversityofSussex, Brighton
EWAN KLEIN, University of Edinburgh
BILL LADUSA w, University of California at Santa Cruz
SCOTT SOAMES, Princeton University
HENRY THOMPSON, University of Edinburgh
VOLUME 39
PROPERTIES, TYPES
ANDMEANING
Volume II:
Semantic Issues
Edited by
GENNARO CHIERCHIA
Dept. of Modern Languages and Linguistics, Cornell University
BARBARA H. PARTEE
Dept. of Linguistics, University of Massachusetts
and
RA YMOND TURNER
Dept. of Computer Science, University of Essex
KLUWER ACADEMIC PUB:tISHERS

DORDRECHT I BOSTON I LONDON
Library of Congress Cataloging-in-Publication Data
Properties, types and meaning.
(Studies in 1inguistics and phi1osophy

v. 38-39)
Based on a conference on property theory,
type theory and semantics he1d in Amherst on
March 13-16, 1986--Pref.
Inc1udes bib1iographies.
Contents: -- v. 1. Foundationa1 issues
v. 2. Semantic issues.
1. Semantics--Congresses. 2. Semantics
(Philosophy)--Congresses. 1. Chierchia, Gennaro.
II. Partee, Barbara HalI. III. 1'urner, Raymond,
1947- IV. Series.
P325.P74 1988 425 88-8820
ISBN-13: 978- 1-55608-070-8 e-ISBN-13: 978-94-009-2723-0
DOI: 10.1007/978-94-009-2723-0
Published by Kluwer Academic Publishers,

P.O. Box 17, 3300 AA Dordrecht. The Netherlands
Kluwer Academic Publishers incorporates the publishing programmes of

D. Reidel, Martinus Nijhoff, Dr. W. Junk, and MTP Press.
Sold and distributed in the U.S.A. and Canada

by Kluwer Academic Publishers,
101 Philip Drive, Norwell, MA 02061, U.SA
In ali other countries, sold and distributed

by Kluwer Academic Publishers Group,
P.O. Box 322, 3300 AH Dordrecht, The Netherlands
AII Rights Reserved

1989 by Kluwer Academic Publishers
Softcover reprint ofthe hardcover 1st edition 1989
No part of the material protected by this copyright notice may be reproduced or
utilized in any form or by any means, electronic or mechanical,
including photocopying, recording or by any information storage and
retrieval system, without written permission from the copyright owner
TABLE OF CONTENTS
TABLE OF CONTENTS OF VOLUME I vi
PREFACE vil
INTRODUCTION 1
JEROEN GROENENDIJK and MARTIN STOCKHOF / Type-

Shifting Rules and the Semantics of Interrogatives 21
DAVID DOWTY / On the Semantic Content of the Notion of
'Thematic Role' 69
GENNARO CHIERCHIA / Structured Meanings, Thematic
Roles and Control 131
GREGORY CARLSON / On the Semantic Composition of
English Generic Sentences 167
LENHART SCHUBERT and FRANCIS J. PELLETIER /
Generically Speaking, Of, Using Discourse Representation
Theory to Interpret Generics 193
HENK ZEEVAT / Realism and Definiteness 269
INDEX OF NAMES 299
INDEX OF SUBJECTS 303
v
TABLE OF CONTENTS
(Volume 1)
PETER ACZEL / Algebraic Semantics for Intensional Logics, 1

RICHMOND THOMASON / Motivating Ramified Type Theory
RA YMOND TURNER / Two Issues in the Foundations of Semantic
Theory
NICHOLAS ASHER and HANS KAMP / Self-Reference, Attitudes and
Paradox
MICHAEL JUBIEN / On Properties and Property Theory
GEORGE BEALER / Fine-Grained Type-Free Intensionality
JOHAN V AN BENTHEM / Semantic Type-Change and Syntactic
Recognition
VI
PREFACE
This collection of papers stems originally from a conference on Property

Theory, Type Theory and Semantics held in Amherst on March 13-16
1986. The conference brought together logicians, philosophers, com-
puter scientists and linguists who had been working on these issues
(often in isolation from one another). Our intent was to boost debate
and exchange of ideas on these fundamental issues at a time of rapid
change in semantics and cognitive science.
The papers published in this work have evolved substantially since
their original presentation at the conference. Given their scope, we
thought it convenient to divide the work into two volumes. The first
deals primarily with logical and philosophical foundations, the second
with more empirical semantic issues. While there is a common set of
issues tying the two volumes together, they are both self-contained and
can be read independently of one another.
Two of the papers in the present collection (van Benthem in volume
1 and Chierchia in volume II) were not actually read at the conference.
They are nevertheless included here for their direct relevance to the
topics of the volumes.
Regrettably, some of the papers that were presented (Feferman,
Klein, and Plotkin) could not be included in the present work due to
timing problems. We nevertheless thank the authors for their contribu-
tion in terms of ideas and participation in the debate.
The conference had a group of invited discussants which included
Emmon Bach, Janet Fodor, Erhard Hinrichs, Angelika Kratzer, Fred
Landman, Richard Larson, Godehard Link, Chris Menzel, Uwe Monnich,
and CarI Pollard. We thank them alI (along with the other participants)
for their stimulating and lively presence.
The conference was made possible by a grant from SDF to Barbara
Partee, along with additional support from the University of Massachu-
setts Vice-chancelIor, Samuel Conti, and from the Hampshire College
School of Communication and Cognitive Science"alI of which is grate-
fulIy acknowledged. Thanks also to Jay Garfield for help in arranging
the facilities at Hampshire College where the conference was held
Vll
viii PREFACE
and ta Kathy Adamczyk, Mary Ann Palmieri and the many graduate
students of the Linguistics Department at the University of Massachu-
setts for their cheerful and eftective help with organization and local
arrangements.
GENNARO CHIERCHIA
INTRODUCTION
It is quite uncontroversial that a semantic theory for natural language

must come to grips with the nature of intensional entities, such as
properties, relations and propositions. For example, it must specify
what it is to attribute a property to an individual and what information
bearing structure results from such an operation. More generally,
semantics must identify what types of entities are needed in a composi-
tional account of the way phrases are interpreted, as well as how such
entities are related to one another.
The research inspired by R. Montague's work provides us with a
quite substantial exemplification of how these questions can be both
fruitfully pursued and shed light on how language works. According to
Montague's original proposal, intensions are built out of individuals and
possible worlds. For example, basic properties can be analyzed as
functions from possible worlds into sets of individuals; propositions, as
sets of worlds. Using, then, the apparatus provided by type-theory, one
can build along similar lines more complex intensional structures (e.g.
properties of properties, relations having propositions as one relatum,
etc.), as they appear to be required for the interpretation of the rich
syntactic constructions one finds in naturallanguages.
Consequently, intensions are rather coarsely grained (any two logi-
cally equivalent expressions turn out to have the same intension) and
neatly encapsulated in some type-theoretic level. The resulting general
picture is that of a system of syntactic categories (that classifies
distributional properties of phrases), systematically associated with a
system of extensions and intensions (that classifies the contribution of
phrases to truth-conditions), where intensions are built along the lines
just sketched. In classical Montague grammar, each phrase is expected
to exist in exactly one such category (even though there can be, of
course, homophonous phrases).
One of the merits of Montague-inspired work has been to push this
picture to its limits, i.e. to see how far one can go with it. Such work has
enabled semanticists to locate the empirical side of problems with a
certain precision, as for example in certain aspects of anaphoric refer-
1
Gennaro Chierchia, Barbara H. Partee, and Raymond Turner (eds.), Properties, Types
and Meaning, Il. 1-20.
1989 by Kluwer Academic Publishers. Alt rights reserved.
2 GENNARO CHIERCHIA
ence, or in the analysis of belief sentences and sentence embedding.

Most of these problems appear to have a common source, which, in a
nutshell, could be characterized as follows. While the notion of possible
world might indeed be fundamental for the proper understanding of
intensionality, much available evidence suggests that its direct marriage
with type-theory is ill-fated. It leads to a theory of meaning which
appears to be not intensional enough and not flexible enough.
The analysis of attitude reports and various forms of embedding
appear to call for intensional entities that are more finely articulated
than functions from possible worlds into extensions. Arguably, they call
for intensional entities that are reflexive, i.e. capable of sustaining forms
of self-reference. Various other phenomena (e.g. the cross-categorial
character of Boolean opera tors such as and and or, the interpretation
of questions, etc.) suggest that several natural language constructions
have a truly polymorphic nature: their contribution to truth-conditions
cannot be uniformly projected into one type-theoretic level.
Volume I of the present collection is devoted to the study of the
logical and philosophical foundations of these two problems ("strong"
intensionality and altematives to type-theory) and their interrelations.
The present volume concentrates on more specifically empirical facets
of such problems. The papers collected here identify and discuss a
number of issues in which the above foundational questions actually
play a part. Such issues include the semantic analysis of questions, the
notion of "thematic role" (such as "agent" or "patient"), and various
problems related, broadly speaking, to anaphora, such as what in the
linguistic litera ture is known as "control" (i.e. the interpretation of the
"implicit subjects" of infinitivals), generic constructions, and the role of
definite NP's (such as the man with the brown hat).
We will now summarize the contributions to the present volume and
discuss briefly their relevance to the general issue of how finely and
how flexibly intensional entities should be structured.
GROENENDIJK AND STOKHOF:

QUESTIONS, INTENSIONALITY AND TYPE CHANGE
Groenendijk and Stokhof investigate aspects of the grammar of ques-

tions relevant in various ways to the issue of what type-systems are
needed in semantics. To get a feeling for'the sort of problems involved,
it may be useful to consider in very broad strokes some of the main
INTRODUCTION 3
strategies that are currently being pursued in connection with the

semantics of questions. Following Groenendijk and Stokhof, we can
classify the main semantic approaches to questions in two groups, the
categorial and the propositional. Let us briefly consider each in turn.
The categorial approach focuses on the conditions that make an
answer well-formed (or "standard"). For example, the knowledge that
(la) is a well formed answer to (lb) but not to (lc-e) seems to be part
of our semantic competence and must be accounted for.
(1) a. with a saw
b. how did Mary cut that?
c. what instrument was used to cut that?
d. did Mary use a saw to cut that?
e. who cut the tree with what?
From this perspective, a plausible starting point is the idea that the
semantic content of a question plus the content of the constituents that
make up its answers must yield the (true and complete) propositional
content of the answer. This leads one naturally to regard the meaning of
a question as an unsaturated stucture, say an (n-place) propositional
function. The meaning of the constituent used for the answer must
supply arguments of the right type. Consequently, to recover, for
example, the true and complete answer to (lb), it is sufficient to apply
its denotation to the denotation of (la). A further consequence of this
line of analysis is that questions and answers end up having meanings of
a non-uniform logical type.
While this approach has many appealing features (in particular, it
embodies a simple account of what makes an answer well-formed), it
faces some difficulties, mainly in two areas. For one thing, questions
can be conjoined and disjoined. E.g.:
(2) John knows who will bring the beer and who will cook what.
But on the categorial approach the first conjunct of the embedded
clause in (2) will denote something like al-place propositional func-
tion, the second, a 2-place one. It then becomes unclear how the
meaning of "and" in (2) relates to the meaning of the ordinary sentential
"and".
The second difficulty that the categorial appro'ach arguably faces has
to do with the relation between questions such as (3a) and (3b).
4 GENNARO CHIERCHIA
(3) a. who went to the rally?

b. did John go to the rally?
Intuitively, if we know the (true and complete) answer to (3a), we
thereby also know the answer to (3b). Our intuitions on the relatedness
of (3a) and (3b) resemble intuitions about entailments. In fact, one
plausible way of accounting for our intuitions relative to (3a, b) is by
generalizing the notion of entailment to questions. In very rough terms,
a reasonable first guess might be to say that a question a entails a
question {3 iff the true and complete answer to a entails the true and
complete answer to {3. This line of explanation, however, is barred from
a pure categorial approach, for according to it, questions and answers
lack a uniform type. This makes it hard to come up with a generaliza-
tion of the notion of entailment such as the one just sketched.
The propositional approach adopts a different point of view. The
meaning of questions is analyzed in terms of conditions for answer-
hood. Roughly speaking, this means in terms of the true and complete
answers to a given question. Thus the content of questions is regarded
as being essentialIy propositional. This offers a better point of depar-
ture for attempting an analysis of coordinatedinterrogative structures.
It furthermore enables one to use the notion of entailment in account-
ing for our intuitions of semantic relatedness as suggested in connection
with (3a, b).
However, these potential strong points of the propositional approach
are also the basis of its weaknesses. In this connection, Groenendijk
and Stokhof point out examples such as the folIowing:
(4) a. Who of Bill, John and Mary will go to the party?
b. Who of Bill, John and Mary will not go to the party?
Evidently, a true and complete (propositional) answer to one of the
questions in (4) also constitutes a true and complete answer to the
other. Hence, in terms of conditions for answerhood, the two questions
are content-synonymous. Yet answering, say, "John and Bill" to the
one conveys different information than it does for the other. How is
this contrast to be accounted for, if the two questions are content-
synonymous?
Thus, we seem to be at an impas se. Both approaches appear to
capture a sound insight into the semantic nature of questions. But
neither of them appears to be able by itself to accommodate alI the
INTRODUCTION 5
relevant phenomena. Groenendijk and Stokhof seek a way out of this

impasse in terms of a more flexible approach to semantic types aud to
their relations with syntax.
They lay out a number of basic phenomena concerning coordination
and entailments that illustrate the extent of the aforementioned prob-
lems. They then develop a variant of a propositional theory of ques-
tions, based roughly on the idea that simple questions can be aualyzed
as functions that at any given world pick out the proposition that
determines their true and complete answers. They then show that it is
possible to provide an enlightening analysis of complex patterns of
coordinated interrogative structures on the basis of simple type-shifting
principles (analogous to those developed in connection with cross-
categorial conjunction aud disjunction - see e.g. Partee and Rooth
(1983), Partee (1986)).
Groenendijk and Stokhof try to push the type-shifting perspective
further in order to attempt an interesting unification of the proposi-
tional perspective with the insights that stern from the categorial one.
Their semantics for questions can be construed as a function that takes
n-place propositional functions as inputs and yields the propositional
content of answers as output. It can, therefore, be regarded as a type-
shifting principle, rather than as a semantic rule of the usual kind.
Interpreting the semantics for questions as a type-shifting principle has
a number of far reaching consequences for the overall organization
of the grammar. Groenendijk and Stokhof discuss some such con-
sequences. (Van Bentham's paper in volume 1 is also relevant in this
connection).
The semantic theory developed by Groenendijk and Stokhof is cast
in terms of possible worlds semantics. It therefore inherits the problems
of the latter, as they point out. Consider for example the embedded
questions in (5).
(5) a. Mary wonders whether second order logic is categorical

b. Mary wonders whether S4.3 is complete.
Clearly, each one of these sentences can be true without the other
being true. Yet, no unaugmented possible worlds analysis of the content
of the embedded questions in (5) will be able to distinguish them.
Examples of this sort could be generated for interrogative structures
of any kind. So, it is clear that one needs, in this respect as well, a more
6 GENNARO CHIERCHIA
intensional notion of semantic content. The work of Groenendijk and

Stokhof paves the way for the development of a more intensional
semantics for questions. It does so in two ways: by isolating the
empirical domain that such a theory has to come to grips with, and by
individuating some of the theoretical lines along which such a theory
will have to proceed.
DOWTY: THEMA TIC ROLES AND EVENTS
Thematic roles (i.e. notions like agent, patient, goal, etc.) have often
proven to be quite useful semantic labels in stating linguistic generaliza-
tions of various sorts. However, their vagueness and the lack of
agreement on their defining criteria have cast serious doubts on their
status. In fact, until quite recently, the possibility of providing a sound
characterization of thematic roles appeared to be beyond the limits of
logical semantics. Dowty's paper is devoted to a discussion of the
formal foundations of a theory of thematic roles. Dowty considers in
particular two views of thematic roles, their relations and their respec-
tive ranges of applicability. Let us consider briefly the two approaches
envisaged by Dowty.
The first theory is based on what Dowty caUs the "ordered-argument
view" of predicate structures. This is the standard view that relations
have a fixed adicity. An n-place relation is an unsaturated structure that
takes an ordered sequence of n arguments to yield something like a
proposition. Systems of relations form complex structures in which the
relata of any given relation end up having properties of various kinds.
For example, suppose that x bears the kill-relation to y (i.e. suppose
that x kiUs y). Then x is the entity that causes the event to take place.
The (relational) property of causing the relevant event is something that
x has in virtue of occupying a certain slot in the kill-relation. Such a
property - or something like it - might be taken as defining agentivity.
This suggests that one can take thematic roles to be properties of
argument slots of relations. Such properties can be detected from the
entailment patterns associated with relations (i.e. "x kills y" entails "a
killing occurs and what x does causes it").
Dowty individuates several features that one would want such a
substantive system of thematic roles to have. Every slot in a (natural)
relation should be associated with one thematic role and no two slots
should be associated with the same one. Furthermore, the properties of
INTRODUCTION 7
a certain slot in a relation should be definable independently of that

relation itself (this is to rule out trivial properties like "being the nth
slot of relation a").
Such criteria are desirable, if one wants thematic roles to be useful
classificatory devices for the kind of relations that can be lexicalized in
the languages of the world. Any system of classificatory criteria should
be as far as possible exhaustive and non redundant. If a system of
thematic roles with such characteristics could be found, we would gain
powerful insights into the organization of the lexicon and the concep-
tual system underlying it.
The second theory that Dowty considers he dubs "Neo-Davidsonian"
for its similarity to Davidson's well-known theory of action sentences.
On such an approach, events are taken as primitives and thematic roles
are viewed as relations between events and individuals. Such relations
satisfy a uniqueness condition (for example, there is at most one agent
for any given event), hence we can regard them as functions from
events into individuals, who are, intuitively, the participants of the
event. So for example, the sentence x kills y, which on the standard
view is represented as in (la), could be analyzed from a Neo-
Davidsonian perspective as in (1 b):
(1) a. kill(x,y)
b. 3e[killing(e) occur(e) agent(e) =
x patient( e) = y] .
Intuitively, the two formulae in (la) and (lb) have the same truth-
conditions. This fact can be exploited in various ways. We can take
thematic roles as primitives, and define (la) in terms of (1 b), which
would constitute a radical interpretation of the Neo-Davidsonian
approach. Or we can take the opposite route, i.e. as sume (la) as basic
and define thematic roles as functions on the events that a certain
relation individuates (in the spirit of the ordered argument approach).
In fact, Dowty suggests that both the ordered argument and the Neo-
Davidsonian approach might be needed in different domains. The
ordered-argument theory maintains that relations with too few (or too
many) arguments yield ill-formed structures. This seems to correspond
to the way verbs work in languages like English. It is a commonplace
observation that arguments cannot be freely a"dded or deleted from
the argument structure of a verb. A natural explanation for this fact
8 GENNARO CHIERCHIA
would be that verbs correspond to predicates that are typed as to the

number of arguments they can take. an the other hand, on the Neo-
Davidsonian approach, predicates do not have an intrinsic adicity and
the number of thematic roles defined for a particular event-type can be
left unspecified. This fits well with the behavior of nominals, where
arguments appear to be basically always optional (contrast, e.g., ?? lohn
gives vs. lohn's gijt, etc.). One can try to explain this fact by regarding
nominals as referring to events whose arguments are specified using
thematic roIes as in the Neo-Davidsonian approach.
Dowty considers various evidence in favor of this view that comes
from the behavior of by-phrases and from anaphoric reference to
events. Both these phenomena work different1y in the nominal and in
the verbal domain and their differences might be explained in terms of
the different ways in which verbs and nouns take arguments.
Dowty then considers several problems with the idea that relations
lexicalized in natural languages can be exhaustively classified in terms
of thematic roles. Such problems arise in various areas. For example,
many generalizations stated in terms of thematic roles like, say, agent,
turn out to involve different and sometimes contrasting notions of
agentivity. Moreover, it appears to be difficult to extend the most
commonly invoked thematic roles beyond a limited class of core cases
so as to cover the entire lexicon. Dowty discusses various ways in which
the requirement that thematic roles form a complete and exhaustive
classificatory system could be weakened, and how various psycho-
linguistic issues bear upon this.
Dowty's discussion of thematic roles clarifies their link to the notion
of event and to the roIe that relations play in classifying event-types.
This raises interesting issues conceming, for example, the relations
between events and propositions in natural language. One such issue,
widely discussed in the literature, is the nature of the relation between
the denotation of a sentence such as lohn walks and the denotation of
the corresponding derived nominal lohn's walk. Saying that the former
is a proposition and the latter an event does not take us very far, unless
we spell out the structural characteristics of events vs. propositions. A
better understanding of this relation is crucial to Dowty's hypothesis
that verbs and nouns take arguments in different ways, and hence that
thematic roles have to be understood different1y in these two domains.
To see this point, consider for example the following problem. We
INTRODUCTION 9
have pointed out above that a relation of the kind shown in (2) should
hold between event-denoting and proposition-denoting structures:
(2) VX D [3e[[walk]N(e) /\ Agent(e) = x] .... [walk]v(x)].
Now, questions that arise are these: Do we have to list axiom
schemas such as the one in (2) for alI derived nominals? Or can they be
derived as theorems from a general hypothesis on the nature of the
relation between events and propositions?
Further related questions having far-reaching consequences for the
overall structure of semantic theory come up in this connection. For
example, a consequence of Dowty's Neo-Davidsonian approach to
thematic roles is that events have in some sense an internal articulation
akin to the articulation in constituents of sentences. Thematic roles in
Dowty's sense may be viewed as functions that select the various
"constituents" of an event. Now on the classical possible worlds
construal, propositions do share certain structural characteristics of
sentences (e.g. they can be conjoined or disjoined and they enter
entailment relations). They lack, however, an articulation into con-
stituents. Hence, on the basis of the possible worlds approach, thematic
roles could not be defined direct1y on propositional structures as such.
However, as we have seen in the preceding sections, various
considerations suggest that levels of propositional structure finer than
the one characterized in possible worlds terms might well be needed.
This opens up the possibility of defining thematic roles directly on
propositional structures, which might affect in various ways the picture
that Dowty offers us. Such a possibility is explored in Chierchia's paper.
CHIERCHIA: THEMATIC ROLES AND PROPOSITIONS
Chierchia's starting point is the notion of "structured meaning". Struc-

tured meanings are appealed to by various authors mainly in connec-
tion with the problem of logic al omniscience. The basic insight, which
goes back to Carnap (1947) and has subsequently been investigated by
several authors, is that certain forms of belief and other mental attitudes
appear not to be closed under any non-trivial notion of equivalence,
and thus seem to crucially involve structures that are as fine-grained as
the structure of sentences (or, perhaps, as the structure of formulae in
some "logical form"). One way of pursuing this Strategy is by assuming
10 GENNARO CHIERCHIA
that sentences can be interpreted at some level as structured objects,

say as trees isomorphic to syntactic phrase markers whose terminals are
not words or morphemes but "meanings" (Le. individuals and relations).
The resulting propositional structures (Le. the structured meanings)
have, thus, an internal articulation in terms of constituents and bear
some resemblance to Frege's and Russell's conception of the nature of
propositions.
Chierchia argues that something like structured meanings supports a
very simple view of thematic roles. Thematic roles can be viewed as
functions that select a constituent of an (atomic) propositional structure
on the basis of the entailments associated with that constituent. Thus,
for example, the structured meaning associated with, say, Pene Iope kills
Arabella is something like (kill', P, A), where kill' is the kill-relation,
P is Penelope and A is Arabella. Agent( (kill', P, A )) will select
the constituent of (kill', P, A) which, if Penelope kills Arabella
(Le. if (kill', P, A) occurs), causes the killing to take place; thus
Agentkill', P, A)) = P.
This captures the main insight of Dowty's ordered-argument approach,
since thematic roles are defined in terms of the semantic properties (Le.
entailments) associated with the slots of a relation. But thematic roles
are defined direct1y on propositional structures and depending on how
the issue of the relation between propositional structures and events is
settled, Chierchia's view might also have some points in common with
the Neo-Davidsonian approach, where thematic roles are construed as
functions from events into their participants.
Chierchia tries to motivate this particular approach to thematic roles
on grounds different from Dowty's (even though most of the points
Dowty makes could be made also on the basis of Chierchia's approach).
Chierchia looks in particular at the semantics of what in the linguistic
literature is called "control". This phenomenon is observable in sen-
tences like Mary promised John to leave or Mary forced John to lea ve.
The infinitive to leave in these examples lacks an overt subject but is
understood as if it had one. The implicit subject of the infinitive is the
matrix subject with promise-type verb and the matrix object with force-
type verbs. One question that arises is what principles, if any, determine
the selection of the "controller" (Le. the understood subject) of infini-
tives. It does not seem a priori implausible to maintain that tlie
selection of the controller is somehow linked to the difference in
meaning between promise and persuade, as it manifests itself in the
INTRODUCTION 11
different entailments that these verbs associate with their argument

slots. Thematic roles are designed as ways of classifying such entailment
patterns. Thus, they might be of help in stating generalizations concern-
ing controller selection. This is an idea put forth by Jackendoff (1972),
and Chierchia tries to reconstruct, explore and defend aspects of it
within a model-theoretic setting.
Chierchia's theory of thematic roles (and several details of his theory
of control) seem to depend crucially on the viability of the hypothesis
that propositional content can be given, at some level, a high degree of
articulation. This links the question of thematic roles rather direct1y to
the issue of fine grainedness in semantics, an issue that theories of
properties and propositions have undertaken to address.
CARLSON : GENERIC PROPERTIES AND GEN E RIC PREDICATION
Genericity constitutes a traditionally difficult area of analysis and has

immediate relevance to many foundational issues in semantics. What is
genericity? As Carlson puts it, from a notional point of view, generic
sentences express a regularity, tendency or habit, as opposed to the
reporting of an event or the description of an episode (contrast, e.g.,
cats meow, vs. cats are meowing because they want ta be fed). The truth
of a generic statement, furthermore, is not normally specified relative to
a circumscribed spatiotemporal region. It somehow involves looking at
a certain pattern across sets of circumstances.
From a linguistic point of view, generic sentences can be individ-
uated in terms of a cluster of syntactic and semantic properties. In some
languages (though not in English), they are marked by overt "nomic"
morphemes. Generally, their verbal aspect is stative. In languages that
have a perfective/ imperfective distinction, genericity is expressed by
means of imperfective forms . From a semantic point of view, they
systematically create intensional contexts (ef. e.g. John eats with a knife.
For this sentence to be true, there need not be any particular knife with
which John eats).
The intensional character of generic constructions is perhaps the
main focus of Carlson's approach. In previous work, Carlson (1980)
has strongly argued for the fruitfulness of a type/token distinction that
cuts across the whole domain of individuals. Ordinary individuals like,
say, John, have spatiotemporally bounded marufestations (or stages).
Individuals as such are what tie individual stages together as manifesta-
tions of one and the same entity. They are therefore intensional entities,
as they have varying extensions (Le. manifestations) at different spatio-
temporal locations. Carlson has further argued that besides ordinary
individuals (and their stages), there are kinds. These are the denotata of
noun phrases like cats, blue striped suits and white gold. Unlike those
of ordinary individuals, the spatiotemporal manifestations of kinds can
be scattered or discontinuous. This is so, of course, because kinds have
ordinary individuals as their instances.
So the domain of entities is articulated into individuals and kinds, on
the one hand, and their stages on the other. This classification leads to a
strikingly simple notion of genericity. A generic statement is about an
"intensional" entity, namely an ordinary individual or a kind. A non
generic statement is about stages, i.e. specific, spatiotemporally located,
"extensional" entities.
Carlson (1980) develops this view in detail and produces an impres-
sive amount of evidence in favor of it. Ris work has generated a lively
debate on the semantics of generics that permeates much ongoing
research.
Let us reflect briefly on the nature of the instantiation relation that
determines the special status of kinds by linking them to ordinary
individuals. A long standing semantic tradition analyzes common nouns
like cat as properties (or propositional functions). Carlson's point is
that each such property (including complex ones like those correspond-
ing to blue striped suits) has a special sort of individual correlate,
namely a kind. Now, for any common noun denotation P, let k(P) be
the corresponding kind. According to Carlson, the kind and the
property related to it are connected as follows:
(1) VPVxD[R(k(P),x)'" P(x)]
where R ("realizes") is the instantiation relation (Le. R(k(P), x) is to be
read "x realizes or is an instance of kind k (P)".
What should be noted concerning (1) is that it clearly links the
instantiation relation involving kinds to predication (or, if properties are
analyzed as propositional functions, to functional application). Principle
(1) makes R and predication coextensional in every world. Thus kinds
are individuals that in this respect behave just like properties.
The parallelism with property theories could hardly be more striking.
Certain property theories (e.g. Bealer's or Jubien's, ef. their papers in
volume 1) take properties to be individuals and link them to their
INTRODUCTION 13
instances via a predication relation, just as Carlson links kinds to their

instances. Other property-theories analyze properties as propositional
functions and provide ways in which such functions can be "nomi-
nalized", i.e. embedded in the domain of individuals. This, again, gives
rise to a relation between nominalized propositional functions and their
instances, structurally isomorphic to the one in (1). This suggests that
property-theory might be a prime candidate for the analysis of the
notion of kind. The interesting point in this connection is that Carlson
arrived at the notion of kind and the principles that govern its structure
on empirical grounds, to account for genericity. The structures hypothe-
sized by Carlson, however, turn out to be closely related, if not plainly
isomorphic to, property-theoretic structures (cf. on this Chierchia
(1982,1984)).
An important aspect of Carlson's proposal concerns the specific
hypothesis of how genericity compositionally arises from the meaning
of the constituents involved. In this regard, the analysis of Carlson
(1980) assigns to generics a subject-predicate structure, where the
subject provides the key ingredient for genericity, namely an individual
of the right sort. In the present paper, Carlson points out several
inadequacies of such an analysis, while maintaining the overall picture
sketched above.
What leads to a different view of the compositional structure of
generic constructions is the observation that there are generic sentences
that seem to involve no individual at all, or at least no subject. Weather-
verbs with expletive subjects are a prime example.
(2) a. It rains 30 a year, here

H
b. It snows in Florida once every decade or so

(Carlson's (15a-b)).
Clearly, the sentences in (2) are generic. Yet their compositional

semantic structure cannot be easily analyzed as predicating something
of a generic subject.
Carlson considers a variety of constructions of this sort (involving
multi-place verbs and adverbials of various kinds) and argues that
genericity should be analyzed as a relation between a complex predi-
cate and some other constituent.
Part of the evidence in favor of this relatioIiul view of genericity
comes from examples such as those in (3).
(3) a. It rains
b. it snows.
The sentences in (3) lack a generic interpretation. This suggest that
the availability of a generic interpretation in (2) is specifically linked to
the presence of adverbials. Thus in this case, genericity arises from a
relation between the predicate rains and the adverbial that modifies it.
Carlson argues that this generic relation creates contexts that are
intensional in varying degrees. The predicational view of generics
developed in his previous work is recast as a special case of the
relational view.
Carlson's work on generics raises a wide variety of interesting issues
for semantic theory, both in terms of the breadth of the phenomena
involved and of the intrinsic interest of the theoretical apparatus
developed to deal with them. The latter, in particular, leads one to ask,
as we have seen, various central questions concerning the very nature of
properties and predication, as they manifest themselves in the semantic
systems of the languages of the world.
SCHUBERT AND PELLETIER:

GENERICS AND DISCOURSE REPRESENT A TION
Pelletier and Schubert develop a fairly detailed proposal concerning the

way generic sentences should be logically represented and interpreted.
Such a proposal is based on a "dynamic" view of how discourse is
interpreted and bears in various ways on the issue of fine-grainedness
in intensional structures.
They discuss a wide range of constructions involving genericity, from
simple subject predicate sentences like
(1) John is an excellent marksman
to sentences with indefinites and bare plurals, like
(2) a. if a cat is dropped, it usually lands on its feet
b. cats usually land on their feet.
The main aspect of their proposal is that the interpretation of
generics requires a set of cases that provides the background with
respect to which generic statements are'evaluated. They call the back-
ground assumptions relevant to the semantics of generics the "reference
INTRODUCTION 15
ensemble". Sometimes, such a reference ensemble is explicitly provided,

as with conditionals (ef. (2a. The role of the antecedent in (2a) is to
set up the relevant reference ensemble. Intuitively, something like (2a)
is true if most of the situations in which cats are dropped, are situations
where they land on their feet. In certain cases, such as (1) and (2b), the
reference ensemble is implicit and has to be reconstructed from the
context.
In trying to spelI out this approach, one immediately runs into
several well known problems having to do with conditionals and the
anaphoric links (involving pronouns and tenses) that they license. For
example, in the consequent of (2a) we find a pronoun anaphoricalIy
linked to the indefinite NP a cat, which occurs in the antecedent.
NormalIy in matrix clauses, indefinites receive an existential interpreta-
tion. However, if we try to represent a cat in (2a) in terms of an
existential quantifier, we get the wrong results. We have essentialIy two
possibilities, represented by the farmulae in (3).
(3) a. 3x [cat(x) A drop on the floare x)1 -> falI on its feet (x)
b. 3x[cat(x) A droponthefloor(x) -> fallonitsfeet(x)l.
In (3a) the scope of the existential quantifier is trapped within the
antecedent of the conditional. Hence the third occurrence of x (i.e. the
one that corresponds to the pronouns) is not bound and its value is
unrelated to the value of the first two occurrences. In (3b) the
existential quantifier does bind alI the variables, but this sentence would
be true in any situation that had something which is not a cat in it, even
if in such a situation every cat that was dropped on the floor fell on its
head. This doesn't give us the right truth-conditions for (2a).
This problem is known in the Iiterature as the problem of "Donkey
anaphora", from a famous illustration of it due to Geach. Donkey
anaphora is being actively investigated within several frameworks. Two
particularly important related attempts are represented by File Change
Semantics, developed by Heim (1982) and Discourse Representation
Theory (developed by Kamp (1984.
The problem of donkey anaphora also concerns tense, as (4)
illustrates:
(4) When a cat drops on the ground, it usuaUy lands on its feet.
Here we have a Iink between the temporal occasions of the drop-
pings and the occasions of the feet first landings. These constructions
have been studied in Partee (1984), within the framework of Discourse
Representation Theory.
Schubert and Pelletier's approach to this is the folIowing. They
represent (2a) as folIows:
(5) [(3x donkey (x drop on the floor (x)] -+ falI on feet (x).
Here we have an implication, and the scope of the quantifier is
limited to the antecedent (as in (3a . However, part of the contextual
information with respect to which sentences are evaluated, is made up
of previous discourse. In particular, the antecedent in (5) can be taken
to provide the context with respect to which the consequent is evalu-
ated. What the antecedent does in this connection is restrict the range
of assignments to the variable x in the consequent. Such assignments
are restricted to those values that satisfy the antecedent. In this way,
value-assignments to variables can, in a sense, be carried over beyond
their actual scope.
TechnicalIy, this is accomplished by recursively defining a context-
changing function O, which (in simple cases) can be viewed as mapping
pairs of formulae and contexts into new contexts. So, in a discourse
~1' . . , ~n' we first interpret ~1 with respect to a context c. We then
interpret ~2 with respect to the context as modified by ~1 (Le. 0(~1' c,
and so ono
This idea is quite similar to the one of Kamp and Heim. However it
is implemented in such a way as to yield an interesting empirical
difference. Consider sentences such as the folIowing:
(6) It I find a dime in my pocket, 1 will put it in the meter.
The Kamp-Heim semantics is based on the idea that conditionals are
linked to Lewis-style uns elective universal quantifiers and that indefi-
nites are interpreted essentially as free variables. Thus indefinites in the
antecedent of a conditional end up being universally quantified. While
this appears to be right for many cases, it seems to be intuitively wrong
for examples such as (6). On the approach proposed by Schubert and
Pelletier, (6) comes out as existentially quantified, while generic sen-
tences with conditionalsend up picking up universal force. ,
This is because generics involve iterated evaluations over the non
generic constructions that are used to setbp the reference ensemble.
Thus on Schubert and Pelletier's theory, indefinites are interpreted
INTRODUCTION 17
as existentially quantified. But they can be used to constrain the context

on which subsequent discourse builds, and in particular, they constrain
the interpretation of pronouns that fall outside their scope. Genericity
consists of iterated evaluations of the constructions so obtained.
Schubert and Pelletier also discuss how their approach can be used
to link readings of bare plurals where they denote kinds to generic
quantifications over instances of such kinds.
An interesting problem for Pelletier and Schubert comes from
sentences like the following:
(7) A cat is usually intelligent, if it has blue eyes.
Roughly speaking, (7) means that most cats with blue eyes are
intelligent. However, notice that the position of the pronoun and its
NP-antecedent are reversed here compared with (2a). So the indefinite
NP a cat ends up being in the consequent of the conditional and one
would expect that it could not be used to suitably restrict to cats
quantification over the relevant cases.
The problem is, of course, that somehow we want to use the
information contained in the NP a cat in the antecedent. But how can
we do this compositionally? Pelletier and Schubert argue that a solution
calls for more richly articulated notions of intension. We need to be
able to recover from the intension of a sentence its articulation into
functions and arguments. We would then be able to use at the right
level the information contained in the consequent of (7). They sketch a
proposal along these lines, based on the notion of structured meaning.
Thus, the semantics of generics provides yet another empirical testing
ground for strongly intensional theories of propositions.
While many problems are left open, Schubert and Pelletier's paper
presents an interesting perspective on the variety of issues that are
involved in this complex area of inquiry.
ZEEV AT: DEFINITES AND DISCOURSE REPRESENTA TION
Many papers in the two volumes of the present work address the
problem of the relation between propositions and representations, a
problem that becomes compelling as propositions become more and
more fine-grained and sentence-like. The problem can of course be
viewed as an instance of the more general issue of "the intention of
intensionality" (to borrow Hintikka's words) or, in rough terms, the
aboutness of information bearing structures. Zeevat's paper addresses

these issues from a specific point of view. Re discusses the notion of
familiarity and its role in a theory of definiteness. Building on the work
of Reim (1982), he investigates various uses of definite descriptions
and the extent to which their contribution to meaning can be viewed as
an instruction to select a referent that is "familiar" to the interpreter in a
communication act. In particular, he argues that uses of definites can be
classified, as a first approximation, in six main groups that he calls
convention, deJinition, demonstration, experience, communication and
relation. Re concludes that in spite of the wide range of ways in which
definites are employed, their functioning does indeed appear to be
govemed ultimately by familiarity. Re sketches an implementation of
this view within a Kamp-style theory of discourse grammar.
A comparison is then drawn between the resulting approach and
more classical ones such as Frege's and Russell's. A central aspect of
both such classical approaches, according to Zeevat, is that the noun-
part of a definite description is taken to provide a concept that
identifies uniquely a referent (if the definite description is successful).
Thus, purely logical notions (such as that of an object uniquely falling
under a concept) suffice, on the classical view, to explain the behavior
of definites. In contrast with this, the familiarity theory calls for an
essentially psychological notion: "familiar" means accessible to the
mind. Consequently, the general picture of discourse dynarnics changes.
According to the familiarity theory, as discourse unfolds we build
representational structures articulated into concepts. These concepts
are what we are familiar with and which get reidentified and retrieved
in subsequent information exchanges. Some such concepts encode
information about extramental entities, other about concepts in other
minds.
As the latter remarks suggest, such a perspective has a further
consequence: it naturally carries with itself a general strategy for dealing
with outstanding puzzles concerning reference to fictional entities and
identities across belief contexts, such as those exemplified by Geach's
famous witch-mania reports (e.g. "Rob believes that a witchi stole his
donkey and Nob that shei killed his cow"). In rough terms, what must
be involved here, if the familiarity theory is right, is reference to (and
identity of) shared representations of some kind. Per contra, a classical
Frege-Russell approach leads semantics io expand the ontology (Le. the
domain of individuals) by introducing something like non actual
INTRODUCTION 19
possibilia, indeterminates, abstract objects and the like. In this connec-

tion, the point that Zeevat tries to make is that theories of the latter sort
can provide a satisfactory semantics for definite NP's only by taking
into consideration discourse-based psychological factors, just like those
that a familiarity theory of definiteness is built ono
So, the idea of a representation anchored to an object (be it another
representation or an extramental entity) plays a central role in a theory
of discourse. Zeevat caUs it an "intention". An attempt at further
spelling out this notion and its role in discourse is sketched in a formal
appendix.
Thus, if the familiarity theory of definiteness is right, semantics
appears to be committed to a form of representationalism. Now, there
are a number of standard problems commonly associated with repre-
sentationalism. For example, in virtue of what can arepresentation
carry information about the world? What is it that makes representa-
tions intersubjective and thus viable tools for communication? Zeevat
addresses some of these central knots. The key to their solution has its
basis in the correspondence between mental representations (i.e.
thoughts) and facts. This correspondence is defined, elaborating on
Kamp's work, in terms of a mapping that embeds representations in a
model, and thereby determines their truth-conditional import.
Propositional structures, as they emerge from Zeevat's proposal,
appear to be sentence like entities whose building blocks are linked to a
semantic value, itself possibly (though not necessarily) a representa-
tional structure. This appears to be very close to Cresswell's (1985)
notion of structured meaning (discussed for example in the papers of
Chierchia and Schubert and Pelletier), and makes more urgent a close
comparison of the two approaches. Zeevat doesn't engage in this
difficult task, but the issues he addresses will be helpful for future work
in this area.
Classical Montague semantics and the variety of related approaches

that have subsequently been developed have led the field to a state
where quite sophisticated formal tools of logic can be fruitfully applied
to complex empiric al domains. Many problems in natural language
semantics can now be formulated and pursued in ways that would
otherwise not be open to us. The papers collected in the present
volume provide, we believe, a good overview of empirical problems
whose solution calls for redesigning intensional structures and semantic
categories. They also give a good idea of the main strategies that are
being explored in this connection. The search for the right texture of
information bearing structures and the role of semantic types (together
with a more dynamic understanding of how discourse unfolds), of
which the research presented here and in volume 1 of the present work
provides a good sample, is likely to remain at the heart of an exciting
theoretical debate for the next several years.
REFERENCES
Carlson, G.: 1980, Reference to Kinds in English, Garland, New York.

Carnap, R: 1947, Meaning and Necessity, Chicago.
Chierchia, G.: 1982, 'Nominalization and Montague Grammar', Linguistics and
Philosophy 5, pp. 303-354.
Chierchia, G.: 1984, Topics in the Syntax and Semantics of Infinitives and Gerunds,
unpublished Ph.D. Diss., University of Massachusetts at Amherst.
Cresswell, M.: 1985, Structured Meanings, MIT Press.
Heim, 1.: 1982, The Semantics of Definite and Indefinite Noun Phrases, unpublished
Ph.D. Dissertation, University of Massachusetts at Amherst.
Jackendoff, R: 1972, Semantic Interpretation in Generative Grammar, MIT Press,
Cambridge.
Kamp, H.: 1984, 'A Theory of Truth and Semantic Representation', in J. Groenendijk,
T. Janssen, and M. Stokhof (eds.), Truth, Interpretation and Information, Foris,
Dordrecht.
Partee, B.: 1984, 'Nominal and Temporal Anaphora', Linguistics and Philosophy 7, pp.
243-286.
Partee, B.: 1986, 'Noun Phrase Interpretation and Type-Shifting Principles', in J.
Groenendijk, D. de Jongh, and M. Stokhof (eds.), Studies in Discourse Representa-
tion Theory and the Theory of Generalized Quantifiers, Foris, Dordrecht, pp. 115-
144.
Partee, B. and Rooth, M.: 1983, 'Generalized Conjunction and Type Ambiguity', in R
Bauerle, C. Schwarze, and A. von Stechow (eds.), Meaning, Use and Interpretation
of Language, de Gruyter, Berlin, pp. 361-383.
JEROEN GROENENDIJK AND MARTIN STOKHOF
TYPE-SHIFTING RULES AND THE

SEMANTICS OF INTERROGATIVES
O. INTRODUCTION
The aim of this paper is a modest one. In what foHows, we will argue
that if one takes into consideration certain constructions involving
interrogatives, a flexible approach to the relationship between syntactic
categories and semantic types may be of great help. More in particular,
we will try to show that if one uses something like an orthodox inten-
sional type theory as one's semantic tool, a more liberal association
between syntactic categories and semantic types becomes imperative.
However, we will also see that such flexibility is by no means easily
introduced into the grammar, and that it needs to be properly checked
in order to avoid undesirable consequences.
The paper tries to make both a descriptive and a methodological
point. First of aH, we want to demonstrate that type-shifting rules, when
combined with general notions of coordination and entailment, are
useful tools in the semantic description of various constructions involv-
ing interrogatives. And second, we hope to show that they are impor-
tant methodological tools as well, which can guide us in finding the
proper semantic types for interrogatives, and in arriving at a 'unifica-
tion' of the two major approaches to the semantics of interrogatives: the
categorial approach and the propositional approach.
The constructions involving interrogatives which we will be con-
cerned with in this paper, are mainly coordination of interrogatives and
entailment relations between them. Coordinated interrogatives, i.e.
conjunctions, sequences, and disjunctions of interrogatives, may appear
to be pretty rare phenomena and not be worthy of too much attention.
Similarly, entailment between interrogatives may seem a questionable
thing. Entailment is defined in terms of truth (conditions), and aren't
questions the prime example of sentences that are not true or false?
True, but there are many other kinds of expressions that, as such,
cannot be said to be true or false either, but of which we can meaning-
fully say that the one does (or does not) entail the other. In fact, this
hoIds for aH conjoinable expressions, i.e. alI expressionsof a semantic
21
1989 by Kluwer Academic Publishers. AII rights reserved.
22 JEROEN GROENENDIJK AND MARTIN STOKHOF
type of the form ( ... t). For aU such types one can define in a general
schematic way, what coordination, conjunction and disjunction, within
such types amounts to. In a similar way, a general definition can be
given for entailment which teUs us for any two expressions of any
particular conjoinable type under what conditions the one entails the
other. The inductive basis of this definition is, as is to be expected, that
of entailment between expressions of type t, entailment between indica-
tive sentences.
Entailment is a fundamental semantic notion. Other basic semantic
notions, such as synonymy, antinomy and meaning overlap, can be
defined in terms of it. And in descriptive semantics, one of the major
goals is to account for semantic phenomena in terms of these and
similar notions. This holds for interrogative constructions as much as it
does for the more familiar indicative ones.
Being the fundamental semantic notion that it is, entailment, espe-
cially when it is combined with generalized notions of coordination, is
also an important methodological and heuristic too1. Semantic theories
can be evaluated with its help, and this holds for theories of the
semantics of interrogatives, too. It a particular theory assigns a certain
kind of semantic object, of a certain semantic type, to interrogative
sentences, we can test it by applying these general definitions, and see
whether the interpretation it gives to coordinated interrogatives and the
predictions concerning entailment relations it makes on the basis of
these definitions, stand to reason. Consequently, entailment and gener-
alized coordination will help to find the right semantic types for
interrogative sentences, and the right kind of semanticobjects within
these types to serve as their interpretation.
We wiU argue that the most adequate theory will assign a number of
different semantic types to interrogatives, depending on the syntactic
construction in which they occur. Type-shifting rules will play an
important role in incorporating the results in the grammar. One of the
most striking features of type-shifting is that it allows for flexibility in
associating semantic interpretations with expressions. With the help of
generally defined semantic operations, a basic interpretation of an
expression can be lifted and shifted to derived interpretations. So, one
and the same expres sion can have a wide variety of possible interpreta-
tions, which can be chosen from in different contexts.
Type-shifting can be put to different'uses. E.g., as a descriptive tool,
it plays a role in the analysis of coordination. Let us give a familiar
THE SEMANTICS OF INTERROGATIVES 23
example. For reasons of simplicity and elegance, it is attractive to assign

to proper names, and possibly certain other NP's as well, a semantic
object of type e as their basic interpretation. This cannot be the only
type for proper names, however, since they can also be conjoined with
other proper names, and with other kinds of NP's. Type e is not the
right type to apply coordination to. Therefore, in the context of
coordination proper names should rather be interpreted as denoting the
set of properties of an object of type e. I.e. we need a second interpreta-
tion of proper names, that of objects of type (( e, t >, t >, which is also the
lowest possible type for quantified NP's. We will see that, for similar
reasons, such shifting in meaning is also required for coordination of
interrogatives.
Apart from this rather 'standard' use of type-shifting in the semantics
of interrogatives, involving well-known lifting and shifting principles,
there is something more. Among the proposed semantic theories for
interrogatives, two approaches can be distinguished: the categorial
approach and the propositional approach. One major difference
between the two is that they assign different types to interrogatives. So,
one may rush to condude, '(at least) one of them must be wrong'. No,
not necessarily so, according to a flexible, type-shifting methodology. If
the types employed by each of the two approaches can be related to
each other by means of a significant uniform semantic operation, both
might prove to be (at least) partially right. We want to argue that there
are reasons to look upon things this way. We will show that the
successes and failures of the categorial approach and those of the
propositional approach are complementary, and that by providing a
more flexible theory that combines the two, we can add up their
successes and eliminate their failures. However, although we will see
that such a unification of the two approaches is possible, the question
remains whether the semantic operation that is needed to get from the
categorial type of interpretation to the propositional type, can really be
viewed as a general type-shifting procedure. It certainly is not an
orthodox one, and one might say that rather than adducing further
support for a flexible approach, it raises foundational questions. If a
flexible approach is to be more than a mere technical descriptive
device, i.e. if it is to be part of a substantial theory about the relation-
ship between syntax and semantics, it has to be based on restrictive
principles. "-
The paper is organized as follows. In section 1 we give a rough
sketch of the ideas underlying the categorial and propositional

approaches, outline why the two cannot be straightforwardly combined,
and indicate how this probIem can be solved. Section 2 deals with
coordination of interrogatives and entailments between them, and
discusses the various types in which interrogatives should be anaIyzed.
In section 3, a flexible approach is developed which deals with the facts
discussed in section 2 and which overcomes the difficulties indicated in
section 1. Section 4, finalIy, sums up the results.
A final remark in this section concems terminology. In what fol1ows,
we shall use the phrase interrogative to refer both to interrogative
complements and to interrogative sentences. Further, we shall discrimi-
nate between sentential and constituent interrogatives meaning expres-
sions such as 'Does John Iove MaryT and 'whether John Ioves Mary' by
the former, and constructions such as 'Who ate the cakeT, and 'who
bought which books where' by the former (using the phrase n-
constituent (interrogative) to indicate the number of wh-phrases that
occur in an interrogative). Finally, it should be noted that interrogative
shall denote Iinguistic expressions, while question is reserved for the
semantic objects they express. 1
1. CATERPILLARS AND BUTTERFLIES
1.1. Introduction
In this section we will outline two approaches to the semantics of
interrogatives and the question-answer relation. Each of these two
approaches, we will argue, solves some important issues, yet, on the
face of it, the two are incompatible. However, we will show that if we
take a flexible view, the conflict may be an apparent one, and that a
type-shifting process may serve to unify the insights of both.
The situation we will sketch, bears a striking resemblance to the
situation one finds in the semantics of noun phrases. Concerning the
Iatter, Barbara Partee writes in her 1986 paper (which was a source of
inspiration for the present paper):
The goal [. . 1. is to attempt a resolution of the apparent conflict between two
approaches [... 1. 1 believe that the most important insights of both sides are basically
correct and mutually compatible. To try to show his, 1 will draw on and extend the idea
of general type-shifting principles [... 1.
[Partee (1986) pp. 1151
THE SEMANTICS OF INTERROGA TIVES 25
The two approaches we will discuss, can be dubbed the categorial and
the propositional approach. In the former much emphasis is placed
on the differences in syntactic category and semantic type between
different kinds of interrogatives, whereas in the latter the postulate of a
uniform, propositional, type is the starting point. Our own analysis, if it
is successful, will be one that covers both, in this sense that it will allow
us to treat interrogatives in a variety of types, which are systematically
related to each other. Such an analysis would provide additional
support for the kind of use of type-shifting that was made for the first
time by Partee in her discussion of NP-interpretations, a kind of use
that considers type-shifting as an explanatory device, rather than as a
descriptive too1.
1.2. The Categorial and the Propositional Approach

If we restrict ourselves to the (model theoretic) semantics proper of
interrogatives, two main approaches can be distinguished: the categorial
and the propositional approach. Disregarding details of concrete imple-
mentation (at least for the moment), they can be characterized as
follows.
On the categorial view, the main semantic property of an interroga-
tive is that it is in some sense an 'incomplete' object. This object
requires for its completion an answer. Different kinds of interrogatives,
it is observed, call for different kinds of answers. Sentential interroga-
tives, for example, are characteristically answered by 'Yes.' or 'No.', and
constituent interrogatives are typically followed by constituents such as
'John.', 'In the park.', 'John, by Mary.', and so on. These constituents do
not form a homogeneous category. Some are terms, others adverbs, and
others again, like 'John, by Mary.', are of a category not ordinarily
found in sentence grammar.
StiH, in the context of an interrogative, alI these different kinds of
constituents are meant to convey information, to express a proposition.
And, of course, which proposition a characteristic linguistic answer
expresses depends on the interrogative it is meant to answer. On the
categorial approach this is accounted for by analyzing interrogatives
and answers in such a way that they fit together and make up a proposi-
tion. Hence, since constituent answers are of all kinds of different
categories, different kinds of interrogatives are to be of different
categories as well.
Taking the orthodox view, on which there is a fixed category-to-type

correspondence, this means that the folIowing general principle under-
lies categorial theories: the syntactic category and the semantic type of
interrogatives are determined by the category and type of their charac-
teristic linguistic answers. This general idea leaves room for many
different implementations, but alI theories have in common that inter-
rogatives are treated as relational expressions, expressing n-place
relations, and that constituent answers serve to fiU in the argument
places.
The other kind of approach, the propositional one, takes a different
view on the semantic content of interrogatives. Answers to interroga-
tives, it is observed again, convey information, hence they are taken to
express propositions. Consequently, the answerhood conditions of an
interrogative are a determination of which proposition(s) count(s) as
answer(s) to it. From this point of view, the semantic content of alI
kinds of interrogatives can and must be analyzed in a uniform way, viz.
in terms of propositions. In view of this, there is no reason not to
consider interrogatives of different kinds to form a homogeneous
category.
So, the gist of the propositional approach can be formulated in the
folIowing general principle: interrogatives are of a uniform syntactic
category and a uniform type, the semantic interpretation giving the
answerhood conditions. Again, this idea can be worked out in a number
of different ways. In most cases, the meaning of an interrogative is
taken to be a function which determines for each possible world a (set
of) proposition(s) which constitute(s) the true semantic answer(s) to
that interrogative in that world. The differences between the various
individual theories mainly reside in what true semantic answers are
taken to be.
As we said above, the situation we are confronted with in developing
a semantic theory for interrogatives resembles the situation conceming
the semantic interpretation of noun phrases which Partee analyzes in
Partee (1986). There are two radicalIy different approaches, each one
based on an intuitively clear idea, and each one capable of explaining
an interesting and important class of phenomena. Each approach makes
predictions about the kind of semantic object that an interrogative
represents, and these predictions are incompatible, if, that is, one takes
the orthodox view on the relationship between syntactic categories and
semantic types. It one assumes that to each syntactic category there
corresponds a unique semantic type, the two approaches are incom-

patible in two ways: the propositional one postulates a uniform
semantic type, whereas the categorial one assumes interrogatives to be
of a large number of different types; and even taking only one kind of
interrogative into consideration, the two will not meet, since on the
categorial approach an interrogative expresses an n-place relation,
whereas in the propositional approach it determines a (set of)
proposition(s).
However, if we take a closer look at the phenomena that each of
these approaches deals with successfully, it can be observed that these
are largely complementary. Hence, there is good reason to suppose that
the incompatibility between the two is only an apparent one which
originates from the assumption that there is such a thing as the
semantic type of an interrogative, and that once this assumption is given
up, the two can fruitfully be combined. In effect, this is what we want to
argue for. So, let us first turn to the alleged complementarity of the two
approaches.
1.3. A Puzzling Situation
In order to get a clearer picture of what exact1y is going on, let us start
by formulating two intuitively plausible requirements that a semantic
analysis of interrogatives should meet. (This is not to suggest that this is
alI there is to such an analysis, but it suffices for our present purposes.)
The first requirement concerns the question-answer relation as a
linguistic relation, i.e. as a relation between an interrogative and its
characteristic linguistic answers. It is the demand that the semantic
content of the interrogative, and the semantic content of the constituent
that forms a linguistic answer, together determine the semantic content
of that linguistic answer.
The second, equally plausible, requirement is that a semantics of
interrogatives should give a proper account of systematic semantic
relationships that exist between interrogatives (and between interroga-
tives and indicatives). Especially in the case of interrogatives, where
intuitions about the kind of semantic object that is their proper
interpretation are slim, meaning relations are the prime data to be
accounted for. A central relationship between interrogatives is the one
that holds if every complete and true answer to the first also gives a
complete and true answer to the second. In effect, one might dub this
'entailment' between interrogatives. This relation holds, e.g., between
'Who will go to the party? And what will they bring along?' and 'Who
will go to the party?', and between the latter interrogative and 'Will
John go to the party?'.
It will need no argumentation that a categorial theory will be able to
meet the first requirement, at least in principle, since it assigns to an
interrogative a semantic type which, when it is combined with the type
of its characteristic linguistic answers, 'cancels out' to t. And it will also
be clear that, again at least in principle, a propositional theory will be
able to meet the second requirement, for it identifies the semantic
content of an interrogative with its answerhood conditions, and stipu-
lates a uniform semantic type, to which a generalized notion of
entailment may be applied in a straightforward way. And it is exactly
the feature of a categorial theory that enables it to meet the first
requirement that makes it doomed to fail on the second. For in a
categorial approach a multiplicity of types of interrogatives is postu-
lated that matches the multiplicity of types of constituents that form
their characteristic linguistic answers. And it is this multiplicity of types
that prevents the application of any standard notion of entailment, since
entailment is typically a relationship between expressions of one and
the same type.
We can illustrate the rather paradoxical situation we find ourselves
in as follows. Suppose there are two interrogatives that are equivalent
under the notion of entailment indicated above, i.e. for which it holds
that each complete and true answer to the first gives a complete and
true answer to the second, and vice versa. And suppose further that
there is a characteristic linguistic answer that as an answer to the first
interrogative conveys different information, expresses another proposi-
tion, than it does as an answer to the second. It such a situation exists, it
is clear that neither a propositional nor a categorial theory will be able
to deal with it. For the first assumption implies that on the proposi-
tional theory the semantic content of the two interrogatives is the same.
Hence, combining it with the semantic content of one and the same
constituent cannot but give the same result in both cases. On the other
hand, a categorial theory might very well cope with the second assump-
tion, but only in virtue of failing to de\;ll with the first. For accounting
for the fact that the constituent answer expresses different propositions
in each of the two cases, requires giving the two interrogatives a
different interpretation, thus failing to account for their assumed

equivalence.
Examples of such pairs of interrogatives are not only theoretically
possible, they actually exist. A simple case, involving almost no
assumptions about the details of an actual propositional or categorial
theory, is the following. Take the two interrogatives: 'Who of John, Bill
and Mary will go to the party?' and 'Who of John, Bill and Mary will
not go to the party?'. These two are equivalent in the sense that they
have the same answerhood conditions. Each proposition which com-
pletely settles the first question, also fully answers the second one, and
vice versa. However, a constituent answer like 'John and Bill.' expresses
a different proposition according as to which interrogative it is used to
answer. In the first case it expresses that John and Bill are the ones that
will attend the party, whereas in the second case it conveys the
information that John and Bill are the ones that won't go to the party.
As we said, this situation is rather puzzling. We have formulated two
reasonable requirements on semantic theories for interrogatives, and we
seem to have found out that a semantic analysis that meets the one
cannot at the same time meet the other. So what are we to do?
There are many ways in which one might react to this predicament.
Before briefly discussing three of them, we want to point out the
following. It should be bome in mind that we are not discussing actual
theories here, but overall approaches. And we take it for granted that
the insights on which the two approaches are founded are basically
sound. In fact, the soundness of the ideas underlying the two
approaches is reflected, we feeI, in the plausibility of the two require-
ments we have singled out and discussed. Of course, both kinds of
theories are wrong in so far as they take their respective starting points
to say ali there is to say about the meaning of interrogatives. That is
exactly what the paradox shows. But, we think that this should not lead
one to reject the underlying ideas as basically correct insights about
aspects of the meaning of interrogatives.
Now, we can envisage (at least) three different reactions. The first
one runs along the following lines. It hooks on to the failure of the
propositional approach to meet the first requirement. Logical equiva-
lence, so it goes, is simply not a sufficient condition for sameness of
semantic content (sameness of meaning). Rath((y, meaning is a more
fine-grained notion, and what the first requirement amounts to is that in
the case of interrogatives it should be at least so finely structured that
within the overall meaning of an interrogative, which in the proposi-

tional approach gives the answerhood conditions, we can distinguish as
a distinct 'part' the n-place relation that the categorial approach con-
siders ta be the semantic interpretation. Sa, instead of the usual
unstructured notion of a function from worlds ta (sets of) propositions,
one should use structured meanings, interpreted derivation trees, ar
what have you.
We feeI that the use of structured meanings that this reaction
proposes ta make, is improper, ar, at least, is not in line with the usual
motivation for using structured meanings. In the analysis of proposi-
tional attitudes, some have proposed the use of structured meanings,
because they feel that in such contexts, which, an their view, are
essentially tied up with mental representations, we need not just the
semantic content of an expression, but also its semantic structure,
assuming that this structure and aur mental representation bear enough
resemblance ta let the one go proxy for the other. However, such use of
structured meanings differs essentially from the one proposed aba ve.
There, no use of the structure of the entire meaning is made as such, it
is only used ta get at a certain part of the meaning that helped ta
generate it. Once you've gat hold of the relevant part, the rest of the
structure can be discarded. Ta aur minds, this goes against what aur
two requirements actually state about the meaning of an interrogative.
They are both requirements an one and the same notion of meaning.
For consider what will happen if we follow this strategy in the case at
hand. According ta the proposed strategy, we need the meaning of a
predicate (ta meet the first requirement), and we need the meaning of a
sentential structure (ta meet the second one). In both cases, the
meanings we use, are 'normal' unstructured meanings, i.e. intensions. It
is only by means of a trick that the two are unified. The two separate,
unstructured intensions are taken together in one 'structured meaning',
but ta aur minds, this is just a cosmetic move, for no structure of the
meaning as such is used in any essential way (in fact, we just use a pair
of intensions as the meaning).
Sa, we feel, there are theoretical reasons ta be dissatisfied with this
appeal ta the notion of structured meaning for this particular problem.
On the practical side, it may be remarked that it may lead ta a theory
that, extensionally, sa ta speak, meets t~e two requirements. However,
structured meanings are no sure cure for any propositional theory. It
depends an the way in which such a theory derives its function from
wor1ds to (sets of) propositions whether, taken as structured objects,

they do contain the required relations as retrievable parts. For example,
quantificational propositional theories, such as Karttunen's and Bennett
& Belnap's, may structure their meanings any way they like, the
required relations just ain't in there.
The second possible reaction we want to discuss, starts from the
categorial point of view, i.e. it takes interrogatives basically as express-
ing n-place relations. The diagnosis it gives for the failure of this
approach to meet the second requirement, that of accounting for
entailment between interrogatives in a general, non ad hoc way, is that
it lacks a uniform type to associate with different kinds of interroga-
tives. Now, property theory is designed to provide such a uniform type,
for it allows for the possibility of analyzing expressions which are of
different types in the ordinary view, as being of one and the same type,
viz. that of entities. This suggests that the two semantic objects we need
in our semantic analysis of interrogatives can be gotten as special
instances of the general relationship that exists between abstract objects
and the corresponding relational 'entities'.
However, a uniform type is one thing that is needed in order to be
able to satisfy the second requirement, but it is not sufficient. What is
needed on top of it, is an entailment structure on (the relevant part of)
the domain of objects. And the main question is how to get the proper
structure. One kind of structure we need to impose on the domain of
objects anyway, is the structure that is inherited from the original
domains of the respective relational types of entities. For example, we
can view propositions as objects, and these objects will bear structurally
the same relations to one another as their propositional counterparts.
And the same goes for one-place properties, two-place relations, and so
ono However, it must be clear that this kind of structure of the
respective parts of the domain of objects will be of no use at alI for
accounting for meaning relations between interrogatives. First of aH, the
structures in question remain restricted each to their own subdomain. If
we identify these subdomains with sorts, we can express this by saying
that these relationships are essentialIy 'intra-sortal'. But, and this is the
important point, entailment relations between interrogatives are cross-
categorial relationships, and hence would have to be 'cross-sortal'
relationships on the entity domain in this approach. And second, the
intra-sortal relationships we do get, are not the proper ones to account
for entailment between interrogatives of the same kind, For example,
sentential interrogatives are not related by entailment to each other

(e.g., 'whether f and 'whether ~ and 1jJ' do not entail one another), but
the corresponding propositions (in the example, 'that f and 'that ~ and
1jJ'), and hence the corresponding propositional objects, have a very
rich entailment structure. It seems that the only way to get the proper
cross-sortal relationships on the one domain of objects, is through an
analysis (at some level) of interrogatives as objects of a propositional
type.
Of course, this does not show that interrogatives can't be, or
shouldn't be, analyzed as entity denoting expressions. On the contrary,
it can be argued that in certain constructions and relations in which
they enter, it is profitable to analyze them as denoting an object. But
what it does show is that such an analysis wiH not solve our present
problem. We stiH need the two kinds of semantic objects that the
categorial approach and the propositional approach postulate. Property
theory wiH enable us to analyze both (also) as abstract objects, and this
may be useful, but it does not enable us to avoid postulating a
propositional type of semantic object, besides a variety of relational
types, as an interpretation of interrogatives.
The third reaction is the one that we think is most adequate. It
analyzes the situation in terms of type shifting. The paradox occurs, so
it goes, because in both requirements mention is made of 'the meaning
of', or 'the semantic interpretation of' interrogatives. The propositional
approach assigns a uniform type to all interrogatives, and, disregarding
ambiguities, in that type each has a unique semantic interpretation. The
categorial view postulates various semantic types, but each kind of
interrogative occurs in one type only. And again, in that type it has a
unique semantic interpretation. So, both approaches take it for granted
that each particular interrogative belongs to a unique type and, in that
type, has a unique interpretation. It we want to stick to that, the
paradox is unavoidable. Or, to put it differently, the paradox shows that
this is something we should not take for granted. What the paradox
indicates is that interrogatives are among those natural language
expressions which do not have a unique interpretation in a unique type.
Rather, taking different perspectives, such expressions can be said to
have different (but related) meanings, that are of different types.
So, the third strategy proposes to solve the apparent paradox by
introducing a relativization to a perspective. In this case, it claims that
the two requirements are equally reasonable., but are made from
different perspectives, taking different constructions as their starting

point, and hence are requirements on different domains. Interrogatives
have to be analyzed in (at least) two different domains, as expressions
of (at least) two different types. On the one hand, they have a clearly
relational meaning, as is most prominently shown in the way in which
they interact with their characteristic linguistic answers. On the other
hand, they also behave as propositional objects, and it is as objects of
the latter type that they enter into systematic relationships, such as
entailment, to each other. (In section 2.3 we wiII see that interrogatives
belong to other domains as weII.)
Within a certain conception of how to incorporate such flexibility
into the grammar, about which we wiII say some more in section 3.3,
this implies that the one major syntactic category of interrogatives has
to be associated with different semantic types. And each individual
interrogative will have to be given an interpretation in a suitable
relational type, and also an interpretation in a uniform propositional
type. An additional requirement is that these two interpretations be
systematically related.
Giving up the assumption of a unique interpretation in a unique type
means that the two intuitive requirements on the semantics of interroga-
tives have to be rephrased along the following lines. The first require-
ment now reads that an interrogative has to be analyzed as being of
(among others) such a type that its semantic content as an expres sion of
that type and the semantic content of the constituent that forms a
linguistic answer together determine the proposition expressed by that
linguistic answer. And the second requirement wiII now state that
interrogatives also have to be analyzed as expressions of one uniform
type in which a proper account of their systematic semantic relation-
ships, in particular of their entailment structure, can be given. And the
concept of a flexible grammar adds to these the additional requirement
that these two should be systematically related.
In order to get a clearer view on what a flexible analysis of interroga-
tives amounts to, we will first concentrate on an area where the use
type-shifting and flexibility is more familiar, viz. coordination. We
discuss various facts and their consequences in section 2, and outline a
flexible framework in sections 3.1 and 3.2. In section 3.3, we will return
to the possibility of implementing the third "strategy to solve our
puzzling situation.
2. COORDINATION, ENTAILMENT AND TYPES
2.1. Coordination
One has to live with a lot of questions, and sometimes one cannot wait
to have them answered only one by one. In such situations, one may
use a conjunction (or sequence) of interrogatives. An example of such a
conjunction, and of the way in which it can be answered is given in (1):
(1) Whom does John love? And whom does Mary love?
- John loves Suzy and Bill. And Mary loves Bill and Peter.
In this example a simple conjunctive sequence of two interrogatives is
given, which, as the answer that follows it shows, in fact poses two
separate questions: the speaker wants to know both whom John loves,
and whom Mary loves.
Another example of an interrogative that involves conjunction is (2):
(2) Whom do John and Mary love?
Example (2) is ambiguous between what we call a direct reading, on
which it is equivalent with (3):
(3) Who is such that both John and Mary love him/her?
and what we call its pair-list reading, on which it means the same as (1)
above, i.e. on which it asks for a specification of the individuals that
John loves, and for a specification of those that are loved by Mary.
A similar ambiguity can be observed in interrogatives such as (4):
(4) Whom does every man love?
This example, too, has a direct reading and a pair-list reading, as the
following paraphrases, and the corresponding answers, illustrate:
(4 ) (a) Who is such that every man loves him/her?
- Peter and Mary.
(4) (b) Whom does Peter love? And whom does Billiove? And .. .
- Peter loves Mary. And Billioves Suzy and Fred. And .. .
An interesting point to note is that on its pair-list reading, as para-
phrased in (4) (a), (4) behaves like (5). :rhe latter is a two-constituent
interrogative, i.e. an interrogative containing two wh-phrases. Although
(4) on the relevant reading contains only one wh-term, it is answered is

the same way as (5):
(5) Whom does which man love?
What (5) asks for is a specification of list of pairs of individuals x and
y, where x is a man and y an individual such that x loves y. The same
holds for (4) on its reading (b), which is why it is called what it is cal1ed.
An example of a disjunction of interrogatives is given in (6):
(6) Whom does John love? Or, whom does Mary love?
- John loves Suzy and Bill.
- Mary loves Bill and Peter.
- John loves Suzy and Bill, and Mary loves Bill and Peter.
Disjunctions of interrogatives, like their conjunctive counterparts,
formulate two separate questions, but, unlike conjunctions, they pose
only one: they leave the hearer a choice as to which one of the
formulated questions she wants to answer. As the answers in (6) show,
a disjunction of interrogatives may be answered by answering either
disjunct or both.
Disjunctive interrogatives need not consist of two separate interroga-
tives, as (7) shows:
(7) Whom does John or Mary love?
Like its conjunctive counterpart (2), (7) is ambiguous between a direct
reading and what we call a choice reading. On the latter (7) is
equivalent to (6), on the former it can be paraphrased as (7) (a):
(7) (a) Who is such that John or Mary (or both) loves himlher?
- Suzy, Bill and Peter.
As we saw above, pair-list readings are not restricted to interrogatives
with overt conjunctions. In the same way choice readings can occur
without overt disjunctions, as a simple example like (8) shows:
(8) What did two of John's friends give him for Christmas?
This interrogative is ambiguous. It has a direct reading, on which it asks
for a specification of the presents that two of his friends gave him. And
it has a choice reading, on which it invites the addressee to choose any
two friends of John's and specify for each one" of them what he/she
gave him for Christmas. Obviously, it is a matter of the internal

semantic structure of a term phrase whether it will give rise to a pair-list
or a choice reading or not. 2 Again, it should be noted that choice
readings of interrogatives are like two-constituent interrogatives, as is
evident from the way in which they are answered.
Like ordinary interrogatives, coordinated ones can be embedded
under extensional and intensional verbs, such as know and wonder
respectively. Also, the ambiguity between a direct reading and a pair-
list reading, and between a direct reading and a choice reading is
preserved in such contexts. As for the distinction between extensional
and intensional embedding verbs, it should be noticed that there is a
difference when disjunction is involved, as (9) and (10) show:
(9) Peter knows whom John loves or whom Mary loves

(10) Peter wonders whom J ohn loves or whom Mary loves.
Sentence (10) is ambiguous, allowing for the disjunction in the comple-
ment to have either wide or narrow scope with respect to wonder. The
wide scope reading occurs when the speaker knows that Peter wants to
know the answer to one of the two questions, but she herself does not
know which one this is. On the narrow scope reading (10) expresses
that Peter will be satisfied when he gets an answer to either one of the
questions involved, no matter which one.
A last observation that should be made here, is that coordination of
interrogatives goes across kinds. It is not restricted to expressions of the
same kind, i.e. to sentential interrogatives, single constituent interroga-
tives and multiple constituent interrogatives, but combines them freely,
as the foHowing examples show:
(11) Who went to pick up John? And are they back already?
(12) Peter knows who went to pick up John and whether they are
back already
(13) Which woman does which man admire most? Or do they aH
detest each other?
This fact, too, can be used to argue for uniformity in assigning types to
these different kinds of interrogatives.
So much for coordination, let us now turn to the second part of our
empirical domain, that of entailment.
2.2. Entailment
Let us first of all recall a familiar fact concerning entailment relations

between coordinated indicatives and their coordinates: a conjunction
entails its conjuncts, a disjunction is entailed by its disjuncts, and a
conjunction entails the corresponding disjunction. Analogous facts hold
for coordinated structures in general, and properly generalized notions
of coordination and entailment should account for them.
Considering interrogatives, we can observe that someone who asks
(11) also asks (14), and that someone who answers (15) also answers
(16):
(14) Who went to pick up John?

(15) Where is your father?
(16) Where is your father? Or your mother?
In section 1.3 we used a notion of entailment between interrogatives
which might be described informally as follows:
An interrogative A entails an interrogative B iff
whenever a proposition gives a complete and true answer to
A,
it gives such an answer to B.
It is easy to check that this description conforms with the observations
just made, and that it likewise predicts that (17) entails (16):
(17) Where is your father? And your mother?
These examples of entailments between interrogatives depend on their
coordination structure. There are also other types of entailments to be
observed. Let us give two more examples. The single constituent
interrogative (18) entails the sentential interrogative (19):
(18) Which men does Mary love?

(19) Does Mary love John?
Getting a complete answer to (18) implies getting a complete answer to
(19). Notice that in this case entailment is a relation between different
kinds of interrogatives, a one-constituent interrogative and a sentential
interrogative. Another example is provided by (18), (20) and (21). A
complete answer to both (20) and (21) gives a complete answer to (18)
as well:
(20) Whom does Mary love?
(21) Who are the men?
Notice that (20) on its own does not entail (18), for knowing the
answer to (20) is knowing which individuals (within the relevant
domain of discourse) Mary loves, and this entails knowing which men
Mary loves only in conjunction with knowledge of which individuals are
men.
In line with recent work, we assume that coordination and entail-
ment are general syntactic and semantic processes. Elements of all
major categories can be coordinated, and a number of people have
proposed general definitions to account for this. 3
Entailment, too, is a relation that holds between elements within any
major category: indicative sentences, of course, interrogatives, as we
have seen above, but also termphrases (every man entails John),
verbphrases (to walk entails to move), nouns (woman entails human
being), and so ono In alI cases it is the same relation that is at stake, viz.
that of the denotation of one element being included in all models in
that of the other. To put it differently, employing a semantic meta-
language based on set theory brings along a definition of entailment for
all categories: inclusion of denotation in all models.
The following definitions of generalized conjunction and disjunction
are based on the work referred to above. First, the notion of a
'conjoinable type' is defined:
CT, the set of conjoinable types, is the smallest set such that:
(a) t E CT;
(b) if b E CT, then (a, b) E CT.
Then generalized conjunction is defined as follows:

X n Y = X /\ Y, for X, Y of type t
X n Y = AX [X(x) n Y(x)], for any other conjoinable type.
The definition of disjunction, U, is analogous.
Entailment, ~, can be defined generally as follows:
"
X ~ Y = X ..... Y, for X, Y of type t
X ~ Y = Vx[X(x) ~ Y(x)], for any other conjoinable type.
It should be noted that employing such general notions of coordination

and entailment means that one is kept to assign semantic interpretations
to expressions in such a way that the entailment relations that can
be observed are accounted for by these independently defined and
motivated notions. Exceptions to this should be well-argued for. 4
It should be stressed that this is a methodological requirement. Of
course, a theory that uses different and unrelated notions of entailment
or coordination for different domains, may very well be empirically
adequate, in the sense that it makes the right predictions. The point we
want to make, is that a theory that makes the same predictions but does
so on the basis of generalized and uniform notions, is to be preferred
on methodological grounds. It provides a simpler account of the
relevant facts, and, hence, has greater explanatory force.
In the next section, we will show how this requirement can be used
to evaluate theories that propose a certain type of semantic object as
the interpretation of interrogatives.
2.3. Types for Interrogatives
In this section we will investigate which types are to be assigned to

interrogatives. In a flexible framework, there need not be a unique
proper semantic type for alI expressions of a certain category. Inter-
rogatives are no exception to this rule. However, for every construction
in which interrogatives occur, there is, as we shall see, a key type: the
type in which the intuitive entailments between interrogatives in that
construction are accounted for by the general definition of entailment
that our framework provides. 5
AIso, we will argue in this section that employing general notions of
coordination and entailment will enable us to evaluate various pro-
posals within the two main approaches to the semantics of interroga-
tives which were discussed in section 1.
As we saw above, the main characteristic of the categorial approach
can be summed up as follows: 6
The syntactic category and the semantic type of an inter-
rogative are uniquely determined by the category and type
of its characteristic linguistic answers. "-
The idea is that the type of an interrogative and the type of its charac-
teristic linguistic answers should cancel out, by functional application,
to that of sentences, i.e. to type t. Let us illustrate this with a few

examples. First a single constituent interrogative:
(22) Whom does John love?

-Mary.
Applying the criterion just mentioned, it folIows that the (simplest
possible) type of a single constituent interrogative is that of a property
of individuals (a one-place relation). Next, consider a multiple con-
stituent interrogative:
(23) Which man does which woman love?

- Mary, Bill; and Suzy, Peter.
Here the resulting type is that of a two-place relation between individ-
uals. The last example is that of a sentential interrogative:
(24) Does John love Mary?

-Yes.
It we apply the criterion in this case, the outcome is not unique, but the
simplest solution is to give sentential interrogatives type t, and hence
consider 'Yes.' and 'No.' as expressions of type (t, t), which is one of
the solutions we tind in the literature. Considering t to be the type of
zero-place relations, we can view sentential interrogatives as zero
constituent interrogatives. Generalizing from these examples, we con-
clude that in the categorial approach, n-constituent interrogatives are
interpreted as n-place relations.
AIthough this approach has attractive features, for one thing, it leads
to a simple and intuitive analysis of the interpretation of characteristic
linguistic answers, it also has its shortcomings. These concern coordina-
tion and entailment, as we shall see.
First of alI, the approach as such does not account for coordination
and entailment across different kinds of interrogatives and it is hard to
see how it could, without giving up its fundamental characteristics. For
entailment and coordination require a uniform type, which the cate-
gorial approach simply does not provide.
Moreover, even if we limit ourselves to interrogatives of the same
kind, in which case the general definitions are in principle applicable,
"-
we find that the wrong predictions are made. E.g. it is predicted that
(25) and (26) are equivalent, which is not the case:
(25) Who walks? And who talks?

(26) Who walks and talks?
The conjunction of interrogatives (25) asks to specify both the individ-
uals that walk and the individuals that talk, whereas (26) asks to specify
the individuals that both walk and talk (so, (25) entails (26), but not the
other way around).
A second example. Analyzing one-constituent interrogatives as
properties, predicts that (27) entails (28), which again is not the case:
(27) Who walks?
(28) Who moves?
It one is told which individuals walk, one is not thereby told which are
alI the individuals that move.
A straightforward condusion that can be drawn, is that if one wants
to employ general definitions of coordination and entailment, then, first
of alI, one has to analyze, at some leve1, alI interrogatives as being of
one and the same type, and, secondly, within this type one has to
associate them with the right kind of object. 7
As we saw above, theories in the second main approach, the
propositional one, do assign one single type to alI interrogatives. We
characterized the main idea of this approach as folIows:
The meaning of an interrogative is given by its answerhood
conditions.
Within intensional semantic theories answers are of a propositional
nature, hence interrogatives are of a 'propositional' type. Here, several
choices are stiH open. The best-known analysis, that of Karttunen,8
makes them expressions of type s, t), t). Le., on this analysis an
interrogative denotes a set of propositions. Karttunen interprets this set
as consisting of those propositions which jointly constitute the true and
complete answer.
Two things should be noted. First of alI, Karttunen's theory is, what
Belnap calIs, a 'unique answer theory', i.e. a theory that assumes that
each interrogative has a unique true and complete answer. Why this is
relevant wiH become dear shortly. Second, since Karttunen's theory
employs a uniform (conjoinable) type, it ma~es predictions about
coordination and entailment generally, also across different kinds of
interrogatives. Let us consider some of these predictions.
The schema of generalized conjunction tells us that the conjunction

of two interrogatives is interpreted as the intersection of the sets of
propositions denoted by each of the conjuncts. Given the interpretation
of these sets of propositions on Karttunen's theory, the result is that a
conjunction of interrogatives (almost) never has an answer. The follow-
ing example illustrates this:
(29) Does John walk? And does Mary walk?
Suppose it happens to be the case that John walks and that Mary
doesn't. Then the first conjunct denotes the set consisting of the
proposition that John walks, and the second denotes the set consisting
of the proposition that Mary doesn't walk. The intersection of these two
sets is empty, which means that (29) cannot be answered. A similar
result holds for interrogatives on pair-list readings.
Disjunction corresponds to taking the union of sets of propositions.
Again, the prediction that the Karttunen analysis makes is not in
accordance with the facts. Consider the disjunction of interrogatives
(30):
(30) Does John walk? Or does Mary walk?
Taking the union of the set denoted by each of the disjuncts results in
the set of propositions which joint1y constitute the complete and true
answer to the conjunction (29), rather than to the disjunction (30).
For entailment, too, the results which we get when we combine the
general schema with Karttunen's interpretation of interrogatives, are
not correct. A simple example is the entailment relation between (31)
and (32):
(31) Whowalks?
(32) Does John walk?
In the intuitive sense, (31) entails (32). But the set of propositions that
is the denotation of (31) in Karttunen's theory is not generally a subset
of the set denoted by (32). Hence, the theory fails to account for this
entailment.
Providing a semantic account of interrogatives which deals with
coordination and entailment adequatelx, then, is not just a matter of
finding a uniform and proper type, but also of associating each
interrogative with the right object of that type. One might think that
Karttunen found the right type, but hit the wrong objects within that
type. However, as our discussion of a second proposal intends to show,

there are reasons to doubt whether this is indeed the case.
Bennett and Belnap have developed an analysis of the semantics of
interrogatives that is explicitly set up to deal with those constructions of
interrogatives on which they allow for more than one complete and true
answer, such as disjunctions and choice-readings. 9 They assign the same
semantic type to interrogatives as Karttunen does, i.e. they, too, take
interrogatives to denote sets of propositions, but they interpret these
denotations in a different way. In their analysis, each of the proposi-
tions in the set denoted by an interrogative on its own constitutes a
complete and true answer. For 'ordinary' interrogatives, i.e. for those
which have a unique answer, this means that they denote a unit set.
Here we have an analysis which differs from Karttunen's, not in the
type that it assigns to interrogatives, but in the objects of that type that
interrogatives are taken to denote. And we might ask whether this
change overcomes the difficulties we noted earlier.
As is to be expected, the Bennett and Belnap approach does well
with respect to those interrogatives for which it was designed, viz.
interrogatives which have more than one unique answer. Sticking to our
general definitions, disjunction still comes down to taking the union of
the denotation of the disjuncts. However, given the kind of set of
propositions that an interrogative denotes on their theory, the result is
correct. Consider the disjunction (30) again. Each of the disjuncts now
denotes a unit set, and taking the union thereof results in a set with two
elements, each of which is a proposition which is a complete and true
answer to the disjunction. It is also dear that on the Bennett and
Belnap analysis a disjunction is entailed by each of its disjuncts.
On this score, Bennett and Belnap do better than Karttunen. But this
does not mean that now we have the right objects of the right type, at
least not in all cases, as the following considerations show. Take the
conjunction (29). Here we still have the same kind of problems as we
met in Karttunen's theory. Given the general definition of conjunction,
the denotation of a conjunction is the intersection of the (unit) sets of
propositions denoted by the conjuncts, and this still results in the empty
set (at least in most cases). Also, we do not get the desired entailment
between a conjunction and its conjuncts. So, we must condude that the
Bennett and Belnap approach is not satisfactory either.
Let us take stock: we have seen that atomic interrogatives, i.e. non-
coordinated interrogatives which are not embedded and are not given a
pair-list or a choice reading, conjunctions of interrogatives and disjunc-
tions of interrogatives behave differently with respect to types and

entailment. An atomic interrogative has a unique true and complete
answer (in each possible world). This means that the simplest denota-
tional type for atomic interrogatives is type (s, t) (giving it a sense of
type (s, (s, t). As for conjunctions, if we disregard their relations with
disjunctions, they could be analyzed at the same level. Since a conjunc-
tion, too, has a unique true and complete answer: the conjunction of the
propositions that answer the conjuncts, also answers their conjunction.
For disjunctions, however, things are different. They do not have a
unique complete and true answer, hence they simply cannot be of type
(s, t). It we look at entailment relations between disjunctions on the one
hand and conjunctions and atomic interrogatives on the other, we see
that in order to account for them we need a uniform type for all, since
generalized entailment requires a uniform type for all elements
involved. The need for such a uniform type is underscored by the
observation that in order to construct disjunctions in accordance with
the general procedure, atomic interrogatives should (also) be of the
same type as the disjunctions which are constructed from them.
Such considerations, by the way, constitute a general argument
against the type that Karttunen, and Bennett and Belnap employ. For,
although the objects of type s, t), t) that the latter associate with
interrogatives give a proper interpretation for disjunction, it simply
cannot be the uniform type which is required, as the examples dis-
cussed above have shown. So here we do have a case against the type as
such.
The question that now arises, is what this uniform type is, and
whether it is sufficient to account for all entailments. The situation we
find ourselves in with regard to interrogatives, resembles that of term
phrases in some important respects. The lowest type for a proper name
is type e. Looking at disjunctions of proper names in isolation, we tind
that we can analyze them as being of type (e, t). For conjunctions this
will not do. There we need a more complex type, viz. the familiar
e, t ), t) (disregarding intensionality). Once we take entailments into
account, we see that the latter is the uniform type we need, hence that
all proper names should have type e, t), t) at a certain level of
analysis. Traditionally, this is achieved by 'generalizing to the worst
case' and treating all proper names in all contexts as expressions of that
type. Within a flexible approach, however, we take e, t), t) as one of
the derived types that proper names can have, a type that they must
have e.g. when occurring in a coordinate structure.
With respect to interrogatives we can folIow the same lead. The key
type of atomic interrogatives, i.e. the type in which the entailments
among them can be accounted for, is type (s, t). Looking at disjunction
in isolation suggests s, t), t) as the proper type (cf. Bennett and
Belnap), but taking a broader view we see that the level at which
coordination and entailment can be accounted for is that of type
(s, t), t,), t). And within a flexible frame of mind, the relation between
the basic type (s, t) and the latter is a familiar one: we get from the one
to the other by the type-shifting rule of 'lifting', the same procedure we
use in analysing term phrases.
The flexible approach is not motivated by reasons of elegance and
simplicity alone. As is argued e.g. in Partee and Rooth (1983), the
strategy of generalizing to the worst case is not only unnecessarily
complicated in many cases, sometimes it is also empiricalIy inadequate.
The 'wide scope or-cases' they discuss, show that there is no a priori
worst case to generalize to. A similar argumentation can be distilled
from the semantics of sentences containing an intensional verb with a
disjunction of interrogatives as its complement (see (10) in section 2.l.
above. We return to this example later on).
But the semantics of interrogatives provides yet another argument
for the necessity of flexibility. To be able to account for entailment
relations between atomic interrogatives, such as hold e.g. between (31)
and (32), we need to analyze them in the key type (s, t). Ii we lift them
to type (s, t), t,), t), we loose entailment relations that hold at the
basic level (s, t). But in order to be able to account for entailment
relations between coordinated interrogatives, or between such inter-
rogatives and atomic ones, we do need the lifted level to get the right
results. So, we cannot assign alI interrogatives a uniform type in alI
cases. What the proper type is, in terms of the predicted entailment
relations, depends on the context (e.g. on the construction in which an
expression occurS).lO
Summing up, we have found that there is no uniform key type for alI
interrogatives. Rather, there is a key type for each of the various
constructions and relations that involve interrogatives. But these types
do not constitute a heterogeneous set, they are related to each other in
a systematic fashion. It is our purpose in the next section to sketch a
theory in which this is accounted for.
3. A FLEXIBLE APPROACH
3.1. Questions as Partitions

Let us now sketch the out1ines of a theQry that satisfies the three
requirements which we formulated at the end of the section 1, and
which accounts for the various observations made in section 2. We start
by giving the general idea on which the theory is based.
The theory stays within the possible worlds framework. FolIowing
Stalnaker, who formulated this view on possible worlds in various
places,ll we view the set of possible worlds that is given with the model
as the set of alI possible altematives, as the set of alI situations which, in
that model, are distinguished from one another. In this view, sincerely
uttering a proposition, or accepting it as true, is restricting oneself to a
subset of some initial set of altematives.
In the same vein, a question can be viewed, not as a restriction on
the set of altematives, but as a division of it, as a grouping together of
altematives from a certain perspective. Each question has a particular
subject matter, and it makes a division of the set of altematives by
grouping together those which do not distinguish themselves with
respect to this subject matter. Each such group is a set of altematives,
i.e. it is a proposition. In each of the worlds within such a proposition,
the answer to the question is the same. Hence, the proposition they
make up together can be viewed as a possible answer to a question.
Let us give some examples. Consider the simple sentential interroga-
tive 'Is Amsterdam the capital of the Netherlands?'. The question
expres sed by this interrogative divides a given set of altematives (which
need not be the entire set of alI possible worlds) into two, depending on
the truth value of its subject matter in those altematives. The one group
of altematives consists of those worlds in which it is true that
Amsterdam is the capital of the Netherlands, the other of those in
which this is false. Hence, the first group forms the proposition that
Amsterdam is the capital of the Netherlands, and the second the
proposition that it is not. If the interrogative in question has any
presuppositions - say regarding the existence of Amsterdam and the
Netherlands, and the existence and uniqueness of capitals - the set of
altematives that the question makes a division on, is restricted to tho~e
worlds in which these presuppositions are true.
As a second example, take the interrdgative 'Which city is the capital
of the Netherlands?'. This question, too, makes a division of the
relevant set of alternatives. In this case the division need not be in two,
it can have many members, as many as there are cities that could be the
capital of the Netherlands. Again, the alternatives within such a group
are indistinguishable as far as the subject matter of the question, i.e. the
extension of the property of being the city that is the capital of the
Netherlands, is concerned. Together, they form a proposition that
asserts of a certain city that it is the capital of the Netherlands, i.e. they
specify a possible extension of the property in question. And each such
proposition is a complete answer to the question.
From these two examples, the following picture emerges. Each
interrogative in naturallanguage expresses a question the subject matter
of which is the extension of an n-place relation (sentential interroga-
tives being the limit case where n = O). Each such question is a
partition of the set of alternatives, Le. divides this set up into a certain
number of mutually exclusive propositions.
This general characterization of the notion of a question, of the
meaning of an interrogative, is made from the propositional perspec-
tive, i.e. from the point of view of answerhood conditions. In fact, the
description of the meaning of an interrogative that we just gave, is
nothing but a statement of its answerhood conditions, i.e. a statement of
the conditions under which a proposition gives a complete answer to it.
But notice that in our general characterization of these answerhood
conditions the subject matter of a question plays an essential role. This
subject matter is, generally speaking, the extension of some n-place
relation, and this brings in the second perspective on the semantics of
interrogatives, that of the categorial approach.
It is also possible to describe the meaning of an interrogative in
terms of the relation that is its subject matter. And in fact, as we saw
above, this is what we need to do in order to be able to account for the
relationship between interrogatives and their characteristic linguistic
answers. Of course, the two perspectives are systematically related:
each possible denotation that we can distinguish from the categorial
point of view corresponds to a unique proposition that we distinguish
from the propositional point of view. We get the latter by collecting
those alternatives where the former is the same. In this sense, we can
say that a theory which gives interrogatives interpretations both of a
relational and of a propositional type, but which links these two in the
way just described, still gives them a unified meaning.
Let us now turn to the formal details of a theory which is based on
this idea. We have concluded above that the key type for atomic
interrogatives is type (s, t). But fixing a type is not enough, we must
also say which objects of this type interrogatives denote. Again,
observations concerning entailment relations will give us a clue. Under
the assumption that we talk about a fixed domain of individuals and
that proper names are rigid designators, it holds that for every name A,
(33) entails (34):
(33) Who walks?
(34) Does A walk?
Given our characterization of entailment between interrogatives (see
section 2 above), this means that every proposition that gives a
complete and true answer to (33), also gives a complete and true
answer to (34). Given that atomic interrogatives such as (33) and (34)
are of type (s, t), we should take them to denote the proposition that is
the complete and true answer, which means that an atomic interrogative
A entails an atomic interrogative B iff in every situation the proposition
denoted by A entails the proposition denoted by B. For that is in
complete accordance with the general definition of entailment.
Since (33) entails (34) for every name A, this implies that the
proposition denoted by (33) gives a complete specification of the
extension of the walking-property. Hence, a single constituent inter-
rogative will denote in each wor1d the proposition that gives a complete
specification of the extension of a property in that wor1d.
This generalizes to n-constituent interrogatives. For example, the
two-constituent interrogative (35) entails for every two names A and B
the interrogative (36):
(35) Who loves whom?
(36) Does A love B?
The two-constituent interrogative (36) asks for, Le. denotes, a complete
specification of the extension of the relation of loving.
In general, a complete answer to an n-constituent interrogative gives
a complete specification of a possible extension of an n-place relation,
and the propositions that express these specifications are its possible
complete and true answers.
This tells us which object of type (s, t), i.e. which proposition, an
atomic interrogative denotes. At the same time, it also determines what
constitutes the sense of such an interrogative: it is a function from

possible worlds to such propositions.
We conclude that every interrogative is based on some underlying
n-place relation (where we take sentences, which underlie sentential
interrogatives, to be zero-place relations). Every such relation has a set
of possible extensions. To each possible extension corresponds a
possible answer, the proposition which specifies this extension. Such a
proposition is the denotation of the interrogative in the world in which
the underlying relation has that extension. And the sense of an
interrogative is a function from possible worlds to possible answers.
The latter object we call a questian. Schematically, we end up with the
following analysis of atomic interrogatives:
n-place relatian questian
r: W ---> pow(Dn) qr: W ---> {O, l}W
where qr( w) = that p s.th. p( w') <=> r( w) = r( w').
This means that questions can be viewed as relations between worlds of
a special kind. They are equivalence relations between the elements of
W, i.e. they constitute partitions of W. The blocks in these partitions,
sets of possible worlds, are the propositions that are the possible
answers to the questions.
In what follows we will make use of this, and sometimes represent
the meaning of an interrogative, i.e. the question it expresses, as a
partition of W We will use the language of two-sorted type theory, in
which quantification and abstraction over worlds is possible, as our
representation language.
Let us then quickly review how sentential interrogatives and con-
stituent interrogatives are interpreted according to this schema. First,
consider the sentential interrogative (37):
(37) Does John walk?
The underlying zero-place relation (formula) is:
(38) walk(w) (1).
Here, w is a variable of type s, ranging over possible worlds. Obviously,
(37) has two complete answers. In a world in which John walks, this is
the proposition that he does, and in a world in which he doesn't, it is
the proposition that he does not. Le. (37) asks to identify the actual
world as belonging to one of two disjunct sets: those in which John
walks, and those in which he doesn't. This means that (37) partitions
the set of worlds in two:
that John walks
that John doesn't walk

W
The two blocks of this partition are the two propositions which consti-
tute the two possible complete and true answers to (37). The meaning
of (37) can now be represented as follows:
(39) AWAw'[walk(w) (j) = walk(w') (j)].
As a second example, consider the one-constituent interrogative (40):
(40) Who walks?
In principle, this interrogative has as many answers as there are subsets
of the domain that it ranges over. Or, to give a different but equivalent
formulation, each proposition that specifies a possible extension of the
one-place relation of walking is a possible complete and true answer to
(40).l.e., (40) induces the following partition of W:
no onewalks
a is the one that walks
b is the one that walks
a and bare the ones that walk
everyone walks
W
THE SEMANTlCS OF INTERROGATIVES 51
The meaning of (40) can thus be represented as follows:

(41) AWAw'[Ax[walk(w) (x)] = AX [walk(w') (x)]].
A representation of the meaning of the two-constituent interrogative
(42) is (43):
(42) Which man does which woman love?
(43) AWAW'[AxAy[woman(w)(x) 1\ man(w)(y) 1\ love(w)(x, y)] =
AXAy[woman(w') (x) 1\ man(w') (y) 1\ love(w') (x, y)]].
Generally, any n-place relational expression acan be turned into a
question that is the interpretation of the corresponding atomic n-
constituent interrogative, by means of the following schema:
(44) AWAw'[a = (Awa) (w')].
This gives a satisfactory treatment of atomic interrogatives, but, as we
have seen above, we also need to raise the type of atomic itnerrogatives
(s, t), in order to be able to deal with coordinated interrogatives and
pair-list and choice readings.
3.2. Types for Coordination and Embedding

As we saw above, for coordinated interrogatives we need the type that
is the lifted version of the type of atomic ones. I.e. we follow a familiar
procedure: faced with incorrect results when we apply generalized
coordination to expressions of some type a, we go over to the lifted
level, i.e. to type (( a, t), t). The general type-shifting rule of lifting gives
us for every expres sion a of the 'basic' type a a corresponding one
which gives the meaning of a as an expression of the lifted type
((a, t), t):
(45) a =} ((a, t), t)
a =} AX(a,dX(a)].
This is familiar from the analysis of NP's. Let us consider application of
this schema to a simple example of coordination of two one-constituent
interrogatives:
(46) Who walks? And, who talks?
At its basic level each conjunct of (46) is represented as an expres sion
of type (s, t). The first interrogative for example is represented as:
(47) Aw'[.h[walk(w) (x)] = AX [walk(w') (x)]].

Applying the lifting procedure of (45) we get:
(48) A~(s,t),dQ(AW'[x[walk(w)(x)] = Ax[walk(w')(x)]])].
It we apply the same procedure to the second conjunct, and then use
the generalized definition of conjunction, we get (49) as the representa-
tion of the conjunction of interrogatives (46):
(49) A~(s, t), dQ(AW'[AX[walk(w)(x)] = AX [walk(w') (x)]]) /\
Q(AW'[Ax[talk(w) (x)] = AX [talk(w') (x)]])].
The conjunction of the two interrogatives is thus taken to denote a set
of sets of propositions, viz. those which contain the answer to the first
interrogative and the answer to the second one. Obviously, we obtain as
a result that, given the generalized definition of entailment, a conjunc-
tion of interrogatives entails each of its conjuncts, for every set of
propositions that contains a complete and true answer to both con-
juncts, contains a complete and true answer to each conjunct.
Next, consider disjunction. Again, we first lift the disjuncts, and then
apply generalized disjunction. (50) is then represented as (51):
(50) Who walks? Or, who talks?
(51) A~(s, t), dQ(AW'[Ax[walk(w) (x)] = AX [walk(w') (x)]]) V
Q(AW'[Ax[talk(w) (x)] = AX [talk(w') (x)]])].
Applying generalized entailment, we see that a disjunction of interroga-
tives is entailed by each of its disjuncts, and, moreover, that a conjunc-
tion entails the corresponding disjunction. Again, bearing the intuitive
content of entailment between interrogatives in mind, these results are
what we want.
As a matter of fact, it can be noticed that generalized conjunction
and disjunction is defined at type (s, t) (interrogative denotations) and
(s, (s, t (interrogative meanings) as well. A little reflection shows that
conjunction could be treated at this level, but disjunction can't. The
reason for this is simple. As we saw above, atomic interrogatives induce
partitions of W. Pointwise intersection of two partitions (which is what
conjunction would amount to) results in a partition again. That is, we
get an object, not only of the right type, but also of the right sub-type,
"
i.e. one which inherits the defining properties. However, taking the
pointwise union of two partitions (which is what generalized disjunction
does) in general does not result in a partition again. What we get is of

the right type, but not of the right sub-type. So, as a additional
requirement on dealing with coordination in general, we can state that
coordination should be performed at the lowest type that can be
reached from the key type, provided that it respects (i.e. stays within)
the relevant subdomains there,12
Let us now turn our attention to embedded interrogatives. Given
the type ( s, t) of atomic interrogatives, the lowest possible type for

interrogative-embedding verbs such as know is s, t ), ( e, t . Taking
these verbs to denote basically relations between individuals and
propositions has some agreeable consequences.
First of alI, given the kind of object we assign to interrogatives as
their meaning (and some familiar, though not always uncontroversial,
assumptions about the semantics of epistemic verbs) we get an account
of the validity of such schemas as (52) and (53):
(52) x knows whether ~
~
x knows that ~
(53) x knows whether ~

not-~
x knows that not-~

Also, given this type-assignment there is no problem in allowing for
coordination of sentential and interrogative complements, using stand-
ard generalized coordination rules:
(54) John knows that Peter has left for Paris, and also whether
Mary has gone with him.
Notice that the following schemata are intuitively valid:
(55) x knows whether ~ and whether 1/J ~

x knows whether ~ and x knows whether 1/J.
(56) x knows whether ~ or whether 1/J ~
x knows whether ~ or x knows whether 1/J.
Above we observed that conjunction at the (s,t )-Ievel respects sub-
domains in the case of interrogatives. However, the lowest level at
which disjunction respects subdomains is that of type (s, t), t), t).
This means that the type of know when it takes a coordinated
interrogative complement has to be s, t), t), t),(e, t)). We get the
required results when we apply a second general type-shifting opera-
tion, that of 'argument-lifting': 13
(57) (a, c) => ( a, t), t), c), provided c is a conjoinable type
a a (y))]
=> A~(a, t), t)[Q(X, Ya'
where Q(X, y, O) = X(AYO), if O is of type t
= AXd[Q(X, y, O(X d))] ,
if O is of type (d, f),
This type-shifting rule allows us to lift the argument of a functor, and
provides a semantics for the resulting functor in terms of its original
interpretation. The example of lifting know of type s, t), (e, t)) to
s, t), t), t), (e, t)) illustrates this, Application of (57) gives the
following result:
(58) A~s, t), t), dAXe [Q (Ap(s, t)[knows, t),(e, t))(p) (x)])]],
It we apply this translation of know to a disjunction of interrogatives,
such as (50) above, we get the required distributive result.
Summing up, for extensional interrogative-embedding verbs, such as
know, we can employ a key type s, t), (e, t)) for dealing with
embedded atomic interrogatives, and for conjunctions. For dealing with
embedded disjunctive interrogative complements we need the derived
type s, t), t), t)(e, t)), which we get by applying the type-shifting
procedure of 'argument-lifting' defined in (57). The latter procedure
allows us to deal in general with cases where a functor is to be applied
to an argument that itself has been lifted.
Besides extensional embedding verbs there are intensional ones,
such as wonder. What basic type is to be assigned to them? One might
think that a simple intensionalization of the basic type of extensional
verbs would do. But the semantics of coordinated interrogative comple-
ments again provides a counter-argument. Above, in section 2.1, we
observed that whereas extensional verbs distribute over disjunctive
complements, intensional ones don't, at least not always, Consider (59):
(59) John wonders who walks or who talks.
The point is that (59) is ambiguous between a wide scope or and a
narrow scope or reading (with respect to wonder), These different
readings can be paraphrased as (60) and (61 ):
(60) John wants to know who walks or to know who talks.

(61) John wants to know who walks or he wants to know who
talks.
Trying to keep the analogy between extensional and intensional verbs
as close as possible would suggest to give them a basic type s, (s, t,
(e, t. In order to deal with (59), we have to apply argument lifting
again. But then we would get a distributive reading only. In order to get
the non-distributive reading we need another, higher type, and it is
clear what this type should be. On the non-distributive reading of (59)
wonder takes the intension of the entire disjunction as its argument,

hence, in this case it is of type s, (s, t), t), t, (e, t. This, then, is
the key type of intensional interrogative embedding verbs. In order to
account for the wide scope or reading of (59), paraphrased in (61), we
might proceed in two different ways. In the line of Partee and Rooth's
treatment of ordinary intensional transitive verbs, we could apply
disjunction at the level: lift(e, 1) (intension (lift s, t) ). Or, we could first
apply an operation of argument-Iowering, and then argument-lift again
to get back original the type of wonder. For several reasons, we prefer
the latter option. First of all, we think there are arguments against the
function-argument 'flip-flop' that the former strategy involves (see also
section 3.3). Secondly, we need argument-Iowering anyway, in order to
arrive at simple representations of sentences with atomic interrogatives
embedded under intensional verbs.
On the basis of the discussion so far, we can distinguish the following
interpretation domains for interrogatives in naturallanguage:
(s, (s, t (s, (s, t), t), t
II IV
INTt ~ EXT INT t ~ EXT
LIFT
(s, t) .. " (5, t), t), t)
LOWER
I III
In this figure, we see the four interpretation domains for interrogatives

which we discussed above, and the type shifting operations which
connect them. The first domain, that of type (s, t), is the denotational
key type for atomic interrogatives and contains the objects that are the
interpretations of the arguments of extensional interrogative-embedding
verbs. The second domain, that of type (s, (s, t)), is the key type for
meanings of atomic interrogatives, i.e. the level at which entailments
between them are to be accounted for. The third domain, that of type
(s, t), t), t), is the denotational key type for coordination of interroga-
tives. And the fourth domain, that of type (s, (s, t), t), t)), contains the
proper objects to be recognized as the meanings of such coordinated
interrogatives: they are of the proper type to be associated with the
arguments of intensional interrogative-embedding verbs, and they
contain the right structure for an account of entailment between
coordinated interrogatives. The domains I and II, and III and IV, are
related by the type shifting rules of INT (intensionalization) and EXT
(extensionalization). The key type for atomic interrogatives and the key
type for coordinated ones are related by the operations LIFf (lifting)
and LOWER (lowering). Notice that the latter is a partial function.
Notice also that only a proper sub set of each of these domains contains
the right objects to serve as interpretations of interrogatives in their
various roles. These subsets are characterized by the specific semantic
interpretation rule (44) that we gave for atomic interrogatives, which
defines the characteristic ('partition') properties which are 'preserved'
by the general type shifting principles.
Are these four all the interpretation domains for interrogatives?
Probably noL One domain one might also want to use is De> which is to
serve as the domain for nominalized interrogatives (e.g. as in 'Whether
~ is a difficult question'). And others might be distinguished as well.
Prominent among them, at least in the context of this paper, are the
relational types that the categorial approach uses. Do they, too, form a
possible interpretation domain for interrogatives that can be fitted into
a flexible framework such as outlined in this section?
That is the topic of the next section, the possible unification of the
categorial and the propositional approach.
3.3. Type Shifting as Unification?

The discussion in the preceding sections was largely aimed at finding
THE SEMANTlCS OF INTERROGATIVES 57
the proper types of semantic objects for the interpretation of interroga-

tives in the contexts of coordination, entailment and embedding. We
found that no one unique type serves as the proper type in alI contexts,
and that we need to pursue a flexible approach in which various
domains, connected to each other by general type shifting procedures,
areused.
In this subsection we want to consider another (set of) type(s) for
interrogatives, the relational ones, which the categorial approach uses
to give an account of another construction into which interrogatives
enter, viz. interrogative/answer pairs. This consideration willlead us to
illustrate yet another aspect of the use of type shifting, viz. that of
unifying equally well-motivated but different semantic approaches
dealing with different parts of some empirical domain. Above we saw
that there are two main approaches to the semantics of interrogatives:
the categorial and the propositional one. The first assigns different
relational types to different kinds of interrogatives, the latter postulates
a unique, propositional type. Also, we saw that arguments in favour of
each can be given, arguments which by and large are complementary.
This suggests that at some level of analysis the two approaches need not
be in conflict. The semantics we outlined above may well contain the
elements that such a unification needs. Recall that it is based on the
following rule:
n-place relation question

r: W -> pow(Dn) qr: W -> {O, 1} w
where qr(w) = that p s.th. p(w') ~ r(w) = r(w').
On the left hand side we find the kind of semantic objects that the
categorial approach typically associates with interrogatives. And on the
right hand side, we have a propositional type. So the basic rule of our
semantics might also be looked upon as turning a categorial analysis
into a propositional one. Couldn't we, then, view this rule, too, as a type
shifting rule, i.e. add to our stock of semantic domains for interrogatives
that of n-place relations, and postulate the rule as an additional type
shifting tool?
Let us first indicate what would be the advantages of such a move.
As we saw above, the categorial approach is inspired by the semantics
of characteristic linguistic answers to interrogatives. And it deals with
them in a natural way. Consider the following example:
(62) Which man walks in the park?

(63) Who walks in the park?
(64) Hilary.
(65) Hilary walks in the park.
Given the sex-neutral status of the proper name 'Hilary', this example
clearly shows that the semantic interpretation of a linguistic answer
depends on the semantic interpretation of the interrogative it answers.
The information that (64) and (65) convey differs according to whether
they answer (62) or (63).
Exactly which semantic property of an interrogative it is, that is
needed for the interpretation of a linguistic answer, we illustrated
above, in section 1.3, with an example like the following:
(66) Who will come to the party?
(67) Who will not come to the party?
(68) John and Mary.
Above we noticed that on a propositional approach, there tends to be
no semantic difference between (66) and (67). The proposition (or
propositions) that give a complete specification of the positive exten-
sion of some property or relation are the same as the one(s) that give(s)
a specification of its negative extension. However, the meaning of (68)
differs depending on whether it answers the positive or the negative
question. From this we drew the conclusion that the semantic inter-
pretation of characteristic linguistic answers essentially involves the
re/alion that underlies the question expressed by an interrogative.
On the other hand, we have seen that there is ample reason for a
propositional level as well. So, it seems that there are two complemen-
tary semantic analyses, each accounting for different aspects in the
meaning of interrogatives and their answers. Unifying them could be
done by postulating n-place relations as a possibIe interpretation
domain for interrogatives and by regarding the rule specified above as a
type shifting rule.
It should be remarked at the outset that we are entering largely
uncharted territory here. One reaction to the afore going question,
whether our basic semantic ruIe caii be viewed as a type-shifting
principle, might be one of distrust: it certainly does not look like the
ones we are familiar with. But another reaction might be: 'Why not, ifit
does the satne kind of work as the others, and does that properly?'
What we seem to lack is a theory of type shifting rules. Although
investigations have been made into the formal properties of various
conglomerates of type shifting rules,14 a body of general and intuitive
constraints characterizing the notion of a type shifting rule as such stiH
remains to be formulated. Unfortunately, we do not have anything to
offer on this score. We just want to point out that there may be a
relation between what one wants to consider as a bona fide principle
and the view one takes on their place in the grammar. It one considers
them to be part of the syntax one's attitude might be just a little more
conservative then if one takes them to play a role in the relation
between syntax and semantics.
Without taking a very firm stand on the matter, we suggest that the
discussion so far has provided evidence for the claim that it is possible
and profitable to take the rule in question to be a type shifting
principle. However, there is a potential problem that such a move
meets. And this problem raises some further reaching questions
regarding the place of type shifting principles in the grammar. The
problem is that of potential overgeneration of meanings of expressions,
and it occurs not only with this particular type shifting rule (if such it
is). In order to discuss this problem, let us first give a very rough
indication of our view on the place of type shifting in a grammar.
Very roughly speaking, we might distinguish two ways of incor-
porating flexibility in the grammar. On the first one, what we have
called type shifting rules are in fact considered to be category changing
rules, which form an integral part of the system of syntactic rules and
categoriesY This approach is orthodox in sa far as it adheres to a rigid
and unique category-to-type correspondence, and consequently to strict
compositionality. For example, accounting for scope-ambiguities by
means of category-changing rules leaves unchallenged the principle that
non-lexical ambiguity in the semantics should be based on derivational
ambiguity in the syntax. However, the view in question also has some
unorthodox features, the most surprising and interesting, perhaps, being
the willingness to give up the traditional notion of constituent struc-
ture. 16 In view of what follows, it should be noted that in a categorial
syntactic framework giving up constituent structure means giving up a
notion of syntactic function-argument structure.
Another view on the place of type shifting rules in the grammar is
more semantic. On this approach, one of the uses of type shifting is to
keep the syntax free from unnecessary complications, such as syntac-
tically unmotivated derivational structures. The notion of constituent

structure, with its associated function-argument structure, is retained. In
fact, as we will argue shortly, it can be used to deal with one of the
problems that the incorporation of type shifting in the grammar posits,
viz. that of overgeneration. The unorthodox aspect of the semantic view
on type-shifting resides in the attribution of meanings to syntactic
structures. In giving up a rigid and unique category-to-type correspond-
ence, it also gives up strict compositionality. Flexibility does not play a
roIe in the syntax, nor in that part of the semantics that consists of the
abstract theory of semantic objects that serve as meanings, but it
concerns the relationship between syntactic structures and meanings.
Of course, this does not imply that there may not be any need for
flexibility in the other components of the grammar as wellP However,
we are convinced that in many cases, e.g. coordination (including non-
constituent conjunction), scope ambiguities, type/token distinctions,
embedding constructions, and so forth, the semantic approach to
flexibility is the more advantageous one. It keeps the syntax simple, and
it links the phenomenon of flexible interpretation to syntactic construc-
tions and contexts.
So, the basic tenet of the approach to type shifting that we favour,
can be summarized in the following three statements:
No fixed category-to-type assignment is assumed, but a
family of types, generated by type shifting rules from a key
type, is postulated for each syntactic category.
Basic expressions go to a key type in the family associated
with their category, and have potential meanings in the other
types of the family predicted by the type shifting rules.
Interpretation of syntactic structures is liberalized to a
relation: 'anything goes that fits'. Le., a syntactic structure has
as many meanings as can be generated from the potential
meanings of its constituents.
A simple, and familiar, example is provided by the analysis of (exten-
sional) transitive verbs and their arguments. We postulate one syntactic
category for these transitive verbs, TV, and one for noun phrases, NP.
The key type corresponding to the category TV that the grammar
specifies, is (e, (e, t, and that corresponding to NP is e. On these
types, type shifting rules may operate generating new types. Lexical
expressions are given a basic translation (are assigned a basic meaning)

of one of these types, and they obtain derived meanings in various
(though not necessarily all) of the other types which are associated with
their category by the type shifting rules. It a TV occurs with two proper
names alI expressions involved will fit an the basis of their basic type
and meaning. Hence, no type-shifting is called for. It one NP, say the
abject, is a properly quantified expression, which is given e, t), t) as
its basic type, the basic type of the TV is inadequate. However, one of
t,
its derived types is ( e, t), t), (e, the result of the application of
argument-lifting ta its basic type, with which is associated a derived
meaning for the TV within that type. Combining these gives a fitting
result. Scope ambiguities of NP-arguments of TV's can be accounted
for as folIows. It can be argued quite generally that type shifting
principles which operate an arguments of functions, must be able ta
operate at arbitrary depth. 18 Different relative scopes of NP-arguments
then result from lifting argument places in different orders. No deriva-
tional ambiguity is needed in the syntax, the readings we want, simply
arise because generating the relevant type for TV's in two different
ways, generates two different meanings for a TV in that type. Clearly,
this approach does not eliminate the complexity of the syntactic view,
but it places it in a different part of the grammar. This can be motivated
not only by an appeal ta a certain kind of intuition ar ta elegance, but
also by pointing out empirical differences.
Ta see why this can be sa, it is important ta note that adding
flexibility in the form of type shifting principles ta the grammar,
whether in the syntax ar in the semantics, faces a potential problem.
These mechanisms may enlarge the power of the grammar. On the
syntactic approach, this means that expressions may be recognized
which do not belong ta the language. And if we folIow the semantic
view, we run the risk of giving an expression a potential meaning it does
not have, i.e for which no context can be found in which that expres sion
must be assigned that meaning. Ta what extent this actually happens,
depends, of course, on the actual set of type shifting rules one adopts.
For example, in Partee and Rooth (1983) type-shifting principles are
used ta give an account of so-called 'wide scape or' readings of
sentences such as 'The department is looking for a phonologist ar a
phonetician'. The way they proceed differs from the strategy we have
folIowed in the previous section. They use a type shifting rule which
alIows them ta interpret the disjunctive abject NP as a function which
takes the TV as an argument, thus giving it the required scope.

However, the same mechanism will also predict impossible readings in
certain cases. For example, the mechanism employed by Partee and
Rooth also predicts that the sentence 'Every student failed or got a D'
has as one of its readings 'Every student failed or every student got a
D', which it does not. Partee and Rooth do not offer a solution for this
problem.
In the present case, i.e. if we add the question formation rule to our
stock of type shifting principles, overgeneration occurs as well. Adding
the rule in question has the rather unpleasant consequence that our
grammar predicts that any expression that expresses a relation, also,
potentially, has the meaning of the corresponding question. For exam-
ple, any simple indicative sentence also gets assigned the interpretation
of the corresponding sentential interrogative, which is clearly something
we do not wish.
A possible solution can be found along the following lines. We
restrict the use of type shifting in generating meanings by combination.
Suppose that, as usual, functional application of meanings serves as the
interpretation of the syntactic operation of concatenation, i.e. that we
have rule pairs like the following:
syntactic rule: BIA +A ---+ B
fJ a y
semantic rule: y' = FA(fJ', a')
In an unrestricted flexible framework, such a semantic rule is a rule
schema, allowing a' and fJ' to be any possible translation that can be
obtained by means of the type shifting rules, as long as ordinary
functional application applies to such a pair of translations. So, a whole
set of translations y' will be the result of applying the semantic rule. We
propose to put the following constraints on the possible translations of
a and fJ:
fJ' should be a possible translation of fJ which is obtained
from its basic translation by only applying argument shift
rules.
a' should be a possible translation of a which is obtained
from its basic translation by only applying global shift rules.
"
The syntactic function-argument structure should be re-
spected.
(Of course, for other syntactic rules, we might want to formulate other
restrictions on the corresponding translation part.) Thus restricted,
functional application allows us to obtain only certain semantic objects
as meanings of complex expressions. (E.g. Partee and Rooth's treatment
of wide scope or would be prohibited, since it implies a reversion of the
function-argument structure of the VP in question.) Let PM( a) be the
set of possible meanings of a. For basic expressions, this is a unit set
(disregarding lexical ambiguity). For derived expressions, it may contain
more than one element. The possible meanings of a complex expres sion
y built by concatenation from a and {3 can then also be defined as
follows:
PM(y) = {t(b) (g(a I b E PM({3), a E PM(a)} = FA({3', a')
where f is any composition of argument shifts, and g is any
composition of global shifts.
(The difference between 'global' and 'argument' shifts is the difference
between e.g. lifting and argument-lifting, intensionalization and argu-
ment-intensionalization, etc.) Notice that this way of implementing type
shifting in the grammar has a remarkable consequence: it makes the
notion of the meaning of an expression relative to its syntactic context.
The meaning of a as a part of {3 with meaning b is that possible
meaning a of a that is used to derive {3 with meaning b. We think that
this consequence is intuitively appealing. Consider the example of a
proper name. Basically it has just one meaning, that of being the name
of an individual. It is only in certain (syntactic) contexts, such as in
coordination with a quantified term, that we consider giving it a derived
meaning as well. Or, consider the case of an atomic interrogative. In
isolation, they must be given a meaning of the proper relational type. It
is only e.g. in the context of an embedding verb, that we assign them
their meaning in the type of questions. As for entailment, we saw that
that requires this propositional type of meaning, too. But then, entail-
ment is a re/ation of which the interrogatives are arguments.
We want to end this admittedly rough sketch with the following
remark. In view of the fact that overgeneration is a potential problem
both for the syntactic and for the semantic approach to flexibility, the
latter, we think, has this going for it that it can employ independently
motivated and restricted notions, such as the function-argument struc-
"-
ture that is inherent in a suitably restricted account of constituent
structure,19 in dealing with this problem. This seems to square with the
semantic relevance that constituent structure can be assumed to hgve.
But we do not want to suggest that the syntactic view on type-shifting

couldn't be sufficiently restricted too. The entire enterprise of incor-
porating flexibility in the grammar is only just beginning, and it seems
wise therefore to explore various options.
4. CONCLUSION
In this paper we have tried to show that generalized notions of

coordination and entailment can be fruitful means to obtain more
insight into the nature of the semantics of interrogatives. Their useful-
ness, both on a descriptive, and on a methodological plane, has been
demonstrated in the foregoing, we feeI. In the course of doing so, we
have made some critical remarks about existing theories within the
propositional and the categorial approach. We want to emphasize that
the observations and remarks that we have made, in no way pretend to
show that these approaches as such are wrong. On the contrary, we feel
that both are right. Our discussion does show, however, that they cover
only part of what an adequate semantic theory of interrogatives should
account for. We also have tried to sketch a theory that incorporates the
insights of both approaches. And in a more speculative manner, we
have indicated that a flexible way of relating syntactic structure and
semantic interpretation may be of great help in achieving this. The exact
format of a grammar that encompasses these principles, is stiH in need
of further investigation. For one thing, one would like to have some
intuitive and independently motivated constraints on what are adequate
and natural type-shifting mechanisms. Despite the many interesting
contributions one can find in the literature, we feeI that this is stiH
largely an open question.
The last remark we want to make, concerns the necessity of incor-
porating an semantic analysis like the one presented above, into the
framework of a theory of intensional objects. The reason for being
interested in this, is that one would like to regard questions, the
meanings of interrogatives, as constituting a separate category of
intensional objects, in much the same way as properties and proposi-
tions do, and for similar reasons. Notice that, since the analysis is
carried out in a standard possible worlds framework, questions <\re
treated extensionally, in the sense that two questions which everywhere
have the extension (i.e. the same complete and true answers), are
identified. In other words, what the framework provides is only an
extensional identity criterion for questions, just as it only gives exten-

sional identity criteria for properties and propositions. Le., we are able
to give an account of the kind of intensional objects that questions are
(viz. equivalence relations between possible worlds), but we do not have
the means to represent alI of the intensionality that they comprise. Just
as being true of the same individuals in every world is a necessary, but
not a sufficient condition for identification of two properties, having the
same true and complete answers in alI situations is not alI there is to
two questions being identical. There is, of course, a relation between
these two facts. Take any two different properties which, in some
suitably chosen set of alternatives, apply to the same objects in alI
situations. Consequently, the question that is based on the first one, wiH
be extensionalIy equivalent to the question which is formed from the
second. But the questions are not the same, just as the properties are
not. For someone might wonder what the extension of the first property
is without also wondering which objects the second one applies to.
How would one incorporate this fact in something like Chierchia and
Turner's theory of properties? 20 One might think that once one has an
intensional theory of properties and/or propositions one automaticalIy
also has an intensional theory of questions, since questions are defined
in terms of properties and propositions. What one would do, then, is
define possible wOrlds using the notion of a proposition and, given that,
define questions as equivalence relations on them. But this is in fact stiH
an extensional approach to the semantics of interrogatives: it stiU
identifies any two extensionalIy equivalent interrogatives, i.e. interroga-
tives which have the same true answers everywhere.
The proper way to go about, then, is to extend property theory to a
general theory of intensional objects, which recognizes besides proper-
ties and propositions, also relations, individuals and questions. Another
argument to the effect that questions constitute an intensional category
in their own right, can perhaps be taken from the mutual dependence of
questions and propositions, interrogatives and indicatives. It is, at least
so since Frege, a commonplace to regard the sentence (statement) as a
fundamental building block of language. But this is only part of the
truth. One of the main functions of language no doubt is to discriminate
the actual wOrld (state of affairs) among the possible ones. But this
function is triggered only when the question of where the actual world
is located, is raised in the first place. To be sure, the dependence is
mutual, for raising a question clearly presupposes the possibility of
making the discriminations which the question calls for. So functionally,

at least, questions and propositions are mutually dependent, a fact
which we might see reflected in the fact that an extensional derivation
of either category to the other is doomed to faii. Within the context of a
general theory of intensional objects these considerations call for the
introduction of a new basic type in our ontology, that of questions, and
for the concomitant formalization of a new extensionalization relation,
between questions and propositions. This relation is, of course, the
relation of answerhood, i.e. the relation of being a complete and true
semantic answer. But an account of that is another topic.
NOTES
I This paper is a further development of some ideas in Groenendijk and Stokhof

(1984), especially chapter VI. Also, various other aspects of the approach described
here are explained and defended there in more detail. We have refrained from
bothering the reader with detailed references.
2 This terminology may be slightly confusing. Certainly to the Montogovian, the use of
two different names suggests that there are two different underlying derivational
processes at work. However, this is not the case. Both readings are the result of one and
the same derivationa! process. It is the internal semantic structure of the term phrase
that is used that determines which reading is the result. Moreover, notice that the two
readings, in a sense, are not complementary. The result we get if we use a term such as
two of his friends gives a choice reading, but once a certain choice has been made what
is required is a list of pairs. Likewise, a simple proper name can be viewed as a trivial
one-item list. In what follows, we will not go into the details of the derivation of pair-list
and choice readings, since in this paper we are only interested in the relevant types.
3 Cf. Gazdar (1980), Keenan and Faltz (1985), Partee and Rooth (1983).
4 An example is conjunction which functions as 'addition'. See Partee and Rooth
(1983), Partee (1986).

5 Notice that the key type is not necessarily the minimal type, in the sense of the least
complex type, of an expression. For example, the least complex type of proper names is

e but their key type is e, t), t ).
6 See e.g. Hausser (1983).
7 It should be noted that for interrogatives of the same kind, a categoria! theory might
obtain correct results by appealing to the same mechanism that we will propose to use,
viz. lifting (see below). Two remarks are in order. First, in a sense such a move goes
against the nature of the approach. Second, this observation does suggest an adjustment
of the use of coordination and entailment we are making here. As an evaluation
measure it works if we constrain the use of such type-shifting procedures as lifting in
order to account for coordination. The following seems intuitively justified, and
prevents the move just mentioned: coordination,should be accounted for in the lowest
common type in which it respects 'subdomains' (see section 3.2).
8 See Karttunen (1977).
9 See Bennett (1979) aud Belnap (1982). What is said here about their approach is a
kind of rational reconstruction of just one aspect of it. The reader is urged to consult
their papers for more information.
10 Notice that something similar would hold for expressions of type e if the domain De
would have an entailment structure defined on it.

II See e.g. Stalnaker (1984).
12 Again, it should be noted that this is not characteristic for coordination of interroga-
tives. The same applies to other domains that are structured by entailment. Cf. also
note 10.
13 See Partee aud Rooth for another application of this rule. There is a difference in
the way they account for wide scope Of readings aud the way in which we proceed.
On their analysis, there is what they call 'function-argument flip-flop'. We keep the
function-argument structure intact. For a motivation, see section 3.3.
14 See e.g. van Benthem (1986).
15 See e.g. Ades aud Steedman (1982), van Benthem (1986), Dowty (to appear).
16 A clear and well-argued case is presented by Zwarts (1986).
17 For example, Moortgat (to appear) argues that we need flexibility in the mor-
phology, and the 'right node raising' constructions discussed in Dowty (to appear) may
be presented as arguments for some kind of flexibility in the syntax.
18 In fact, the distinction is rather particular to a functional formulation of type-theory.
If we were to use a relational version (see Muskens (1986) for an exposition aud some
arguments in favour of using such a theory), we would simply say that argument-lifting
may operate on any argument of a relation.
19 The restricted framework developed in Landmau aud Moerdijk (1983) seems to
offer a good starting point.

20 See Chierchia and Turner (1987).
REFERENCES
Ades, A. and Steedman, M.: 1982, 'On the Order of Words', Linguistics and Phi-
losophy 5, pp. 517-558.
Belnap, N.: 1982, 'Questions and Answers in Montague Grammar', in: S. Peters and E.
Saarinen (eds.), Processes, Beliejs and Questions, Reidel, Dordrecht, pp. 165-198.
Bennett, M.: 1979, Questions in Montague Grammar, Indiana University Linguistics
Club, Bloomington.
Benthem, J. van: 1986, 'The Semantics of Variety in Categorial Grammar', in: W.
Buszkowksi, W. Marciszewski and J. van Benthem (eds.), Categorial Grammar,
Benjamins, Amsterdam.
Chierchia, G. and Turner, R.: 1987, 'Semantics and Property Theory', forthcoming in
Linguistics and Philosophy.
Dowty, D.: To appear, 'Type raising, Functional Composition, and Non-Constituent
Conjunction', in: D. Oehrle, E. Bach and D. Wheeler (eds.), Categorial Grammars
and Natural Language Structures, Reidel, Dordrecht.
Gazdar, G.: 1980, 'A Cross-Categorial Semantics for Coordination', Linguistics and
Philosophy 3, pp. 407 -410.
Groenendijk, J. and Stokhof, M.: 1984, Studies on the Semantics of Questions and the
Pragmatics ofAnswers, diss., Department of Philosophy, University of Amsterdam.
Hausser, R: 1983, 'The Syntax and Semantics of English Mood', in: Kiefer, F. (ed.),
Questions and Answers, Reidel, Dordrecht, pp. 97-158.
Karttunen, L.: 1977, 'Syntax and Semantics of Questions', Linguistics and Philosophy 1,
pp.3-44.
Keenan, E. and Faltz, L.: 1985, Boolean Semantics for Natural Language, Reidel,
Dordrecht.
Landman, F. and Moerdijk, I.: 1983, 'Compositionality and the Analysis of Anaphora',
Linguistics and Philosophy 6, pp. 89-114.
Moortgat, M.: To appear, 'Mixed Composition and Discontinuous Dependencies', in: D.
Oehrle, E. Bach and D. Wheeler (eds.), Categorial Grammars and Natural Language
Structures, Reidel, Dordrecht.
Muskens, R: 1986, 'A Relational Formulation of the Theory of Types', ITLI-pre-
publication 5, Amsterdam.
Partee, B.: 1986, 'Noun Phrase Interpretation and Type-Shifting Principles', in: J.
Groenendijk, D. de Jongh and M. Stokhof (eds.), Studies in Discourse Representa-
tion Theory and the Theory of Genera/ized Quantifiers, Foris, Dordrecht, pp. 115-
144.
Partee, B. and Rooth, M.: 1983, 'Generalized Conjunction and Type Ambiguity', in: R
Buerle, C. Schwarze and A. von Stechow (eds.), Meaning, Use and Interpretation of
Language, de Gruyter, Berlin!New York, pp. 361-383.
Stalnaker, R: 1984, Inquiry, Bradford Books, Cambridge, MA.
Zwarts, F.: 1986, Categoriale Grammatica en Algebra'ische Semantiek, diss., Depart-
ment of Linguistics, University of Groningen.
University Of Amsterdam
DA VID R. DOWTY
ON THE SEMANTIC CONTENT OF THE

NOTION OF 'THEMATIC ROLE'
1. INTRODUCTION
The notion of "thematic roles", a more modern term for Fillmore's

(1968) case relations, Jackendoff's (1972, 1976) and Gruber's (1965)
thematic relations, and Panini's karakas, has been appealed to by
contemporary linguists in the statement of natural language generaliza-
tions about syntax, morphology, and semantics for some 20 years now.
Until recently, thematic roles were invoked only in research that treated
semantics non-formally (i.e. not model-theoretically). The level of
semantic rigor behind these notions in this literature is suggested by the
characterizations in (1) of some of the roles from Andrews (1985) and
Jackendoff (1985), which are very typical of the way roles are intro-
duced in this literature.
(1) a. Agent - a participant which the meaning of the verb
specifies as doing or causing something, possibly intention-
ally. Examples: subjects of kill, eat, hit, smash, kick, watch.
b. Patient - a participant which the verb characterizes as
having something happen to it, and as being affected by what
happens to it. Examples: objects of kill, eat, smash, but not
those of watch, hear, and love.
c. Experiencer - a participant who is characterised as
aware of something. Example: subject of love, object of
annoy.
d. Theme - a participant which is characterized as chang-
ing its position or condition, or as being in a state or
position. Example: object of give, hand, Subject of walk, die.
From Jackendoff (1985):
e. Source - object from which motion proceeds. [e.g. the
box in remove the bookfrom the box -'l)RD]
f. Goal - object to which motion proceeds. [e.g. the box in
put the book into the box - DRD].
69
1989 by Kluwer Academic Publishers. Ali rights reserved.
70 DA VID R. DOWTY
To be fair to these linguists, it should be said that such descriptions

are not intended as final deJinitions of thematic roles but are no doubt
thought of as the kind of interim, rough-and-ready intuitive charac-
terizations that linguists sometimes find it expedient to give of those
concepts that are acknowledged to be real ones in natural language but
whose precise theoretical explication is complex and stiH controversial;
the assumption is that the niceties can safely be postponed until later,
but research can go forward in the interim. The I-can't-define-it-but-I-
know-it-when-I-see-it stance toward the notion "coreference" expressed
in 1970 by Paul Postal (Postal 1970,439)1 also exemplifies this kind of
methodological tack.
It thematic roles are indeed the fundamental categories of linguistic
structure that these linguists assume they are, then we should expect
research involving thematic roles to converge on a consistent system,
attested eventuaHy in many languages; the interim seat-of-the-pants
approach to roles would then have been vindicated. In fact, two
problems with thematic roles remain persistently vexing: (i) lack of
agreement among linguists as to which thematic roles exist, and the
absence of any obvious way to decide this question, and (ii) the lack of
any effective way to independently justify the assignment of noun
phrases to thematic roles in particular sentences. It is unfortunately not
unheardof for two researchers to propose assigning the same NP in an
example to two different thematic roles, the assignments fitting the
respective syntactic generalizations about thematic roles that each
wants to maintain, with no way of resolving the conflict coming to light
except to choose whose syntactic generalization is going to be pre-
served. Doubts remain in some quarters whether the pre-theoretic view
of thematic roles will evolve into a cJear and linguistically well-
motivated theory or whether roles will eventually be replaced by some
other concepts, just as the notion of "coreference" that was often taken
for granted in the 1970's did not evolve into something precise but was
instead abandoned in favor of rather different concepts.
It seems obvious that if any independent justification for a substan-
tive, motivated, and reasonably precise theory of thematic roles is going
to be found, it will have to come from semantics. Characterizations like
those in (1) are manifestly semantic in nature, and 1 know of no
proposed way of characterizing theqatic roles that is not equally
semantic. It is encouraging that the past two years have seen the
appearance of attempts by two formal semanticists (Chierchia 1984,
SEMANTIC CONTENT OF THEMA TIC ROLES 71
Carlson 1984) to construct model-theoretic accounts of thematic roles.

However, I am troubled that some very fundamental questions about
the notion of thematic roles have not yet been explicitly raised. In this
paper, I will try to explain what these fundamental questions are and
why I think they are important, though I cannot begin to really answer
some of them.
In sections 2 and 3 of this paper, I will construct two different
theories of thematic roles from the point of view of model-theoretic
semantics. Both of these are what I call "strong" theories of roles, in
that they attempt to incorporate ideas about roles that have been
advocated by different linguists. In section 4, I will discuss some
reasons to believe that both these theories could apply to natural
language, but in different domains. A broader goal of sections 2
through 4 is to generate some feeling for what the difference in
linguistic predictions of the two semantic theories of roles could be. In
section 5, I briefly survey linguistic evidence that has been cited for
thematic roles, suggesting that it may really warrant something weaker
than the "strong" theories of roles formalized earlier, and section 6
closes with some speculations about the cognitive and developmental
linguistic functions that "weaker" conceptions of thematic role might
serve to explain.
2. TWO THEORIES OF THE PREDICA TE-ARGUMENT RELATION
I will begin by contrasting two theories of how the meanings of

predicates in natural language are to be associated with those of their
arguments in the process of semantic interpretation. The first theory,
embodied in what I will call an ordered-argument system, is, pre-
sumably, the theory of predicate and argument interpretation we have
inherited from formal predicate logic and its model theory, which
persists not just in linguistic theories employing Montague semantics
but also in situation semantics and elsewhere. The second theory is that
of a thematic ro les system, in which thematic roles like Agent, Patient,
Source, Goal, etc. are somehow involved in associating a predicates's
meaning with its arguments in semantic interpretation; the linguistic
theories called lexical functional grammar, or LFG, (Kaplan and
Bresnan 1982), and government-binding theory, or GB, (Chomsky
1981, van Riemsdijk and Williams 1986) claUn to embody some
version of a thematic roles system, as does the less formal research of
72 DAVID R. DOWTY
linguists like Jackendoff. The first problem at hand is to try to under-

stand what fundamental differences there are, if indeed any, between
these two theories.
Let me begin asking what a thematic role system is by stating two
things that it is not. First, it is not one that differs from an ordered
argument system only in that we use notations like (3) in place of (2):
(2) give(x, y, z)
(3) {(predicate give), (agent x), (theme y), (gaal z)}.
That is, a thematic role system is not just a notation in which the
various arguments of a predicate are distinguished from one another by
their order in a sequence but by giving each one an arbitrary but
distinct label, so that one can rearrange them at will, but nevertheless a
theory in which use of the labels "Agent", "Theme", "Goal," etc. implies
no semantic commitment at alI (i.e. a theory which implies none beyond
the necessary but minimal task of distinguishing one argument from
another in the syntax that is interpreted). Now, the notion in (3) may
have advantages for certain purposes - for example, if we are using a
syntactic or parsing framework that makes much use of the operation of
unification (Kay 1985), then a notation like (3) may add a desirable
flexibility to compositional semantic interpretation (as, e.g. in LFG) for
we do not have to combine a predicate with its arguments in any fixed
order (in contrast to categorial grammar for example). This strategy
may or may not turn out to be an important innovation, but it is not the
essence of the thematic role theory that 1 am talking about here.
Rather, the key feature of a thematic role system must surely be
understood to be the claim that the categories "Agent", "Patient" etc.
(or whatever particular roles turn out to be correct) distinguish one
argument from another semantically, not merely as syntactic labels.
In other words, thematic roles somehow "index" the arguments of
predicates semantically, parallel to but independently of the way
the grammatical relations subject, direct object, etc. index arguments
syntactically.
But the distinction between "syntactic" and "semantic" may become
subtle. As used in some linguistic theories, the term semantics does not
necessarily involve the relationship between a language and the objects
or situations denoted by the language (as it does in logic and as 1
as sume, throughout this paper, that it does) but to a component of a
linguistic theory whose relationship to non-linguistic objects is left

undefined. By the proponents of such a theory, notions like Agent,
Theme, Goal will be viewed as "semantic" simply because it was
stipulated in constructing the theory that they belong to the "semantic
component" of the theory rather than the "syntactic component", or
because the terms "Agent" etc. are used to labeI parts of a formal object
called a "semantic representation" instead of parts of one labeled a
"syntactic representation." From our point of view, i.e. logical seman-
tics, one of these two representations is a matter of syntax just as much
as the other is, so it is hard to see how such a view of thematic relations
is really any different at alI from using them as purely syntactic indices
on arguments, at least until more is said about "real" semantics.
To exclude this latter "pseudo-semantic" use of thematic roles, let me
clarify what 1 mean by semantic content of thematic roles as follows:
when 1 say that a thematic roles system distinguishes one argument
from another semantically, 1 mean that it permits (real-world, non-
linguistic) objects to be distinguished from one another by virtue of the
distinctive properties they have as they participate in an event named
by a verb, properties that can be identified ("in the real world")
independently of a language or its "semantic representations." (On this
point, the use of "theta roles" in GB is problematic and requires more
comment below.)
Of course, some linguists will respond that thematic roles have a
cognitive significance (some role in linguistic and perhaps non-linguistic
cognitive processing), and that this "cognitive content" is what keeps
them from being merely abstract indices in some representation (ef.
Jackendoff 1983). This may be true as well, but 1 will follow the
strategy advocated in such works as Barwise and Perry (1983) and
Stalnaker (1984) that one should try to justify and explain the cognitive
semantic significance of language in terms of its referential semantic
significance. 1 offer my own speculations on the cognitive significance of
thematic roles at the end of this paper, in section 6.
2.1. The Ordered-Argument TheoryofPredicates andArguments

An ordered-argument system treats a verb as an unsaturated predicate
wanting a particular fixed number of arguments to form a formula.
Semantically, a verb denotes an n-place relation, and when the denota-
tions of the proper number of arguments are combined with the
74 DA VID R . DOWTY
relation in an appropriate way, then a truth value is denoted, a

proposition is expressed, or a situation is described (depending on
which semantic theory you use); if too few argument denotations are
made available, then no truth value (or proposition or situation) can
result.
In this system, it is important that the syntactic and semantic rules be
constructed to combine the argument-denotations with the verb denota-
tion in a fixed arrangement determined by the syntactic arrangement
(Le. by the grammatical relations the arguments bear in the sentence in
question), though that semantic arrangement is, in a sense, ultimately
arbitrary. Arbitrary, because the choice of relations to be denoted by
verbs is likewise open in this theory. The inverse of a two-place relation
can carry information equivalent to that of the original relation (think
about, e.g. like versus please), and similarly each of the permutations of
an n-place relation can carry the SanlL- information. It doesn't matter
which permutation we chose for a verb's denotation, as long as the
compositional semantic rules match each syntactic argument of each
verb with its appropriate semantic "slot" in the interpretation of that
verb. (Since we want compositional rules to apply systematically to all
verbs, our choice of permutation for one verb will of course fix the
choice of permutation for all other verbs of equal or smalIer adicity.)
In talking about an ordered-argument theory, 1 do not distinguish
between syntactic analyses which combine a verb with alI of its
arguments "alI at once" (as in the syntax of most formallogics) or one
at a time (as in categorial grammars and many versions of Montague
Grammar), thus treating the denotations of verbs as function-valued
functions (so-calIed "curried" functions), both of these are ordered-
argument theories.
2.2. Thematie Roles, from an Ordered Argument Point of View

Whatever a thematic role system is, it is one that presumably differs in
some way or other from this classical, "ordered-argument" picture that
we are familiar with. 1 wilI first try to construct formalizations of what 1
will caII strong theories of thematie ro les, which combine as many of the
desiderata of the various linguistic proposals about thematic roles as
possible, from the possibility of stating thematie hierarehy eonstraints as
in the work of Jackendoff, Gruber and olhers, to the requirement of a
one-to-one correspondence between NP complements of a verb and
thematic role assignments to these arguments, as in GB and LFG. In

section 5 I will take up the question whether there are reasons to relax
"strong" theories of roles in one way or another.
The first question to ask is whether existing "ordered argument"
theories already have the means at their disposal to describe the
semantics of thematic roles, and if so, what form the characterizations
of roles will take there. In fact, it is possible to define thematic roles
quite adequately 'within' a standard ordered-argument theory, I believe,
as long as one has recourse to a description of lexical entailments.
I assume any complete semantic theory of natural language must
permit us to describe the lexical entailments of verbs. For example, if a
sentence x builds y is true, then it is necessarily also true that x
performs purposeful actions, that as a result of these actions a structure
or other artifact y comes into existence, and so ono Whether meaning
postulates, structural constraints (in situation semantics), lexical decom-
positions or some other means are the proper ones to describe lexical
entailments is not my concern here, as long as it is agreed that they be
describable.
It is surely the case that among these entailments of build, one will
find lexical entailments about the subject of that verb which amount to
the criterial properties for being an Agent, and also among them are
entailments about the object of build which amount to criterial
properties for being a Patient. Indeed, it is impossible to imagine how
to describe the meaning of x builds Y without saying something that
directly or indirectly identifies the subject as Agent and the object as
Patient. And of course, I believe this state of affairs obtains not just for
the verb build but for alI verbs, and for every lexical semantic
characteristic one would want to caII a thematic role.
At this point we are faced with a choice: should thematic roles be
determined by and derivative of the lexical meanings of verbs, or
should we say that thematic roles are somehow a separate part of the
semantic description of a natural language from the rest of lexical
meaning - in which case we are, according to the considerations of the
previous paragraph, in some sense saying that they are redundantly
specified in the semantic description of a language. At least Jackendoff
has always asserted that thematic role assignment should be derived
from the semantic representation of lexical meaning (Jackendoff 1972,
pp. 39-43; 1976). The question of the exact relationship between
thematic roles and "the rest of" lexical meaning has received Jittle
76 DAVID R. DOWTY
further discussion in the thematic roles literature, to my knowledge. 1

will here take the more conservative and apparent1y not very controver-
sial position that thematic roles are not redundantly specified and are
therefore determined completely and solely by verb meanings.
From the point of view of the ordered-argument theory, then, a
"thematic role" can and probably must be regarded as a cluster of
entailments and presuppositions shared by certain arguments of certain
verbs; this has already been proposed in the model-theoretic literature. 2
In this paper, however, 1 will not be distinguishing presupposition from
entailment but will speak as if only entailments are involved. For
example, the thematic role Agent might be regarded as a set of
properties including "is a rational and sentient being", "acts volitionally
in this circumstance", and so ono
We reach the formal description of roles in three steps. First, we will
detine an individual thematie rale as in (4):
(4) Given an n-place predicate <5 and a particular argument Xi'
the individual thematie role (<5, i) is the set of all properties
a such that the entailment .
D[<5(x[, ... ,xi , ,xn ) -+ a(xi )]
holds.
Thus an individual thematic role is specific to a particular verb and
to a particular argument-position of that verb. We might refer to the
individual thematic role for the tirst argument of the verb build as the
"builder-role", because it is the set of all the things you can conclude
about X solely from knowing that the sentence x builds y is true; we can
call the individual role of the second argument of build the "buildee-
role", since it consists of all you can conclude about y from this same
sentence.
In (4), 1 have not yet said what counts as a property, a denotation of
a. It we allow properties to be defined quite generally, as for example
in Montague's intensional logic (or, in theories in which some proper-
ties are primitives, if we allow the properties that can appear in (4) to
include properties that are defined by abstraction), then one property
that will be trivially entailed will be, for each argument of each verb <5,
that property represented by the lambda-aQstract in (5):
For exampIe, suppose that in (5) b is the verb build and Xi is the
direct abject position, corresponding ta the build-ee roIe. Then (5) is
the property of being an X such that for some y, y builds x. I wilI call
such a property a trivial individual thematie role entailment. This kind
of property may seem pointless ta talk about, but I will have reason ta
mention such properties later an.
Building an the definition in (4), we proceed ta the definition of a
thematie-rale type, in (6):
(6) Given a set T of pairs ( b, ilJ where b is an n-place
predicate and ia the index of one of its arguments (possibly a
different i for each verb), a thematie roZe type i is the
intersection of alI the individual thematic roles determined
by T.
In view of (6), for example, the particular thematic role-type that we
want ta call the Agent role-type wilI be the set of entailments that are
common ta alI the individual thematic roles of the arguments, of
various verbs, that we identify as "Agent"-arguments.
Of course, (6) defines a thematic-role type as the intersection of any
set of individual thematic roles whatsoever, and what we are interested
in is the possibility that there exists a particular set of thematic role
types that plays a special role in linguistic theory - the set that includes
Agent, Patient, Source, etc. Let us call this distinguished set the L-
Thematie Rale Types, ar, when no confusion can arise, simply the set of
Thematie Rales.
We must digress briefly ta point out that not alIlinguists who use the
terms thematie role ar O-role are committed ta the existence of a set of
L-thematic roles of this sort; for example, van Riemsdijk and Williams
(1986), in describing (their conception of) Government-Binding theory
say "O-theory, as outlined here, is not committed ta ... a system of
argument types ... [implied by] terms such as agent, patient (ar theme)
and gaal." (p. 241); similarly, ef. Marantz (1984), p. 31-32. For these
writers, the total number of O-roles is equal ta the sum of alI the
subcategorized arguments of alI the predicates in the language: "O-role"
in this sense corresponds exactly ta my "individual thematic role". Do
thematic roles in this sense have semantic content? In a literal sense
they do, of course, as they distinguish one argument from another an
semantic grounds. But the kind of "semantic indexing" performed
would be of a totalIy different kind, since in no sense (except the trivial
78 DA VID R. DOWTY
one) would roles constitute semantic categories of arguments, in

contrast to the traditional theory of thematic roles; the questions of
extra-linguistic and cognitive significance of thematic roles discussed in
section 6, Carlson (1984) or Jackendoff (1983) make no sense if
thematic roles are limited to individual roles. (Of course, within the GB
theory, O-roles have tasks to perform which have nothing directly to
do with semantics: the task of indexing the arguments of a predicate
for syntactic purposes, e.g. to insure that a verb does not acquire or
merge arguments between d-structure and s-structure, and so that
"non-thematic" 3 arguments can be distinguished. Although GB would
turn out to be a more empirically interesting theory if there is a class of
L-thematic roles in our sense that could serve this specialized deriva-
tional indexing function simultaneously, apparently little would be lost
that is important to GB theorists if this turns out not to be possible.
This may explain the willingness of GB theorists to abandon the
traditional idea of L-thematic role, leaving the terminological connec-
tion with that earlier view now perhaps rather gratuitous.)
How can we further characterize these L-role-types? Presumably,
this set should be non-empty but finite, and rather small; this much
is obvious, but not very helpful. From the way thematic roles are
employed in linguistic descriptions in the litera ture, 1 suggest that there
are two relationships that the L-thematic role types should have to the
relations denoted by naturallanguage verbs:
(7) (Completeness) every individual thematic role contains some

L-thematic role type (or as we may equivalently say, every
argument position of every verb is "assigned" an L-thematic
role type).
(8) (Distinctness) Every argument-position of every verb is
distinguished from every other argument-position of the
same verb by the L-thematic role types the two argument-
positions are assigned.
(These conditions are of course reminiscent of the (j-criterion of GB,

but the conditions serve a quite different purpose in our enterprise.) 1
believe that something like (7) and (8) may be implicit, if not explicitly
stated, in the use of thematic roles by a variety of linguists who make
"-
reference to roles. Specifically, this literature often proposes that there
exists one or more hierarchies of thematic role types (e.g. Jackendoff's
(1972) hierarehy Agent > Source, Gaal > Theme) and then states
eonditions to the effeet that the NP in the sentenee bearing the highest
thematie role on the hierarehy has sueh-and-sueh properties (ef. for
example Jaekendoff 1972, 1976; Nishigauehi 1984). This sort of
statement would seem to presuppose that al! arguments of a verb ean
be classified by thematie role, which requires the eondition (7), and it
definitely presupposes that they ean be uniquely classified by their
thematie roles, whieh presupposes (8).
There are aetually various versions of (8) that we eould imagine
adopting: (8a),
(8a) No two argument-positions of the same verb are assigned

the same L-thematic role type, and every argument position
is assigned only one L-thematie role type.
whieh is the strong eondition that the GB 8-eriterion adopts, or the
weaker eonditions (8b) and (8e).
(8b) No two argument-positions of the same verb are assigned

the same L-thematie role type.
(8e) No two argument-positions of the same verb are assigned
exaetly the same set of L-thematic role types.
I include the last two beeause some linguists have proposed analyses
in whieh a eertain argument bears two thematie-role-types simultane-
ously: (8b) would for example allow the subjeet of the verb buy to be
both Agent and Souree (as proposed by Jaekendoff 1972), as long as
neither of these two roles are assigned to other arguments of the same
verb. The still weaker (8e) would permit two arguments of a verb to
have one or more roles in eommon las long as they did not share exaetly
the same set of roles. (8e) would be needed if, as is sometimes
proposed, we allow role-types whieh are sub-types of other role-types,
e.g. if the roles Souree and Goal are to be eonsidered subtypes of the
role Direetional, as suggested by Andrews (1985, p. 70).
Before going any further, I should note that I have individuated both
individual roles and role-types above in terms of "argument positions",
rather that in terms of the entities that are denoted by eaeh argument.
This is deliberate, however. To see why, let us take a brief look at the
theory of thematie roles in Chierehia (1984)'. (1 will make a few
simplifieations in deseribing Chierehia's theory that I hope do not
80 DAVID R. DOWTY
obscure anything relevant to our discussion here.) Chierchia starts

with a primitive ontology that consists of individuals and a set of
"natural properties" and their complements (Chierchia 1984, p. 324;
Cocchiarella 1983, ch. 2) (i.e., these are n-place relations). He then
defines what we will call an event as in (9):
(9) (after Chierchia 1984) An event is an n+ 1-tuple (pn, Xl'
... , x n), where pn is an n-place property and Xl' .. , Xn are
individuals.
(This is of course similar to the way situations are built up in
situation semantics, and Chierchia intends events to be used in much
the same way here.) A O-role is then defined as a partial function from
events into the set of individuals:
(10) A O-roIe O is a partial function from the set of events into
the set of individuals such that for any event k, if O( k) is
defined, then O( k) E k.
The intent is that a O-role such as Agent is a function which, when
given an event as argument, selects the participant which plays the role
of agent in that event. For exampIe,
(11) Agent(( "kill', X, y) = X
Patient(( "kill', X, y) = y.
A consequence of this way of defining thematic roles is, as Chierchia
notes, that a requirement corresponding to our distinctness condition
(8) cannot reasonably be imposed. To see the reason, consider what
will happen in a sentence with a reflexive pronoun, such as John kills
himself:
(12) a. Johnkillshimself
b. Agent(( "kill',j, j) = j
c. Patient((" kill', j, j) = j.
Unless we want to rule out such events altogether, we cannot impose
distinctiveness with Chierchia's way of defining thematic roles.
Notice that with our present approach, which individuates thematip
roles types relative to argument-positions, there is no problem with
events like (12); we can still distinguish the entailments that accrue to
John in virtue of his appearance as the second member of the sequence
SEMANTIC CONTENT OF THEMATIC ROLES 81
in (12b, c) from those that result from his appearance as the third
member of that same sequence; thus we can maintain distinctness of
thematic roles. Notice also that by distinguishing thematic role types in
terms of argument positions, we do not commit ourselves to defining
thematic roles in terms of linguistic expressions rather than in terms of
real-wor1d, non-linguistic entities: position in a sequence, where the
sequence is an ordered n-tuple in the denotation of a predicate or is
part of a situation, is just as good for our purposes as argument position
in the linguistic expression itself.
With the definition of thematic role types that I have given, we can
when necessary still identify the individual denoted in an argument-
position that is assigned a certain thematic ro le type - relative of
course to a particular sentence and to a noun phrase in the argument
position in question; thus we really don't Iose what Chierchia's way of
defining of thematic role provides. To avoid confusion, I will refer to an
individual denoted by an argument whose position is assigned a certain
thematic role as the bearer of that thematic role.
However, I believe we still have not achieved the intended semantic
content of L-thematic role types. To see this, note that nothing we have
said so far exdudes the existence of a hypothetical L-thematic-role-type
X that is assigned by the subject of the verb kill, the object of the verb
build, the indirect object of the verb give and the object of the verb
inhabit. We could simply define the (only) entailment in the role X as
the disjunction of all the trivial individual thematic role entailments of
the relevant argument positions of these verbs, that is, a property
equivalent to (13):
(13) Ax[3ykill'(x, y) V 3ybuild'(y, x) V 3y3zgive'(y, z, x) V
... ]
The problem here, I believe, is that we have not exduded the
possibility of referring to meanings of particular verbs in defining L-
thematic role types; hence we might collect arbitrary argument posi-
tions together into one disjunctively defined "thematic role type" as in
(13). But I think it is dear that linguists insist that whatever criterion
individuates thematic roles must be independent of particular kinds of
relations (and of particular verbs denoting such relations). In other
words, whatever criterion identifies an Agent semantically, that criterion
should be a set of properties that are common to agents of ALL verbs,
that is, properties that can be recognized independently of knowing
82 DAVID R. DOWTY
exactly which verb we are identifying an Agent of. Let us describe this
requirement as in (14):
(14) (Independence) The properties in an L-thematic role type

must be characterizable independently of the relations (de-
noted by naturallanguage verbs) that entail them.
Note incidentally that although Chierchia's way of defining thematic
roles treats them as functions on events, events are for him sequences,
one element of which is the particular relation denoted by some verb,
so his definition as it stands does not necessarily rule out "bizarre"
thematic roles such as (13) that depend in an idosyncratic way on
particular verbs, unless it is stipulated in his definition too that thematic
role functions be defined across all situations in terms of some "natural
properties" or other for each role.
Perhaps (14) is really stronger than what we want. For example, it is
sometimes suggested that the role Source has a literal characterization
"place from which/to which something moves" with motion verbs as in
(15), but only an abstract or metaphorical terminus of movement with
verbs like (16):
(15) John rolled the ball to the fence.

(16) Mary explained the idea to John.
Thus perhaps such roles should have one kind of criterion for
motion verbs and a different, "metaphorical" one for other verbs.
However, 1 have no idea at present how to go about constructing a
criterion that permits thematic roles to depend on what we might call
natural classes of verb meanings, as illustrated by (15) and (16),
without permitting quite arbitrary dependence on verb meaning, as in
(13), so for now 1 wiUlet (14) stand in for the independence criterion
we eventually seek.
One way of insuring (14) would be to keep our present approach
and stipulate that the properties contained in L-thematic roles be taken
from the set of "natural properties" suggested by Chierchia (1984) and
CocchiareUa (1983) or from some other designated set. However, now
that we have identified the three criteria of completeness (7), distinct-
ness (8) and independence (14), 1 belie~e we are in a position to see
that there is a quite different way of constructing a thematic role theory
from which these criteria arise, if not automatically in all three cases, at
least quite naturally.
3 . A NEO-DAVIDSONIAN SYSTEM OF THEMATIC ROLES
1 wilI call this other approach to thematic roles a Neo-Davidsonian

System of thematic roles because of its similarity to Davidson's (1967)
familiar theory of adverbs in action sentences. Davidson suggested that
event-sentences quantify over events and that adverbs are essentially
predicates of events, as in, for example, the formula in (17b):
(17) a. Jones buttered the toast at midnight in the bathroom.
b. 3e[buttered(Jones, the-toast, e) & at-midnight( e) & in-
the-bathroom( e)].
What 1 will propose is that to construct a thematic role system, we
should stipulate that not only the modifiers but also the arguments of
verbs are actually predicates of events; more precisely, thematic roIes
are relations between individuals and events. Hence the verb of any
sentence has no arguments of its own, other than an event variable, and
is likewise a one-place predicate of events, rather than something that
denotes a relation. That is, example (17a) is instead to be interpreted as
(18):
(18) 3e[buttered( e) & Agent(Jones, e) & Patient(the-toast, e) &

at-midnight( e) & in-the-bathroom( e)].
If thematic roles relating events to their participants are the only way
we can express what we were formerly thinking of as "relations" in
natural language, the three conditions on thematic-role-types we devel-
oped above will now folIow:
Complete ness - the requirement that every argument entails some
thematic-role - folIows because there is no way in this system to state
that an individual is a participant in an event except by relating it to the
event via some thematic role.
Distinctness - the requirement that all arguments of a verb are
distinguished from one another by the thematic roles they bear, is quite
natural in this approach, for if two participants in an event are to be
given any semantically distinct status at all in this method, it will have to
be by means of different thematic roles which relate them to the event
in question. On the other hand, one might wonder about the possibility
of predicating the same role of two different individuals, as in (19); 1
williater have reason to ruIe such cases out entirely (and, incidentally,
be thereby committed to (8a) rather than (8b) or (8c.
84 DA VID R . DOWTY
(19) 3e[verb(e) & 3y3z [THR] (y, e) & THR](z, e) & y '" zll.
Independence - the requirement that thematic roles be semantically

characterizable independent1y of the meaning of the verb describing the
event - also follows naturally here from the fact that statements have
the form of conjunctions of predications about an event: the truth
conditions of each conjunct will have to be determinable independent1y
of the other conjuncts. True, we still will need to stipulate that the
relations denoted by thematic roles be something like natural proper-
ties, to be absolutely sure of ruling out exotic definitions which amount
to (13), but such a requirement seems more motivated under the Neo-
Davidsonian approach, since thematic roles, not verbs, denote the
compositionally basic relational terms we are working with. (In an
ordered argument system, by contrast, thematic roles are secondary
(and second-order) notions, derived from verb denotations by inter-
secting entailments.)
A brief historical note is in order here. First, some people, for
example Chomsky (1981, p. 35), seem to have thought that the theory
implied by (18), rather than that implied by (17a), is what Davidson
did propose. However, I find no justification for this attribution in
Davidson's writings themselves, in fact Davidson argued against (18) in
Davidson (1967a). The system implied by (18) was, on the other hand,
explicitly proposed as a modification of Davidson's by Terence Parsons
(1980), and the theory of thematic roles in Carlson (1984) is, as far as I
can tell, equivalent to Parsons' proposal and mine. While I certainly
owe a debt to both these papers, 1 might point out that neither Parsons
nor Carlson observe that the three requirements of Completeness,
Distinctness and Independence motivate the Neo-Davidsonian analysis;
also, I will propose that the Neo-Davidsonian analysis may apply to a
different domain in natural languages from what Parsons and Carlson
have assumed.
To fiU out our Neo-Davidsonian system, we will apparent1y need two
more things. First, we may want to as sume that certain kinds of events
are entailed to have a certain collection of participants (i.e. entailed via
the lexical meanings of the verbs denoting them in Montague Seman-
tics, or via Structural Constraints in Situation Semantics or other
theories taking properties as primitives). For example, if events of
giving always have three participants, <l' Source, a Theme and a Goal,
then we would need to be able to state that (20) is an entailment for
events named by this verb, and similarly for other verbs.
(20) Ve o [giving( e) ---> 3xAgent( x, e) & 3ySource(y, e) &

3zGoal(z, e)].
AIso, there will seemingly be entailments about event participants
that are specific to individual verbs, entailments above and beyond
those that follow from the thematic rale predicate itself. For example, if
a person is the agent of an event of singing, then that person's vocal
cords are (mostly) in motion, he is exhaling air, etc. If he is Agent of an
event of whistling, then he is exhaling but his vocal cords are noI
moving. To describe these facts, we apparently need postulates like
(21):
(21) VeVxo [[singing( e) & Agent(x, e)] ---> moving-vocal-cords(x)].
Such entailments as these will correspond, in an ordered-argument
system, to entailments of individual thematic rales that are not part of
any L-thematic-role type. 1 will return to the status of entailments like
(20) and (21) below.
FinalIy, we may want to impose a uniqueness requirement on the
individuals asserted to bear thematic rales in events, to the effect that
no other individual other than the one mentioned bears that same rale
in that same event:
(22) (Uniqueness ofrole-bearer) For alI thematic rales TH;:
VeVxo[TH;(x, e) --+ Vy[TH;(y, e) --+ X = Yll.
Both Carlson (1984) and Chierchia (1984) as sume such a condition
will hold; in Chierchia's formulation, it folIows because thematic roIes
are defined as functions into the set of individuals, rather than relations.
However, note that uniqueness is a different condition from distinct-
ness; distinctness only requires that the set of entailments relative to a
given argument position are distinct from those of other argument
positions. Uniqueness requires that the individual denoted by the
argument in this position is the only one that bears this thematic raIe.
FinalIy, note that uniqueness does not commit us to Chierchia's (1984)
position that thematic rales are functions into the set of individuals:
uniqueness as in (22) means each thematic rale has only one bearer,
but does not rule out the possibility that the same individual may be
bearer of more than one thematic role.
It might seem that this condition would caus'e prablems with plural
sentences, such as John and Mary sang. However, 1 think it need not
do so, if we adopt the theory of plurals praposed in Link (1983),
86 DA VID R. DOWTY
Hoeksema (1983) and elsewhere, which holds that plural NPs consis-
tent1y denote groups (i.e. non-singleton sets), and that distinctions
between collective and distributive interpretations, as in (23), are not
differences in the types of NP denotations but only differences in lexical
entailments of verbs:
(23) a. John and Mary sang, fell asleep (distributive)
b. John and Mary painted the house, bought the car
( distributive or collective)
c. John and Mary met, are alike (collective).
That is, all three examples predicate something about the group
consisting of John and Mary. However, with a distributive predicate,
(23a), it is entailed that whenever the predicate is true of a group, it is
also true of aH the members of the group individually. With a collective
predicate, (23c), it is entailed that the predicate is never true of the
individuals making up the group. Predicates like those in (23b) could in
principle be treated as vague, not ambiguous, between group and
individual readings in this method, but an ambiguity analysis is prob-
ably preferable for this dass: ef. Roberts (1986) for some discussion.
If we assume that the variable x in (22) ranges over a domain of
discourse that contains groups as well as individuals, then it seems to
me to give sensible results with this theory of plurality, if we pay careful
attention to the way events must be individuated when we invoke the
distributivity axioms. It is required that in the case of (23a), three
events are entailed to exist: the event of John and Mary's singing (which
will have a group as Agent), plus the distinct event of John's singing and
the event of Mary's singing: each of the last two has only an individual
Agent. In example (23c), only one event of meeting is involved, and this
one has as Agent the group with John and Mary as members, though
neither John nor Mary can separately be Agents of this event (though
we may be able to condude that they are Agents of other events, e.g.
events of moving to some place, etc.) Though these ways of individu-
ating events may be novel, 1 am not aware of any problems that this
involves us in.
3.1. Expressive Power of the Two RoZe Theories

"
We can now raise the question whether the two theories of predicate-
argument association are equivalent in expressive capacity, or whether

one is stronger than the other. 1 will approach this question by noting
that if both theories were adequate to describe all natural language
verb meanings correct1y, then in principle it should be possible to
state equivalences between relational (i.e. ordered-argument) and Neo-
Davidsonian descriptions of event-types as in (24), for every natural
language verb. (Here give denotes a 3-place relation on individuals
and giving denotes the corresponding Davidsonian one-place event-
predicate.)
(24) \fx\fy\fz D [give(x, y, z) ... 3e[giving( e) & occur( e) &

Source(x, e) & Theme(y, e) & Goal(z, e)].
Whether or not such equivalences will actually hold depends on
three things. First, is it truly the case that all verb meanings of alI
natural languages are such that a finite set of L-thematic role types
meeting the conditions of completeness, distinctness and independence
characterizes their arguments? Note that such a claim about role types
substantively limits the verb meanings that natural languages could
have, so in principle this claim could turn out to be empirically false.
An ordered-argument system involves no such limitation, so in this
sense it is a stronger theory. If verb meanings are such as to be
completely characterizable by thematic role types, then a second
question is whether we permit ourselves to include postulates like (20)
and (21) above in our thematic role system (which entail that events of
a certain type always involve certain role-bearers as participants, or
which entail that role-bearers in an event-type are entailed to have
properties other than just those of the role itself). If we prohibit one or
both types of postulates, then obviously the resulting neo-Davidsonian
system will be weaker in these two additional respects than an
unrestricted ordered argument system. Thus, in order of ascending
expressive capacity, there should in principle be three classes of
systems:
(i) "bare" neo-Davidsonian systems (no postulates of form of

(20) or (21 ,
(ii) neo-Davidsonian systems where the complete argument-set
of a predicate is inferrable (postulates of form of (20)
assumed),
88 DA VID R. DOWTY
(iii) neo-Davidsonian, argument-set inferrable systems with arbi-

trary additional entailments for each argument (postulates of
form of (21)): should be equivalent to ordered-argument
systems, on the assumption that arguments of alI natural
predicates can be semanticalIy indexed by some set of L-
thematic role types.
4. A HYPOTHESIS ABOUT THE DOMAINS OF THE

TWO KINDS OF PREDICA TE-ARGUMENT ASSOCIA TION
IN NATURAL LANGUAGE
At this point, we should observe that there is an important difference

between the syntactic properties of the Ordered-Argument theory and
those of the Neo-Davidsonian theory. Since the Ordered-Argument
theory assumes verbs denote relations of a fixed number of places, it
follows that (under classical interpretations) a predicate with fewer than
its prescribed number of arguments does not have any welI-defined
interpretation. Instead, it is usually assumed (ef. Dowty, WalI and Peters
1981 ; Dowty 1978) that if we want to form a sentence from an n-place
predicate without its full prescribed number of arguments, it is first
necessary to apply an operation to the predicate that syntactically and
semantically reduces it to an n - 1 place predicate. Consider for
example the operator O, as defined in (25):
(25) a. syntax: If o is an-place predicate, then 0(0) is an n - 1
place predicate.
b. semantics: If o denotes an n-place relation d, then 0(0)
denotes that n - 1 place relation R such that for any
n - 1 place sequence of individuals ( Xi' . . . , X n -1 > E R
iff there is some individual y such that (XI' .. , Xn -1' y>
Ed.
That is, O is an operator that "existentialIy quantifies" one argument
position of a predicate.
The Neo-Davidsonian theory of thematic roles is quite different,
because each argument is added to a verb by conjoining, in effec,
another formula. If for example the verb give denotes the kind of event
of which it makes sense to predicate a Source, Theme and Goal, then
there is no obvious reason why ali the formulas in (26) should not be
semanticalIy coherent:
(26) a. 3e[giving(e) & Source(j, e) & Theme(the-book, e) &

Goal(m, e)]
b. 3e [giving( e) & Source(j, e) & Theme(the-book, e)]
c. 3e[giving(e) & Source(j, e) & Goal(m, e)]
d. 3e[giving(e) & Theme(the-book, e) & Goal(m, e)]
e. 3e[giving(e) & SourceU, e)]
f. 3e[giving(e) & Theme(the-book, e)]
g. 3e[giving(e) & GoalU, m)]
h. 3e [giving( e)].
(If we stipulate independently that giving denotes a kind of event

that necessarily has a Source, Theme and Goal, then each formula (b)-
(h) will entail the existence of alI three participants, even though they
may not be mentioned in the formula.)
Whether or not one should expect differences in syntactic form in
natural languages to correlate with differences in the structure of
semantic interpretation such as this depends on one's linguistic meth-
odology. Although the methodological preference for strictly composi-
tional semantic analyses a la Montague seems to have fallen somewhat
out of vogue in some branches of semantics in recent years, 1 still
believe that the general expectation of parallelism in syntactic form and
semantic function has, historically, led us to insightful analyses of
naturallanguages far more often than it has led us astray.
From this point of view, it is interesting to observe that a syntactic
difference in argument-structures of just this sort can be found in
naturallanguages, namely, between the arguments of verbs and those of
event-denoting nouns. Verbs characteristically exhibit subcategorization
- i.e. each verb has one or more fixed sets of complements - and this
subcategorization is independent of the matter of what event partici-
pants are entailed to exist by the meaning of the verb. The verbs dine,
eat, and devaur, for example, alI entail the consumption of some
quantity of food, but the first is obligatorily intransitive, the third
obligatorily transitive and the second belongs to both subcategories:
(27) a. John dined.

b. * John dined his lunch.
c. Johnate.
90 DA VID R. DOWTY
(27) d. John ate his lunch.

e. * John devoured.
f. John devoured his lunch.
It has been argued in Dowty (1978,1979) and Bresnan (1982a) that
the variable polyadicity of verbs like eat should be analyzed by means
of lexical operations that reduce (or in different cases expand) the
number of arguments a verb takes and are interpreted like the operator
O above (and ef. below). Of course, no morphological sign of such an
operation is present on detransitived eat, but it is clear quite indepen-
dently of intransitives that English is a language that often performs
lexical operations without a visible morphological marker. Other
languages do use an affix to mark detransitivization (e.g. Hungarian,
Swedish, Nivkh, Georgian, and for some verbs, Russian; ef. Comrie
(1985), pp. 319-322), and in morphologicalIy richer languages, virtu-
alIy every polyadicity change can be signaled by a visible verbal affix.
By contrast, nouns denoting events evidence no real subcategoriza-
tion at alI on a par with verbs; as a typcial example, alI forms of the
nominal gift in (28) are grammatical:
(28) a. The gift of a book from John to Mary would surprise Helen
b. The gift of a book from John would surprise Helen
c. The gift of a book to Mary would surprise Helen
d. The gift from John to Mary would surprise Helen
e. The gift from John would surprise Helen
f. The gift of a book would surprise Helen
g. The gift to Mary would surprise Helen
h. The gift would surprise Helen
Aside from one small and restricted class of counterexamples (-ing

nominals of transitive or ditransitive motion verbs, such as handing,
sending, etc. (Rappoport 1983, the possibility of omitting one, several
or alI "arguments" seems to obtain for alI event nouns, derived and
non-derived. Adjuncts of verbs (as opposed to their subcategoriz;ed
arguments, or complements) are distributionalIy (and as we shall see,
semanticalIy) like noun arguments in this respect.
At this point, I would like to put forward the hypothesis that both
ways of associating predicate with arguments are employed in natural

languages, but in different domains:
(29) a. The Ordered-Argument method of argument association
is employed to associate verbs with their subcategorized
arguments.
b. The Neo-Davidsonian method of argument association is
employed to associate (i) verbs with their adjuncts, and
(ii) nouns denoting events with their arguments.
I should admit at the onset that the hypothesis is not unequivocally
supported by available evidence. In fact, I suspect that to the extent that
(29) is correct, it is an over-simplification of the tme situation. But I
think the pattern of data that accords with (29) is striking and, I believe,
illuminates some further semantic consequences of the Neo-Davidsonian
theory. Though (29) has independent interest as a claim about nominals,
even if this claim turns out to be wrong it will have served the broader
expository purpose intended for it here.
4.1. Parsons' Analysis of Modifiers andArguments in a

Neo-Davidsonian Theory
The way I have described the Neo-Davidsonian theory, using conjunc-
tions, should not mislead us into thinking that this theory literally
requires a conjunctive syntactic analysis of sentences or noun phrases.
Parsons (1980) showed how a Davidsonian analysis of (some) adverbs
as predicates of events is compatible with a syntactic analysis of
adverbs as VP-modifiers, i.e. in a categorial grammar, of category
VPNP. Parsons' analysis (which as already mentioned also treated
arguments in what I here call a neo-Davidsonian way) mns as follows.
First, let VPs denote sets of events. U sing a lambda-calculus to
interpret English, Parsons translates a verb run as in (30)
(30) run transiates into: e [mn( e)].
A VP-adverb like quickly, of category VPNP, would be translated
as in (31) in Parsons' account,
(31) quickly translates into: Pe [PC e) & quick( e)].
Thus the VP run quickly translates as in (32), equivalent by lambda-
conversion to (33):
92 DAVID R. DOWTY
(32) APAe[P( e) & quick( e)] (Ae[run( e)]

(33) Ae[run(e) & quick(e)].
The rule for adding the subject NP adds the Agent role, and rules for
tense and other rules for completing the clause add the semantic effect
of Davidson's existential quantifier, so John runs quickly has in
Parsons' analysis a translation that, with lambda conversions, reduces to
(34):
(34) 3e [run( e) & Agent(j, e) & quick( e)].
According to hypothesis (29) of course, the subject of a verb would
not be interpreted this way, but the point to be exploited here is that
the "arguments" of event-nominals such as gift can be treated syntacti-
cally as noun adjuncts (of category CN/CN in a categorial grammar)
and interpreted semantically via neo-Davidsonian roles as in (35):
(35) by John (in cat CN/CN) translates into:
APle[P(e) & Agent(j, e)].
Rappaport (1983) takes a similar but not identical position about the
"arguments" of nominals: she argues that they are not syntactically
subcategorized, but she maintains that the set of arguments an event-
noun will take is nevertheless encoded into the lexical entry of the noun
(what is called the "predicate-argument structure" in LFG, a notion that
has no exact correspondent in the kind of syntactic theories I have in
mind), whereas I am suggesting that nothing in the lexical entry for a
derived nominal need mention its arguments. 4
The claim that verbal adjuncts are predicates of events of the Neo-
Davidsonian sort makes very traditional and uncontroversial predic-
tions about how adjuncts differ from subcategorized arguments; I am
thinking here of adjuncts as corresponding not just to adverbs but also
to the oblique grammatical relations of Relational Grammar and
Lexical Functional Grammar, that is, instrumental, benefactive and
locative NPs in their adjunct role, as contrasted with subcategorized
arguments like subject, direct and indirect objects. 5 Such obliques are
not syntactically obligatory and can be used in any sentence where their
meaning does not conflict with what we know about the real world, i.e.
the reason we do not say things like John is talt with a ladder or John is
talt for Mary is to be attributed to such a conflict, not to any syntactic
or compositional semantic incompatibility. It is also predicted that
obliques, unlike direct or indirect objects, have constant meaning, e.g. a

benefactive always implies that its referent "benefits", and so. In fact,
Richard Larson (to appear) has proposed a theoretical distinction
between adjuncts and arguments in the framework of situation seman-
tics which anticipates many of the differences implied (for wrbal
adjuncts and arguments) by (29).
4.2. Carlson 's Arguments for a Thematic-Role Account of Verb

Subcategorization
Contrary to what the hypothesis in (29) says about the arguments of
verbs, Carlson (1980) has argued that a theory of thematic roles like
the neo-Davidsonian one I have discussed gives the better account of
verbal arguments, so we will want to examine his evidence. Carlson's
first argument involves the relationship between examples like (36a)
and (36b):
(36) a. John ate a sandwich.
b. John ate.
It I understand Carlson's point correctly, it is that the Ordered-
Argument theory requires us to state both a syntactic and a semantic
relationship between (a) and (b): to describe the meaning of the
intransitive in terms of that of the transitive, we must say that when the
direct object is omitted, then the two-place relation denoted by eal is to
be understood as if the direct-object argument were existentiaUy
quantified. The thematic roles theory, Carlson claims, needs only the
syntactic statement that eat can be used without an object, but needs
no semantic statement at alI; thus (36b) asserts simply (37),
(37) 3e[eat(e) & Agent(John, e)]
but because of our knowledge about the nature of events of eating, i.e.
our knowledge that (38) holds for aU events of eating,
(38) \fe[eating(e) -- 3xAgent(x, e) & 3yPatient(y, e)]
we know that there nevertheless exists something that John ate on this
occasion. Carlson further contrasts this case with the verb kick,
(39) a. The muIe kicked something.
b. The muIe kicked.
94 DAVID R. DOWTY
where the intransitive form (39b) does not entail that there is something
the muIe kicked. On the thematic roles account, Carlson says, we
merely need to observe that with events of kicking, unlike events of
eating, there may sometimes be a patient, but sometimes not (i.e., a
postulate like (38) does not hold for kick). On the ordered argument
theory, however, we would apparently have to say that two different
relations are denoted here - a two-place relation, where a movement
of the foot brings it in contact with an object, and a one-place relation,
where no contact is entailed. Thus the ordered-argument theory
requires a different treatment of the detransitivization here from that in
(38).
If alI cases of verb detransitivization were like eat or kick, then these
would be indeed be arguments that a thematic role theory is simpler.
But the data is more complicated. First, it should be noted that transi-
tive/intransitive verbs that are semanticalIy like kick are extremely rare;
1 know of no other clear example that works like kick. 6 More impor-
tantly, there are other cases of detransitivization that don't fit this
pattern, and these must somehow be distinguished semanticalIy even in
the Thematic Roles theory. First, consider verbs 1ike dress, bathe and
shave:
(40) a. shaved 1
John I bathed someone.
I
dressed
b. shaved.
John bathed.
dressed.
The sentences in (40b) do not entail the corresponding ones in
(40a); rather, they entail John shaved himself, and so ono Even on the
thematic roles theory, these intransitives cannot be derived from the
transitives without a separate semantic rule (or they may instead be
listed independently in the lexicon with a semantics for this special
relationship to the transitives).
A further class includes verbs like notice and understand (ef.
Bresnan 1978, Grimshaw 1979, Fillmore 1986), which can occ,ur
intransitively, though their "missing" object is always understood
indexically, as referring to some entity mentioned in the previous
discourse or apparent in the context:
(41) a. No one noticed.

b. John understood.
(42) a. John entered, but no one noticed.

(= ... but no one noticed him enter)
(-:F but no one noticed anything/something)
b. Mary nodded, and John understood.
(= ... and John understood what her nod meant.)
(-:F .. and John understood something or other.)
The variety in meanings that a transitive verb can acquire in its

detransitivized use has on1y been touched on here; for a more complete
survey, see Fillmore (1986), which also describes many cases of
polysemous verbs where one sense alIows on1y one kind of detransi-
tivization but another sense alIows a different kind.
A similar problem for Carlson's proposal arises with the omis sion of
one of the objects of a ditransitive verb like serve (ef. Dowty 1978):
(43) a. The waiter served the customer the soup.

b. The waiter served the customer.
c. The waiter served the soup.
Serve alIows its remaining object to be interpreted as either the

Theme or Goal, and if alI ditransitives worked this way, Carlson's
simplicity argument would apply here too. Unfortunately, there are
other types; on the one hand there are verbs like seU, which in their
transitive form require that the remaining object be understood as
Theme, rather than Goal, and on the other hand there is the verb feed,
which requires that its one object be understood as Goal rather than
Theme:
(44) a. Mary sold the man the house.
b. [*] Mary sold the man.
c. Mary sold the house.
(45) a. John fed the children hamburgers.
b. John fed the children.
c. [*] John fed hamburgers.
96 DAVID R. DOWTY
(The bracketted asterisks indicate that the pragmatically "natural"

reading is unavailable, so that only a bizarre reading is possible.) As it
is not possible here that any principle involving a hierarchy of thematic
rale types' could predict how the three kinds of ditransitives are
interpreted, they must have this information identified in their lexical
entries somehow. The general response, then, to Carlson's arguments
about detransitivization is that there exists quite a variety of different
semantic effects of detransitivization that must be listed in lexical
entries of individual verbs in any case, so being able to state the lexical
entries of just one class of these a bit more simply is hardly a notable
advantage.
Finally, Carlson suggests that a Thematic Role theory gives a better
account of the pair (46) and (47):
(46) Martha was left alone.
(47) Someone (or something) left Martha alone.
On at least one interpretation of (46), it is not convincingly para-
phrased by (47) but means rather that Martha was completely un-
bothered, i.e. everyone left her alone. Carlson writes "on this reading
(where Martha is not bothered), if 1 leave Martha alone it does not
follow that she was left alone." 1 fail to follow the logic of this. If
Carlson is implying that the passive be lefi alone has only one inter-
pretation, i.e. "be left alone by everyone", then it should follow that the
phrase by John is redundant in (48), and that (49) is contradictory:
(48) Martha was left alone by J ohn
(49) Martha was left alone by John, but Martha was not left alone
byMary.
However, (48) is not redundant and (49) clearly has a non-contra-
dictory reading. But if Carlson admits that (46) is ambiguous, then 1 fail
to see how the thematic rale theory has any advantage at alI over the
ordered argument theory here; on both theories, the passive be lefi
alone has two interpretations, one of which is predictable fram the
meaning of the active, one of which is not. The ordered-argument
theory and the well-known semantic account of passive that goes with. it
(Bach 1980, Dowty 1978) do not rule Qut the possibility that a passive
form could have an idiomatic interpretation, in addition to its rule-
predicted interpretation, and this theory seems quite indistinguishable
from the other theory in this respect.
4.3. By-Phrases in Verbal Passives and in Nominals

Next, I want to turn to the interpretation of by-phrases in passive VPs
and in nominals. One fact that was noticed early in the days of Case
Grammar (and reported later in Fillmore 1977), but has been periodi-
cally forgotten and rediscovered since then (e.g. Marantz 1981), is that
the object of by in verbal passives is not always an Agent but can in
fact bear any thematic rale that the subject of the corresponding active
bears. For example, this NP is Agent in (50a), Instrument in (b), and
Experiencer in (c); the verbs in (d) and (e) are stative, so whatever the
thematic rales of the by-objects here, they are not Agents, on anyone's
definition of Agent.
(50) a. The house was constructed by Mary.
b. The house was destrayed by fire.
c. The rumbling was felt by the crew.
d. That fact is now known by almost everyone.
e. He is stillioved by his parents.
Obviously, the correct description of by-phrases in verbal passives is
that they can denote whatever the subject of the corresponding active
verb denotes, without regard to thematic rale; consequently, the
compositional semantics of the by-phrase must be stated in terms of the
grammatical relation subject. To invent a new thematic rale THR x for
alI and only the NPs that are object of by in a verbal passive is to make
up the kind of semantically vacuous thematic rale that I was trying to
rule out with the independence condition in (16), for clearly the only
way to characterize the hypothetical THR x semantically would be by
reference to the meaning of the particular passive verb which it
accompanies in each case.
The situation is quite different with by-phrases in derived nominals;
nominals of stative verbs (cf. (51a)-(51e)) cannot take by-phrases, nor
can nominals of non-agentive non-statives (ef. (51 e)-( 51g)):
(51) a. * The love (of him) by his parents.
b. * The knowledge (of that fact) by almost everyone.
c. * The belief in God by the faithful
d. * The hope for peace by alI peoples
e. * The death by Ghandi (i.e. where Ghandi dies)
98 DA VID R . DOWTY
(51) f. * The disappearance by the magician

g. * The awakening by Venus (Le. where Venus awakens).
It is instructive to observe that by-phrases can also occur as adjuncts
of nominals that refer to the result of an event (an eJfected object),
rather than the event itself, as in (52),
(52) a painting by Rembrant
a book by Chomsky
a sonata by Mozart
and can occasionally be used with a noun not directly tied semantically
to an event at alI, as (53):
(53) flowers by Pierre
Mary's children by John.
But to understand by in these cases, we are required to appeal to
some contextually salient event - the creation of the objects in (52),
the delivery or arrangement of the flowers and the conception of the
children in (53) - and understand the object of by as the Agent of that
event.
Rappaport (1983) points out that for some reason the nominals of
the "psych-movement" class of verbs, e.g. (54), are always understood
as stative,
(54) a. Amy's fright
b. The class's boredom
c. Deborah's amusement
d. Sam's annoyance
even though, as is well-known, the corresponding verbs are ambiguous
between a stative and an active interpretation; for example, (55) can
report either a state (Louise felt amusement at Herbie) or event (Herbie
performed some act that caused her to experience amusement):
(55) Herbie amused Louise.
Hence the examples of (56) are anomalous because of the adverbials
which require a non-stative reading:
(56) a. * the boredom of the class that happened ten minutes
after the lecture started.
(56) b. * the unintentional fright of the children that occurred

when they saw the scarecrow.
But by-phrases cannot occur with these nominals either:
(57) a. * Amy's fright by the scarecrow
b. * The class's boredom by the lecture(er)
* The boredom of the class by the lecture(er)
c. * Deborah's amusement by Randy
* The amusement of Deborah by Randy.
Thus we can give a uniform explanation of alI this data involving by
with event nouns - and possibly even the cases (52) and (53) as well -
by hypothesizing that the object in an adjunct by-phrase refers to the
agent or cause of an event, i.e. a possible thematic role; however, the
object of by with verbal passives must be distinguished from the object
of by with nominals, just as the hypothesis (29) requires. 7
4.4. "Linking" in Event-Noun Arguments and Thematic Uniqueness
Having now seen some reasons for distinguishing the argument associa-
tion of event nouns from that of verbs, 1 would like to turn to a
phenomenon that occurs only with event-noun arguments that can be
accounted for with the "uniqueness" postulate (22). Consider (58):
(58) a. The destruction of the city by the Romans
b. The city's destruction by the Romans
c. The Romans' destruction of the city
d. * The Romans' destruction of the city by the barbarians.
As we can see from (b) and (c), the prenominal possessive can
designate either the Agent or the Patient role; since post-nominal of
and by are unambiguous in the role they designate (if we know we are
dealing with the nominal of a transitive verb), we understand that the
possive denotes whichever is not specified elsewhere. Now why is (58d)
unacceptable? By comparing (58d) with (a), we can see that it is very
unlikely that this deviance is syntactic in nattre, e.g. a violation of
syntactic subcategorization, at least within mono-stratal syntactic theo-
ries like GPSG or LFG. But why should it be semantically deviant?
100 DAVID R . DOWTY
Note that both (59a) and (59b) are coherent (and mean different
things); why doesn't (58d) just mean the same as one of these?
(59) a. The destruction of the city by the Romans and the
barbarians
b. The destruction of the city by the Romans and the
destruction of the city by the barbarians.
The uniqueness 8 condition (22), repeated here, is pertinent to just
this sort of case:
(22) (Uniqueness ofrole-bearer) For alI thematic roles TH;:
\fe\fxD[TH;(x, e) ..... \fy[TH;(y, e) ..... x = Yll.
In view of (22), (58d) wiII be deviant, since it must have either two
Agents or two Patients. (22) is consistent with (59a), and with any
phrase designating a conjoined NP, if NPs conjoined with and refer to
groups, as in the analyses of Link (1983) and Hoeksema (1983); thus
(59a) must refer to an event with a group Agent. On the other hand,
(22) implies that (59b) refers to two different events, i.e. two destruc-
tions of the same city, presumably separated enough in time for the city
to be rebuilt in the interim.
Even more interesting is a phenomenon which Roger Higgins in his
dissertation referred to as "linking" (Higgins 1979) (not to be confused
with the different phenomenon caIIed "linking" by Ostler 1979 and
others in the GB literature):
(60) a. We witnessed an attack on the Sabines by the Romans
b. We made an attack on the Sabines
c. * We made an attack on the Sabines by the Romans
(Higgins 1979)
(61) a. John's defeat
b. I suffered a defeat
c. * I suffered John's defeat.
d. I witnessed John's defeat. (Higgins 1979)
Here again, the unacceptability obviovsly does not have to do with
violations of syntactic subcategorization. Rather, the phenomenon has
to do with verbs like make, perform, and undergo when they take NPs
referring to events as their objects. When x makes an attack, then x

must be the Agent of the event referred to as an attack, and when x
suffers a defeat, x must be the Patient of the defeat. That is, equiva-
lences of the form of (62a, b) must hold as a consequence of the
meanings of the verbs make (as in (60)) and suffer, where e ranges over
events and x ranges over individuals:
(62) a. VeVxo [make(x, e) .... [occur( e) & Agent(x, e)ll
b. VeVxo[suffer(x, e)"" [occur(e)&Patient(x, e)ll.
The verb witness, on the other hand, does not "link" its subject to
any thematic role in its event-object. Together with (22), (62) will
insure the anomaly of (60c) and (61c), and will distinguish these two
anomalous examples from acceptable examples like We and the
Romans made an attack on the Sabines. They also insure that the two
mentions of an attack in (63a) must refer to distinct events, while those
in (63b) could turn out to refer to the same event (though of course by
the Maxim of Quantity, the speaker should instead use definite
anaphora here if he already knows the same attack is involved):
(63) a. We made an attack on the Sabines, and the Romans
made an attack on the Sabines too.
b. We witnessed an attack on the Sabines, and the Romans
witnessed an attack on the Sabines too.
Examples can be found of the same phenomenon with the two
"conflicting" NPs more widely separated:
(64) a. An attack on the Sabines by the Romans seems very
likely to have been made at dawn.
b. * An attack on the Sabines by the Romans seems very
likely to have been made at dawn by the barbarians.
An interesting question now is why (64a) sounds better than (65),
since (64a) presumably entails (65):
(65) ? An attack on the Sabines by the Romans seems very likely
to have been made at dawn by someone (or some persons).
Note that (65) is not inconsistent with (22) at alI, nor of course is
(64a). However, (65) and the uniqueness axiom require us to condude
that the "someone or some persons" in question are identical with the
102 DA VID R. DOWTY
Romans; thus perhaps (65) violates one or more Gricean maxims, since
the phrase "by someone" normally implicates that the speaker does not
know any more precise description of the person in question than
this. (The difference between (64a) and (65) could conceivably be a
problem for theories which treat semantics as entirely a matter of the
logical form of a sentence, at least if it tumed out that (64a) had to have
the same logical form as (65) in those theories).
Ivan Sag (1985, ms.) and Janet Fodor (personal communication)
have called attention to discourses like (66):
(66) a. John made a promise. It was to Mary.
b. John made a promise. It was to shave himself.
c. John made a promise. It was to shave *herself.
d. John made a promise, which was to shave himself.
This shows that control of reflexives in infinitival complements can,
at least in some instances, be "remote control." (This kind of control,
incidentally, can be adequately described by combining the so-called
lexical theory of control in Chierchia (1984), Dowty (1985) and
Jacobson and Chierchia (1985) with the thematic roIe analysis of event-
complement verbs like make above, assuming that the events denoted
by event-nouns can be referred to by anaphoric pronouns; see also Sag
(1985, ms.) for a different analysis.) However, note that this is one way
in which the anaphoric possibilities for event nouns, like promise in Sag
and Fodor's data, differ from the corresponding verbs:
(67) a. John made a sale yesterday.1t was to Mary.
b. John sold a house yesterday. *It was to Mary.
(68) a. John made an agreement with Mary yesterday. It was to
perjure themselves.
b. John agreed with Mary yesterday. *It was to perjure
themselves.
b .John and Mary agreed yesterday. *It was to perjure
themselves.
(69) a. John gave Mary a waming then. It was to protect herself
from puffins.
b. John wamed Mary then. *It was to protect herself from
puffins.
To be sure, events can be referred to by pronouns even when they

are not introduced by event nominals:
(70) a. John kissed Mary.1t occurred in the garden. 9
b. John kissed Mary.1t took place in the afternoon.
c. John kissed Mary. It happened so quickly she couldn't
object.
But the point to note about (70) is that these sentences do not
identify an additional thematic role bearer for the event, as the
discourses with nominals in (66)-(69) do. Rather, what folIows the
second verb in (70) is an adjunct, not an argument. (Another difference
is of course that verbs like accur and take place are used in the second
case, but only the copula can be used to "add" a true argument as in
(66)-(69).)
Another thing to be noted is that a need for the uniqueness postulate
(22) does not manifest itself at alI with the arguments of verbs; we
presumably do not have the possibility of violating it by "duplicating"
an argument, because syntactic subcategorization would block such
duplication independently (e.g. *Jahn gave Mary a baak ta Susan).
FinalIy, note the data in (63)-(69) together form an argument that
both an ordered-argument system and a thematic role system are
needed for natural language. Examples (63)-(66) seem to require a
thematic role system, for there is no other obvious mechanism to
explain how arguments can be added to an event "across a sentence
boundary" as it were (as pointed out by Sag (1985. On the other
hand, (67)-(69) show that this evidence exists only for nominals, not
for verbs themselves, so in addition to the other ways in which
argument association must be differentiated for verbs, we must prevent
arguments from being "added" in a subsequent sentence for an event
introduced by a verb. Though details of such a mechanism remain to be
worked out, the claim that event nouns differ from verbs in that only
the former directly refer to events provides a principled basis (and at
present the only obviously motivated one) on which to build a theory of
event anaphora 10 that makes this distinction correctly.
5. THE EMPIRIC AL EVIDENCE FOR A

STRONG THEORY OF ROLE-1'YPES
The main concern of this paper is with the nature and form of a
104 DA VID R . DOWTY
semantic theory of thematic role types, not with the empirical linguistic
evidence for any particular role type or set of them. Nevertheless, 1
believe it would be inappropriate and possibly misleading to condude
without a brief assessment of the state of linguistic evidence for role
types. Since the literature that argues for some role type or other is now
vast and couched within widely varying assumptions about linguistic
theory, such an assessment must necessarily be impressionistic, as space
predudes a systematic survey.
5.1. The Problematic Nature ofthe Data

Evidence for particular role types is generally of one of two forms: (i) it
is daimed that some natural language syntactic phenomenon must be
described as conditioned by a particular thematic role (e.g. the way to
predict which lexical passives are possible in English is to recognize
that they can only be formed from verbs whose objects are Themes), or
(ii) it is daimed that some phenomenon must be analyzed in terms of
some hierarchy of thematic role types (e.g., the nuH subject of purpose
dauses and infinitival relative dauses is, under certain conditions,
controlled by that NP argument of the matrix dause verb which has the
highest role type on the hierarchy Goal > Source, Location, regardless
of the grammatical relation that NP bears to its verb).
At the onset, it must be noted that some of the best known cases of
these two kinds have tumed out, upon doser inspection, not to be
generalizations statable in terms of thematic roles at alI. The hypothesis
that English lexical passives must be formed on Theme objects - put
forth in Anderson (1977), Wasow (1980), Williams (1981), and
Bresnan (1982) - has tumed out to be false: in fact, such passives can
be formed from all and only the lexical monotransitive verbs in English
(Levin, to appear), Le. verbs that can appear with one object NP and no
other complements in their active form, regardless of the thematic role
type of this NP. (This generalization, incidentally, is predicted to exist
by the categorial theory of lexical rules and relation-changing rules in
Dowty (1978).)11
The hypothesis that control of nun subjects of infinitival relatives and
purpose dauses is determined by thematic role or by a thematic role
hierarchy (Jackendoff 1972, to appear; Grimshaw 1975, Williams
1980, Nishigauchi (1984) is shown in I.:adusaw and Dowty (1987) to
have counterexamples in the form of structurally and semanticaHy
parallel sentences which allow different NPs to control the infinitive.

Following Bach (1982), Ladusaw and Dowty argue that extra-linguisic
practic al reasoning determines the control in these cases.
Of course, these two cases do not by any means exhaust the syntactic
motivation for particular role-types or hierarchies of them. But they do
show that even well-studied cases of this kind can turn out to have
purely syntactic or purely pragmatic explanations.
Even after twenty years of study of the phenomenon of thematic role
types in modern linguistics, there is surprisingly little consensus on what
the set of linguistically significant role types is (i.e. the L-thematic role
types of section 2). As Andrews (1985, 70) puts it, "no presently
known system of semantic roles [= our L-thematic role types, DRD]
can be comprehensively applied in a convincing manner." Indeed, there
probably is not even an agreed-upon proper sub set of the L-role types.
The following brief discussion is merely to illustrate the range of
problems that stand in the way of this goal.
A central example of inconsistency is the use of the role type Theme
by Jackendoff and Gruber and their followers to characterize the direct
objects of common transitive verbs like build, kill, cook, move, etc.,
while other writers classify these arguments as Patients (e.g. Talmy
1985, Andrews 1985). The disagreement is not merely terminological:
Themes are defined as (at least) those arguments that the predicate
entails to undergo a change of position or other change of state, thus
change-of-state arguments of intransitive verbs (e.g. subjects of die,
arrive, awaken, collapse, etc.) are Themes but cannot be Patients as that
role is usually conceived; conversely, other direct objects (e.g. in pound
the table, strike the Jence, interrogate the prisoner) are clearly Patients
by the usual definitions but are not Themes (because they are not
entailed to undergo a particular kind of change as a result of the verb's
action). Curiously, the possibility both role types are to be employed,
with these direct objects being assigned both of them, has been
proposed only very recently (Culicover and Wilkins 1986, Jackendoff
to appear).
The role-type Theme itself is not defined consistently: Jackendoff
(1972, 1974) and others, following Gruber (1965), explicitly advocate
the disjunctive definition of Theme as either "the moving object" (as in
The train travelled from Detroit to Cincinnatti) or "the object whose
location is specified" (as in Max is in Africa);" ef. Jackendoff (1974,
93-95). Aside from the dubious strategy of adopting a disjunctive
106 DAVID R. DOWTY
semantic definition for a key semantic concept in one's theory, this step
conflicts with other purposes to which the role Theme has been put,
e.g. Verkuyl's (1978, ms.) attempt to explicate the semantics of telic
predicates (Vendler's accomplishments and achievements) in terms of
the role Theme, which requires that only the first half of Gruber's
definition be used to characterize this role-type. 12
Another problem is that the traditional thematic role types, when
subjected to careful semantic scrutiny, tend to fragment into two or
more distinct notions. If for example one examines the criteria for
identifying Agents in English, it becomes necessary to distinguish two
separate notions of Agent (Cruse 1973, Dowty 1979); Jackendoff's
(1983, 180; to appear) "Agent" versus "Actor" appears to acknowledge
this same distinction.
A somewhat different kind of fragmentation occurs with the com-
mercial transaction verbs (buy, seU, etc.). These cases were at first
analyzed as involving the three traditional roles Theme, Source, and
Goal. For example, in (71) and (72),
(71) Mary sold a book to John.

(72) John bought a book from Mary.
a book is Theme, because it changes possession; Mary is Source,
because she possesses the Theme-role bearer at the beginning of the
sales event, and John is the Goal, because he possesses the Theme-role
bearer at the end of the event. But one realizes this analysis is in trouble
as soon as one considers examples like (73),
(73) Mary sold a book to John for five dollars.

which remind us that selling is a four-place relation: it necessarily
involves not only an item sold but also a quantity of money paid for it.
But since the money changes possession just as surely as the sold item
does, the former would seem to qualify as Theme exactly as well as the
latter does.u And since this second Theme-bearer is originally in the
possession of the seller and ends up in the possession of the buyer, the
seller should be Source (as well as Goal) and the buyer should be Goal
(as well as Source). Apparently, roles such as "Secondary Theme",
"Secondary Source" and "Secondary Goal" will also be needed.
A reply that can be made to such problematic cases (and often has
been made) is that they do not show that the theory of role-types is
necessarily inconsistent or otherwise problematic, but merely that the

set of L-role types must be richer than the existing literature often
assumes. This reply is not very consoling, however, in view of the
overall paucity of linguistic evidence on individuating roles. It for
example the buy/sell case is the only one that even indicates the
existence of the role-types "Secondary Theme", etc. (as it appears to
be), then there would seem to be no hope of arriving at independently
testable criteria for identifying instances of this and similarly ephemeral
roles. And what happens to the claims that thematic role hierarchies
govern certain phenomena, if there exist not merely the four or so role-
types that one inevitably finds in the justification of such hierarchies but
rather dozens (hundreds?) of distinct role-types? Is there evidence to
show, for example, whether all the sub-types of Themes occupy the
same position in the hierarchy, or whether Primary Theme might be
above Goal but Secondary Theme below it in some hierarchy? It is
conceivable that the evidence to resolve all such questions is actually
present in languages like English and has been overlooked, but this now
seems not very likely.
Yet another kind of (familiar) problem is the case of predicates that
do not have any apparent difference at ali in their entailments with
respect to two of their arguments, hence offer no semantic basis for
assigning distinct role-types to these arguments, such as the symmetric
predicates in (74): 14
(74) a. Mary is as tall as John.
b. Line AB is parallel to line CD.
c. The salt water mixes with the fresh water at the mouth of
the river.
d. X is different from y.
These are of course an obstacle to maintmmng (any version of)
distinctness of thematic role types (8) (or at least, to maintaining
distinctness along with completeness (7.
5.2. Inadequacy of L-role-types for Argument Indexing

As mentioned at the beginning of the paper, and as many readers will
have observed, it is possible to view many of the problems with the
notion of thematic role in linguistic theory as resulting from the tension
108 DA VID R . DOWTY
between using thematic roles in an indexing function in theories like

GB and LFG (ef. the "biuniqueness" requirement on theta-roles) and
using thematic roles to formulate semantic generalizations about syn-
tactic and lexical processes in the work of descriptively-oriented seman-
ticists like Jackendoff, Fillmore, Talmy and others. 1 believe problems
like those mentioned in this section indicate the chances are slim
indeed that a characterization of thematic role-types which pays serious
attention to semantic data could ever serve this indexing function at the
same time. This same conclusion has been vigorously championed
recently, though from a very different theoretical perspective, by
Jackendoff (to appear), to which the reader is referred for further
linguistic data arguing for the conclusion.
5.3. Thematic Roles as the Locus of Semantic Generalizations about

Syntax
Suppose then that, along with Jackendoff, we abandon the attempt to
use thematic role types to index arguments semantically and instead
focus on descriptive generalizations. We would no longer require a set
of L-role types in the sense defined in section 2 (satisfying complete-
ness, distinctness and independence), but we would still be interested
in the question whether there exists some unique finite set of (non-
complete, non-distinct) role types that deserve to play a special role in
linguistic theory. This question will undoubtedly not be answered
decisively for some time, but 1 will try to put the issue in some
perspective as follows. At the very least, this is surely agreement among
linguists on (75):
(75) There are many instances in natural language where syn-
tactic and lexical processes are dependent on 15 aspects of
the lexical meaning of predicates (verbs and adjectives).
The hypothesis that role-types have a place in linguistic theory is, as
a minimum, (76):
(76) There are many cases in many (alI?) naturallanguages where
syntactic and lexical processes are dependent on an aspect
of lexical meaning analyzable as one of some universal set of
role-types.
A much stronger hypothesis would be (77); some recent syntactic
litera ture sounds as if (77) is being assumed:
(77) AlI cases in natural languages where syntactic and lexical

processes are dependent on lexical meaning of predicates
are analyzable a dependency on a role-type.
The problem for descriptive linguistic research is to demonstrate
convincingly that (76) or (77), rather than (78), is the case:
(78) Thematic roles per se have no priviledged stams in the
conditioning of syntactic processes by lexical meaning,
except insofar as certain semantic distinctions happen to
recur more frequently than others among natural languages,
for, perhaps, purely cultural or other non-linguistic reasons,
and as a consequence these semantic distinctions tend to
condition syntactic processes slightly more often than others
do.
The difficulty of distinguishing (77) from (78) on sound empiric al
grounds is, I believe, seriously underestimated.
However, I will suggest in the remainder of this paper that concep-
tions of something like "linguistically significant role-types" much more
limited than the familiar ones may be empirically justified by language
data, as are a place for role-types in general linguistic and cognitive
development.
5.4. A More Compelling Kind of Evidence: Subject Selection Principles

and the Agent-Patient Opposition
A kind of data bearing on role-types which has not yet been discussed
in this paper is that which motivated Fillmore's (1968, p. 33) "subject
selection principles". A fundamental insight of Fillmore's, and a key
motivation for his theory of case grammar, was the observation that the
distribution of semantic classes of arguments of natural language
predicates among the grammatical relations Subject, Direct Object,
Indirect Object, etc. was not random but followed apparently consistent
principles within and across languages. These subject selection prin-
ciples specified that if there was (in our terms) an Agent role type
among a verb's arguments, it would be the Subject; if there was no
Agent but an Instrument, that would be the Subject; otherwise, the
"-
Subject was, roughly, the Patient (or Theme). (Some linguists later
made a place for Experiencer in this rule.)
Popular grammar, and occasionally the traditional grammartans,
110 DA VID R. DOWTY
assumed the grammatical notions Subject and Object to be defini-

tionally connected with Agent and Patient, respectively, and in the
1970's the non-formal typological literature explored the idea of
treating Subject and Object as cluster concepts, to which Agent and
Patient were two respective contributing properties (cf. Keenan 1976,
1984). But the theoretical treatments of grammatical relations in alI
current formal syntactic theories are purely formal and are independent
of such semantic associations (even though some theories tacitly rely on
just such associations to identify the subject and object relations in the
"deep" or "initial" syntactic stratum, ef. e.g. Rosen (1984) for Relational
Grammar). Also from the point of view of model-theoretic semantics
(for ordered-argument systems), as mentioned, any permutation of a set
of n-place relations can convey the same information as the original
relations (provided we know whether and which permutations we are
dealing with), so there is no necessary correlation of syntactic position
with role-type here either. Thus any cross-linguistic pattern in the
association of role-type with lexicalization of subject and object is, just
as much today as when Fillmore pointed it out in 1968, an empirical
fact of naturallanguage to be studied and accounted for.
Of course, such grammatical/role-type association principles do not
fulIy determine argument arrangement for alI verbs: in English we have
permutation (near-)synonyms as like vs. please,16 give vs. receive, and
among three-place predicates buy vs. seU, and buy vs. pay. However,
there is one central class of cases where "alignment" of role-types with
grammatical relation is extraordinarily predictable: the case of transi-
tive verbs (i.e. 2-place but not 3-place predicates) in which one
argument is a highly protypical Agent (i.e. acts physically (not just
cognitively) and volitionalIy, etc.) and the other is a highly prototypical
Patient (i.e. is causalIy aftected by the Agent and undergoes some
specific change of state as a result, including coming into existence and
going out of existence.) - in other words, verbs like those in (79),
which Andrews (1985, 68) refers to as primary transitive verbs:
(79) kilI, eat, smash, build, move, break, cook, etc.
Among this class, the Agent is always subject, and we never find,
apparently in any language, a permutation near-synonym for a member
of this class - e.g. a basic (i.e. non-passiv~ verb meaning be killed by.
One point 1 wish to make here about this generalization is that there
is no good reason that we must appeal to a set of discrete role types in
order to describe it adequately. In fact, I believe it can be argued

(Dowty, in preparation) that (i) we only need an opposition between
"Agent" and "Patient", and not other role-types, to describe the subject
selection principles; (ii) for purposes of subject selection, "Agent" and
"Patient" are cluster concepts, not distinct categories, and (iii) these two
"proto-roles" are used by the first-Ianguage Iearner as defaults for the
meanings of the transitive verbs of his language.
My second point here is that the body of data arguing for an Agent-
Patient opposition to align lexical meanings of transitive verbs with
Subject-Object is quite weighty, as it is abundantly observed in most if
not all 16 natural languages. By contrast, the evidence for the effects
of other, discrete role-types in grammar is fragile and comes from
relatively few languages. We will return to the significance of the
subject-selection principles below in section 6.1.
5.5. Arguments of Derived Nominals from an Alternative Viewof

Role-Types
If no distinct and complete system of role-types should exist, then how
can this fact be reconciled with the data in section 4 showing that the
arguments of derived nominals behave in a neo-Davidsonian way and,
in particular, the fact it is possible to "predicate" an additional thematic
rale of an event in a clause independent of the one in which a derived
nominal introduced the event? This possibility exists, it should be
noted, not only for the familiar roIe types Agent, Patient etc. but also
cases like those marked by the prepositions in, ta and of in (80):
(80) a. Mary's odd belief ... It is not in the theory of general
relativity at all but in a very different theory.
b. John's new commitment ... It is to the unordered base
hypothesis.
c. The relevant knowledge he possesses ... It is of secret
dealings with the USSR.
These prepositions are idiosyncratically required by the respective
nominals belief, commitment and knowledge, though I think it is self-
evident that their appearances here do not correspond to role-types
that can be motivated from objective (truth-conditional) semantic
considerations.
112 DA VID R. DOWTY
But I think there is an analogous problem with the gender systems of

natural languages that may shed light on this situation. Gender in
English is an almost pure case of what linguists call natural gender -
that is, masculine forms, the pronouns in English refer to male beings
(ignoring "common gender" uses), feminine forms to female, and neuter
pronouns to sexless entities. In this kind of language, gender can be
analyzed as a straightforward semantic phenomenon (ef. Cooper 1983).
However, grammatical gender (as in familiar Indo-European languages
like French, German and Latin) is usually assumed to have no
consistent semantic basis but is instead analyzed as an idiosyncratic
morphosyntactic property of individual common nouns.
A known problem with this view of grammatical gender is that the
gender of a deictic pronoun is conditioned by the noun(s) that could be
used to refer to the object indicated by the pronoun, even though that
noun may never appear in the surrounding discourse. So for example, if
you were waiting for a delayed streetcar with a crowd of impatient
would-be passengers in a German speaking country, someone might
well say Er kommt gleich, literally "he (will) come soon", meaning that
the streetcar everyone is waiting for will come soon. The masculine
pronoun er must be used because the noun for streetcar, Strassenbahn-
wagen, is masculine in German. It there are two nouns of different
genders that are in common usage to refer to such an object, the
speaker has a choice of two pronouns: in German, one can point to a
new car and say either Er is schn or Es ist schn, because of the
existence of the masculine noun Wagen 'car' and the neuter noun Auto
'auto'.
The question now is whether we are to analyze the gender of deictic
pronouns as being determined by grammar or by semantics. It we select
grammar, then it seems we would have to postulate an inaudible noun
somewhere in the discourse whenever a deictic pronoun is used (e.g.
Strassenbahnwagen in the above example) to "trigger" the gender of the
pronoun; this solution is, I take it, unpalatable and no doubt highly
problematic to execute. The only other obvious alternative is to as sume
there really is a semantic property of the referent of the deictic
pronoun corresponding to gender, so that the pronoun is semantically
inappropriate if it is of the "wrong" gender.
It seems to me not really absurd to think of (grammatical) gender as
a semantic property of objects, though it would of course be a property
defined indirectly in terms of the vocabulary of some language: an
object itself is neuter (masculine, or feminine) just in case there is a

noun in the language in question referring to objects of its kind which is
neuter (etc.). (It is sometimes suggested that grammatical gender should
be treated as a discourse phenomenon "intermediate" between syntax
and semantics, but current studies of discourse anaphora have shown
that semantics is repeatedly and intimately involved in conditions on
anaphora. For example, studies of the choice between singular and
plural discourse anaphors (Root 1986, Sells 1985) show that complex
facts about the semantic interpretation of a whole sentence can
determine whether a subsequent anaphor is singular or plural, even
though the antecedent NP itself is singular. Thus it is not clear to me
that claiming gender is a "discourse phenomenon" really implies that it
is not semantic in nature.)
1 suggest that thematic role-types for derived nominals should
perhaps be thought of, in a similar way, as denoting relations which are
ultimately defined in terms of the semantics of other words. The set of
these indirect role-types should then correspond, almost exactly,17 to
the set of prepositions that mark arguments on verbs and nominals, e.g.
by, ta, from, in, with, etc. Thematic roles would stiH be relations
between events and individuals, but the semantic content of role-types
would not be constant across verbs but would depend on the particular
verb whose nominal named the event. More precisely, a thematic role-
type TH; - say the to-roIe - denotes, for any event named by a
derived nominal NOMj , the relation an individual x stands in to that
event whenever the event is truly described by a sentence containing
the verb from which NOMj is derived and x as denotation of the
argument syntactically marked by ta with that verb. (Note that roIe-
types would then be of the semantically "disjunctive" kind mentioned in
(5) in section 2.) Thus in a way parallel to gender in deictic pronouns,
we could stiH account for thematic-role predication in an independent
clause via semantics rather than syntax.
1 propose that it is an important similarity between preposition-
marking of roles and grammatical gender is that both are clearly
"rooted", in some sense, in semantics (e.g. classes of male individuals
are almost always referred to by masculine nouns even in grammatical-
gender languages, and semantically directional NP arguments of derived
nominals are marked by the preposition ta rather than some arbitrary
preposition, etc.), but these lexical and morphological distinctions have
been "grammaticized" and extended beyond their original semantic
114 DA VID R. DOWTY
range in much of the vocabulary. The proper way to understand the

relationship between the semantic "roots" of role-type marking in
nominals and its ultimate "grammaticized" status will be a subject of the
next section.
6. WHY SHOULD THERE BE THEMA TIC RO LE TYPES

(IF IN FACT THERE ARE ANY)?
If the hypothesis of a fixed and exact set of identifiable thematic roIe

types is so dubious, why should apparent if only partly correct
generalizations in terms of them be seen so often in natural language?
What good could a system of thematic roIe types be to a natural
language, anyway, and why should it appear so attractive for attaching
the arguments of event-nominals and adjuncts? What I suggest as
answers to these questions will be speculative but, I hope, not implau-
sible; hote I make no attempt to argue here that mine are the only
defensible answers that might be given to such questions.
6.1. Role- Types and the (First-Language) Learning of a Grammar

As discussed in section 5.3 above, the most widely attestable evidence
of role-types across languages, as weB the c1earest evidence within lan-
guages, seems to come from the "subject selection principle" for transi-
tive verbs with prototypically Agent-like and Patient-like arguments.
Consider now the dilemma of the child acquiring her native lan-
guage, in particular, at the stage at which she first begins to figure out
how grammatical relations are indicated in the language. More specifi-
caBy, when the child recognizes a sentence as containing a verb and two
nouns, e.g. (81 ),
(81) [aIN [;SIN [1filv
how will she figure out how her language tells her which is the
grammatical subject and which is the grammatical object? It could turn
out that word order indicates this, but it might be that in her language it
indicates nothing of the sort. It might be that the two nouns have affixes
which distinguish nominative from accusative case, a system of affixes
which might be complicated to figure out if the language has many
dec1inational c1asses and several cases. It might be that neither of these
determines subject and object; instead, agreement affixes on the verb
could be the only signals of grammatical relations, and in that case she
must learn the categories of nouns relevant for agreement, as well as the
verbal affixes. AH three of these "cod ing" systems are of course
common among languages of the world. She will have to divine her own
language's system by implicitly comparing a number of different
sentences in order to discern patterns.
Obviously, this task would be much more straightfoward if there
were some independent way of knowing, when given a sentence and the
situation which the sentence is used to describe, which of the two nouns
IS the grammatical subject and which is the object. But if there are
consistent semantic subject selection principles that must hold in some
central class of sentences (such as "Agent is subject, Patient is object"),
the child can exploit this regularity in learning this part of the grammar.
Once this fundamental grammatical distinction is mastered, then the
child, relying on it, can go on to learn kinds of verbs where no such
semantic principle applies, such as the psychological predicates like and
please, three-place verbs such as buy and seU, and of course, begin to
distinguish active from passive constructions. (Alec Marantz (1985) has
independently suggested this possible significance of subject-selection
principles.)
This scenario requires several things to be true of the language
acquisition process: first, it requires that the semantic categories
"Agent" and "Patient" be operative cognitive categories for the child at
the time that grammatical relations are first identified, and it requires us
to suppose that a number of prototypical Agent-Patient verbs like
make and hit are regularly learned before psychological predicates
such as please, before three-place predicates, and before passive
constructions. As far as 1 know, all these assumptions are consistent
with what is known about the acquisition process.
Is there any positive psycholinguistic evidence that this sort of thing
really happens? 1 so far know of two possibilities. Alec Marantz himself
has conducted an acquisition experiment (Marantz 1982) in which he
attempted to teach children made-up verbs, including both verbs that
conform to the Agent-Patient selection principles and some that do not,
and he found that indeed children seem to have significantly more
difficulty learning the items that violate this principle. The second kind
of evidence comes from an observation Dan Slobin (Slobin 1966;
similar data in Gvozdev 1949) has made about the acquisition of
Russian. Russian, like many Indo-European languages, has transitive
116 DAVID R. DOWTY
verbs that take various cases for their second argument; to some extent,
this variation depends on semantic conditions, e.g. what thematic role
this second argument bears, but to some extent the variation is
idiosyncratic; the case for some verbs must be learned individually. It
seems that children do not always use exact1y the case patterns in their
parents' speech, but rather make "errors" that seem to follow a
consistent semantic pattern: they use accusative case only for Patient-
like direct objects and not for other, non-Patient direct objects which
are accusative in adult speech.
It is also interesting to compare this hypothetical scenario with a
procedure actually recommended in a recent textbook for linguists
when beginning to analyze the grammatical system of an unknown
language (Andrews 1985, 68-69). The author recommends that the
linguist first try to elicit from the native informant a representative set
of what Andrews calls "primary transitive sentences", sentences with
transitive verbs that have clear instances of Agents and Patients. From
this list, the linguist should be able to discover how the grammar
distinguishes subjects from objects (and whether the language is
accusative or ergative - cf. note 16). Only then is the linguist advised
to go on to verbs such as psychological predicates, which can be
examined, using the grammatical criteria already established, to see
which arguments the language treats as subjects and objects in these
less predictable cases.
Rowever, 1 would like to further speculate that the function of
thematic roles may be much more general than an aid to grammar
learning. 1 suggest that thematic roles may be a part of how children -
and perhaps adults too, on some occasions - make preliminary
identifications and preliminary classifications of events. Carlson (1984)
has some very interesting things to say about this matter. Following
suggestions by Jackendoff (1983) and ter Meulen (1984), he questions
the common idea
that model-theoretic semantCs must assume that the domain of discourse has a fixed,
absolute, individuated structure, independent of and prior to the interpretation assigned
to expressions in language. Rather, language should be viewed as capable of projecting
a structure onto a domain with little or very different inherent structure. Given this
view, a principle of individuation can be looked upon as a means of projecting ontQ a
domain an individuated structure it may not 'have' of its own right.
Re goes on to suggest that event-denoting descriptions "portion out

certain sections of perceived reality as events" and that
... thematie roles play a crucial role in the theory of [the principles of individuation for
events] - that thematie roles are among the factors used in discriminating events from
one another. Thematie uniqueness forms the basis; if there is a proposed event with,
say, two themes, then there are (at least) two events and not one. (Carlson 1984, 272-
273)
Though I am not sure I agree with Carlson and ter Meulen on the
way primitives of a model-theoretic structure should be decided, I do
think Carlson has made a promising suggestion about the function of
thematic rules in the cognitive individuation of events. But 1 think that
in addition to distinguishing one event from another event, another
equally important, perhaps more important function of thematic role
types is in distinguishing events that are cognitively and linguistieally
important for human beings from a the enormous baekground of what
goes on around us that we do not talk about or attend to.
Consider a child sitting in a room with objeets and several other
people. From the point of view of a physicist or a philosopher, there
are surely a multitude of events going on around the ehild, an infinite
number. But the ehild cannot, and should not, try to attend eognitively
to alI of them, but rather has to beeome attuned to (to use the
Gibsonian phrase) those events that matter. Suppose someone says
Mary just gave a book to Susan, and suppose the child knows who
Mary is, doesn't know what the verb give means, has a only a vague
idea what a book might be, and doesn't know Susan. Where, among her
pereeptions of happenings in the room over the last few minutes, does
the child search for the event the sentenee might have deseribed? It will
help, I suggest, if the child looks for (or perhaps has already singled
out) spatio-temporal regions like these: regions where there are one or
more partieipants such that one participant has performed an agent-like
aetion (assuming the ehild is attuned to intentional actions of others), or
has made at least some movement, or may have experienced some
perception or motion, or where something may have changed location
or possession, and in these cases, the child might attend to the place
from which and to which it changed location or possession, and so ono
Such regions are not only likely to be referred to linguistically by
sentenees the ehild hears at this age, but identifying such regions as
events and classifying them should be important for general cognitive
reasons: to move in and manipulate the environment sueeessfully, to
interact with other humans, to avoid danger, and so ono
Another way of describing this idea is to say that thematie role types
may form a system of prototypes for classifying events, that is, a set of
118 DA VID R. DOWTY
typical ways in which the various participants of cognitively and

culturally relevant types of events interact in these events.
Of course, part of the genius of language is that "pratotype" criteria
for ordinary noun denotations do not prevent us fram eventually
adopting more precise criteria for these denotations, where relevant
and necessary, e.g. replacing criteria for kinds of plants or animals
based on prototypes with criteria based on a more scientific taxonomy,
as a consequence of our more sophisticated understanding of biology,
chemistry, etc. Even in the stages of acquisition of noun meanings by
young children, recent language acquisition research by Keil and
Batterman (1984) has shown that the child's first understanding of
common nouns such as unele, island, museum, etc., which is based on
prototypes, is replaced very soon by a criterial definition of the kind
adults use - thus a young child may believe that any adult male who
visits its family regularly and behaves in a friendly way is an unele,
without regard to being a parent's sibling or spouse thereof, but very
soon the child replaces this eluster of prototypical praperties with the
kinship criterion adults use to elass people as uneles.
Likewise, 1 think it very plausible that an initial identification of the
participants in an event based on something like thematic rale types
may eventually give way to more sophisticated criteria for identifying
types of events and their participants that are not limited by these
prototypes. In the pracess, the various entailments about the individual
participants that were once treated simply as either pratypical "Goals"
or "Themes" (or whatever) may so fragment and diverge in this pracess
of refinement that one can no longer identify a consistent set of
semantic entailments that characterize "Goals" (or other role types)
across the entire vocabulary of verbs. It so, then it seems to me that we
might expect just the kind of natural language data that we observe -
tantalizing glimpses of what looks a system of rale types, but no
genuine comprehensive inventory of rale types that can be precisely
defined in semantic terms.
It something like this picture is right, why should argument associa-
tion methods for nouns and verbs differ in the way 1 suggested earlier?
First, keep in mind that the ordered argument method is the richer one
(establishing this was part of my motivation for comparing the expres-
sive power of the two theories in section 2); it can distinguish more
relations from one another than a Neo-Davidsonian system. Also, note
that verbs and their system of subcategorized arguments are not only
learned earlier by children than event nouns with their arguments are
learned, but also that even in the fully adult grammar, verbs have a
syntactically more varied and complex system of arguments than nouns
do - a fact which Chomsky amply demonstrated for English in
"Remarks on Nominalization" (Chomsky 1970). Most event-denoting
nouns are morphologically derived from verbs. And of course, refer-
ence to events by verbs is simply more common than reference to them
by means of event-nouns, at least in colloquial speech. If the thematic
role argument system is indeed a more preliminary or primitive way of
indexing arguments than the subcategorization system, we should
expect verbs to acquire the more complex system at an earlier stage of
language acquisition than event-nouns would need to acquire them.
Hence, if verbs and event-nouns differ at alI in their argument-marking
systems, then it is reasonable to expect the nominal system to retain
more of the outward characteristics of the "primitive" system than verbs
do.
If my suggestions in section 5.5 are on the right track - that role-
types for noun arguments are, like grammatical gender, a frozen and
somewhat idiosyncratic "grammaticalization" of an originally semantic
system - then this grammaticalization should be the very last stage of
the acquisition process involving roles, the state at which true thematic
role types have fulfilled their acquisitional function and disappeared
from the compositional semantic interpretation principles of the lan-
guage altogether. 18
Notice that if my general line of speculation in this section is correct,
it may be that we won't really arrive at a complete understanding of the
semantics of thematic roles from studying fully-formed adult semantic
judgments about the meanings of sentences, but rather from psycho-
linguistic studies of how children and adults make initial individuations
and recognitions of events and from how they process sentences.
Nevertheless, it will undoubtedly require much more precise semantic
descriptions of adult judgments about this data than we present1y have,
to decide whether the psycholinguistic speculations in these paper or
other such psycholinguistic hypotheses are on the right track or nol. 1
believe that the initiation of the model-theoretic study of thematic roles
by Carlson (1984) and Chierchia (1984) (and hopefully, this paper too)
is an important foundation for these goals.
ACKNOWLEDGEMENTS
1 would like to thank Greg Carlson, Gennaro Chierchia, and Frank

120 DAVID R. DOWTY
Trechsel for their very helpful comments on this paper, without of

course implying that they agree with any or alI conclusions in it. Bill
Ladusaw has contributed substantively to my thinking about this topie.
In addition to the University of Massachusetts Property and Type
Theory conference, earlier versions of this material were presented at
the Berkeley Linguistics Society, Stanford University, and the Micro-
electronies the Computer Technology Corporation, and I would like to
thank the members of those audiences for their comments. This
research was supported by grants from the Ohio State University, the
Center for Advanced Studies in the Behavioral Sciences, and by
National Science Foundation Grant number BNS-8306067.
NOTES
1 " the phenomenon of coreference in human language is so vast, complicated and

obscure as to easy require several huge volumes even for an introduction to its
mysteries .... Nonetheless, let us emphasize that no definition will be offered. The
linguistic study of coreference does not depend on an a priori understanding of
reference. It depends rather on the ability to recognize clear instances of the manifesta-
tion and nonmanifestation of coreference in the sentences of some human language.
This condition is met."
2 At least Chierchia (1984, ms.), Dowty (1985), and Ladusaw and Dowty (1985) have
explicitly proposed that thematic roles be regarded, in a model-theoretic semantics
theory, as sets of entailments of lexical items.
3 The "non-thematic" arguments are the so-called dummies there and it, as in There is
a unicorn in the garden and It is true that John snores; the assumption of many linguists
is that such NPs must be excluded when the arguments of a predicate are interpreted
semantically. However, in Dowty (1981), it is shown how the semantics of these NPs
can be treated in a model-theoretic rule-to-rule semantic theory operating systemati-
cally off the ("surface") syntactic structure: although such NPs would have real enough
denotations, the predicates in these cases can be treated as simply generating no
entailments whatsoever with respect to the denotations of arguments in their "non-
thematic" position, hence it matters not at alI what their denotations are; semantically,
the result is indistinguishable from one in which such predicates were analyzed as
having one fewer arguments. See Dowty (1981,137-147) for more details. To be sure,
this possibility does not automatically argue that this "pure" semantic solution to the
phenomenon of dummy NPs is ultimately preferable to some quasi-syntactic one (i.e.
prohibiting dummies from being interpreted, by one means or another), merely that it is
not obviously any less adequate; this issue has many ramifications and may well remain
unresolved for some time.
4 Strictly speaking, this comment is intended to ~pply only to arguments of nominals
marked with the "contentful" prepositions by, ta, with, [rom, an, etc. and is (probably)
not correct for arguments marked with of, which though optional, appear to be the
exceptional case where grammatical relations of the verb play a role in determining
what argument is designated. As is well-known to generative linguists, of with a derived
nominal marks the argument that would be subject of the corresponding verb if it is
intransitive, but the object if the verb is transitive. Thus in the familiar examples below
(Chomsky 1957: 88), of indicates Agent in (i) (intransitive verb), Patient in (ii)
(transitive verb), and is only ambiguous in (iii) because the verb shoot can be either
transitive or intransitive:
(i) The growling of lions
(ii) The raising of flowers (similarly for the raising ofJohn)
(iii) The shooting of the hunters.
It might also seem that the prenominal genitive depends for its significance on the
subcategorization of the verb: note that The Klingon's must designate Agent in (iv),
Patient in (v):
(iv) The Klingon's defeat of the Federation
(v) The Klingon's defeat by the Federation.
However, this behavior follows from effects of the uniqueness of role bearer
condition (22) discussed in 4.4 below and from the independently verifiable fact that
the prenominal genitive denotes any contextually salient identifying relation (i.e. it is
context-dependent and vague as to thematic role designated, if it designates a thematic
role bearer at all, and not ambiguous among two or more roles).
5 In referring to "oblique grammatical relations" 1 exclude the case of those higher
order verbs that take as a subcategorized argument what would be an oblique with
simpler cases, e.g. parallel to the instrumental oblique in cut the salami with a knife is
the direct object in use a knife to cut the salami, the temporal extent adverbial in cut the
salami for an hour is paralleled by by the object in spend an hour cutting the salami,
etc. The Bantu languages are well-known for having productive "relation-expanding"
rules (Dowty 1982), known as "applied forms" to Bantu scholars, for deriving a verb
that takes as direct object an NP that would otherwise be an instrumental, benefactive
or locative adjunct to the simple verb. These derived Bantu direct objects are also to be
treated as arguments, not event predicates.
6 It might appear that one large class of cases of this sort would be altemations like
The water boiled vs. John boiled the water, i.e. where the second might be analyzed as
simply the added specification of an Agent role (whereas the first had only Theme).
Indeed there will be a multitude of such examples if this analysis is correct, for the
second is simply the causative form of the first, and many languages have a widely used
and sometimes completely productive process for deriving causative verbs from non-
causatives (or even from other causatives, recursively). However, this proposal encoun-
ters a problem with thematic uniqueness when the non-causative already has an Agent,
as in John walked the dog from The dog walked. Even if we allowed thematic
uniqueness to be be violated, Carlson's added-role-only analysis would seem to give an
inadequate semantics for such cases: we need to say not m~ely that both the dog and
John are Agents, but also that they are unequal Agents: John is the controlling Agent
and the dog is the controlled Agent (otherwise we risk confusing John walked the dog
122 DAVID R. DOWTY
with The dog walked John or John and the dog walked). Obviously, the solution to
both these problems for the Neo-Davidsonian is to say that causative formation
semantically introduces a new event: John's walking the dog involves an event of the
dog's walking (of which the dog is Agent) and an event of John's causing the first event
(of which John is the Agent): this gets the asymmetry right and preserves thematic
uniqueness too, but of course it means that there is more to the semantics of causative
formation than the addition of an Agent role. Though causatives lexically-derived from
verbs that already have Agents are not too common in English, many languages have
productive causative-deriving processes that as far as 1 know work equally well with
verbs having Agents (like walk) and verbs not having agents (like boii), so an attempt
to use the two-event analysis for the first class but still appeal to the added-thematic-
role analysis for the second class would run afoul of the linguistic generalization that
the two processes seem semantically and morphologically the same in these languages.
Thus the two-event analysis should be chosen for both classes, the argument would go.
7 Again, the situation looks more complicated than this when examples with of are
also taken into account: see note 4.

8 It is sometimes said that locatives and temporals are exceptions to thematic unique-
ness (e.g. Levin 1982,615) because of examples like (i) and (ii):
(i) Mary arrived at 4 AM on a Tuesday in the first week of January in 1978
during a snowstorm.
(ii) Mary wrote the book on a table in our living room in Portsmouth in
southern Ohio in the United States.
But such a statement can only arise from a superficial and incorrect view about the
semantics of temporals and locatives, for such examples do not violate the uniqueness
in (22) at all. Example (i) does not specify that Mary arrived at five times but rather at a
single time which is partially identified by each of four indefinite descriptions: a
formalized semantics for time adverbials (Dowty 1979, ch. 7) makes this perfectly
explicit. The same is true of the locative expressions in (ii). It two temporal (or locative)
expressions which cannot be satisfied by a single time (place) appear in the same
sentence, the sentence is ungrammatical, as (22) predicts, (and cannot, e.g. refer to a
temporally discontinuous event, so (iii) is bad and cannot mean (iv:
(iii) *Mary worked on Monday on Tuesday.
(iv) Part of Mary's work took place on Monday and part on Tuesday.
9 There are of course examples like John kissed Mary, and then she did it to him. But
there are two differences between this and (66)-(70): (i) The it here does not refer to
the same event as that in the previous sentence (as it does in (66)-(70, but to a
distinct event of the same type as the first, and (ii) the it refers not really to an event or
event-type, properly speaking, but to what should probably be called an action-type,
and a 'transitive' one at that - a function from pairs of individuals to event-types. Cf.
. . . and she did it on a number of occasions, exhibiting reference to several events, in
spite of the "definite" anaphoric form it. 1 am not sure how, if at all, the action-type
anaphora should be related to the two kinds of event anaphora discussed in the text.
10 Richmond Thomason (personal communication) has suggested to me that a very
natural way to construct a theory of event anaphora woilld be via Discourse Repre-
sentation Theory (Kamp 1981). That is, event sentences would be represented with a
kind of "free variable" over events, just as indefinite NPs are realized with a free
variable over individuals in this theory, and Davidson's existential quantifier in event
sentences would be unnecessary. Discourse anaphora for events would then proceed
exact1y like ordinary NP anaphora with indefinite antecedents.
II In this theory, the domain of the English lexical passive rule, as well as that of the
syntactic passive rule, is the category TV ("transitive verb phrase"). But in a categorial
grammar, syntactic categories may have both lexical and phrasal members. According
to the theory of lexical rules in Dowty (1978, 1979) a syntactic rule applies to phrasal
as well as lexical members of its input category, while a lexical rule applies only to
lexical members of its input category. Since e.g. seU to Mary and seU a house are
phrasal TV's, we predict syntactic passives such as be sold to Mary and be sold a house,
but since seU is lexically (and idiosyncratically) reduced from TVIT to TV in such a
way as to retain its Theme argument (John sold a house) but not its Goal argument
(* John sold the customer), only the former TV is predicted to give rise to lexical
passives (ef. an unsold house, *an unsold customer); the data needed to distinguish this
prediction from one in terms of thematic roles (e.g. ef. the unfed children, *the unfed
soup with He fed the children, *He fed the soup) is found in Dowty (1979a) as well as
in Levin and Rappaport (to appear).
12 Since the semantics of aspect and aktionsart is now so much better understood than
the semantics of thematic roles - in particular the semantic entailments of a telic
predicate with respect to its "Theme" argument have been characterized in detail in
algebraic frameworks for analyzing both events and NP denotations (Bach 1986,
Hinrichs 1986) - it is hard to escape the conc1usion that it should be more profitable
to try to analyze Theme in terms of aspect that to try to analyze aspect in terms of
Theme. Of course, such an identification assumes that Theme should be a role-type
satisfying only the first half of Gruber's disjunctive definition, but this is obviously a
desirable step in any serious semantic analysis of role-types.
13 lackendoff (to appear) points out this same case as a problem for the O-criterion of
GB.
14 One suggestion (Talmy 1978, 1985) is that such arguments are distinguished by the
role-types Figure vs. Ground, both in the case of symmetric predicate (i) and that of
asymmetric, converse stative pairs (ii):
(i) Bill (Figure) resembles lohn (Ground).

lohn (Figure) resembles Bill (Ground).
(ii) The lamp (Figure) is over the chair (Ground).
The chair (Figure) is under the lamp (Ground).
But notice that in contrast to ali previous examples of thematic roles, such pairs are
not distinguished by any objective feature of the situation described but at best by the
"point of view" from which it is described. This naturally raises the question whether
such distinctions as may exist between these pairs appropriately belong with other
thematic role phenomena or rather with other notions of discourse reference (definitel
indefinite, referential/attributive, topic/comment), given that the grammatical subjectl
non-subject distinction is independently known to correlate with the referential topicl
non-topic distinction in English (e.g. Keenan 1976).
124 DA VID R. DOWTY
15 By using the phrase "dependent on", 1 of course do not mean to imply any
necessarily directionality in an analysis of such syntactic-semantic correlation. For
example, the fact that impersonal passive constructions in many languages cannot be
formed on intransitives whose subjects are so-called "unaccusative" predicates (Le. in
terms of role-types, predicates that have Themes instead of Agents as subjects) is
described in some analyses by restricting the impersonal passive rule so as not to apply
to unaccusatives. By contrast, Nerbonne (1984), employing a Montague semantics (in
which syntactic rules can have non-trivial semantic interpretations), analyzes the
German impersonal passive construction as conventionally implicating that the action
predicated by the verb of its argument is intentional, thus attributing the absence of
"unaccusative" impersonals to a semantic clash between the rule and the lexical
meaning of unaccusatives. Pre-theoretically, there is no reason for preferring one kind
of analysis over the latter.
16 A possible counterexample to this generalization would seem to be the case of
"Deep" ergative languages, such as the Australian language Dyribal (Dixon 1972) and
the Central American Mayan language Mam (England 1983), if the arguments in
Schmerling (1979), Trechsel (1982) and some descriptive linguists (ef. Plank 1980) are
correct that the Patient arguments of transitive verbs (e.g. their absolutive arguments)
are, grammatically, actually the subjects of these verbs, while the Agent (ergative)
arguments are something analagous to the direct objects of more familiar languages.
However, what is remarkable about such languages is that there is no evidence in the
descriptive literature on them that they are anything less than perfectly consistent about
their "inverse" association of Agent/ Patient with Subject/Object in transitive verbs.
That is, there is no report known to me of any language - ergative, accusative or even
split ergative - in which some percentage of prototypical Agent-Patient transitive verbs
align Subject arguments in one way while the rest of such verbs in the language align
the other way (but N.B. that "split ergative" case marking is not a counterexample to
this, because it is dependent on NPs or on tenses, not on choice of verbs). Thus, "deep"
ergative languages can be viewed as making use of the very same semantic basis for
subject selection as the more familiar ones, but effecting the inverse grammatical
alignment; in this sense they further confirm the universality of the Agent/Patient
opposition in lexical semantic organization.
17 Minor deviations from this one-to-one relationship arise from the fact mentioned in
note 4 that the preposition of and possessive determiners can denote the argument
encoded as either the subject or the object of the corresponding verb, depending on the
transitivity of the verb and the presence of others arguments; this is a complication for
the view under discussion here but, as far as 1 can tell, not an obstacle. And note that in
view of the arguments in section 4.3, by must denote a "true" role-type, not one defined
indirectly in terms of grammatical relations of the corresponding verbs. Note also that
infinitive complements would constitute an indirect role-type unto themselves, in view
of examples like John made a promise. It was to leave.)
18 However, one should perhaps not neglect to consider also the possibility that even
though a Neo-Davidsonian system does not give a fully adequate semantics, some
cognitive counterpart of such a system might stil! play a role in "on-line" processing
strategies, at least in the initial steps of such processing, presumably to be supplemented
later (or if necessary) by other mechanisms to "correct" and/or expand details of the
semantic interpretations of sentences, details such as those of the problematic data in
section 5.
REFERENCES
Anderson, Stephen: 1977, 'Comments on the Paper by Wasow', in Formal Syntax,

edited by Peter Culicover, Thomas Wasow, and Adrian Akmajian, New York:
Academic Press, pp. 361-378.
Andrews, Avery: 1985, 'The Major Functions of the Noun Phrase', Language Typology
and Syntactic Description, voI. I: Clause Structure, edited by Timothy Shopen,
Cambridge: Cambridge University Press, pp. 62-154.
Bach, Emmon: 1980, 'In Defense of Passive', Linguistics and Philosophy 3, pp. 297-
342.
Bach, Emmon: 1982, 'Purpose Clauses and Control', The Natureof Syntactic Repre-
sentation, edited by Pauline lacobson and Geoffrey Pullum, Dordrecht: D. Reidel,
pp.36-57.
Bach, Emmon: 1986, 'The Algebra of Events', Linguistics and Philosophy 9, pp. 5-16.
Barwise, Ion and Perry, lohn: 1983, Situations and Attitudes, Cambridge: Bradford
Books.
Bresnan, Ioan: 1982, 'Control and Complementation', The Mental Representation of
Grammatical Relations, edited by Joan Bresnan, Cambridge: MIT Press.
Bresnan, Joan: 1982a, 'Polyadicity', The Mental Representation of Grammatical Rela-
tions, Cambridge: MIT Press, pp. 149-172.
Carlson, Greg: 1984, 'On the Role of Thematic Roles in Linguistic Theory', Linguistics
22, pp. 259-279.
Chierchia, Gennaro: 1984, Topics in the Syntax and Semantics of Infinitives and
Gerunds, University of Massachusetts dissertation (to be published in revised form
as Predication in Natural Language, Cambridge: MIT Press).
Chierchia, Gennaro: 1985, 'Non-Existent Subjects', ms., Brown University.
Chierchia, Gennaro and Jacobson, Pauline: 1985, 'Local and Long Distance Control',
paper presented at the 1985 Annual Meeting of the Linguistic Society of America.
Chomsky, Noam: 1957, Syntactic Structures, The Hague: Mouton.
Chomsky, Noam: 1970, 'Remarks on Nominalization', Readings in English Transforma-
tional Grammar, edited by Roderick lacobs and Peter Rosenbaum, Waltham: Ginn
and Company, pp. 184-221.
Chomsky, Noam: 1981, Lectures on Govemment and Binding, Dordrecht: Foris
Publications.
Cocchiarella, N.: 1983, Logical Investigations of Predication Theory and the Problem of
Universals, Naples: Bibliopolis.
Comrie, Bernard: 1985, 'Causative Verb Formation and Other Verb-Deriving Mor-
phology', Language Typology and Syntactic Description, edited by Timothy Shopen,
Cambridge: Cambridge University Press, pp. 309-348.
Cooper, Robin: 1983, Quantification and Syntactic Theory, Dordrecht: D. Reidel.
Culicover, Peter W. and Wilkins, Wendy: 1986, 'Control, PRO, and the Projection
Principle', Language 62, pp.120-153.
Cruse, D. A.: 1973, 'Some Thoughts on Agentivity', Joumal of Linguistics 9, pp.
1-204.
Davidson, Donald: 1967, 'The Logical Form of Action'- Sentences', The Logic of
Decision and Action, edited by Nicholas Rescher, Pittsburgh: University of
Pittsburgh Press, pp. 81-95.
Davidson, Donald: 1967a, 'Reply to Comments', The Logic of Decision and A.ction,
126 DA VID R. DOWTY
edited by Nicholas Rescher, Pittsburgh: The University of Pittsburgh Press, pp.

115-120.
Dixon, R M. W.: 1972, The Dyirbal Language of North Queensland, Cambridge
Univer~ity Press.
Dowty, David R: in preparation, 'Thematic Roles as Defaults in Lexical Semantic
Acq uisition'.
Dowty, David R: 1978, 'Governed Transformations as Lexical Rules in a Montague
Grammar', Linguistic lnquiry 9, pp. 393-426.
Dowty, David R: 1979, World Meaning and Montague Grammar, Dordrecht: Reidel.
Dowty, David R: 1979a, 'Dative 'Movement' and Thomason's Extensions of Montague
Grammar', Linguistics, Philosophy and Montague Grammar, edited by Steven Davis
and Marianne Mithun, Austin: University of Texas Press, pp. 153-222.
Dowty, David R: 1982, 'Grammatical Relations and Montague Grammar', The Nature
of Syntactic Representation, edited by Pauline Jacobson and Geoffrey Pullum,
Dordrecht: Reidel, pp. 79-130.
Dowty, David R: 1982a, 'More on the Categorial Theory of Grammatical Relations',
Subjects and Other Subjects: Proceedings of the Harvard Conference on the Repre-
sentation of Grammatical Relations, edited by Annie Zaenen, Bloomington: Indiana
University Linguistics Club, pp. 115-153.
Dowty, David R: 1985, 'On Recent Analyses of the Semantics of Control', Linguistics
and Philosophy 8, pp. 1-41.
Dowty, David R, Wall, Robert, and Peters, Stanley: 1981, lntroduction to Montague
Semantics, Dordrecht: Reidel.
England, Nora C.: 1983, A Grammar of Mam, a Mayan Language, Austin: University
of Texas Press.
Fillmore, Charles: 1968, 'The Case for Case', Universals in Linguistic Theory, edited by
Emmon Bach and Robert T. Harms, New York: Hoit, Rinehart and Winston, pp.
1-90.
Fillmore, Charles J.: 1977, 'The Case for Case Reopened', Syntax and Semantics 8:
Grammatical Relations, edited by Peter Cole and Jerrold Sadock, New York:
Academic Press, pp. 59-82.
Fillmore, Charles J.: 1986, 'Pragmatically Controlled Zero Anaphora', Proceedings of
the Twelfth Annual Meeting of the Berkeley Linguistics Society, edited by Vassiliki
Nikiforidou, Mary VanClay, Mary Niepokuj, and Debora Feder, Berkeley: Berkeley
Linguistics Society, pp. 95-108.
Grimshaw, Jane: 1975, 'A Note on the Interpretation of Subjects of Infinitival Rela-
tives', ms., Amherst: University of Massachusetts Linguistics Department.
Grimshaw, Jane: 1979, Complement Selection and the Lexicon, Linguistic lnquiry 10,
pp.279-326.
Gruber, Jeffrey S.: 1965, Studies in Lexical Relations, Doctoral Dissertation, MIT,
distributed by the Indiana University Linguistics Club.
Gvozdev, A. N.: 1949, Formirovanie u rebenka grammaticheskogo stroya russkogo
yazyka (Formation in the Child of the Grammatical Structure of the RussiC!n
Language), Parts 1 and II, Moscow: Pedag. Nauk RSFSR
Higgins, Roger: 1979, The Pseudo-Cleft Construction in English, MIT dissertation [also
published by Garland Publishing Company, Outstanding Dissertations in Linguistics
Series].
Hinrichs, Erhard: 1986, A Compositional Semantics for Aktionsarten and NP Reference

in English, Ohio State University Dissertation.
Hoeksema, Jack; 1983, 'Plurality and Conjunction', Studies in Model-Theoretic Seman-
tics (Groningen-Amsterdam Studies in Semantics, v. 1), edited by Alice G. B. ter
Meulen, Dordrecht: Foris Publications, pp. 63-84.
Jackendoff, Ray: 1972, Semantic lnterpretation in Generative Grammar, Cambridge:
MITPress.
Jackendoff, Ray S.: 1976, 'Toward an Explanatory Semantic Representation', Linguistic
Inquiry 7, pp. 89-150.
Jackendoff, Ray S.: 1983, Semantics and Cognition, MIT Press.
Jackendoff, Ray S.: 1985, 'The Role of Thematic Relations in Linguistic Theory', paper
presented at the Symposium on Thematic Roles, 1985 Annual Meeting of the
Linguistic Society of America, December 27-30,1985.
Jackendoff, Ray S.: To appear, 'The Status of Thematic Relations in Linguistic Theory',
Linguistic Inquiry.
Kaplan, Ronald M. and Bresnan, Joan: 1982, 'Lexical-Functional Grammar: A Formal
System for Grammatical Representation', The Mental Representation of Grammati-
cal Relations, edited by Joan Bresnan, Cambridge: MIT Press, pp. 173-281.
Kamp, Hans: 1980, 'A Theory of Truth and Semantic Representation', Formal Methods
in the Study of Language, edited by J. Groenendijk, T. Janssen and M. Stokhof,
Amsterdam: Mathematisch Centrum, pp. 277-322.
Kay, Martin: 1985, 'Parsing in Functional Unification Grammar', Natural Language
Parsing, edited by David R Dowty, Lauri Karttunen and Arnold Zwicky, Cam-
bridge University Press, pp. 251-278.
Keenan, Edward L.: 1976, 'Towards A Universal Definition of "Subject"', Subject and
Topic, edited by Charles N. Li, New York: Academic Press, pp. 303-334.
Keenan, Edward L.: 1984, 'Semantic Correlates of the ErgativelAbsolutive Distinction',
Linguistics 22, pp. 197-223.
Keil, Frank, and Batterman, N.: 1984, 'A Characteristic-to-Defining Shift in the
Development of Word Meaning', Journal of Verbal Leaming and Verbal Behavior
23, pp. 221-236.
Ladusaw, William, and Dowty, David R: To appear 1987, 'Towards a Non-Gram-
matical Account of Thematic Roles', On the Nature of Thematic Roles, edited by
Wendy Wilkins, New York: Academic Press.
Larson, Richard: To appear, 'Adjuncts and Implicit Arguments in Situation Semantics',
Linguistic and Philosophy.
Levin, Beth: To Appear, 'The Formation of Adjectival Passives', Linguistic Inquiry.
Levin, Lori S.: 1982, 'Sluicing: A Lexical Interpretation', The Mental Represetation of
Grammatical Relations, edited by Joan Bresnan, MIT Press, pp. 590-655.
Link, Godehard: 1983, 'The Logical Analysis of Plurals and Mass Terms: A Lattice-
Theoretical Approach', Meaning, Use and Interpretation of Language, edited by R
Baeuerle, Christoph Schwarze, and Arnim von Stechow, Berlin: de Gruyter, pp.
302-323.
Marantz, Alec: 1982, 'On the Acquisition of Grammatical Relations', Linguistische
Berichte 80-82, pp. 32-69. "
Marantz, Alee P.: 1984, On the Nature of Grammatical Relations (revision of 1981
MIT dissertation), Cambridge: MIT Press.
128 DAVID R. DOWTY
Nerbonne, John A: 1984, German Temporal Semantics: Three-Dimensional Tense

Logic and a GPSG Fragment (Ohio State University Working Papers in Linguistics
30) Columbus: Ohio State University Department of Linguistics.
Nichigauchi, Taisuke: 1984, 'Control and the Thematic Domain', Language 60, pp.
397-414.
Ostler, N. D. M.: 1979, Case-Linking: A Theory of Case and Verb Diathesis Applied to
Classical Sanskrit, MIT doctoral dissertation.
Parsons, Terence: 1980, 'Modifiers and Quantifiers in Natural Language', Canadian
Joumal of Philosophy Supplementary Volume VI, pp. 29-60.
Plank, Frans, ed.: 1980, Ergativity: Toward a Theory of Grammatical Relations,
London: Academic Press.
Postal, Paul M.: 1970, 'On Coreferential Complement Subject Deletion', Linguistic
Inquiry 1, pp. 439-500.
Rappaport, Malka: 1983, 'On the Nature of Derived Nominals', Papers in Lexical-
Functional Grammar, edited by L. Levin, M. Rappaport, and A Zaenen, Blooming-
ton: Indiana University Linguistics Club.
Riemsdijk, Henk van, and Williams, Edwin: 1986, Introduction to the Theory of
Grammar, Cambridge: MIT Press.
Roberts, Craige: 1986, Modal Subordination, Anaphora and Distributivity, University
of Massachusetts Dissertation.
Rosen, Carol: 1984, 'The Interface between Semantic Roles and Initial Grammatical
Relations', Studies in Relational Grammar 2, edited by David M. Perlmutter and
Carol Rosen, Chicago: University of Chicago Press, pp. 38-80.
Sag, Ivan (ms.): 'A Semantic Theory of Obligatory Control', Ms., Stanford University.
Sag, Ivan: 1985, 'On the Semantics of Control', paper presented at the 1985 Annual
Meeting of the Linguistic Society of America.
Scha, Remko: 1981, 'Distributive, Collective and Cumulative Quantification', Formal
Methods in the Study of Language, edited by J. Groenendijk, T. Janssen, and M.
Stokhof, Amsterdam: Mathematical Centre Tracts, pp. 483-512. (Reprinted in
Truth, Interpretation and Information (1984), ed. J. Groenendijk, T. Janssen, and
M. Stockhof, GRASS voI. 2, Dordrecht: Foris, pp. 131-158.)
Schmerling, Susan F.: 1979, 'A Categorial Treatment of Dyirbal Ergativity', Texas
Linguistic Forum 13, pp. 96-112.
Sells, Peter: 1985, 'Restrictive and Non-Restrictive Modification', CSLI Report No.
CSLI-85-28, Stanford: Center for the Study of Language and Information, Stanford
University.
Slobin, Dan 1.: 1966, 'The Acquisition of Russian and a Native Language', The Genesis
of Language, edited by Frank Smith and George A Miller, Cambridge: MIT Press,
pp.129-148.
Talmy, Leonard: 1978, 'Figure and Ground in Complex Sentences', Universals of
Human Language, Volume 4: Syntax, edited by J. H. Greenberg et al., Palo Alto:
Stanford University Press, pp. 625-649.
Talmy, Leonard: 1985, 'Lexicalization Patterns: Semantic Structure in Lexical Forms',
Language Typology and Syntactic Description, edited by Timothy Shopen, Cam-
bridge: Cambridge University Press, pp. 57-149.
ter Meulen, Alice: 1984, 'Events, Quantities and Individuals', Varieties of Formal
Semantics, edited by Fred Landman and Frank Veltman, Dordrecht: Foris, pp.
259-280.
Trechsel, Frank R: 1982, A Categorial Fragment of Quiche, Texas Linguistic Forum 20,
University of Texas Department of Linguistics.
Verkuyl, Henk J. (ms.): 'Aspectual Asymmetry and Quantification' (ms. University of
Utrecht).
Verkuyl, Henk 1.: 1978, 'Thematic Relations and the Semantic Representation of Verbs
Expressing Change', Studiesin Language 2, pp. 199-233.
Wasow, Thomas: 1980, 'Transformations and the Lexicon', Formal Syntax, edited by
Peter Culicover, Thomas Wasow and Adrian Akmajian, New York: Academic
Press, pp. 327-360.
Williams, Edwin S.: 1980, 'Predication', Linguistic Inquiry Il, pp. 203-238.
Williams, Edwin S.: 1981, 'Argument Structure and Morphology', The Linguistic
Review 1, pp. 81-114.
Department of Linguistics
Ohio State University
GENNARO CHIERCHIA
STRUCTURED MEANINGS,
THEMATIC ROLES AND CONTROL *
O. INTRODUCTION
One of the salient features of current work in semantics has been the
search for fine grainedness. The strategies that are most actively being
explored can perhaps be classified in two groups. The first group of
proposals tries to come up with a theory of logical space capable of
weighting information content in subtle ways. Data semantics, for
example, (Veltman (1983), Landman (1986)) and situation semantics
(Barwise and Perry 1983) falI within this group. The second group of
proposals centers around the idea that at some level propositional
content must be represented in terms of sentence-like structures, as
Frege and Russell suggested. The structured meaning approach (a
variant of Carnap's intensional isomorphism approach), developed in
Cresswell (1985) and related work falIs in this second group.
In principle, approaches that take propositions as primitives (e.g.
Thomason 1980) could go either way. But taking propositions as
primitives might perhaps be more useful in connection with the first
task. For if the structure of propositions is as fine grained as the
structure of sentences, then it is hard to give to propositions any
content but in terms of something analogous to sentence-like structured
objects.
In the present paper, I will try to argue that whatever theory of
logical space one wants to adopt, something like structured meanings is
called for in semantics. However, 1 will not try to defend this view in
the context of a theory of mental attitudes. The arguments that 1 will
give are based on the status of thematic roles (i.e. traditional notions of
non model theoretic semantics such as agent, patient, etc.) and the
semantics of what linguists call control. 1 will try to show that a model
theoretic characterization of thematic roles and a simple and powerful
semantics for so called control structures can be obtained using
something like structured meanings. If this is S".o, structured meanings
would not be needed only to deal with propositional attitudes. They can
be motivated independently. A view of propositional content developed
131
andMeaning, II. 131-166.
in connection with the analysis of mental attitudes turns out to have an

empiric al coverage broader than expected.
In section 1, I will outline a version of CresswelI's approach that I
think has the advantage of being more compositional than the one he
develops. In section 2, I wilI sketch a theory of thematic roles based on
it. In section 3, I will articulate my semantics of control structures and
in section 4, I will extend such semantics to purpose clauses.
The development of an explicit formal theory of structured meanings
is a complex matter. Non-trivial domains that contain structured
meanings unavoidably end up having certain reflexive characteristics
that make the construction of such domains within standard set theory
very hard, if possible at alI. Theories of properties and propositions
suggest themselves as natural alternatives in this connection. However, I
will not attempt here a property-theoretic reconstruction of structured
meanings, a task which I believe can be successfulIy pursued in more
than one way. My concern is rather to provide some additional
semantic motivation for structured meanings. If property theories are
then required to make such an approach really work, we will also have
a further indirect argument in favour of a property-theoretic semantics.
1. STRUCTURED MEANINGS AND COMPOSITIONALITY
In the present section we are going to review very briefly the motivation
for the structured meaning approach that derives from a theory of
propositional attitudes and to sketch a version of such approach.
If we present the information content associated with a sentence 1/J
as the set of worlds in which 1/J is true, then any sentence ~ 10gicalIy
equivalent to 1/J will have the very same content (i.e. will express the
same proposition) as 1/J. But of course even if two structures are linked
to truth in the same set of circumstances, we may well fail to see that.
This is at the basis of the so-called "logical omniscience" problem, a
central one for a theory of propositional attitudes. Given that the
source of the problem lies in the interconnection of structures with
truth and falsity, it is tempting to deal with it by granting at some level
finer structure to propositions, perhaps even an articulation in con-
stituents, akin to the one of sentences.
Adopting this position does not auto~aticalIy entail discarding the
notion of propositional content as a set of possible worlds. There may
well be different levels of propositional content that need to be
STRUCTURED MEANINGS 133
recognized. What 1 mean by "level of propositional content" should

become clear as we go along.
Cresswell (1985) argues, in this vein, that the meaning of that-clauses
should be represented not only in terms of regions in a logical space,
but also as structures of a certain kind. Such structures are built up out
of the meaning of the constituents of a sentence and are isomorphic to
syntactic structures (perhaps isomorphic to the syntax of some linguis-
tically motivated logic al form). For example, clause (la) can be
interpreted as shown in (lb).
(1) a. that Mary hugged Bill
b. (hug', M, B).
To be a bit more explicit, let us as sume that hug' is the translation of
the verb to hug in an intensional logic. Hug' is interpreted as a 2-place
propositional function, where propositions are just sets of worlds. l Let
us assume, furthermore, that things like those in (lb), .e. expressions of
the form ( ... ) are terms of our logic. The interpretation of such terms is
the one transparently suggested by the notation. A term of the form
(al' ... , an) is interpreted as the ordered n-tuple (l aJ, ... , [an]).
By interpreting that-clauses in this way, we can analyze mental
attitudes as relations of agents to structured meanings. Let us call
structured propositional meanings "information units", and let us keep
the term "proposition" for sets of worlds. A sentence such as the one in
(2a) can be interpreted as in (2b) or as in (2c).
(2) a. John believes that Mary hugged BiU

b. believe'(J,hug'(M, B))
c. believe'(J, (hug', M, B))
d. John believes that Mary hugged Biti and that 54 is
complete
e. (1\, (hug', M, B), (complete', S4 )).
On the reading in (2b), the belief-relation is stid to hold between
John and a proposition. This representation is arguably adequate for
the sense in which belief is a disposition to act in certain ways
(discussed in e.g. Stalnaker 1984). On this reading, however, sentence
(2a) will entail the sentence in (2d), since the propositions associa~ed
with the embedded clauses are identical. There is clearlya sense of (2a)
in which this is wrong. One can claim that such sense is represented in
(2c). On the reading in (2c), (2a) will not entail (2d), since the
information unit associated with that Mary hugged Bill and that S5 is
complete is as shown in (2e), and thus different from the one associated
with the embedded clause in (2a).
For the sake of explicitness, we shall make the following assump-
tions. ' 1\ ' is interpreted as a function that maps pairs of sets of worlds
into their intersection. Other connectives are interpreted in a parallel
way (e.g., negation as a function mapping sets of world into their
complement, etc.). The universal quantifier is interpreted as a function
V from propositional functions into propositions such that if r is a
propositional function and w a world, w E V(r) iff for every u in the
domain of individuals, w E reu). We as sume that formulae of the form
n.
Vx1jJ are interpreted as V([ Ax1jJ We assume moreover that e is the
type of individuals, p the type of propositions and (a, b ) is the type of
functions from a into b.
Before moving on, let me recall some general features of the
structured meaning approach that arguably make it an interesting
solution to the problem of logical omniscience.
First, any non trivial theory of logical space wiU map distinct
sentences 1jJ and ~ into the same region, i.e. wiU precipitate some
notion of equivalence. But then the problem we want to solve will
reproduce itself, for however strong our notion of equivalence, we stiU
might not see that 1jJ and ~ are associated with the same region of the
logical space and hence we might have different attitudes towards them.
This seems to suggest that at some level of content sentence-structure
has to be preserved.
Second, our attitudes can be de re, even when substitution of
equivalents is unwarranted. The structured meaning approach has a
strikingly simple account for this. It, for example, the names in (2a) are
sufficiently vivid (or under whatever condition on de re belief we might
want to assume), our semantics wiU validate inferring from (2c) the
foIIowing:
(4) 3x[believe(l, (hug', x, B )]
where (4) has the usual interpretation, namely that some individual 11
makes theformula believe (l, (hug', x, B)~ true.
Third, we can make structured meanings as finely structured as we
like. For example, we might assign to the complement of (3d) the

following:
(5) ( /\, hug'(M, B), complete' (S4).
The information unit in (5) is much more coarse grained than the
one in (3e), as its constituents are sets of worlds. And one could easily
imagine other ways of structuring the meaning of the complement of a
believe clause, each of which can be potentially relevant to a particular
context.
A lot of hard questions arise in connection with this approach, but
we cannot address them here. There is however a problem that has to
do with compositionality and should perhaps be discussed.
How is it that readings in terms of information units come about?
Consider, for example, a standard bottom up interpretive procedure,
such as the one illustrated in (6).
(6)
S
S, hug'(M, B)
NP,M VP, x[hug'(x, B)]
~
V,hug' NP,B
I I
that Mary hugged Bill
By the time we reach the S node, we will have a set of worlds from
which we will not be able to recover the meanings of the constituents.
Cresswell suggests in this connection that the semantics of that-
clauses is perhaps bound to be uncompositional. To assign the right
structured meaning to the top node in (6), we have no other option but
to go down the tree till we are able to recover the meaning of the parts;
then we can recompute the meaning of the clause by bracketing the
meanings of its parts together. So a potentially promising approach to
the semantics of propositional attitudes seems to induce a substantive
complication in the syntax-semantics map.
However, a perhaps more compositional alternative that comes
to mind is resorting to type-shifting principles. Let us illustrate it
informally.
Let us imagine having a type-shifting operator of the following sort:
(7) a. hug'* = AXAy(hug', x, y)
b. Rn* = AX1 , , Axn(Rn, Xl' . .. , Xn).
By interpreting hug as in (7a), we would compositionally get the
right meaning at the right level. (7b) is the general definition of our
type-shifting operation, denoted as "*". * maps propositional functions
into functions from individuals into information units (which are also
individuals). We might imagine this operator as just being freely
available for the interpretation of verbs, along the lines discussed in,
e.g., Partee and Rooth (1983). Verbs can be freely shifted from
propositional functions to information units functions.
In fact, one can generalize such type shifting-operator to other
functors, such as in particular, the ones associated with connectives and
quantifiers. This is illustrated below:
(8) a."1 * = At("1, t), where t is of type e or type p.
b. /\ * = AtI At2( /\, tI, t2), where tI' t 2 are of type e or type
p.
c. "ih = Af(V, f), where fis either of type (e, p) or of type
(e, e).
The type shifting operator * is, thus, polymorphic. It applies to
entities of various types. In particular, (8c) creates an operator that
looks for either a propositional function or an information unit
function. This enables us to associate with sentences like (9a) the
information units in (9b, c).
(9) a. every man carne
b. (V, AX ( ..... , (man', x), ( come', x)
c. (V, AX [man'(x) ..... come'(x)).
The reading in (9c) is more coarsely grained than the one in (9b). To
illustrate our approach, let us show how (9b) comes about.
(10) S, (, (V, x( -+, (man', x), (come', x ), (l~ave, M
S,
(V, X( -+, (man', x), (come' x) conj, * S, (leave', M)
~
NP, VP, come'
/\
NP, M VP,leave'*
P ( V, x ( -+ , (man " x ), (P, x )))
~
everyman carne and Mary left
Now, there is an obvious correspondence between information units

and propositions. Structured meanings can be embedded into proposi-
tions. The embedding takes the form of a simple homomorphism
defined, in our logic, as fol1ows:
(11) a. if {3 is a proposition or a propositional function, E ({3) =
{3; if (3 is a function from entities of type a into
structured meanings, E({3) = Xa [E({3(x))]
b. It {3 is an atomic information unit (Le. if {3 = (Rn, X 1, ,
xn), E({3) = Rn(x1, ... ,xn)
c. if {3 = (a, t1, t2 ), E({3) = E(t1)aE(t2)
d. if {3 = (a, a), E({3) = aCEea)).
We will call atomic information units "eventualities". It (3 is an
information unit, E({3) is a proposition that can be read as 'it is the case
that (3'. The definition in (11) induces in an obvious sense a definition
of truth on information units. We say that an information unit (3 is true
in a world w iff w E [E({3)].
The view just sketched enables us to incorporate the structured
meaning approach in an extended categorial grammar that makes
allowance for type-shifting principles. By allo\ving free use of the
operator *, each sentence will be associated with multiple interpreta-
tions. So, in a sense, we are making the syntax semantics map multi-
valued. But multivalued is better that uncompositional. It is one way of
reconstructing formally the idea that there are severallevels of proposi-
tional structure.
There are several known formal problems lurking behind the
assumption that structured meanings are individuals. For example, our
embedding operator E resembles dangerously a truth-predicate. A
further problem, pointed out in Cresswell (1985, pp. 85ff), arises in
connection with iterated beliefs. Consider for example (12a).
(12) a. John believes that Mary believes that Bill is home

b. believe'(J, (believe', M, home'(B.
One of the readings that our type shifting principles associate with
(12a) is (12b), where a propositional function, believe' applies to
something that contains that very function. Under the standard set
theoretic understanding of the notion of function, this is impossible.
This can be taken as evidence that the standard set-theoretic
construal of the notion of propositional function is unsuited to our
goals. New approaches are called for. The challenge is to come up with
a notion of property (or propositional function) that enables us to stick
as closely as possible to the above simple view of semantics and syntax-
semantics map. 1 believe that several current theories of properties are
capable of meeting this challenge.
Notice, in particular, that in the picture sketched above we are
making crucial use of two ways of applying a function to an argument.
One is direct functional application, the other is via *. The * operator
that generates the structured meanings is strongly reminiscent of what
Aczel (1988) calls "structure building" predication. The two forms of
functional application we are adopting here appear to be closely related
to Aczel's notion of internal vs. external predication. The above
considerations point towards what looks like a specific (and rather
central) semantic problem to which Aczel's distinction might be directly
relevant.
A serious pursuit of these remarks would involve showing how to
articulate this proposal by using Frege structures or some othyr
property theory. But this is not what we set out to do here. 2 We shall
turn instead to a discussion of different areas of semantics where the
notion of structured meaning might play an important role.
2. A THEORY OF THEMATIC ROLES
Notions such as theme, agent, source, goal, etc. have played an

important part in linguistics, particularly since the work of Gruber
(1965) and Jackendoff (1972). However, while their usefulness as
descriptive labels is unquestionable, they have for a long time appeared
to be too vague to be analysed in model theoretic terms. 3 1 believe that
building on the notion of eventuality, as defined in the previous section,
a simple and useful characterization of the notion of thematic role (or
O-role) is within reach. Thematic roles may be viewed as partial
functions from eventualities (Le. atomic information units) into one of
their constituents. Let me give an illustration of this idea, by providing
some tentative definitions.
(1) For any eventuality {3:
a. Ag({3) = Xi' for that unique Xi E {3 whose action causes
{3 (or, if you prefer, the event classified by {3)
to occur. It there is no such Xi' Ag({3) is
undefined.
b. Th({3) = Xi' for that unique Xi E {3 such that whenever
{3 occurs, Xi moves, changes possession or is
acted upon by Ag({3). It there is no such Xi'
Th({3) is undefined.
c. Go({3) = Xi' for that unique Xi E {3 such that whenever
{3 occurs, Th({3) moves towards Xi. It there is
no such Xi' Go({3) is undefined.
d. So({3) = Xi'for that unique Xi E {3 such that whenever
{3 occurs Th({3) moves from Xi. It there is no
such Xi' So({3) is undefined.
Analogous definitions might be given for "Experiencer", "Benefi-
ciary", etc.
Thematic roles are therefore viewed not as primitives but as partial
functions defined in terms of the entailments associated with even-
tualities. 4 Given what eventualities are, the crucial defining entailments
turn out to be those associated with the argument slots of the con-
stituent-relation of an eventuality, much in lin~ with the first theory
described in Dowty (this volume).
The empirical claim is that the relations lexicalized in natural

language display recurrent patterns of entailments in their argument
structures. Such patterns of entailment can therefore be used to classify
relations, state generalizations about their argument structures and
characterize the types of events of which relations are part.
One of the clearest generalizations based on O-roles concerns
parametric variations in case assignment. For example:
(1) In non ergative languages the agent-argument is marked
nominative and the theme-argument accusative. In ergative
languages, the theme-argument is marked absolutive and the
agent argument ergative.
The statement in (1) is a very rough approximation. But any way of
refining it that 1 am familiar with makes crucial appeal to thematic
notions. In the next section, we shall consider other generalizations that
appear to be optimally statable in terms of thematic roles.
In principle, any cluster of entailments can define a legitimate
thematic role. 4 Notice for example that we are using a disjunctive
definition for the notion of "theme" that subsumes what is often called
"patient". Consequently, while the present approach does seem to
provide a way of making precise "traditional" O-roles, it is by no means
committed to them. Suppose for instance one needs to define a
thematic role that selects the argument of a relation that is responsible
for the outcome of the proceedings described by the relation (as
proposed in unpublished work by C. Jones and in Farkas (forth-
coming)). There would be no difficulty in accomodating such O-role.
Our theory provides us with a device in terms of which O-roles can be
made precise, but does not automatically determine the content of
specific O-roles.
What O-roles the grammar is actually sensitive to is a difficult
empirical question. Ideally, one would like to be able to isolate a small
set of O-roles in terms of which alI relations lexicalizable in natural
languages (and the eventualities that they determine) could be classified.
Jackendoff (1983) states a strong form of this hypothesis that he calls
the "thematic relations hypothesis". He argues that relations pertaining
to various semantic fields can be classified in terms of the O-roles
relevant to spatial expressions (such as source-goal-theme). This is so,
Jackendoff argues, because "the mind does not manufacture abstract
concepts out of thin air. It adapts machinery that is already available,
both in the development of the individual organism and in the evolu-

tionary development of the species" (ibid. pp. 188-189).
In fact, I think that we are stiU far from such a goal, which probably
in this strong form is unattainable (again, see Dowty (this volume) for
discussion). But this should not discourage us from pursuing O-roles in
semantics. Certain pattems of entailments (such as those associated
with the agent-patient or source-theme pairs) are fairly clear and
recurrent, at least in certain domains. Various grammatical processes
seems to be sensitive to them. Even if they do not extend to every
relation and every semantic field, they may stiU be useful in stating
subgeneralizations (such as the one pertaining to case marking men-
tioned above or those pertaining to control to be discussed in the next
section). And they might stiU be regarded as salient components of
our capacities to collect and classify eventualities. It seems that our
approach to thematic roles supports this view and enables us to pursue
it with the sharp formal tools of model- (and property-) theory.
Several thematic notions are at present often used in the literature
under conflicting definitions. The notion of "theme" as extended
beyond verbs of motion is a prime example. This is of course to be
avoided. But even if we try to sharpen our defining criteria by pinning
down exactly the relevant entailments along the lines proposed here, a
residue of vagueness wiU in general stiU remain. Our defining entail-
ments might adscribe to arguments properties that are intrinsically
vague. Vagueness in natural language is widespread and affects many
phenomena, from the meaning of lexical items to comparatives and
conditionals. While vagueness stiU is not completely understood, vari-
ous promising approaches to it do exist. I think this should warrant
some hope also with respect to the present construal of O-roles.
The formal theory of O-roles outlined here crucially depends on the
possibility to recover the "constituents" of propositional (or, perhaps,
event-theoretic) structures. It thus supports a view that at some level
grants such a structure to propositional content, in the way the
structured meaning approach does.
3. THE SEMANTICS OF CONTROL
In this section I wiU discuss the semantics of a large number of control

structures. Such a semantics wiU be cast in terms " of O-roles, and
hopefully iUustrate the fruitfulness of the present perspective.
In general descriptive terms, control is the syntactic and semantic

relation that holds between a Noun Phrase (NP) and an infinitive or a
gerund in sentences like those in (1 ).
(1) a. John; tries [; to win]
b John forces Mary; [; to win]
c. John; promised Mary [;to win]
d. John recommended [;reading War & Peace] to Mary;.
In alI of the examples in (1) the control relation is indicated by
coindexing: the infinitive (or gerund) is understood as being predicated
of the NP coindexed with it. Our goal is to see what it takes to spelI out
what this implicit predication amounts to semantically.
We assume that the kind of entities that can be predicated of
something are properties. 5 Thus, whatever the syntactic constituency of
infinitives might be, we take them to be semanticalIy properties. It
folIows, then, that the semantics for the constructions in (1) will be as in
(2).
(2) a. try' (j, leave ')
b. force' (j, m, leave ')
c. promise(j, m,leave')6
d. recommend ' (j, leave', m).
Since none of the arguments in (2a-d) are propositional structures,
our semantic intuitions that the embedded clauses in (1) have an
understood subject must be accounted for in some indirect way. A
natural hypothesis within this set of assumptions is that such intuitions
are simply grounded on the entailments associated with the relevant
constructions. For example, part of knowing what try means is to know
that if I try to P, then I try to bring about a situation where I have
property P.7
It the semantics of control rests on entailments of this sort, then the
pattern of such entailments across various relations might be relevant in
determining which argument of a certain relation is selected as the
controlIer. We are adopting the view that recurrent patterns of entail-
ments correspond to O-roles. So it is not unreasonable to expect that
O-roles, i.e. the way we classify eveiltualities, might be central in
determining the controller of a property-argument.
In fact, this was noted as early as Jackendoff (1972). For example,

by looking at the constructions in (1) we see that the theme argument
generally controls, if present. It no theme is available (as, arguably, in
(lc, d, then the agent- or goal-argument takes over.
Traditional thematic labels might not provide the best criteria in this
connection. Perhaps notions such as "responsible for the outcome" are
called for, as pointed out above. But 1 would like to see here how the
relation between O-roles (in my extended sense of the term) and
controller selection can be represented in general, rather than pursuing
how to exactly define the relevant O-roles, even though the latter is
clearly very important for the overall success of the enterprise. 8
3.1. Strict Obligatory Control

In order to be in condition of predicting the selection of the controller
in terms of the thematic structure of a given eventuality, we need a
simple definition. Consider the following examples.
(3) a. John promised Mary to leave
a'. <promise', 1, M,leave')
b. John promised to leave
b'. <DROP 2(promise'), 1, leave)
(4) a. J Ohn forced Mary to leave
a'. <force', J, M,leave')
b. Mary was forced to leave
b'. <PAS(force'), M,leave').
The expressions in (3a', b') and (4a', b') represent the eventualities
corresponding to (3a, b) and (4a, b), respectively. The idea is that (4b)
is related to (4a) by an operation that deletes the second argument of
promise, represented here as DROP 2 Similarly, the passive in (4b) is
related to the active by an operation that deletes the first argument of
force, i.e. the one that corresponds to the subject of the active form. 9
Intuitively, we would like to say that the eventualities in (3a') and (3b')
are eventualities of the same type, namely "promises". Similarly, we
would like to say that the eventualities in (4a') and (4b') are both of
the force-type. It is easy enough to give a definitlon that does exactly
that. E.g.:
(5) The type of an eventuality {3 is the relation r such that g(r)

is the first member of {3. g is either an argument manipu-
lating operation (such as PAS or DROP2) or the identity
map. We represent an eventuality of type ras {3r.
Finally, we shall use the notation {3 [a] to represent an eventuality
that has a as one of its constituents. We shall furthermore use {3[aly]
for the eventuality that is exactly like {3 except that y is uniformly
substituted for a (the parallel with variable assignments should be
obvious).
At this point, the semantics of control structures simply becomes the
following:
(6) a. E({3r[P)) E({3r[PIP*(O({3))))
-+
b. Th > Go > Ag ....

The notation in (6a) is compact but, I hope, not too hard to unpack.
It says: suppose that an r-eventuality that has property P as one of its
constituents is the case; then an r-eventuality obtained from {3 by
replacing P with the eventuality (P, x) (where x is the bearer of the
role O with respect to {3) is also the case. lO
"O" in (6a) is variable whose values can be determined in terms of
the hierarchy in (6b). In general, if the basic relation r has a theme-
argument, then Th will be the value of O. Otherwise, it will be the next
higher thematic role, and so ono This procedure for filling in values of
"O" is to be taken as a default procedure, which can be overriden, at a
cost.
The principle in (6a) is, thus, a schema that generates constraints on
meanings. In (7a-d) you can see a simple illustration of how it is
supposed to work.
(7) a. John tries to swim
b. Etry', 1, swim ')
c. E ({3try[P)) -+ E ({3try[P/P*(Ag({3))))
d. E try', 1, (swim', J).
The formula in (7b) represents the meaning of sentence (7a). (7c) is
the instance of the principle in (7a) relevant to try-eventualities,
generated by the procedure just described. Given (7c), (7b) entails (7d),
i.e. the information associated with (7b) carries with itself the informa-
tion that, roughly, John tries to bring about a situation where he swims.
This is, 1 propose, how intuitions about what controls what come about,
in the unmarked case. In marked cases, one might have to stipulate
what the controlling O-role is going to be.
1 think that the semantics in (6) can be improved substantially by
studying more in depth the entailment patterns that can be relevant to
control. But even in its present preliminary form, (6) shows us
something non trivial. Namely, it shows us that we can come close to
actually predicting how the controller in a control structure systemati-
cally depends on the way we classify eventualities.
The principle in (6) has further consequences that are worth pointing
out. First, it embodies a certain view of how "local" the control relation
is. It says that a property can be controlled only by a coargument. It
also says that a property can be controlled by any of its coarguments.
Consequently, one would expect, among other things, control into
subject position from some argument internal to the VP to be possible.
The following appear to be relevant cases:
(8) a. [i to leave] bothered John i
b. [icorrere] piace a Gianni i
to-run pleases to Gianni
Gianni likes running.
It is well-known that bother in English admits also a form of long
distance control (although piacere in Italian does not):
(9) a. John i told me that [imaking a fool of himself] bothered
Mary
b. * Gianni i mi ha detto che [iarrivare ubriaco] non e'
piaciuto a Maria
Gianni told me that to arrive drunk (mas.) did not
please to Maria.
This forces me to the conclusion that the bother in (8a) is different
from the one in (8b). 1 think that there is some independent evidence
that this indeed is so, but 1 will not go into it here (see Chierchia and
Jacobson (1986)).
A second consequence of the principle in (6) is that the controller
must be uniquely identified among the arguments of a relation. There
cannot be two distinct bearers of the same O-roles in an eventuality, for
otherwise O-roles would not be functions. This seems to be roughly

correct. Cf. the ungrammaticality of sentences like:
(10) * John forced Mary to make fools ofthemselves.
A third consequence of (6) is the overt obligatory presence of the
controller. If an eventuality {3 lacks role O among its constituents, O({3)
will be undefined and (6) cannot be met. It follows, by contraposition,
that if a relation satisfies (6), its corresponding eventualities must
always have the bearer of the designated O-role among their con-
stituents. This might be at the basis of the ungrammaticality of the
following structures:
(11) a. * it was tried to win
b. * John forced to win
c. * to leave bothers
d. * correre piace
to-run pleases.
To illustrate, the eventuality corresponding to (lla) would be the
following:
(12) (PAS(try'), win').
But (12) has no agent. Thus if try satisfies (6), PAS(try) must be
undefined, at least when try takes a property (Le. at least for the
property-taking try). Notice that it was tried to leave by John is also
ungrammatical. This follows if we assume that by-phrases are predicate
modifiers. The function argument structure of the VP tried by John
would be by(J) (PAS(try)). But PAS (try), as we have just seen, is
undefined, and hence by(J) (PAS(try)) will also be. A similar account
can be given for the other cases in (12). The fact in (11a) exemplifies a
general tendency, noted by Visser (roughly: subject control verbs lack
passive counterparts). The fact in (11b) also exemplify a tendency
which is very general in English, as noted in Bach (1979) (roughly:
object control verbs do not detransitivize).
Thus on the present approach, alI the facts in (11) receive a uniform
account: certain predicates do not tolerate that the control relation be
affected by argument manipulating operations. It is not clear that this
realIy identifies the right dimension of the problem. Many of the
predicates in question appear not to tolerate tinkering with their
argument structures quite independently of the presence of a control-

seeking constituent. Cf. far example the ungrammaticality of (13).
(13) *John persuaded that she should leave.
However, some other data pulIs in the opposite direction and does
suggest that the prohibition in question might be specificalIy triggered
by something related to control. For example, (lla) and (lld) have
grammatical counterpart, when no control is involved:
(14) a. everything was tried

b. le nuove collezioni sono piaciute molto
the new collections pleased a lot.
Cf. also the paradigm in (15).
(15) a. John taught us that the earth is flat

b. John taught us to speak Latin
c. John taught that the earth is flat
d. * John taught to speak Latin.
Again, here it is only when control is involved that detransitivization
yields ungrammaticality.
We might explain these facts as folIows. The lack of detransitiviza-
tion is always triggered in the way we are assuming, also for verbs like
persuade. Le. if a relation R takes properties, it satisfies (6). This yields
a clash with any argument manipulation O that would delete the
controlIer. Consequently, we assume that O is altogether undefined for
R. For example, we assume that, say, DROP2 (persuade') is generalIy
undefined, for alI the types of arguments that persuade can take.
However, more markedly, we can make O(R) undefined only for those
arguments that actualIy yield the violation of (6), namely properties. So,
for instance, DROP2 (teach') lacks a value for property-arguments (i.e.
DROP2 (teach') (x, P) is not a proposition) but has a value for
eventuality-arguments (i.e. DROP2 (teach') (x, (P, x is a proposition).
The learning strategy that this view supports is as folIows. When we
find an infinitival-taking verb R, we assume that (6) applies to it. From
it we infer that certain argument manipulations f (namely those clashing
with (6 must be undefined for R. Encountering'data such as (14) or
(15a) will farce us to revise our assumption for the relevant verb and
make us adopt the weaker hypothesis that feR) is only undefined for
property-arguments, i.e. those that actually trigger the clash between
feR) and (6).
A further problem should be addressed, before moving on to a
consideration of other structures. The domain condition built into (6),
namely coargumenthood, is fundamentally similar to the one proposed
in Williams (1987). Recent work (such as Belletti and Rizzi (1986) and
Pesetzky (1987 has revived the idea that in the constructions in (7)
the controllers are superior to the embedded clause, at some level. If
that line turns out to be right, then the domain condition on strict
obligatory control should presumably be tighter than the one built into
(6). The line 1 am pursuing does not exclude that stricter structural
conditions (such as immediate C-command at some level of structure)
might be operative in certain cases, thereby obviating for such cases the
need for a direct appeal to thematic roles. This would perhaps limit the
scope of my proposal. The evidence currently available, however,
suggests that purely structural domain conditions will not work in
general. A large class of control structures, besides those considered in
this section, seems to call for a principle that appeals directly to O-roles.
The following sections are devoted to a discussion of such cases.
3.2. Control by Implicit Arguments
There are verbs that clearly admit implicit controllers. For such verbs
(6) is too strong and must be slightly modified.
Consider, for example:
(1) a. it was decided to leave (by John)
b. Mary helped (Bill) do the dishes
c. to leave was rude (of Bill).
(la) contrasts with (IOa) in the preceding section, (lb) with (lOb)
and (lc) with (lOc, d). Constructions like (lb) are very rare in English,
but quite common in Italian:
(2) a. AlIa manifestazione di ieri, le brutalita' della polizia e'
stata tale che ha costretto a reagire con decisione. '
In yesterday's manifestation, police brutality was such
that it forced to react with decision.
(2) b. L'atteggiamento di Giolitti in quell'occasione lascio'

davvero perplessi. 11
Giolitti's attitude on that occasion let really perplexed.
It tums out that a minimal change in our semantics is sufficient to
accommodate the constructions in (1) and (2). AlI we need is the
following:
(3) E({3r[PJ) /\ O(jJ) = xi -+ E({3r [PIP*(O ({3)))).
The only difference between (3) and the previous version our
semantics for control is in the underlined condition. Formula (3) says:
suppose that an r-eventuality (with P as an argument) is the case and
suppose furthermore that {3 has an argument which is a O; then that
argument will control P.
So predicates that select (3) will be expected to have the same
properties as the predicates that select our previous version of the
control principle. In particular, the same domain condition holds and
the fact that the controller is uniquely identified as the bearer of a
particular O-role also holds. However, nothing stands in the way of
applying to the predicates that satisfy (3) operations that result In
elimination of the controller from their argument structure.
To illustrate, consider the sentence in (4a) and its meaning in (4b).
(4) a. it was decided to leave
b. E(PAS(decide'),leave')
c. E ({3decide[PJ) /\ Ag({3) = Xi -+ E ({3decide[PIP*(Ag({3)))).
Formula (4b) lacks an agent (the designated controller) and thus the
relevant instance of (3) (given in (4c)) does not apply to it. However,
any analysis of passive will have as a consequence that whenever
something is decided, it is decided by someone. I.e. the semantic
definition of PAS will validate the conditional in (5a), or, equivalent1y
in (5b).
(5) a. E(PAS(decide'), leave') -+ 3y[E(decide', y, leave'))
b. PAS(decide') (leave ' ) -+ 3y[decide'(y,leave')).
Now (4c) will apply to the consequent in (5a), for the relevant
argument (the agent) is present there, and it will trigger the desired
entailment. So, for this kind of predicates, our intuitions about control
are obtained by chaining together the entailments licensed by argument

manipulating operations (such as PAS or DROP 2) with the entailments
licensed by (3). This is, 1 submit, the semantic content of the notion of
"control by an implicit argument".
The minimal difference between the two versions of our semantic for
control enables us to actually consider (6) in the previous section and
(3) above as a unique principle. The underlined part of (3) can be
regarded as a parameter that classes of predicates in various langauges
are free to set. The version of (3) where the underlined conjunct is
absent is the most restrictive one (Le. it rules out more structures).
Hence it should be considered the unmarked form for acquisition
purposes (ef. Berwick's (1982) subset principle).
It is now time to say something about how the present view of
control relates to syntax. At the very least, the syntactic si de of the
control relation must account for the pattern of agreement one finds in
control sructures, e.g. it must account for the ungrammaticality of, say
(7) a. * John did not persuade Mary to wash himself
b. * John did not promise Mary to wash herself.
It is in principle conceivable that the deviance of (7a) is grounded on
a semantic fact, namely the (indirect) attribution to Mary of contra-
dictory properties. However, semantic treatments of gender become
extremely artificial (if viable at alI) for languages with grammatical
gender. In German, for example, the word for "girl" is neuter. In Italian,
the word for "egg" is masculine in the singular and feminine in the
plural, and so ono
Now, if syntax should take care of these phenomena, a minimal way
of dealing with them is by ensuring that controllers and controllees are
syntactically coindexed at some level, where coindexing entails same-
ness of feature specification. This leads to the following condition:
(9) NP () must be coindexed with Pred, if they are coarguments.
We want, of course, the relevant NP to be the argument that bears
the right O-role in the relevant argument complex. One way of ensuring
this is in terms of the following principle of correspondence:
(10) Two items are syntactically related iff they are semantically
related.
A and Bare "syntactically related" iff they are coindexed by (9).

They are semantically related iff their logical translations are linked by
the control principle (i.e. (3.
Condition (9) is essentially a variant of what Williams (1987)
proposesP It is interesting to note that Williams has arrived to a
formulation of control in terms of thematic domains in a totally
independent way. 1 think that the semantics developed here can be seen
as a way of determining the truth-conditional content (i.e. the model-
theoretic interpretation) of a Williams-style approach to control, via the
simple principle of correspondence given in (10).
While much more needs to be said in this connection, these
considerations suggest that a principled map between the syntax and
semantics of control is within out reach (and can be implemented in
various ways).
To sum up, we have developed a parametrized principle of control
based on O-roles that accomodates strict obligatory control and control
by implicit arguments. Many empiric al properties of these constructions
seem to follow from our semantics. However, such a semantics cannot
stand alone. An autonomous syntactic dimension must be granted to
the control relation and dealt with. We have tried to indicate how. To
test the soundness of the line we are taking, we now turn to a
consideration of a further, rather recalcitrant set of control structures.
4. PURPOSE CLAUSES
Purpose clauses (pC's, henceforth) are infinitival structures such as

those occurring in (1 ).13
(1) a. Mary invited Joan [_ to talk to the dean]
b. The manuscript is available [for the referees to go
over _]
c. Mary bought those books [_ to read_ during the brake].
Following Jones (1985), 1 will call PC's such as those in (la), where
the gap is in subject position "Subject Purpose Clauses" (SPC) and use
the name of "Object Purpose Clauses" (OPC) for the constructions in
(lb, c), which have a gap in non subject position. These are meant as
purely descriptive labels, with no theoretical import. OPC's can have
one or two gaps. If they have two, one of them occurs in subject
position. I will have very little to say about the syntactic constituency of
PC's and the syntactic nature of their gaps. I will just assume that PC's
are attached within the VP. This much seems to be fairly clear, while it
is not so clear whether they are VP-complements or VP-adjuncts.
My main concern will be the semantics of PC's as it relates to the
general issue of control, about which I believe that something can be
said, while remaining neutral on several details of their syntax.
The main descriptive generalizations conceming control in PC's
appear to be the following. The gap in SPC's and the non subject gap in
OPC's have to be controlled by an overtly present argument of the
matrix verb, as paradigms such as the one in (2) show:
(2) a. I am now hiring
b. I am hiring him
c. I hired him [_ to go over the reports] (SPC)
d. * I hired [_ to go over the reports]
d. I hired him [for John to work with_] (OPC)
f. * I hired [for John to work with_].
As (2a, b) illustrate, hire has a detransitive variant. However, when it
is construed with a purpose clause, the object of hire has to be present.
This suggests that SPC and the object gap in OPC need an overtly
realized controller.
However, by comparing the data in (1) and (2), we see that the
controller can be either the subject or the object. Considerations such
as these, led Bach (1982) to a dual classification of PC's. In cases like
(lb), Bach takes PC's to be arguments of the predicate (i.e., the
adjective available). In the rest of the cases, Bach argues that they are
Transitive Verb Phrase-modifiers (i.e. functions from TVP-meanings
into TVP-meanings).14 This leads to logical forms ofthe following kind:
(3) a. This is available to read

a'. available (x, [to read]')
b. He bought that to read
b '. [[to read] , (buy')] (x, y).
Control of the non subject gap in (3a') should reduce to ordinary
cases of control (such as, say, try). Control in (3b') should follow from
the object-oriented character of TVP-modifiers, which arguably can be
motivated on independent grounds in relation with "passive sensitive"
adverbials (ef. Keenan and Faltz 197 8).
What about the control properties of the second gap in OPC, i.e.
cases like (le)? In this connection, Bach (building on previous work by
Jackendoff) argues that we must resort to the notion of "resultant
state". Consider the pair in (4).
(4) a. Johnj bought iti [_j to read_;]
b. John gave iti to Maryj [_j to read _;].
If John buys x, the state which results is one in which he owns x.
This means that he will be in control of the object which is to be read.
But if John gives something to Mary, Mary has x as a result, and she
will be in control of it. It seems, therefore, that the argument which
ends up "having" the object is construed as the understood subject of
OPC's. Notice however, that such an argument doesn't have to be
obligatorily present:
(5) This book i is available [_ to read_;].
Using O-roles, one could describe this pattern by saying that the
subject gap in OPC's is controlled by the goal-argument of the matrix
verb, if such argument is present.
So we seem to have the following situation. The obligatory gap of
PC's (i.e. the subject gap in SPC's and the non subject one in OPC's)
needs an overtly realized controller and is controlled by the object, or,
for intransitive verbs, by the subject. The non obligatory gap in OPC's
(i.e. the subject one) is controlled by the goal argument, if present.
In fact, however, I don't think that the above generalization is quite
ac curate as it stands. Consider the paradigm in (6).
(6) a. I gave that i to Johnj [_j to read_;]
b. I gave Johnj that i [_j to read -d
c. Johnj was given that i [_j to read_;].
We see from (6b) that in double object constructions it is not the
direct object (i.e. the one that passivizes, ef. (6c that controls the non
subject gap of the Pc. The indirect object does: This seems to be a
difficulty for Bach's view that PC's are TVP-modifiers. Here we would
have to regard them as "Ditransitive Verb"-modifiers. Moreover, by

contrasting (6a) and (6b) we see that shifts in grammatical relations do
not affect what the controller is going to be: in both (6a) and (6b) it is
the entity that changes possession that is understood as the object of
the Pc. These considerations suggest that a better description of the
facts could be along the lines of (7).
(7) a. The obligatory gap of PC's is controlled by the matrix
theme
b. The non obligatory gap of PC's is controlled by the
matrix goal, if there is one.
So, the control properties of purpose clauses seem to be thematically
governed. 1S Thematic roles seem to provide us with the theoretical
vocabulary that most simply enables us to state the relevant generaliza-
tions. This makes sense in view of the fact that the control properties of
these constructions seem to depend on the way in which we classify
actions and situations, which is just what O-roles purport to do. Our
general approach to control developed in section 3 is designed to deal
with thematic governance and we would expect it, therefore, to extend
fairly naturally to control in PC's.
The ingredient that we need is the theory of adverbial modification
developed in McConnell-Ginet (1982). McConnell-Ginet argues that
the passive sensitivity phenomena that TVP-modification was designed
to capture receive a more general account by taking adverbials to be, in
some sense, arguments of the verbs they modify. Consider the contrast
in (8).
(8) a. Lisa rudely departed

b. Lisa departed rudely.
Sentence (8a) has a reading roughly paraphrasable as "Lisa acted
rudely in departing". This says nothing about the manner of Lisa's
departure. In contrast, (8b) says specifically of the manner of her
departure that it was rude.
The interesting problem here is what to pin this ambiguity ono
Clearly, appealing to TVP-modification will not help, for there is no
transitive verb there to be modified. Moreover, rudely lacks the
properties that characterize S-modifiers: So a shift in perspective seems
to be called for. Somehow, adverbs must be able to be arguments of
verbs. We could then say with respect to the facts in (8) that in one case
rudely is an argument of the verb, in the other a predicate modifier
systematically related to the former.
McConnell-Ginet suggests that this be cast in a general view of how
"optional arguments" function. Her idea is roughly the following. While
properties and relations have a basic adicity (i.e. are typed as to the
number of arguments they can take), they also have what she calls
"natural extensions" or "augmentations". Consider, for example, the
pairs in (9).
(9) a. John spoke
a'. John spoke to someone
b. John bought it
b'. J ohn bought it from someone.
We can regard speak as a one place property that has a 2-place
natural extension and buy as a 2-place relation that has a 3-place
natural extension. Notice that the primed variants entail the non primed
variants in (9), but the converse does not hold in general, for (9a) does
not entail (9a '). Thus while by taking the J-place buy as basic, we could
define the 2-place one, we cannot always follow this strategy. The
"natural extension" strategy can fill in where the former fails.
By applying this line to adverbs, we can say that for example depart
in (8) has a natural augmentation, which is a 2-place relation between
agents and manners. The logical form of (8b) would thus become
something like the following:
(10) depart'(L, e) /\ rude'(e).
McConnell-Ginet calls these kinds of adverbials "Ad-verbs". Their
role is to select the relevant natural augmentation of the verb they
modify and to simultaneously supply the argument. Ad-verbs have
predictable predicate-modifiers counterparts, generated by the follow-
ing schema:
(11) For every ad-verb , * is a function from properties into
properties, such that for any property P and any x,
[*(P)] (x) is the case iff x acts -Iy in doing P.
The predicate-modifier counterpart of rudely ls what is involved in

(8a). This gives a principled account of the contrast in (8).
For a detailed discussion of the consequences of this view (and, in

particular, for how it extends to passive-sensitivity phenomena) we will
have to refer the reader to McConnell-Ginet's paper.
From the perspective of McConnell-Ginet's theory of verb-modifica-
tion, it is very natural to regard purpose clauses as arguments of verbs,
in the same extended sense in which adverbials in general can be
arguments of verbs. Some verbs can have natural augmentations that
contain purposes. This solves the problem with regarding them as TVP-
modifiers pointed out in connection with (6b) above. Furthermore, it
alIows for a unified treatment of alI cases of PC-modification: they can
always be analyzed as arguments in this extended sense. FinalIy, it
accounts nicely for the fact that PC's can only modify verbs of a certain
kind. Compare:
(12) a. Mary built that board for the children to play with
b. * Mary destroyed that board for the children to play with.
Destroy, unlike build, simply lacks a natural augmentation that can
take purposes. 16 This does not entail, of course, that destroying cannot
be a purposeful activity, in some more general sense. We often act as if
it is. The term "purpose" is used here in a technical way for the what
PC's denote (although I do believe that the semantic analysis of PC's
may constitute a lead in understanding what purposes are).
What do purpose clauses denote? I am not sure, but I think we can
set forth an hypothesis conceming their logical type. In the same
general spirit of section (3), we can regard them as open, unsaturated
structures. We are modelling unsaturated structures as properties and
relations. Consequent1y, we as sume that PC's with one gap are 1-place
properties, while PC's with two gaps are 2-place relations. Recall that
we are not taking a stand on the syntax of PC's and consequent1y on
how the gaps in question come about. It is pretty clear, however, that
non subject gaps in PC's must be licensed by some kind of operator-
binding mechanism. For example, Chomsky (1981) argues for a
structure along the folIowing lines.
(13) 0i[PRO to read ti to the children).
On the view adopted here, the operator Oi would have to be
interpreted as an abstractor (just like PRO in general would have to
be).I.e., we interpret (13) as
(14) AX),Xj[Xj to read Xi to the children).
Under these assumptions the logical forms associated with, say, the
examples in (1) above become the folIowing:
(15) a. invite;o(M, J, .h[x talk to the dean'))
a'. (invite;o, M, J, AX [x talk to the dean '])
b. available;o(the manuscript', AX [for the referees to go
over 'x))
x
b'. (available;o' the manuscript', Ax[for the referees to go
over'
c. buY;o(M, those books', AXAY[Y to read x during the
brake))
c'. (buy;o' M, those books', AXAY[Y to read x during the
brake.
The primed version of the pair of formulae in (15) specify the
eventuality corresponding to the non primed versions. Furthermore, for
any relation R, we indicate its PC-taking natural augmentation as Rto.
At this point it should become clear that control in PC's falIs rather
squarely within the domain of our principle of control, reported here as
(16).
(16) E({3r[P)) 1\ 0(/3) = Xi ~ E({3r[PIP*(O({3)))).
Let us pursue this. Consider first SPC's. Principle (16) applies to
them in its unmarked form, i.e.:
(17) E({3to[P)) ~ E({3to[PIP*(Th({3)))).
The controller is taken to be the highest element in the O-hierarchy,
i.e. the theme. This has the effect that, for example, (18b), the semantic
representatIon of (18a), entails (18c) and it constitutes alI we need to
interpret PC's with one gap.
(18) a. Mary invited John to talk to the dean
b. E(invite;o, M, J, AX[X talk to the dean'])
c. E(invite;o> M, J, (AX[X talk to the dean'], J).
In the case of purpose clauses with two gaps, the non subject gap
works in the same way. The relevant instance Qf (17) is given in (19)
along with an example.
(19) a. E({3to[R)) ~ E({3to[RIR*(Th({3))])
(19) b. John gave War and Peace to Mary [_ to read_]

c. E(give to , J, W &P, M, AXAY[Y reads x]
d. E(giveto , J, W &P, M, AZ(AXAY[Y reads x], z, W &P)
e. E(give to , J, W &P, M, Az(read', z, W &P).
The meaning of (19b) is (19c). Via (19a), we infer (19d) (which is
equivalent to the more simple (1ge. Thus if John brings War and
Peace to read, he brings it so that it can be read. The formula in (19d)
expresses a relation between John, Mary and an eventuality function.
Now a problem seems to arise, which has to do with what we take to
be the range of the variable P in (17). It we understand P as being
restricted to propositional functions (as opposed to eventuality-func-
tions), then (17) would not apply to it, thereby leaving the subject of the
ope without controHer. It, on the other hand, we let P in (17) range
over both propositional and eventuality-functions, then by going over
again the process just illustrated, we would obtain the following:
(20) E(give to ' J, W &P, M, (read, W &P, W &P.
But this is wrong. Intuitively, the theme argument has been already
"used up". We want to apply to (1ge) an instance of our control
principle that uses the next highest O-role.
A solution can be achieved by generalizing further our control
principle (Le. (16. The ba sic idea is to let the variable P in it range
over any kind of function (i.e. any n-place propositional or eventuality-
function), while simultaneously adding a requirement that, intuitively,
the same O-role cannot be used twice.
It the reader can bear through a little more idiosyncratic formalism,
here is how one might reformulate our control principle. The notion
that we need is that of eventualities with some (possibly aH) of their
participants in common. In our formalism, this notion can be defined as
in (21a).
(21) a. {3=OY=ctfO({3)=O(y)
b. a =0 Y =ctf \fx[a(x) =0 y].
'=0' is an equivalence relation that can be read as "has the same O
as". It can be put to various uses and extended in obvious ways. For
example, we can extend it to eventuality-functions. Thus in particular, if
a is, say, al-place eventuality-function, we can extend '=0' to it as
shown in (21b). Intuitively, (21b) says that the eventualities that result
from a have the same O and y. The latter is what we need to state the
necessary constraint on (16). But it might facilitate our grasp of such a
reformulation to take a look first at what we want to get out of it in the
specific case at hand. Essentially, we want to replace (17) with
something like (22).
(22) E (f3to[P]) 1\ P* =f Th 13 --+ E (f3to[PIP*(Th(f3]).
This says the following. Suppose a to-eventuality occurs with P as its
purpose. Suppose, furthermore, that eventuality that results from
applying P* to some x has a theme different from the theme of 13. Then
the controller of P is the theme of [3.
The disjointness requirement in (22) can be seen as a general
condition goveming our control principle, which could be restated
accordingly as follows:
(23) a. E(f3r[a)) (1\ 0(13) = x) --+ E(f3[ala*(O([3])
b. where a is any function and O is the highest O-role on
the scale in c such that O(r, Xl' . . . , xn is defined and
a=fof3
c. Th > Go > Ag > ...
d. the parenthesized part may be selected by specific types
of eventualities.
The formulation in (23) incorporates a formalization of the notion of
"highest available O-roIe".
This modification of our original principle may appear ad hoc. 1
think, however, that the restriction it introduces can be viewed not as
an independent stipulation built into (23) but rather as a consequence
of the syntax-semantics interface. Recall that controller and controllee
must be coindexed in the syntax and satisfy what we have labelled the
correspondence principle (ef. the end of sec. (3. It follows then that if
the semantics of control were allowed to link one and the same
argument with two slots in a relation, the corresponding syntactic
structure would have to look something like:
(24) John brings War and Peace; [_; to read _;] (or O;[PRO;
read t;]).
But any reasonable condition on binding will rule (24) out. It follows
that arguments cannot be used twice by our control principle, if such a

principle is to be non-vacuous. Whence the restriction in (23) ensues.
Our modified control principle has as a consequence that the
controller of the subject gap in OPC's cannot be the theme, and is going
to be the next available O-role. Let us illustrate. Consider again (19b)
and its meaning, reported here:
(25) a. John gave Mary War and Peace [_ to read_]

b. Egive;o"!, M, W &P, read').
(23) (in its unmarked form) generates a constraint on meaning that
applies to (25b) using the highest available O-role, namely theme, and
yields:
(26) Egiveta> J, W &P, M, Ax(read', x, W &P).

Now, (23) generates an instance that applies to (26), but it has to
select the next available O-role, namely goal. We thus obtain, eventually:
(27) Egiveto , J, W &P, M, (read', M, W &P).

The entailment from (25b) to (27) does justice to our intuitions as to
what the understood subject of these structures is. Notice that principle
(23) must apply to (26) in its marked form, as the controller in these
cases is optional. Thus, the instances of (23) relevant to purpose clauses
become:
(28) a. E (j3to[ a]) -+ E (I3to[ al a*(Th(j3))])

a any function such that a "# th 13
b. E (j3to [a ]) /\ Go(j3) = x -+ E(j3[ al a*(x)])
a any function such that a =th 13.
The constraints in (28) are instances of (23). The only thing that
needs to be determined, as in general is the case, is whether the highest
available O-role has to be overtly realized in the relevant argument
structure or not (i.e. how the parameter in (23) is to be set).
That (28a, b) follow from (23) in a principled manner can be empha-
sized, perhaps, with the help of the following acquisition metaphor.
Imagine a child trying to figure out the pattern of control in purpose
"-
clauses. We assume that her general strategy is the following. When
confronted with unsaturated structures in argument position, she tries
to saturate them by means of the control principle (23) .. The control
properties of SPC's follow squarely from the unmarked instantiation of

(23) and so do the control properties of the non subject gap of OPC's.
The output of the application of (23) to the non subject gap in OPC's
can still yield an unsaturated structure (Le. a property) in argument
position, and so the child tries to saturate it as well in the same way.
However, theme is no longer available. Interaction of control and
binding, prevents using twice the same thematic role. So, the child
switches to the next O-role in the hierarchy. This yields the goal-
oriented character of control into the subjects of OPC's. Finally, from
positive evidence (i.e. upon hearing sentences like this book is to read),
she learns that the form of the control principle relevant to the latter
cases is the weaker one (Le. the one that allows to leave the controller
implicit). A complex array of facts is thus essentially deduceable from
one conceptually simple principle.
What does not follow from this account is which of the two gaps in
OPC's calls for the strongest form of control. As far as our theory goes,
things could turn out to be the opposite of what they are (Le. with the
subject gap of OPC's, rather than the object one calling for the overt
presence of the controller). This is probably wrong and 1 believe that an
explanation for it should be sought in the syntactic differences between
the two gaps. As pointed out above, on anyone's theory the non subject
gap must be licensed by some kind of variable-binding device, while the
subject has a different status. Presumably, the conditions on the
antecedents of variable binding devices are stricter than conditions on
the subjects of infinitives. E.g., in terms of the syntactic structure in (13)
above, the conditions on 'O' are stricter than those on PRO.18
To summarize, at worst we have a precise and simple characteriza-
tion of the truth-conditions of purpose clauses (provided by (28a, b.
At best, we have an explanation of their control properties that derives
them from an independently motivated theory of adverbial modification
and an extremely general principle of control, in interaction with other
aspects of grammar.
There is a further consequence of the present theory that it is worth
pointing out, as it hinges on the claim that purpose clauses are
semantically unsaturated structures.
As it turns out, PC's can be wh-moved. Le. we can form what-for
questions such as the following:
(29) a. What did John bring that wine for?
b. to enjoy with dinner.
Moreover, PC's can also be questioned "across the board", e.g.:

(30) a. what [did John buy this wine_l and [Bill take cake
for -l?
b. to enjoy with dinner.
(30b) is clearly a well-formed answer to (30a). Let us ask ourself
what its semantic value could be. Suppose that PC's denote saturated
structures (i.e. propositions). Then the propositional structure that
constitutes Bill's goal is Bill's enjoyment of the wine. The one that
constitutes John's goal is John's enjoyment of the cake. They are not the
same structure. Which of the two possible propositions would (30b)
denote?
On the present theory the answer is: none. (30b) is an unsaturated
structure, something that yields a proposition only if we supply two
arguments. On the propositional view of the meaning of PC's, the
meaning of the answer in (30b) appear to be a mystery.
There is a clear sense, 1 believe, in which Bill's and John's purpose in
(30) is the same one. Our theory explains how that can be, while
allowing for a straightfoward view of what the meaning of answers such
as (30b) can be. This seems to constitute further evidence in favour of
the line we are taking.
5. CONCLUSIONS
Our starting point was the elaboration of a notion of structured

meaning, building on current theories of propositional attitudes. This
notion has been put to a variety of semantic uses.
For one thing we have tried to provide a precise characterization of
the nature of O-roles (agent, theme, etc.). We observe in the argument
structure of predicates recurrent patterns of entailments. The existence
of such regularities indicates the presence of general criteria at the basis
of the way in which predicates are conceptualized. Our suggestion is
that these criteria, the O-roles, can be viewed as higher order functions
that select arguments on the basis of the entailments associated with
them. The proper notion of "argument structure" by means of which w
can link O-roles to entailments appears to be naturally definable in
terms of structured meanings. "
Our second major goal has been the investigation of a large class of
control struc1ures. In particular, we have focused on strict obIigatory

control, control by implicit arguments and control in purpose clauses.
We have developed one very general principle of control. Such a
principIe is a schema that licenses semantic constraints (akin to
"elucidations of meaning" or "meaning postulates"). The proposed
schema, which uses the notion of "highest available O-roIe", has been
seen to have a broad empirical coverage and seems capable of
accounting for several key properties of the relevant constructions.
Various of our definitions have made crucial use of the notion of
argument complex as a structured object.
Many aspects of my proposal are tentative and provisional. And
many interesting alternatives on the (formal) semantics of control are
currently been pursued. However, as far as I can see, none of them
manages to cover as much empirical ground in terms of a uniform
mechanism.
The leading idea in the present attempt is to systematically Iink the
way controllers are selected to the way eventualities are classified and
seek in this link reasons for some of the choices that grammar makes.
This idea is not new and has been seriously pursued within other
frameworks (e.g. Jackendoff's). The foundational shift that is taking
place in truth-conditional semantics, i.e. the powerful theories of
intensionality that are being investigated might shed new light on this
enterprise. Or, at least it is worth trying.
NOTES
* Since I started thinking about these issues, I have been helped by more people than I
can possibly acknowledge. The present work has been most directly influenced by D.
Dowty, P. Jacobson, R. Larson and E. Williams. For the many remaining problems,
mea culpa.
1 Here and throughout this paper, tense will be disregarded for simplicity.
2 See Chierchia and Turner (1987) for discussion of an explicit property-theoretic
grammar of a fragment of English.
3 See Dowty (this volume) for extensive discussion of the issues involved.
4 In fact, it should be possible to extend the notion of presupposition to information
units. In this way one might be able to use also presuppositions in defining thematic
roles.
5 See Dowty (1985), Chierchia (1984, 1987) for some arguments.
6 Bach (1979), following previous proposals of Thomason's and Partee's, argues that
the function-argument structure of promise-type verbs actually is promise'(j, leave', m).

1 am inclined to believe that Bach's proposal is correct, but as far as our present
concerns go we can remain neutral on this point.
7 In Chierchia (1987), it is argued that relations of individuals to properties are indeed
the way in which natural language expresses attitudes de se, in the sense of, e.g., Lewis
(1979).
8 Relevant discussion can also be found in Ladusaw and Dowty (1987). 1 think that the
theory developed here is compatible with the points they make, even though 1 would
disagree with their use of the term "non-grammatical" for thematic roles.
9 See e.g. Bach (1980) and Dowty (1982) for relevant discussion of the semantics of
the operations involved.
10 Thanks to Wyn Chao for pointing out an inadequacy with a preceding formulation
of this principle.
II Rizzi (1986) makes a strong point to the effect that null objects in Italian are
instances of pro. He also argues, however, that pro in object position must cooccur with
generic tense and that its semantic value cannot be pragmatically (i.e. contextually)
specified. Constructions such as those in (2) cooccur with non generic tense and their
value is clearly determined by the context. Thus by Rizzi's own criteria we should not
be in presence here of a pro-drop phenomenon in his sense.
12 An approach along these lines accounts also for binding of reflexives and non
reflexive pronouns in control clauses. One attempt to speli out a Williams-style
approach to the syntax of predication and control (within a categorial setting) can be
found in Chierchia (1985).
13 Extensive discussion of purposive and related constructions can be found in Faraci
(1974) and Jones (1985). The (model theoretic) semantic of purpose clauses is studied
in Bach (1982). We rely heavily on such works, from which most of the examples in the
text are drawn.
14 Bach actually argues that certain transitive verbs, such as choose, take PC's as
arguments.
15 Jones (1985) reaches these conclusions, on the basis of arguments more detailed
than those presented here.
16 Bach and Faraci describe the general characteristics that a verb must have to be
compatible with PC-modification. It should also be noted that the view that PC are
arguments in this special sense is not incompatible, as far as 1 can see, with the view
that they are syntactically adjuncts.
17 For simplicity, let us assume that if P is an eventuality-function, P* = P. Le. *
maps propositional functions into eventuality functions and eventuality functions into
themselves.
18 See, e.g. Aoun and Clark (1986) for a relevant proposal.
REFERENCES
Aczel, P.: 1988, 'A1gebraic Semantics for Intensional Logic, I' volume 1 of the present
collection. "
Aoun, J. and Clark, R.: 1986, 'On Non-overt Operators', ms., USC and UCLA.
Bach, E.: 1979, 'Control in Montague Grammar', Linguistic lnquiry 10, pp. 515-531.
Bach, E.: 1980, 'In Defense of Passive', Linguistics and Philosophy 3, pp. 297-341.
Bach, E.: 1982, 'Purpose Clauses and Control', in P. Jacobson and G. Pullum (eds.),
The Nature of Syntactic Representation, Reidel, Dordrecht, pp. 37-57.
Barwise, J. and Perry, J.: 1983, Situation and Attitudes, MIT Press.
Belletti, A. and Rizzi, L.: 1986, 'Psych-verbs and O-theory', ms., Pisa and Geneva.
Berwick, R: 1982, 'Loca1ity Principles and the Acquisition of Syntactic Knowledge',
Ph.D. Diss., MIT, Cambridge.
Chierchia, G.: 1984, Topics in the Syntax and Semantics of Injinitives and Gerunds,
Ph.D. Diss., UMass, Amherst.
Chierchia, G.: 1985, 'Aspects of a Categorial Theory of Binding', to be published in E.
Bach, D. Wheeler, and R Oehrle (eds.), Categorial Grammar and Natural Language
Structures, Reidel, Dordrecht.
Chierchia, G.: 1987, 'Anaphora and Attitudes De Se', ms., Cornell U.
Chierchia, G. and Jacobson, P.: 1986, 'Local and Long Distance Control', in S. Berman,
J. Choe, and J. McDonough (eds.), Proceedings of NELS 16, GLSA, Amherst, pp.
57-74.
Chierchia, G. and Turner, R: 1987, 'Semantics and Property Theory', forthcoming in
Linguistics and Philosophy 11 (3).
Chomsky, N.: 1981, Lectures on Government and Binding, Foris, Dordrecht.
Cresswell, M.: 1985, Structured Meanings, MIT Press.
Dowty, D.: 1982, 'The Categorial Theory of Grammatical Relations', in P. Jacobson
and G. Pullum (eds.), The Nature of Syntactic Representation, Reidel, Dordrecht.
Dowty, D.: 1985, 'On Some Recent Analyses of Control', Linguistics and Philosophy 8,
pp.1-41.
Dowty, D.: (this volume) 'On the Semantic Content of the Notion of "Thematic Role"'.
Faraci, R: 1974, 'Aspects of the Grammar of Infinitives and FOR-phrases', Ph.D. Diss.,
MIT, Cambridge.
Farkas, D.: 1987, 'On Obligatory Control', forthcoming in Linguistics and Philosophy.
Gruber, J.: 1965, Studies in Lexical Relations, Ph.D. Diss., MIT, Cambridge.
Jackendoff, R: 1972, Semantics Interpretation in Generative Grammar, MIT Press.
Jackendoff, R: 1983, Semantics and Cognition, MIT Press.
Jones, C.: 1985, The Syntax and Thematics of 1njinitival Adjuncts, Ph.D. Diss., UMass,
Amherst.
Keenan, E. and Faltz, L.: 1978, Logical Types for Natural Language, UCLA Occasional
Papers in Linguistics 3.
Landman, F.: 1986, Towards a Theory of 1nformation, Foris, Dordrecht.
Ladusaw, W. and Dowty, D.: 1987, 'Towards a Non-Grammatical Account of Thematic
Roles', in W. Wilkins (ed.), On the Nature of Thematic Roles, New York, Academic
Press.
Lewis, D.: 1979, 'Attitudes De Dicta and De Se', The Philosophical Review 88, 513-
543.
McConnell-Ginet, S.: 1982, 'Adverbs and Logical Form', Language 58, pp. 144-184.
Partee, B. and Rooth, M.: 1983, 'Generalized Conjunction and Type-Ambiguity', in R
Bauerle, C. Schwartze, and A. von Stechow (eds.), Meaning, Vse and Interpretation
of Language, de Gruyter, Berlin.
Pesetzky, D.: 1987, 'Binding Problems with Experiencer Verbs', Linguistic Inquiry 18,
126-140.
Rizzi, L.: 1986, 'Null Objects in Italian and the Theory of pro', Linguistic Inquiry 17,
pp.501-558.
Stalnaker, R: 1985, Inquiry, MIT Press.
Veltman, F.: 1983, 'Data Semantics', in J. Groenendijk, T. Janssen, and M. Stokhof
(eds.), Truth, Interpretation, InJormation, Foris.
Thomason, R: 1980, 'A Model Theory for Propositional Attitudes', in Linguistics and
Philosophy 4, pp. 47-70.
Williams, E.: 1987, 'Implicit Arguments, the Binding Theory and Control', in Natural
Language and Linguistic Theory 5, pp. 151-180.
Cornell University
GREG N. CARLSON
ON THE SEMANTIC COMPOSITION OF

ENGLISH GENERIC SENTENCES
One part of the task in presenting a semantics for a natural language is

ta show how the meanings of the constituents of a sentence are
combined with one another ta build up a meaning for the whole
sentence. Ta a certain degree, this part of the task can be worked an
without committing oneself too specifically ta any particular theory of
meaning, and that is what I am going ta try ta do here. The construction
of interest is that of generic sentences. Such sentences are found in alI
natural languages (though not in alI artificial languages), but I wilI
confine my remarks ta how English generic sentences are built up from
constituent meanings, ever hopeful that what I have ta say about
English will shed light an similar constructions in a wider range of
natural languages. 1 wilI begin by giving a series of descriptions of what
generic sentences are, working from the most national and towards the
most linguistic, ta first stake out the domain of inquiry. I then turn ta
the semantic composition of such sentences, arguing that they are not
uniformly of subject-predicate form, but that a wider variety of internal
relations must be recognized. The overall purpose of the following
discussion is ta provide a set of general considerations that any formal
semantic theory incorporating generics must contend with.!
1. GENERIC SENTENCES
Notionally, a generic sentence is one expressing a regularity, as

opposed ta an instance from which one infers a regularity. For example,
the generalization "The sun rises in the east" expresses a regularity,
while "The sun rase this morning in the east" expresses an instance
from which, along with other such instances, one infers a regularity.
Epistemologically, a generic sentence is one expressing a truth (ar
falsehood) the true value of which cannot, in general, be ascertained
solely with reference ta any particular localized time. For instance, the
present tense sentence "Dogs bark" is true, even though, at the present
time there may be no dogs barking. Or, take the assertion "Bears
hibernate" said in midsummer; ta check an it one must wait for winter.
167
and Meaning, II. 167-192.
1989 by Kluwer Academic Publishers. Al! rights reserved.
168 GREG N. CARLSON
In contrast are the specific instances, which in general can be "localIy

determined". If 1 say "Jenny is watching TV in the den," this is
something 1 can, at least in principle, check on right now, and other
times become irrelevant to the truth or falsity of this assertion. Or if 1
assert that Bruno is in hibernation, 1 check on it now; 1 do not wait until
midwinter.
LinguistiealIy, generic sentences have no formal distinguishing fea-
tures, at least in English and many (though not alI) other languages (see
Dahl (1985) for an overview). Surface syntax, morphology, and phonol-
ogy do not distinguish generies from non-generics. They do, however,
have some distinguishing semantic features: generic sentences are (i)
stative sentences (ii) based on lexically non-stative predieates,2 and (iii)
they are intensional and (by alI appearances) non-monotonie. For
instance, the aspectual category of ''fetch newspapers" is non-stative (a
process), yet the sentence "Dogs fetch papers" passes alI the standard
tests for stativity (see for example Dowty (1979.
2. A SUBJECT-PREDICATE ANALYSIS OF GENERICS
Let us now turn to another semantic definition of generics, proposed in

Carlson (1980a). 1 wilI first simply present it, without any attempts at
motivation. Then 1 will discuss some ways in which it is inadequate, and
then what a more adequate analysis might look like. What wiH emerge
is a slightly different view of what generics are, and a signifieantly
different view of their semantic composition.
This analysis of generics, undertaken within a model-theoretic
(Montague grammar) framework, depends on making some assump-
tions about the' denotations of certain kinds of noun phrases. In
particular, a proper name like "John Smith" denotes (the property set
of) an individual. Individuals are not to be regarded as extensions
proper; rather, individuals are intensional objects. They are the sorts of
things that can appear at different times and places (and in different
worlds) and stiH be "the same" even if the instances themselves are
distinguishable. These instances are called (following Quine) "stages,"
spatialIy and temporalIy bounded instances of individuals. What 1 hav~
in mind here corresponds in a general way to what Barwise and Perry
(1983) call "located individuals," and m'ore closely to the "stages" of
Gabbay and Moravcsik (1979). (The idea is roughly this: take your
ENGLISH GENERIC SENTENCES 169
favorite individuation puzzle, ask what it is that gets individuated in the

first place, and you have the extensions proper.)
Another sort of individual, on this account, is kinds; a noun phrase
such as "bears" or "rickety railroad cars" denotes, like a proper name,
an individual as well, though a kind by virtue of its being able to have
multiple simultaneous stages at distinct locations in the same world.
Bears, unlike lohn Smith, can appear simultaneously in the Bronx Zoo
and the Denver Zoo; Smith can be in but one of those places at any
given time.
Given a basic distinction between stages, conceived of as basic
extensional elements, and individuals, conceived of as basic intensional
elements subtyped into regular individuals and kinds (at least), one can
then differentiate types of predicates, if we take extensions of predicates
to be sets of basic elements. 3 Some predicates may have as extensions
individuals (subclassified into kinds and regular individuals), and some
may take as extensions stages. It seems to be an empirical fact, given
this perspective, that no predicates "mix" the intensional and the purely
extensional; the denotation of a predicate with regard to appropriate
points of reference must either consist entirely of individuals and no
stages, or entirely of stages and no individuals.
Even though there are (at least) these two domains of predicates,
expressions denoting members of each domain may be formally related
to one another. This is at the heart of the view of generics under
consideration. The proposal is that we have predicates which denote
sets of stages (at points of reference), which can be operated on to yield
a predicate that denotes individuals expressed by a formally identical
linguistic form. For instance, consider the predicate "fetch my news-
paper" (as in "last night Fido fetched my newspaper"). This predicate
denotes stages, and this example sentence is true iff among those stages
last night is a stage of Fido. But now consider the generic sentence,
"Fido fetches my newspaper" (e.g. every morning). There is a semantic
operation (corresponding to a nuH syntactic operation), Gn, which
operates on the stage-denoting predicate "fetch my newspaper" to yield
a generic predicate "fetch my newspaper" which now denotes (at points
of reference) individuals, expressed by the same form. And the generic
version is true just in case the individual Fido is to be found in that
denotation (or, alternatively, that the meaning of the generic VP is truly
predicated of Fido). Given this typing of entiues and predicates, we
arrive at a tentative definition of a generic sentence: a generic sentence
170 GREG N. CARLSON
is one in which a predicate (Le. the meaning of a syntactic VP) is

applied to an individual, rather than to a stage of an individual, to yield
a sentence meaning (e.g. a proposition). In part, the purpose of this
analysis is to provide a unified account of when bare plural NP's
("rabbits," "unhappy students", etc.) and indefinite singular NP's ("a
cat", "a friend in need") appear be interpreted "universalIy" rather than
existentialIy. It a non-generic property is predicated of them, they have
existential interpretations (e.g. "A cat ran across my lawn"), whereas the
predication of a generic property gives rise to the apparent ''universal''
reading ("A cat chases mice") (see Carlson, 1977). Thus, whether
subject bare plurals and singular indefinites appear to be interpreted
universally or not becomes a diagnostic for whether one has a generic
or a nongeneric sentence, with the "universal" reading indicative of the
generic, and the existential indicative of a non-generic.
The definition of a generic sentence under this analysis is profoundly
tied up with the view that alI such sentences have a subject-predicate
form; the subject of the sentence provides the key ingredient for the
generic - the individual - and the generic predicate denoted by the
VP is predicated of the denotation of the subject NP.
3. SOME DIFFICUL TIES
This analysis, though, suffers some difficulties. The first, something of a

minor embarrassment, is that such a definition makes alI sentences
generics once the details of the proposed analysis are fulIy presented.
Though there are various ways out of this little dilemma, we won't
consider them here as the discussion would be narrow and unenlighten-
ing. And in any case there are much stronger reasons to think this
particular view is in need of even more major repair. To these more
important matters I now turn.
3.1. Locatives
Let us consider an example from Gary Milsark's dissertation (Milsark,
1974):
(1) Hurricanes arise in this part of the Pacific.
This sentence has at least two salielt interpretations. The first is
factually false - that hurricanes in general have a common origin in a
certain part of the Pacific. On this reading we are talking about "aU"
hurricanes. This is the generic reading predicted by the subject-predi-
cate analysis of generics.
But there is another salient reading for this sentence, one that says
nothing about "alI" hurricanes, but only teUs us that fram time to time a
hurricane occurs in these parts. This reading is accurately paraphrased
as (2a) or (2b):
(2) a. There arise hurricanes in this part of the Pacific.
b. In this part of the Pacific arise hurricanes.
But is this reading in fact a generic reading of (1)? By aU tests, it
appears to be. Consider whether it expresses a regularity (it does);
whether it is epistemologicaUy determinate with regard to the present
moment (it is not); whether it is stative (it is); whether it is based on a
nongeneric (it is; there is an eventive reading, more salient in the past
tense - "Hurricanes arase in that part of the Pacific" (so aU boats
avoided the area)). It is also intensional; consider substituting the
phrase "where 1 am pointing my finger" for the phrase "this part of the
Pacific." The intersubsitution is not automaticaUy licensed even if where
1 am pointing is the part of the Pacific 1 am talking about. (3) has a
different possible reading, in which 1 am possessed of power over the
forces of nature:
(3) Hurricanes arise where 1 am pointing my finger.
So, it appears that this latter reading of (1) is indeed a generic
reading, in spite of the fact that the subject NP has a clear existential
interpretation of the same sort it would have in non-generic sentences
such as (4a) (equivalent to (4b)):
(4) a. Hurricanes slammed against the Texas coast last week.
b. (=) Some hurricanes slammed against the Texas coast
last week.
Note that addition of "some" to (1) on the reading of interest yields
an accurate paraphrase, in the same way (4b) paraphrases (4a):
(5) Some hurricanes arise in the part of the Pacific.
If the subject NP is interpreted "universaUy" in a generic sentence,
though, this substitution very clearly fails to pravide a paraphrase:
172 GREG N. CARLSON
(6) a. Hurricanes have high winds.

b. (#-) Some hurricanes have high winds.
There are many similar generic sentences in which the subject NP is
interpreted existentially, alongside another generic reading in which the
subject NP is interpreted "universalIy". AlI the folIowing, with locative
phrases folIowing the verb, have these two generic interpretations:
(7) a. Flowers grow out behind the old shed.
b. A guard stands in front of the queen's palace.
c. U nfriendly tribesmen dwell just over those hilIs.
d. Productive farmland lies to the south of Des Moines.
e. A statue of Shakespeare sits near the halI window.
As Dowty (1979) points out, such sentences as those in (7) are best
construed as generics; yet they alI have "existential" subjects (e.g. in (7b)
no general characteristic of guards is being presented). One may at first
sight think these lack the epistemic independence of the present; but
this is not so. One can say (7a) in midwinter when there are no flowers,
or continue (7b) by noting that there is no guard there now, or (7c) can
be said even if no inhabitants are there now, or note in (7e) that
the statue has been moved in order to vacuum the area it normalIy
occupies (note that one could have a different statue every day). Only
the immobility of farmland prevents similar things from being said
about (7d).
Thus, we find our first set of apparent counterexamples to the claim
that generic sentences necessarily have a generic property predicated of
the denotation of the subject. Here were examples of sentences which
are evidently generic, but which do not distinguish the subject of the
sentence.
3.2. Transitives
Consider the folIowing example, suggested to me by Barbara Partee:

(8) A computer computes the daily weather forecast.
Again, there are a couple of sali~nt generic readings for this
sentence. One is quite implausible, attributing a weather-forecasting
function to "alI" computers. The other, the same as the more salient
reading of (9), tells us how the daily weather forecast gets figured out:
(9) The daily weather forecast is computed by a computer.
Again, the predicational analysis of generics fails to account for this
reading of (8). In (10) are more examples of generic sentences like (8):
(10) a. Robots build the new cars.
b. Cows give the milk that feeds this colony.
c. Vicious dogs protect Ft. Knox's gold.
d. Slaves work Lower Slobbovia's salt mines.
e. An oriental rug covers the dining room floor.
Each has a generic reading in which the subject NP is apparently
interpreted existentialIy, contrary to the predictions of the Carlson
(1980a) analysis.
3.3. Adverbials
Let us set these examples aside, and move on to another set of

examples in which generic sentences have existentialIy interpreted
subjects. Stump (1981, 1985) and Farkas and Sugioka (1983) note that
the Carlson (1980a) analysis is also unable to deal adequately with
certain adverbial constructions.
Consider example (11):
(11) A cat runs across my lawn every day.
There is an implausible "universal" reading in which it is a general
characteristic of cats that they run across my lawn. But there is also a
much more plausible reading, one where daily some cat or other
crosses my lawn; this, too, is a generic interpretation of the sentence. In
general, adverbials such as time and conditional adverbials (though not
manner adverbials), alIow for generic sentences with existentialIy-inter-
preted subjects. The folIowing have such generic readings, alongside the
usual "universal" readings:
(12) a. Mice squeak loudly if you pound on this walI.
b. Fire engines come when you dial 911 and ask for help.
174 GREG N. CARLSON
(12) c. A little red light comes on when you push this button.
d. People mutter in dis gust every time Bert appears on the
news.
e. Once in a while, volcanoes erupt and cause much
damage.
A related difficulty for the subject-predicate analysis is also dis-
cussed by Stump, and Farkas and Sugioka. They note that some
predicates, such as "widespread" and "rare" are not associated with a
generic "type-lifting" operator Gn, being basic predicates that apply to
kinds of things. Now one account of the source of the generic nature of
such examples as (13a) is that the adverbial clause combines with the
non-generic interpretation of the VP to form a non-generic predicate;
this predicate is then operated on by the generic operator to form a
generic predicate, which is then predicated of the subject. Thus, an
analysis of (13a) would, schematically, be as in (13b) (the adverbial is
taken to be a mapping from predicates to predicates):
(13) a. John vanishes when there is work to do.
b. John([Gn[Nongen([vanishes]) [when there is work to do)]]).
But example (14) allows for no such analysis, as there is no
opportunity for placing a Gn operator in the predicate of the sentence:
(14) Small fish are widespread when big fish are rare.
Yet (14) can clearly be understood generically.
3.4. Dummy Subjects

Another difficulty with the subject-predicate analysis of generics,
perhaps the most obvious, is that some sentences lack denoting subjects
but can be generic nonetheless. The clearest examples are cases of 'it'
with weather verbs; these can participate in generic sentences:
(15) a. It rains 30' a year here.
b. It snows in Florida once every decade or so.
c. It gets foggy if there's an easterly wind and cool tem-
peratures.
d. It is hottest when summer high pressure stagnates over-
head.
In order to rescue the subject-predicate analysis of generics from

examples like (15) one must posit some individual as the denotation of
"it". If this were the only difficulty the subject-predicate analysis faced,
something to this effect might be worth the effort (see e.g. Bolinger
(1970) for one such discussion), but on top of the other difficulties the
examples of (15) merely seem to make matters worse.
4. A SENTENTIAL-OPERATOR ANALYSIS
We continue to assume that generics are systematically related to their

nongeneric counterparts, and that the former are to be derived from the
latter, rather than the other way around. But the evidence presented
above suggests that a predicational analysis, in which a generic predi-
cate is applied to the denotation of the subject, is either too restrictive
an analysis Of else simply wrong. One reasonable alternative would be
to hypothesize that generic propositions are derived from nongeneric
propositions, rather than generic predicates being derived from non-
generic predicates. In practical terms this means postulating in the
semantics (and perhaps the syntax) an operator, GN, which applies
to sentence denotations (e.g. propositions) to derive other sentence
denotations (e.g. other propositions). So, for instance, the analysis of
"Dogs bark" would not be as in (16a), but instead as in (16b):
(16) a. [dogs' ([Gn([bark ')))] (where the "bark" portion is a
nongeneric VP)
b. [GN([dogs bark'])] (where the "dogs bark" portion is a
nongeneric S).
Such an analysis (actually, a program for an analysis) has the
immediate advantage of being able to meet the objections to the
predicational analysis evaluated above, at least in principle.
However, the sentential-operation analysis, even presented at this
level of generality, raises some immediate and pressing questions. One
problem it does not appear to cast much light on is the relationship
between the existential and universal readings of bare plurals (and mass
terms) on the one hand, and the generic or nongeneric nature of the
sentence, on the other. Whereas the predicate analysis makes some
incorrect predictions, the sentential analysis simply makes no predic-
tions at alI. This is not a step in the right direction.
There are stronger empiric al reasons for doubting the sentential-
176 GREG N. CARLSON
operator analysis. Probably the biggest puzzle is why there are no

generic sentences consisting solely of a subject and an intransitive verb
as a predicate, in which the subject is interpreted existentially instead of
universally. That is, alongside the ambiguity of the subject in (17a-
20a) there is no corresponding ambiguity in the examples in (17b-
20b):
(17) a. Slaves work Lower Slobbovia's salt mines.

b. Slaves work.
(18) a. Mice squeak when you pound on this wall.
b. Mice squeak.
(19) a. Flowers grow out behind the old shed.
b. Flowers grow.
(20) a. A robot cooks my moming coffee.
b. A robot cooks.
It is as if there must be something else in the sentence besides just a
subject and a simple predicate in order to have a generic with an
existentially-interpreted subject bare plural or indefinite singular.
Nor does the sentential analysis suggest a natural means of dealing
with the difficulty of interpreting empty subjects predicated of simple
intransitives. The examples of (21b-23b), in contrast to examples
(21a-23a), appear to be thoroughly vapid generalizations, leaving one
hunting for some kind of natural interpretation.
(21) a. It rains hard here.

b. (??) It rains.
(22) a. It snows in Florida once every decade or so.
b. (??) It snows.
(23) a. It gets foggy when there is an easterly wind and cool
temperatures.
b. (??) It gets foggy.
The (b) examples appear to be in need of something else in order to
make them understandable generics. Of course, as non-generics, these
are quite natural:
(24) a. It rained.
b. It snowed.
c. It is getting foggy.
So one cannot offer as a reason for the vapidity of the (b) versions
the lack of a nongeneric base from which to derive a generic.
5. A RELATIONAL ANALYSIS
We can summarize these observations as follows. Generic sentences

require two elements to be related to one another in order to have a
complete generic interpretation. The first element is a nongenerically
interpreted matrix sentence or predicate containing the sentence main
verb; then "something else" is required to complete the generic inter-
pretation. In example (1) it is the locative phrase (or, perhaps, just the
NP in the PP); in example (8) it is the direct object; in example (11) it is
the time adverbial; and in an example like "Cats meow" it is the subject.
Subjects, it tums out, appear to be the favored "something else", but
this by no means exhausts the possibilities. For want of better terminol-
ogy, I will call this "something else" "the related constituent" (and
sentences in which the subject plays the related constituent are "stand-
ard generics").4
What is wrong, then, with the vapid generalizations of (21b-23b) is
that there is no related constituent in the sentence to support a generic
interpretation (assuming 'it' to be nondenoting); why the examples of
(17b-20b) have only "universal" readings for their subjects is that the
subject is the only candidate within the sentence to supply the related
constituent necessary for a generic interpretation. But before we
explore this relational account further, we need to mention one other
source of "related constituents."
Some element understood as a part of the discourse, and not a part
of the syntactic representation of the sentence itself, can also condition
a generic reading - it appears that the discourse itself may supply
the related constituent necessary to complete the interpretation. For
instance, example (25), suggested to me by Emmon Bach, has no
natural generic reading if understood in isolation:
(25) John builds a cabin.
The only available reading seems to be a present-tense "announcer's"
178 GREG N. CARLSON
sense. Yet, if one understands a previous discourse in which the topic

of conversation is how John works off his frustrations, or what he does
every time he wants to make a little extra money, then it can clearly be
understood as a generic. The key element in making (25) a generic,
then, is contributed by the discourse in which it is embedded. Below are
some more examples:
(26) a. (Topic: What happens in the event of a fire). An alarm
sounds.
b. (Topic: What happens in a recession). People get laid off.
c. (Topic: Late-summer small-town events). A circus arrives
in town.
d. (Topic: Winter weather around these parts). Oh, it does
snow, you know.
One might reasonably question whether examples like those in (25)
and (26) have generic meanings; perhaps they are non-generics that
contribute to a generic proposition upon being integrated into a
discourse. What few facts bear on this question (assuming it's a
coherent one in the first place) favor treating them as generics rather
than non-generics. For instance, the unequivocal marker of a generic in
English - the "used to" construction - is quite natural under these
same circumstances. For instance, (27) sounds quite natural in appro-
priate discourse:
(27) (What Aristotle did when frustrated.) Aristotle used to write
a book. (cf: 'A. used to write a book' without supporting
context)
So, 1 as sume that something has been integrated from discourse into
the "local" interpretation of the sentence in such cases as (25-6). (The
questionnaire of Dahl (1985) elicits similar generic markers in such
contexts in a variety of languages.) 1 have little to say about these
contextually-interpreted generics apart from noting them and the way
they lend support to a relational treatment of generics. 1 strongly
suspect there is a direct connection between the discourse elements
required here, and those necessary to account for many interpretatiol).s
of pronouns in discourse (e.g. the cases of "modal subordination" noted
in Roberts (1986) and references citell). The "conversational back-
ground" postulated by Kratzer (1977) to appropriately interpret modals
and used by Stump (1981, 1985) to interpret free adjuncts and

absolutes may also provide a very promising point of departure.
6 . WHAT DO GENERICS RELATE?
In this section we will explore the question of what sorts of things can
be generaUy related to one another (the further question of the precise
nature of this relation lies beyond the scope of the present work). From
a syntactic perspective, NP's and certain soTts of adverbials can be
related to predicates on the one hand, and propositions (or perhaps
predicates as weB) on the other. 1 am not going to be concemed about
how to deal with the disjunction predicate or proposition - there are
notions available such as 'situation type' (Barwise and Perry, 1983)
which seem capable of unifying the two. Instead, 1 would like to focus
on the related constituent, making the point that intensionality of the
related constituent is a crucial ingredient in the interpretation of generic
sentences.
Consider the interpretation of the NP "the daily weather forecast" in
example (8), repeated here:
(8) A computer computes the daily weather forecast.
In this context, "the daily weather forecast" requires an intensional
interpretation, where its meaning cannot be taken as rigidly referring to
the present weather forecast, e.g. the one appearing in today's copy of
the Times predicting light rain and highs in the upper thirties. In non-
generic sentences like (28), though, the present example of the forecast
can be understood as the reference of the NP:
(28) a. Today, the daily weather forecast caUs for snow.
b. John got aU depressed today after he read the daily
weather forecast.
c. The daily weather forecast appears on p. 7 of today's
Times.
Here we find a dear connection between intensionality and generic-
ness, with only the intensional reading co-occurring with the generic
interpretation.
Now consider the interpretation of the adverbial in (11), repeated
here, recaUing the reading of interest:
180 GREG N. CARLSON

A phrase like "every day" is, as many have observed, ambiguous
between a reading which quantifies over some (normally contextually-
defined) finite set of elements, or else it may have an "unbounded"
reading. In (11) and other such examples, only the latter "unbounded"
reading appears. Consider, for instance, what occurs when there is an
overt restriction of the range of quantification to a finite set, e.g.:
(29) a. A cat runs across my lawn every day this week and last.
b. A cat runs across my lawn every day from last Monday to
next Friday.
One can perhaps interpret these as stage directions, but one clearly
cannot interpret (29) in the same straightforward generic way that one
can understand (11) (in languages in which the simple present is
routinely used nongenerically, such examples as (11) are judged
ambiguous). Similarly, substitution of other quantifiers with bounded
readings, such as numerals, yields the same results in the English simple
present.
(30) a. A cat runs across my lawn (on) four days.

b. A cat runs across my lawn several times.
The examples of (30) have the same difficulties as the examples of
(29), on their more salient reading. 30a) clearly has other acceptable
generic readings we will comment on below.)
So why regard the "unbounded" readingof "every day" as inten-
sional, rather than simply quantifying over all possible days? There are
several reasons. First, it is simply not quantificationally unbounded at
alI in the interpretation of (11). (11) can clearly be regarded as true
even if there are some days no cat has, or will, run across my lawn. For
instance, (11) is perfectly compatible with the occurrence of cats
dashing across my lawn having begun only last month, and its present
truth certainly does not require that this go on forever. Rather, it is a
generalization that presently holds. So it really cannot be a matter of
simply quantifying over all days, real or possible.
Secondly, even if there is a restriction of days you are talking abaut
to an unlimited number, it seems to he'lp very little; (31) seems as bad
as (29) and (30), and certainly (31 b) is no paraphrase of (11):
(31) a. A cat runs across my lawn (on) an unlimited number of

days.
b. A cat runs across my lawn every day of an infinite
number of days.
Rather, the phrase "every day" in (11) is "unbounded" in some other
sense, rather than quantifying over an unlimited number of instances.
The beginnings of a satisfactory analysis would treat the meaning of the
phrase in (11) as a function from contexts to aH days in that context,
and it is this meaning that is related in the generic interpretation of
(11).5
Why can't such a "bounded" interpretation be found for the adver-
bial phrase "every day"? The answer is, put simply, that a day is not an
intensional entity, whereas a part of an individual appears to be. A day,
in this view, is an extension; it is a unique, unrepeatable token, in no
sense able to appear and reappear as a whole at different times, nor
appear as a whole in different locations at the same time.
There is of course a sense in which "day" is intensional. If by
"Monday" I mean a recurring period of time (the first or second day of
the week, depending on how you count) rather than a certain particular
day, it is intensional. There are just seven days in this sense. Now,
reconsider (11), repeated below:
Now, there seems to be a reading of this on which (11) is equivalent
to a conjunction of generic propositions, "A cat runs across my lawn
(on) Sunday, ... ," A cat runs across my lawn (on) Saturday", with each
day-name interpreted intensionally. The reading that is missing, though,
is the one where some bounded set of token days is quantified over
(e.g. equivalent to a conjunction of propositions "A cat ran across my
lawn yesterday, and ... , and the day before that ... ").
In support, consider a case with the NP "every day" as the subject of
a standard generic. Note, in (32), that the name of the day must be
interpreted intensionally as the subject of a generic (1 revert to the
generic past "used to" construction to force an unambiguously generic
interpretation):
(32) Sunday used to go by too quickly.
In (32) one cannot mean, for instance, last Sunday, but instead one
182 GREG N . CARLSON
means Sundays in general. In similar fashion, consider (33a) in contrast

to (33b):
(33) a. Every Sunday used to go by too quickly.
b. ?? Every Sunday in the past month used to go by too
quickly.
The reason (33b) seems strange is that it means that if x was a
Sunday in the past month, x used to go by too quickly; this is the
'bounded' interpretation. But since x here does not take as values
intensional entities (token days are not intensional), it cannot be the
subject of a standard generic. (33a), on the other hand, is quite natural,
though only on an unbounded intensional reading.
Turning to the clausal adverbials, a reasonably clear "bounded/
unbounded" distinction is found there as well, and here, too, it is quite
easy to understand the distinction as extensional/intensional. A tem-
poral when clause can have two interpretations; on the one hand it can
point to a single instance, as in:
(34) Johnjumped when the fire alarm went off.
On the other hand, it can mean, roughly, "whenever", the interpreta-
tion most salient in (35):
(35) John eats when he gets hungry.

Some languages formally distinguish these two senses of "when"
clauses (e.g. Scandinavian languages). In (34) a single token occurrence
or instance is designated by the adverbial, whereas in (35) one could
regard the adverbial as some function that picks out, with respect to a
context, instances of John's being hungry. It is this latter, intensional
reading that is the only one possible in the (generic) interpretation of
(12c) repeated below - with the subject understood existentially:
(12) c. A little red light comes on when you push this button.
Other clausal adverbials work in similar fashion. It is significant that
those adverbials which do not exhibit this same ambiguity in the simple
present (e.g. "because, although, since" clauses, and examples of E.
Williams (1975) 'level IV' adverbials, do not allow for generic readings
in such cases. For example, (35a) is strange, while (36b), in the past
tense and hence more readily interpretable as a nongeneric, seems quite

natural:
(36) a.?? A little red light comes on because you push this
button.
b. A little red light carne on because you pushed this
button.
(O ne must take care not to confuse two distinct notions: that when
(etc.) clauses are interpreted either extensionally or intensionally, aud
that the clauses themselves cau be interpreted generically). See ter
Meulen (1985) for an insightful discussion of conditionals and generics
in a situation semautics framework.
Similarly, one finds apparently variable interpretation of gerunds aud
infinitives, discussed in Chierchia (1984). As subjects of nongeneric
sentences, these are interpreted as designating token occurrences, as in
(37):
(37) a. Riding on the merry-go-round is making Bill sick.
b. Repairing Janet's bike took several hours.
c. Getting a hand stuck in the snow-blower cost Bob two
fingers .
In each case some particular token occurrence is understood as
being designated by the subject of the sentence. In generic sentences,
though, the interpretation takes on a more general quality, and does not
attribute generic properties to token occurrences; instead, properties
themselves seem designated:
(38) a. Riding on a merry-go-round makes Bill sick.
b. Repairing Janet's bike takes several hours.
c. Getting a hand stuck in the snow-blower requires very
little concentration.
This is the same pattern as the adverbials.
Before moving on to consideration of nonadverbial NP's, let us
summarize momentarily. In each of the examples discussed so far, it
was argued that in generic sentences the related constituent had to be
interpreted intensionally. While extensional interp'retations are available
184 GREG N. CARLSON
for each related constituent so far examined, they cannot function in the
capacity of a related constituent if so interpreted. The reason proposed
is that the meanings of gerunds and infinitives, of adverbial clauses and
of adverbials like 'every day' at points of reference pick out pure
extensions (token occurrences, token periods of time), and those
extensions are not the kinds of things that can be generically related to
something else.
When we turn to consideration of argument NP's, much the same
pattern can be observed in certain instances. Consider examples of
purely intensional NP's that arguably have no standard extensions, like
the average American family (the one with 1.8 kids, for instance). Such
NP's are impeccable as subjects of generics:
(39) a. The average American family eats at MacDonald's twice

a week.
b. The average American family spends 17.3% of its
income on food.
c. The average American family watches television after
dinner.
But these are highly questionable as subjects of nongenerics:
(40) a.? The average American family joined me for breakfast

this morning.
b. ?The average American family moved in next door.
c. ? The average American family is sitting over there in the
corner. 6
So again, it is intensionality that seems essential for the related
constituent in generics.
Consider bare plural and unmodified singular indefinite subjects of
generics and episodics. Once we isolate them as the crucial elements in
the generic, there is a clear relationship between genericness and
intensionality. The "existential" reading of bare plural and singular
indefinites corresponds to a pure extension (i.e. a stage) and thus
cannot be the related element of a generic; but the "universal" reading
corresponds to the "unbounded" or intensional reading here, and as
such is suited to participating in the generic relation.
It we examine generics where the direct object is the related
constituent, we can make a similar argument for a wider range of NP's.

Some of the examples of object generics above require an intensional
reading of the definite direct object, as in:
(41) a. A computer computes the daily weather jorecast.
b. Cows give the milk that jeeds this colony.
c. Robots assemble the new cars.
d. A master craftsman builds the houses in this subdivision.
The related nongeneric sentences, though, do not require such an
intensional interpretation of the direct object NP.
When we examine stiH a wider range of NP's in generic sentences,
though, the ingredient of intensionality appears no longer essential. If
intensionality is so important in generic sentences, then it is natural to
ask why this is not readily apparent from the examination of standard
generics and the interpretations of their subjects. 1 think the reason is
quite simple. If one holds that individuals are not intensional constructs,
then many examples of what 1 have been calling 'generic' sentences
appear to have extensional subjects.
(42) a. John votes for Socialists.
b. Mary smokes.
c. The mayor of Cleveland drives a Ford.
However, if you make the (possibly non-standard) assumption that
individuals are intensional constructs, and do not in and of themselves
constitute extensional entities, the analysis can be maintained.
In order to sustain this general line of analysis, though, it is necessary
to recognize two 'levels' of intensional analysis in the interpretation of
certain NP's, such as universally quantified NP's. Consider the case of
(43), adapted from Dahl (1975).
(43) a. Every friend of John's votes for Socialists.
b. Every friend of John's smokes.
On the intensional ("unbounded") reading of the subject NP, the
sentence expresses a requirement for anyone who might be a friend of
John's, and it applies to alI John's potential friends as well as his present
ones. This is the reading we fully expect if inten'Sionality is a require-
ment for genericness. But then there is another reading, the "bounded"
186 GREG N. CARLSON
reading, in which it describes a (possibly accidental) common charac-

teristic of alI John's present friends (or some contextually-circum-
scribed subsetof them). Consider a case where John has three friends,
a, b, and c, and each votes for Socialists. This condition alone is
sufficient to assert the bounded sense of (43). This condition, however,
is not sufficient to assert the intensional or unbounded sense of (43),
because voting for Socialists might simply be a sampling accident, and
does not apply to John's potential friends. Even so, the interpretation is
equivalent to the conjunction of three generic sentences, each asserting
that x votes for Socialists, with x assigned a, b, and c as values.
Thus, one can establish that one generic interpretation of (43) has
the so-called "bounded" reading of the subject NP as the designated
element of the generic 43b) assures us of this). However, the reason a
generic reading is possible here is that each of John's friends, a, b, and
c, is individually an intensional entity.7
Let me briefly outline what 1 would take to be the 'logic al form' of
the two generic readings of (43b). The "bounded" sense involves
predicating the generic predicate 'smokes' of each of John's (present)
friends in turn; this is the standard (restricted) quantificational analysis
in (44).
(44) Vx :friend of John's (x) [[Gn(smokes)) (x)).
The "unbounded" generic reading, on the other hand, ascribes the
generic property of smoking to the meaning of the NP "every friend of
John's":
(45) [Gn(smokes)) (Vx: friend ofJohn's (x)).
These two readings are possible because there are two "levels" of
intensionality in NP's such as these: the meaning of the NP itself, and
the individuals quantified over at given points of reference. On the
other hand, adverbials, gerunds and infinitives, and NP's like "every
day" (on the token reading of "day") and proper names, in the examples
above, exhibit but one "level" of intensionality, so similar boundedl
unbounded ambiguities in the generic reading do not appear.
In some instances, the nature of the predicate of the sentence
precludes the individual-Ievel ("bounded") intensional reading of the
generic. For instance, the object-generics of (41) above have only the
reading where the generic property is related to the "unbounded"
readings of the abject NP's - where the NP meaning is the related
constituent. But this does not hold for alI object generics. Consider the
examples of (46).
(46) a. Vicious dogs protect Fl. Knox's gold.
b. Slaves work Lower Slobbovia's salt mines.
c. An oriental rug covers the din ing room floor.
In each example of (46) an intensional reading is of course possible,
but alongside we also find possibility of the "bounded" individual
reading. What makes for the difference between the examples of (41)
and (46)? Plainly, it is aspectual. The examples of (41) alI contain verbs
which are accomplishments, in Vendler's terminology (Vendler, 1967).
These verbs, in the generic, do not readily take direct objects which
denote single individuals, as the folIowing examples aUest:
(47) a. J ohn builds a cabin.
b. A cow gives this quart of milk.
c. A computer computes today's weather forecast (the one
before us)
d. Vincent paints a picture.
Each example in (47) requires some context if it is to be interpreted
genericalIy. (See H. Verkuyl (1985) for an analysis which deals with
examples like (47).) From a cross-linguistic point of view, the examples
of (41) are very much in line with the fact that one finds a consistent
formal association between generic interpretations and imperfective
aspect (as opposed to perfective aspect). It is quite common for
imperfects, as opposed to perfects, to be the preferred if not the sole
vehicle for expressing genericness (Carlson (1980), Oahl (1985); see
also Oeclerck (1985)).
Let us now examine the example of (48):
(48) A master craftsman builds every house in this area.
This does have a clear generic reading (the preferred one being the
one intended - with the subject NP interpreted existentially apparently
within the scope of the universal in the object NP). Now let us suppose
that there are, at present, three houses in the area: the Smiths', the
Ooes', and the Nelsons'. Yet, it does not seem that (48) entails the
generic sentence (49):
188 GREG N. CARLSON
(49) A master craftsman builds the Smiths' house.

Note further that (48) makes it sound like there are going to be more
houses built in this area. Consider a case where the entire area has been
developed - a completed house on each lot, with no more space for
houses; under those circumstances (48) would be decidedly odd,
making it sound as if more houses are planned. The point is that (48) is
not to be analyzed as bounded universal quantification over individual
houses, suggested by the representation in (50):
(50) Vx:house in this area(x) [Gn(Ay[a master craftsman builds
y)] (x)ll
This would entail the generic sentence (49), which (48) does not (in
contrast, see (44) above). In contrast to (44) and the examples of the
locatives in (7) (which are uniformly states and not accomplishments),
the aspect of the predicate in (48) rules out the possibility of a
"bounded" generic reading ascribing a generic predicate to individuals.
Thus, anything along the lines of (50) is not appropriate. The "logical
form" of (51) or perhaps (52) would appear much more promising:
(51) Gn(Ay[a master craftsman builds y] (Vx:house in this
area (x
(52) Gn(A.9'[a master craftsman builds .9']) (Vx:house in this
area(x (where .9' is a variable over NP meanings).
In constructing a semantics for (51) or (52), one needs to keep in
mind that (48) does in fact entail sornething or other about individual
houses; setting aside an important qualification, (48) entails the non-
generic sentence (53):
(53) A rnaster craftsrnan built the Srniths' house.
(The qualification is that the generalization expressed in (48) had to
be "in force" at the time the Smiths' house was built; it is not a
contradiction to say that a rnaster craftsman now builds every house in
this subdivision, but that Srniths' house, which is in this subdivision, was
built before that policy took effect).
So, although many NP's have two "levels" of intensionality available
for serving as the related constituents of generics, at least two factors
rnay corne into play to eliminate one of those possibilities. The cornmon
noun portion of the NP may not be of the appropriate sort (the token
interpretation of "every day"), or the aspect of the predicate in the

sentence may rule out the "bounded" individual reading.
6 . SUMMARY AND CONCLUSIONS
Precisely how to implement these observations into a formal semantic

theory remains well outside the scope of this paper. Nor have we even
attempted to consider the entire range of "logical forrns" that generic
sentences might take on, concentrating on a only few of the variants
without trying to map out the whole territory (see note 3). Nevertheless,
we can draw some reasonable conclusions from this discussion about
the shape any detailed analysis would have to take.
First, generics are not of uniform subject-predicate form; rather, a
wider array of intensional elements may come into play instead of just
the grammatical subject of the sentence, and even there the subject NP
may combine to form a generic in two distinct ways; elements from
discourse and possibly other facets of context may also be brought into
play. This suggests quite strongly that generic interpretation is not
associated with any single syntactic rule of the language, but at best
with a range of such rules. And if the semantic theory is a strongly
typed theory and closely associated with the syntax, it would appear
that a range of types would enter into the generic relation, and not a
single one.
Second, we have cause to hold that the meaning of a generic
sentence stems fundamentally from a relalion between intensional
elements: any generic operator would have to be dyadic. It does not
relate purely extensional elements to one another or to intensional
elements, but only intensions to intensions. This entails that individuals
be regarded as intensional, and that such NP's as "every friend of
John's" be analyzed as having two levels of intensionality associated
with it, one for the "unbounded" reading, and the other for the
"bounded" reading. Such NP's as "every day," on the other hand, may
have but one level of intensionality, depending on how "day" is
interpreted.
And third, the generic relation would not appear to be a predica-
tional one. The fact that adverbials can play the same type of role in
generic interpretation as argument NP's suggests that the relation
between matrix (S, VP) and related element" (e.g. adverbials, NP,
perhaps PP's) is not that of predication (i.e. being in the extension of a
190 GREG N. CARLSON
property), and that we should seek another relation that can hold for
NP's and adverbials like "every day" and "when you pound on this
wall" indifferently. In other words, if we as sume that the extension of a
property is a set of individuals, we cannot hold that a generic sentence
involves predicating a generic property of an individual - not unless
we are prepared to define "individual" so broadly as to virtually void
the term of content, to include adverbials, NP-meanings, denotations of
infinitives, etc. Precisely what this relation consists of if it is not
predication, though, remains an open question.
NOTES
1 For an excellent overview and critique of formal theories of generics, see Schubert
and Pelletier (1986).
2 This property is not essential as generics may be based on stative nongenerics; here,
though,1 deal with those based on nonstative predicates.
3 Terminology here should be clarified. The extension of an expression may consist of
intensional elements at a given point of reference, and is an abbreviated way of ta1king
about the denotation of an expression with respect to (wrt) given points of reference; on
the other hand, if a construct is (ontological1y) extensional or a pure extension, it
cannot be an intensional construct as well.
4 Extending the list further typically requires dealing with more complex constructions
which make sure judgment of meaning difficult; complexity also makes it increasingly
difficult to factor out the discourse factors. So, for instance, in (i) an indirect object NP
is the focus of our attention. The sentence seems to have a generic reading, but is this
due to the NP itself, or is it due to the contextual interpolation of an implicit time
adverbial? Are the subject and object NP's interpreted existentially? Is the indirect
object interpreted intensionally?
i. A familiar person gives every ehild a present.
While factors such as these do not make investigation impossible, it becomes difficult to
consider a wide variety of such examples.
5 John Barwise pointed out the interest of such an analysis to me, but 1 do not know if
he would endorse it. See Croft (1986) for further relevant discussion.
6 The example "The ave rage American family watched the president on TV last night"
does not seem so bad, though a nongeneric; whether such examples show the presumed
analysis incorrect remains uncertain.
7 One might ask whether it is fair to call such examples as (19) 'generic'; terminologi-
cally, it is customary to distinguish such examples from those in which kinds or less
controversially intensional NP's function as subjects by applying the labeI "habitual" o
the former and "generic" to the latter only. 1 prefer to emphasize the similarities
between "habituals" and "generics" by applying 'ilie same labei; 1 am not particularly
satisfied with the terminological imperialism inherent in my use of "generic" in its
presently broad sense, but there is no appropriate alternative from either linguistics or
philosophy for me to draw ono As for the substance of the distinction, in Carlson (1980
pp. 97-98) 1 offer evidence that the two should be classed together. Furthermore,
while numerous languages morphologically distinguish what I am calling "generics" here
from nongenerics, few if any distinguish habituals from "generics" in the narrower (and
more common) sense (see Dahl, 1985, and Carlson, 1980b). See Smith (1975) for a
detailed lament over terminology, and for a wealth of data on generics.
REFERENCES
Barwise, Ion and Perry, Iohn: 1983, Situations and Attitudes, Cambridge: MIT Press.
Bolinger, Dwight: 1970, 'The Lexical Value of "It''', Working Papers in Linguistics,
University of Hawaii, pp. 57-76.
Carlson, Greg N.: 1977, 'A Unified Analysis of the English Bare Plural', Linguistics and
Philosophy 1,413-56.
Carlson, Greg N.: 1980a, Reference to Kinds in English, New York: Garland
Publishing.
Carlson, Greg N.: 1980b, 'Nomic Morphemes', paper presented at the Fourth
Groningen Round Table, Groningen, the Netherlands.
Chierchia, Gennaro: 1984, 'Topics in the Syntax and Semantics of Infinitives and
Gerunds', unpublished University of Massachusetts Ph.D. dissertation.
Croft, William: 1986, 'Universal Quantifiers and Generic Expressions', Stanford
University ms.
Dahl, Osten: 1975, 'an Generics', in E. Keenan (ed.), Formal Semantics of Natural
Language, Cambridge: Cambridge University Press, 99-111.
Dahl, Osten: 1985, Tense and Aspect Systems, Oxford: Basil Blackwell.
Declerck, Renaat: 1985, 'The Manifold Interpretations of Generic Sentences', Catholic
University of Leuven ms.
Dowty, David: 1979, Word Meaning and Montague Grammar, Dordrecht: D. Reidel
Publishing.
Farkas, Donka and Sugioka, Yoko: 1983, 'Restrictive If/When Clauses', Linguistics and
Philosophy 6, 225-58.
Moravcsik, Iulius and Gabbay, Dov: 1973, 'Sameness and Individuation', Joumalof
Kratzer, Angelika: 1977, 'What 'Must' and 'Can' Must and Can Mean', Linguistics and
ter Meulen, Alice: 1985, 'Generic Information, Conditional Contexts, and Constraints',
in E. Traugott et al. (eds.), On Conditionals, Cambridge: Ambridge University Press.
Milsark, Gary: 1974, 'Existential Sentences in English', Unpublished M.I.T. Ph.D.
dissertation.
Roberts, Craige: 1986, 'Modal Subordination, Anaphora, and Distributivity', Unpub-
lished University of Massachusetts Ph.D. dissertation.
Schubert, Lenart and Pelletier, F. Jeffry: 1986, 'Problems in the Representation of the
Logical Form of Generics, Plurals, and Mass Nouns', in E. LePore (ed.), New
Approaches to Semantics, New York: Academic Press. "-
Smith, N. V.: 1975, 'an Generics', Transactions of the Philological Society 1975, 27-
48.
192 GREG N. CARLSON
Stump, Gregory: 1981, 'The Formal Semantics and Pragmatics of Free Adjuncts and
Absolutes in English', Unpublished Ohio State University Ph.D. dissertation.
Stump, Gregory: 1985, The Semantic Variability of Absolute Constructions, Dordrecht:
D. Reidel Publishing.
Vendler, Zeno: 1967, Linguistics in Philosophy, Ithaca: Cornell University Press.
Verkuyl, H. J.: 1985, 'Nondurative Closure of Events', in J. A. G. Groenendijk et al.
(eds.), Information, Interpretation, and Inference. Proceedings of the Fifth Amster-
dam Colloquium. Dordrecht.
Williams, Edwin: 1975, 'Small Clauses in English', in J. Kimball (ed.), Syntax and
Semantics 4, 249-74.
University of Rochester
LENHART K. SCHUBERT AND FRANCIS JEFFRY PELLETIER
GENERICALL Y SPEAKING, OR,

USING DISCOURSE REPRESENTATION
THEORY TO INTERPRET GENERICS
1. INTRODUCTION
In this paper we discuss a proposal for representing the logical form

of generic and habitual sentences. The proposal is a development of
Schubert and Pelletier (1987), although it stands on its own and can be
comprehended without having first read the earlier paper. (Most of
what would be missed would merely be a discussion of background
litera ture, and reasons to want a different logical form from those
extant in the literature).
In this earlier paper, three different "sophisticated" theories of
generics were compared: those of Carlson (1977a, 1977b, 1979, 1982),
Chierchia (1982a, 1982b), and Farkas and Sugioka (1983). Various
faults were found with alI of them. Some faults were minor, specific
defects; but more importantly we found that alI of them ignored (what
we there calIed) "the most salient feature of habitual and generic
statements". At the end of that paper we surveyed two directions for
further work. One was a "conservative" proposal which we do not wish
to develop further, at least not at this time.! The second proposal we
mentioned was more radical. It involved the notion of an "ensemble of
cases". It is our view that habitual and generic statements rely on a
reference to such an "ensemble" for their semantic evaluation. For
example, a habitual sentence with an explicit adverb of quantification
("usualIy") like
1. John usually beats Marvin at ping pong
does nof say that most of the time John is beating Marvin at ping pong.
Rather, the usually gathers a certain class of "reference situations",
namely alI those situations in which John and Marvin play a game of
ping pong, and the usually is evaluated with respect to this class of
situations, counting whether most such situations are such that John
beats Marvin. If so, then (1) is true; otherwise it"s false. Similarly, when
the sentence is generic, such as
193
1989 by Kluwer Academic Publishers. AU rights reserved.
194 LENHART K. SCHUBERT AND FRANCIS JEFFRY PELLETIER
2. Cats land on their feet
it is not evaluated as if it said that at alI or most times cats are landing
on their feet, but rather a certain class of "cases" or "situations" is set
up - such as alI those cases where cats drop to the ground - and the
sentence is evaluated with respect to those cases. (To account for the
nomic import of a sentence like (2), cases in certain "nearby" possible
worlds need to be considered along with actual cases - see Sections
VIII, XI, and the Appendix, clause (16).) If most of these situations are
such that the cats involved land on their feet, then (2) is true; otherwise
it's false.
Thus it is our view that semantic evaluation of habitual and generic
statements depends on reference to these "ensembles of cases". We
think that the relevant ensemble is determined in part or entirely by
context and presuppositions, or in part or entirely by restrictive clauses
and adverbials. We think there are two kinds of such reference
ensembles, closely paralleling the two kinds of uses of adverbs of
quantification: ensembles of situations 2 (corresponding to "temporal"
uses of adverbs of quantification) and ensembles of objects (corre-
sponding to "atemporal" uses of adverbs of quantification).3
This view of habituals and generics being evaluated with respect to
some pre-given ensemble (or set) of situations (or objects) will probably
bring to mind a certain group of semantic theories genericalIy known as
"Discourse Representation Theories".4 These theories too attempt to
give a semantic evaluation of sentences based on some pre-given
context: for them, the pre-given context is previously "proces sed" text.
The goal of such theories is to be able to account for how the "context",
according to which one semantically evaluates a given sentence,
changes in the sentence-by-sentence (indeed, clause-by-clause) reading
or "processing" of an entire text. The crucial test in such theories (so
far, at any rate) has been to account for pronoun anaphora, both within
and across sentences. Another issue under investigation has to do with
the "sequencing of reported events" in the text so as to account
correctly for the observed use of tenses and tense adverbs in the text.
StiH further developments would be to investigate how discourse can
allow pronoun reference to, for example, "plural objects" in a reason-
able manner. 5
Our theory wilI be in this general vein. " However, we disagree on
various details, especialIy on the interpretationof "donkey sentences";
GENERICALL Y SPEAKING 195
and we hope to give a somewhat better account of tense than those

which have thus far been given. Finally, of course, we will try to show
how our discourse theory accounts for various generic sentences. Here
we will be partially successful, particularly for episodic generics whose
indefinites or bare plurals are confined to the restrictive clause. After a
discussion of some deep-rooted difficulties, we sketch a potentially
more comprehensive account. One overarching difference between our
theory and the most popular of the others (Kamp 1981, Heim 1982) is
that ours will be compositional; at least in the weak sense that
syntactically identical phrases are uniformly translated, and logical
expressions are in turn interpreted independently of the expressions
embedding them (relative to a given "context"). (The theory is not
"strictly" compositional, in part because it allows quantifier raising and
in part because the interpretation of an expres sion may depend on
more than just the intensions of its top-Ievel constituents.) In this, our
theory resembles that of Groenendijk and Stokhof (1987).
Let us therefore start with an exposition of the general background
for our discourse representation theory, and follow this with a discus-
sion of "donkey sentences" (which provide a clear arena that allows
our theory to be distinguished from other discourse representation
theories). AlI of this will be carried on at the intuitive level, after which
we introduce some small bit of formalism so as to be able to show to
some degree how we view our method as working. (A somewhat more
formal account of the inner workings of the theory is given as the
Appendix to the paper.)
II. DISCOURSES AND DISCOURSE REPRESENTATION THEORIES
Consider a "discourse" like

3. John is at the door. He is behatted.
Is the second sentence of (3) true? Well ... it depends upon (a) what
the referent of he is, and (2) whether this referent is wearing a hat. We
have not much to say about how one decides whether he denotes John,
or some deictically indicated or otherwise salient person, or (perhaps)
some previously-mentioned person from earlier in the "discourse".
Instead we would recommend a "representation" of this discourse as
(ignoring various fine points):
3'. at-door(j) . behatted(he)

leaving it to some other "module" to replace the he with appropriate
(logical form) items. One might call this some kind of "pragmatic
disambiguation" - but then again, maybe one wouldn't. But once a
choice for he has been made, it becomes the duty of the semantic
theory to decide whether it is true (in our model or interpretation). One
choice for he presumably is j, and then the question of whether the
second sentence is true should amount to whether [behatted] ([j ] = 1
- ignoring issues abut time and modality. It is this latter step, the
semantic evaluation, that we wish to explore here.
The treatment suggested above for intersentential anaphora should,
we think, be given also for (certain) intrasentential anaphora. For
instance, we think that the same treatment should be accorded such
sentences as
4. a. John is at the door and he is behatted.
This should be translated as
4. b. at-door(j) & behatted(he).
As is well-known, there can also be such pronominal reference to a
quantified noun phrase:
5. a. A man is at the door. He is behatted.
b. A man is at the door and he is behatted.
For these we would recommend representations like 6
5. a' .(3x: man(x))at-door(x) . behatted(he)
5. b '.3x :man(x))at-door(x)) & behatted(he).
Again there may be numerous choices for "disambiguating" he. We
draw attention to the possibility of using x (in Sa'):
6. (3x: man(x))at-door(x) . behatted(x).
Of course, this leaves a free x in the second sentence. How shall our
semantic evaluation treat this? Two choices come to mind: (a) Do a
further adjustment on this "logical form" and make the scope of the
quantifier found in the first sentence be" the pair of sentences ("quan-
tifier raising"), (b) Evaluate the second sentence as if the x denoted the
value of some previously quantified x (which made the sentence it
occurred in true - that is, a value which satisfied the open formula to
which the quantifier applied). Here we would make the x of the second
sentence take a value which satisfied the existentially quantified first
sentence; that is, we would make it be a man at the door.
The present method, in common with Kamp (1981), Heim (1982),
Groenendijk and Stokhof (1987), among others, takes the second
strategy. Reasons for this have to do with the facts that not all
quantifiers can be "raised" so easily as 3, e.g., no, few, and the like
(although it must be admitted that these quantifiers do not readily
supply pronoun antecedents either); and also that even 3 cannot be
"raised" over every sentence connective (part of the point of "donkey
sentences"). So, the question arises: why not just adopt Kamp's or
Heim's theory? As we mentioned before, we find two basic disagree-
ments with them. First is the non-compositionality inherent in their
proposals. And second, we find that their interpretation of the crucial
"donkey sentences" not to be in accord with our intuitions. (This latter
shortcoming is also in Groenendijk and Stokhof). Since the present
proposal differs from the others most clearly in the interpretation of
"donkey sentences", we now turn to look at them.
III. DONKEY SENTENCES
This issue of "carrying over variable bindings" from one quantified NP

to be used in evaluating an occurrence of a pronoun in another clause
or sentence has been extensively discussed in the literature, usually with
reference to "donkey sentences" such as 7
7. If Pedro owns a donkey, he will ride it to town tomorrow.
The discussions of this phenomenon have, in our opinion, been marred
by a concentration on (what we take to be) unlikely "readings". Here
are seven possible "readings" of (7): 8
7. a. [Deictic Reading] If Pedro owns a donkey, Pedro will
ride [some otherwise specified object, not any donkey
mentioned in the antecedent] to town tomorrow
7. b. [Generic Reading] Pedro has the habitual disposition to
generally ride donkeys Pedro owns to town tomorrow
7. c. [Universal Reading] For any donkey" x, if Pedro owns x,
Pedro will ride x to town tomorrow
7. d. [Specific Existential Reading] If Pedro owns a certain

donkey [e.g., the speaker might "mean" Annabelle], Pedro
will ride it to town tomorrow
7. e. [Non-Specific Existential Reading] There is some donkey
x such that if Pedro owns x, Pedro will ride x to town
tomorrow
7. f. [Definite Lazy Reading] If Pedro owns a donkey, Pedro
will ride the donkey Pedro owns to town tomorrow
7. g. [Indefinite Lazy Reading] If Pedro owns a donkey, Pedro
will ride some donkey Pedro owns to town tomorrow.
The Deictic Reading occurs when the it in the consequent refers to

some item referenced by previous discourse or by the speaker (point-
ing). We shall not discuss this further. The Generic (or habitual)
Reading - which is perhaps unavailable with the adverb tomorrow in
the sentence - predicates some disposition of Pedro. It does not imply
that Pedro necessarily will ride to town, but rather that riding-to-town-
on-his-donkey is something that Pedro is in the habit of doing -
something he generally does. The Universal Reading is often (pre-
emptively) called the "standard" reading (e.g., by LePore and Garson
1983), and many theorists take their task to be to show how to deliver
it (Hintikka and Carlson 1979, Hornstein 1984, Kamp 1981, LePore
and Garson 1983, Heim 1982, Roberts 1987, Groenendijk and
Stokhof 1987, among others 9 ). We find this a pretty implausible
meaning for (7) - for one thing it entails that Pedro will ride every
donkey he owns into town. It might be counter-argued that this
discrepancy arises not from the Universal Reading being a mistransla-
tion of (7), but rather from a Gricean implicature in (7) to the effect
that Pedro owns at most one donkey. However, while a situation in
which Pedro owns two donkeys and rides one into town does intuitively
violate some presumption here, it intuitively renders (7) true aU the
same. Certainly if the speaker had wagered on the truth of (7), he
would feel entitled to coUect, and it is hard to imagine the counter-
claim holding sway, that Pedro after aU didn't ride his second donkey
into town. (We discuss possible implicatures below). The Specifi::
Existential Reading treats the indefinite a donkey more or less as a
disguised name: the speaker has some specific donkey in mind, say
AnnabeUe, does not commit him- or herself to Pedro's ownership of
Annabelle, but claims that if Pedro does oWll Annabelle he will ride it
to tOWll. Such a reading, it seems to us, cannot readily be represented
by any quantificational analysis of the sentence - at least not on the
standard semantics of quantification, which does not permit reference
to any particular donkey (but see Section XII). Semantically such
indefinites behave much like a definite, as argued by Fodor and Sag
(1982). Some theorists, not wauting the entailments of the Universal
Reading, but yet thinking that some "ordinary" quantificational reading
is correct, have suggested that the Nonspecific Existential Reading is
appropriate. They want Pedro to perhaps not own any donkeys, but if
he does own any he will ride one of them to town. But the N onspecific
Existential Reading does not say this. Rather, it is true (quite vacuously)
if there is a donkey (any donkey) that Pedro doesn't own. Possibly this
is a legitimate meaning of (7), but it surely isn't a natural meaning. The
inappropriateness of the Nonspecific Existential Reading shows that a
"quantifier raising" type of analysis is mistaken even in these non-
quantified-subject donkey sentences. lO
The Definite Lazy Reading 11 suggests that we can somehow find a
sui table replacement for the it in (7) by considering the statement made
in the antecedent. Our theorists differ in how they think the replace-
ment is to be found. Partee (1972)12 suggests that there be some
syntactic way to find it; Evans (1977) claims that it can be got by
"semantic" considerations 13; aud Cooper (1979) and Partee (1978)
claim that it is due to "pragmatics" wherein the logical form translation
has a free (pragmatic) variable whose value "will be determined by the
context of use. According to some contexts of use [this pragmatic
variable] will represent the property of being the donkey which [pedro]
owns and thus will correspond to an anaphoric reading of [it]" (Cooper
1979: 84). We have called this (aud the next reading) "lazy" readings as
a way to distinguish the meanings attributed to the donkey sentence it
from deictic or habitual or bound variable readings; aud we call these
accounts "lazy readings" in spite of the fact that each of Cooper, Evaus,
and Partee (1978) are at pains to distinguish their accounts from (what
they call) "the pronoun of laziness account". This shows that what
we call Lazy Readings are not what this term of art has always meant to
each previous investigator. For instance, we do not intend to insist that
a lazy reading be able to be read off directly from the words used.
There is a reason, however, to lump the preceding authors together:
they alI agree that, however it is to be done, we get a reading by
constructing something from the preceding text. And in particular, for

the donkey examples (and this is why it is called the Definite Lazy
Reading), the it is to be read as definite: "the donkey Pedro owns" (if
he owns one at alI). But surely this is going too far ... is (7) false when
Pedro owns two donkeys? - Regardless of how many he rides into
town?
We think not. We do agree that there is some kind of implicature in
(7) which suggests that Pedro owns at most one donkey. But this surely
comes from the particulars of donkey-ownership amongst the kind of
people who have the name 'Pedro'. Typically, think North American
academics, people named 'Pedro' who are in the running for donkey-
ownership are poor farmers who certainly wouldn't own more than one.
This feeling is re-enforced by suggesting that Pedro wilI ride it into
town, rather than putting it in his new horse trailer with three others
and driving into town. That the implica ture comes from such social
facts as these and not from any syntactic-semantic features of the
sentence under consideration can be seen by considering other, similar
sentences:
8. Everyone who has a donkey must donate its services for one
day during the festival.
Does anyone seriously think that, if (8) were ordered by the local
government, wealthy farmers with two or more donkeys could plead
that they were exempt on that ground alone? And against the Universal
Reading, surely these wealthy farmers are not required by this order to
donate the services of ali their donkeys. To us, the order (8) seems to
say that alI donkey owners are required to donate a donkey-day's
service. Consider also
9. If 1 have a quarter in my pocket, I'1I put it in the parking
meter.
Surely I'm not lying if 1 have two quarters in my pocket, especially if 1
put one into the meter. But the Definite Lazy Reading would have it
thus: 1 do have a quarter so the antecedent is true, but since the it in
the consequent in effect means "the one and only quarter in my
pocket", the consequent becomes false. And the Universal Reading
fares equally badly: according to this reading, unless 1 am lying when 1
uHered (9), 1 must empty my pockets 01 alI quarters and put each of
them into the parking meter!
We think that the Indefinite Lazy Reading comes closest to capturing

the intuitions behind these examples, and indeed of an examples of an
indefinite NP being the antecedent of it in cases like the ones
mentioned. For example, we think that in a sentence like
10. Every man who owns a donkey wi11 ride it to town
tomorrow
is true just in case every male donkey-owner will ride at least one of
his donkeys to town tomorrow (again in contrast to Kamp-Heim-
Groenendijk and Stokhof, who require such a person to ride them ali
to town tomorrow). We stiH think there are cases of deictic pronouns,
of generic readings of pronominal sentences, and of bound variable
anaphora. But for the present type of pronoun we think that this
Indefinite Lazy Reading expresses the correct truth conditions (under
certain assumptions about constraints imposed by the context of
utterance), though our own translation (17b below) will not invoke any
ad hoc phrase as the translation of it (such as "a donkey that Pedro
owns"). Although we think that there are some types of readings for
donkey sentences other than the Indefinite Lazy Reading, we think that
the Universal Reading is just plain wrong. What makes it seem correct
in certain cases is its confounding with a "generic" or "habitual" or
"gnomic" understanding of these sentences. But everyone agrees that a
universal analysis of generics/habituals/gnomics is certainly incorrect;
and hence it is incorrect as a reading for donkey sentences. Let us
consider
11. If 1 find a quarter, 1'11 give it to you.
A gnomic understanding of (11) would have the sentence expressing a
rule of (typical) behaviour on my part. We would in such a case wish to
avail ourselves of a Carlson-like Gn operator which has the effect of
attributing to me the habit of quarter-giving-to-you-when-I-find-one.
Note that this "gnomic, rule of behaviour" reading does not imply that 1
will give you every quarter 1 find, anymore than the truth of Birds fly
entails that every bird flies. This gnomic reading is certainly a possible
reading of (11), but not the only one. Perhaps more natural is the
Indefinite Lazy Reading, wherein you desire a quarter and 1 commit
myself to giving you one if 1 find any. But certainly 1 do not commit
myself to giving you every quarter 1 find! Nor aIn 1 attributing a rule of
behaviour to myself - merely that 1 undertake to give you one quarter.
It was mentioned at the end of Section II above that one strategy to

try to get the correct truth conditions for such sentences would be to
"do a further adjustment on the logical form which had been initially
generated". One strategy (which we rejected) was simply to "raise" the
quantifier so as to allow it to have wider scope. Another strategy might
be to perform a "scope widening", wherein existential quantifiers are
changed to universal quantifiers (with an eye towards preserving logical
equivalence). But there is compelling evidence against this attempt to
do such a further alteration on the "logical form" with the adjustment of
the scope of the quantifier. Such a "scope widening" 14 of the existential
quantifier in the antecedent, replacing it with a wide-scope universal
quantifier, would yield the Universal Reading analysis. Even if, contrary
to our claims above, this were the correct understanding of the donkey
sentence, it is stiH not always right to do this "scope widening".
Consider a generic sentence such as
12. Usually, if a cat drops to the ground it lands on its feet.
Scope widening for the ostensibly existential a cat is thwarted by the
presence of the quantificational adverb usually; Le., the sentence simply
cannot be paraphrased as "Usually, every cat is such that if it drops to
the ground, it lands on its feet".1s
One attempt to deal with the difficulty mentioned in the last
paragraph is to translate existentials like a donkey in contexts like (7)
or a cat in contexts like (12) as free variables, to be bound at the
sentence level by an unselective quantifier. This is the approach taken
by Lewis (1975) and his followers (especially Farkas and Sugioka
1983, McCord 1981, Heim 1982). Despite the specific difficulties we
enumerated for the proposals of the first two in our earlier paper, the
appeal of such an approach lies in the prospect it offers of accounting
uniformly for alI generic sentences. The approach is problematic,
however, from the perspective of compositional semantics, since it
interprets indefinites as free variables in some syntactic contexts
(especially in restrictive clauses) and as existentially quantified in
others. 16 For example, a tail in
13. Every dog has ataiI.
would be existentially interpreted, while a bone in (14a) and (146)
would be interpreted as a free variable (with restrictions):
14. a. When a dog has a bone, if often buries it

b. A dog often buries a bone if it isn't hungry enough to
eatit.
Note that in (14b) a bone occurs in the main clause, yet would be
treated "as if" it occurred in the restrictive clause. We will be very
much occupied with this sort of difficulty in Sections XI and XII, which
ultimately confronts any attempt at a compositional analysis. 17
IV. INTRODUCTION TO CONTEXTS, SOME SIMPLE EXAMPLES
In this section we wish to introduce some rudiments of our formal

theory (again, for the details, consult the Appendix) and show how it
operates on the two examples given above. This will involve riding
roughshod over various fine points in our theory, such as its treatment
of modality and tense, but it wilI give the essentials (and symbolism) of
the theory.
Instead of the usual "interpretation and valuation function" we wish
instead to use the notion of a context, and invoke the idea of a context
change. 18 We write contexts using square brackets, as [ ]. Such a context
(completely) specifies values for variables, constants, predicates, indexical
terms, now, yesterday, and the like. Sometimes we wish to consider a
''variant context", for example the result of forcing a variable to take
on some particular value from the domain. Such variant contexts are
written: ~ ]x:d' which stands for the result of (re-)interpreting the vari-
able x so that it is required to denote d, but that alI other features of [ ]
remain the same. Should we wish, for example, to force the context to
talk about some particular time j as being now, we would write [ ]now:j'
Since we are taking now as one of our "indexical constants", this
amounts to saying "let j be now". It we want to say that a sentence is
true in a context [ ], we would say [] = 1. (We will later relativize
truth, and context change, to time and wor1d indices, but avoid this
refinement in our preliminary exposition.) In addition to evaluating the
truth of a sentence in a context, we wish to consider how a context
changes by having just "processed" or "evaluated" a sentence. We use
[ ] to indicate this new context that has resulted from the "processing"
or "evaluating" of sentence in the (old) context [ ]. So the picture of
processing a text 1 2 .. n is that the context continualIy
changes from [] (the initial context) to n(n-'( . .. (,[]) ... ). Along

the way we evaluate each sentence for truth or falsity with respect to
the context as thus far developed, and suitably alter the context in order
to evaluate the next sentence. So, the formal development of this theory
must say how to evaluate a sentence in a context and also say how this
"processing" wilI effect an alteration in the previous context. Since we
wish to look at some simple examples, we must give some account of
those matters that wilI alIow us to treat these examples. We give here
just enough to look at the examples, without considerations of tense
and intensionality, and without showing how the parts we give fit in
with the rest of the theory.
The examples we wish to consider contain indefinite NPs, which we
uniformly translate with an existential quantifier (in our restricted
quantification manner). To evaluate an existentially quantified sentence
for truth/falsity, we appeal to this rule:
[(3x: x)W x] = 1 iff there is a (variant) context [lx:d such that
[x ]x:d = 1 and x[W x ]x:d = 1,
where the order of operations in the last expression is understood to be
(x([ ]x:d (Wx), i.e., x is (re-)interpreted as d first, then the variant
context is transformed by applying x, and finalIy 'II x is evaluated in
the transformed context. Keep in mind that in the original context, [ ],
alI values are specified. In particular, a value for x has been given. For
an existential quantifier which uses x as its variable of quantification,
we will want to find a "re-interpretation" of the context in which x is
made to denote some object which makes the unquantified sentence
true. If we can find such a context then the existentialIy quantified
sentence was true. Note however that we have added a slight embellish-
ment: when evaluating the "matrix" of the sentence (that is, the portion
of the sentence which is not the quantified NP), we alter our context in
a certain way, namely by saying that we have just processed the
quantified NP. Depending on just what the quantified NP was, this may
or may not have any effect on the context in which we are evaluating
our "matrix". We think, for example, that atomic sentences (e.g.,
man(x do not cause a further change in the context except possibly
for the value of certain reference times (to be discussed later) as well ~s
for the value of [now] which arguably should get incremented as we
process a text sentence-by-sentence (w~ shall not consider this slight
complication here); but more complex sentences - that is, more
complex subject NPs (e.g., (man(x) & (3y:donkey(y))owns(x,y)),

which is our rendition of man who owns a donkey) will contribute to a
change in the context. So, that is how we would evaluate a sentence
with an indefinite NP as subject. The context that is generated from this
evaluation is then "passed along" to the next sentence, to be used in
evaluating it. Let us return to our previous example:
15. a. Arnan is at the door. He is behatted.
translated as
15. b. (3x: man(xat-door(x) . behatted(he).
and "disambiguated" as
15. c. (3x:man(xat-door(x)' behatted(x).
We start the evaluation of this "discourse" with the context [ 1, and ask
whether the first sentence is true in that context. According to our rule,
this is
16. [(3x:man(xat-door(x)1 = 1 iff there is a context [ Ix:d such

that [man(x)1x:d = 1 and which is also such that man(x) [at-
door(x)1x:d = 1.
So we try to find some d E D (the domain) which, if it were the
interpretation of x, would make man(x) true (that is, we try to find a
man in the domain). If we succeed, we look at a slightly altered context
- namely, the context in which we just "processed" man(x) - and ask
whether in that context our d that we found can be used as the value of
x in at-door(x). That is, we ask whether this d (which we have already
ascertained to be arnan) is also at the door. Now, we have remarked
that we do not think that atomic sentences cause a change in the
context, and so the context, man(x) []x:d' will just be [L:d' In this
context, which is just like our original starting context except that x is
forced to denote d, we wish to know whether at-door(x) is true. WeB, it
is if the man d we picked is at the door; otherwise it is false.
We now wish to evaluate the second sentence in the context
generated from having already "proces sed" the first sentence. So the
question is, how does a context change when one has just "processed"
an existentially quantified sentence? The answer we propose is this:
(3a : cI> a )'P a[ ] = (cI> & 'P) [ ]a:d'
That is to say, the new context will force a to be interpreted as d, and

there may be further modification by (<1> & W). Since in the case under
consideration <1> and W are atomic, there is no further modification,
and so the new context is just [ ]x:d' It is this new context in which we
wish to evaluate behatted (x). But since x is forced to be a man at the
door, this second sentence is true in this context just in case d, the man
who is at the door, is wearing a hat. (And since this second sentence,
behatted (x), is atomic it will not induce any further context changes
which need to be considered in evaluating a third sentence. 19 )
Let us now turn to our evaluation of our "donkey sentence" 20
17. a. Ii Pedro owns a donkey, then he will ride it to town
translated as
17. b. ((3x: donkey(x))owns(p, x) --+ rides(he, it))
which might be disambiguated to
17. c. ((3x: donkey(x))owns(p, x) --+ rides(p, x)).
We wish to evaluate the truth of this in the context [ ]. Our rule for
conditionals is
[--+W]=1 iff []=Oor<l>[W]=1.
That is, a conditional is true in a context if and only if either the
antecedent is false in that context or if the consequent is true in the
context which results from modifing the original context by having
"processed" the antecedent. Let us suppose the interesting case, that the
antecedent is not false in the original context. As in the previous
example, this means that the new context will be just like the old
context except that x will be forced to denote some d which made
donkey(x) and owns(p, x) be true. So, in such a context is rides(p, x)
true? That is, does [rides(p, x)k" = 1? Well, yes - if the d picked out
before by the evaluation of the antecedent was a donkey Pedro rides to
town. No, otherwise.
V. TRUE-IN-ENGLISH vs. TRUE-IN-A-CONTEXT

The notion of truth (in a context) which was partially explicated in the
preceding section deviates from a more intuitive notion of truth, a
notion that we will labeI "true-in-English". For example, in our discus-
sion of the man at the door, if the initial context had two men at the
door only one of whom was behatted, then the second sentence would
depend for its truth or falsity upon which of these men was "selected"
to be the value of x in the "processing" of the first sentence. That is, our
rules of context-change induced by an indefinite NP arbitrarily pick one
of all the possible satisfiers of the indefinite, and use that one to
evaluate pronomial reference in the following sentences. (Similar
remarks also hold for our treatment of times, which we have not yet
introduced.) Intuitively though, in the imagined case, both sentences
should be judged true - unless there be some other, "pragmatic"
reason to rule out one of the two men as a possible antecedent for he. 21
The position we take here is that in the ordinary notion of "true in
English", the conventions of language (both semantic and pragmatic)
allow for a good deal of leeway in the interpretation of predicates and
terms in any particular context of utterance. In interpreting an utter-
ance, hearers avail themselves of this leeway, i.e., they seek an inter-
pretation that renders it true. In other words, we identify "truth in
English" with the existence of an idealized context that satisfies
conventional constraints on meaning (including constraints imposed by
the context of utterance) in which the sentence is true (in the idealized
sense). So the question of whether the second sentence of (15) is true-
in-English in the imagined situation, amounts to the question of
whether there is an idealized context, admitted by the actual utterance
situation, which renders it true (in the idealized sense). We shall shortly
try to make this more precise.
Consider again the sentences in (15), and suppose we are going to
make a bet on the truth of the second sentence. If we both agree that
the utterance situation is one in which 'he' refers to Adam (who might
not be at the door at all), then the bet amounts to whether Adam is
wearing a hat; that is, as to whether there is an idealized context in
which all objects are assigned either to the set of hat wearers or to its
complement, where this assignment agrees with ordinary beliefs (and so
on) in those cases in which we have clear feelings about the matter,22
and in which 'he' refers to Adam. You win if Adam is in the set of hat
wearers in that context, and I win if he is in its complement.
Other times it is clear that we are using the pronoun in a way to talk
about men at the door. Nonetheless, in some utterance situations, the
choice of referent may be narrowed to the man the speaker intends to
refer to. Again, suppose I utter (15) on the grounds that I answer the
see just one hatless man, and we both agree that it is this person to
whom 1 am referring. 1 might reasonably be said to have spoken falsely-
in-English despite the other, behatted, man at the door whom 1 do not
notice. The reason here seems to be that, in such a case, (15) is being
taken as a report of observations made by the speaker, and so the
situation precludes unobserved men as possible referents of he (or
indeed, even of a man in the previous sentence) in the utterance. If, on
the other hand, there is some aspect of the utterance situation such that
(15) need not be regarded as a report of direct speaker observations, or
if 1 in fact see both men, then the second sentence would be true-in-
English, since there is an idealized context among those admitted by the
actual utterance context that renders it so. So in general we say that in
an idealized context ali meanings/denotations are fully determinate.
And in such a context we are merely to look at whether the denotation
of he is behatted; and if so, then the sentence is "true in that (idealized)
context", otherwise ''false in that (idealized) context". On the other
hand, the notion "true in English at (time) i and (world) w" amounts to
saying that there is an idealized context which (i) satisfies the preceding
discourse (in a way to be spelled out later), (ii) makes the values of the
indexical expressions now, here, etc., be the time and place of speech,
etc., and more generally (iii) is "consistent with English" (which means,
roughly, that the extensions it assigns to constants and predicates at i
aud w, as well as predicates' intensions, are consistent with English
speakers' understanding of the original English terms). So a formula <1>
(translating an English sentence) is true-in-English at i, w - relative to
a preceding text <1>1' <1>2' .. <1>n - if and only if there is a context space
(see Section VII, or the Appendix) containing some context [ ] consis-
tent with English at i, w and this context is such that (<1>1 <1>2 .
<1> n )i, w[ <1> ]i, w = 1.
VI. SOME ELEMENT ARY COMMENTS ON

TENSE AND TIME ADVERBS
We now turn aur attention ta the issue of interpreting tenses and time
adverbs within the intuitive theory we have sketched above. The
proposed treatment will necessarily be incomplete, and unspecific ,at
certain points, even in the more formal account of the Appendix. Here,
though, we merely highlight the salient f6atures that allow us to treat the
generics which we are interested in. The treatment we give involves
two main ideas: (1) time adverbs take narrow scope relative to tense,
and they do not involve any "shifts" in the time index of evaluation (ef.
Richards and Heny 1982); (2) in keeping with most modem theories of
tense, we avail ourselves of a "double indexing" method to simultane-
ously keep track of the time of evaluation and certain reference times,
including "now". Our innovation in (2) is that the index which keeps
track of the reference times is assumed to be a vector or sequence of
times, one of which can be "in focus" during any specific evaluation of a
sentence. We use r for this "reference time vector", whose elements are
created or replaced in a context-dependent way when a sentence
describing an episode is proces sed. At any stage of sentence evaluation,
one of the elements of r, written as r, is "in focus", and the semantic
value of the clause under consideration may depend on the value of r.
We consider the following time adverbs: YEST, AT-THAT-TIME,
THEREUPON. If <1> is an untensed sentence, then YEST(<I, AT-
THAT-TlME(<I, THEREUPON( is a temporal sentence then PRES(<I, gPRES(<I,
PAST(<I, PROG(cI, PERF(<I are (tensed) sentences. PRES serves as
simple present, while gPRES is the "generic present", which can occur
only in conditional and temporal clauses, or in combination with
frequency adverbials. The distinction between PRES and gPRES is
needed if one assumes (as we do) that simple present forces evaluation
at or near [now], the time of speech; for, in generic/habitual present-
tense sentences such as (1), (2), (12) or (14), the cases considered are
clearly not confined to the time of speech. As set down in the
Appendix, the semantics of alI five operators involves a shift in the time
of evaluation to some other, indefinite (existentially quantified) time,
constrained in certain ways relative to the given time of evaluation. In
the case of PRES, the new time must nearly coincide with [now], in the
case of gPRES, it lies in [present], an extended time stretch reaching
into the past, and in the case of PAST, it lies before [now]. The
intended use of THEREUPON and AT-THAT-TIME is as "default
adverbials" in tensed sentences in a narrative. For example, in
(18) The alarm rang. John woke up.
the default adverbial in the second sentence is assumed to be THERE-
UPON, which places a reported event immediately or shortly after a
previously reported event. In
(19) John entered the room. Mary was asleep.

the default adverbial in the second sentence is assumed to be AT-
THAT -TIME, which places a reported state of ajfairs at the end of a
previously reported event, or at the same time as a previously reported
state of ajfairs. (Of course, states of affairs may exist prior to and
beyond the stipulated time; to affirm a state at one time is not to deny it
at others.) We assume that the distinction between predicates describ-
ing events and those describing states of affairs is correlated with
explicit syntactic features (e.g., a dynamic/static pair of features).
We would translate (20a) as (20b):
(20) a. Mary left yesterday.
b. PAST(YEST(leave(m))).
In the evaluation of this formula at time i, the PAST operator "checks"
whether i is the current value of [now] and also whether the embedded
formula holds at some time j before i. Thus there is a "backward shift"
from i to j. But in the evaluation of the embedded formula at j, the
operator YEST merely checks whether j lies within the time interval
designated in the current context by yesterday (viz. the day before now)
... it does nof try to tind a time which is the day before j.
The second idea was the use of the reference time sequence r to
provide temporal connections within and between sentences. As we
remarked, one element of r, denoted r, can be in focus. It we were to
"look inside of" r we would find a sequence such as (il' ~, ... ), where
the boldface ~ indicates that it is "in focus", i.e., it is the value of r.
As we also remarked, the evaluation of a sentence might add elements
to the sequence, and it might change which of the elements is in focus.
We think that we need only consider cases where the new element in
focus is the one immediately to the left of the current one in focus
(which corresponds to "shifting forward in time one notch") or where
the new element in focus is the one immediately to the right of the
current one in focus (which corresponds to "shifting backward in time
r
one notch"). We indicate this by and r, respectively. It might happen
that evaluation of a sentence requires a shift to the right (for example),
but there is no i at that location. We will allow this to happen, and say
that the focus is undetined.
The way this works, intuitively, can be seen by considering sentences
(21):
(21) John entered the room. He called Mary.

The first sentence creates a reference time sequence consisting of the
time of speech, which is in focus, and the time of John's entering. The
second sentence here, translated and disambiguated, is PAST(THERE-
UPON( cal! (i, m))), since there is no explicit time adverb in the English
sentence. The PAST operator, before doing its evaluation as described
earlier, shifts the focus of r back in time, so that the THEREUPON
clause will be evaluated with the time of John's entering in focus.
THEREUPON then places the event it embeds immediately or shortly
after that focal time r. Similarly, the second sentence in
(22) John entered the room. Mary was asleep.
would be translated as PAST(AT-THAT-TIME(asleep(m))); AT-

THAT -TIME affirms the state predication for the moment of time at
the end of the reference time r (which is again the time of John's
entering). This method deals successfully not only with these sentence
pairs, but also with such donkey sentences as
(23) If Pedro bought a donkey yesterday, he sold it shortly
afterwards.
Here, not only is there a "pronomial connection" to be "carried

forward" from the antecedent to the consequent, but also a "temporal
connection" needs to be made which places the selling shortly after the
buying. We deal with this "temporal connection" in essentially the same
way as we outlined above for the "pronomial connection" - the
reference time is brought forward by the context change induced from
having proces sed the antecedent, to be used in evaluating the con-
sequent. A more complex example is considered in the Appendix; for
now we will simply allow the informal intuitive correctness of the
method carry us along to our consideration of generics.
We should also say here, since it will figure in our later treatment,
that we find certain when-sentences ambiguous between a "simple
report of an event" and a "general" statement. If, for example, the if of
(23) were replaced by when, we would say that the "antecedent" could
be taken in either of these two ways: a report of a single event that
happened (perhaps the most natural understanding), and statement
about a series of happenings yesterday of Pedro"s buying a donkey. In
either case we are called upon to evaluate whether immediately or
shortly thereafter he sold it. The difference is whether there is one such
event, ar a series of pairs of correlated events. This ambiguity can be
eliminated by putting a frequency adverb (always, often, ...) either in
the consequent clause ar as an entire-sentence adverb. There are no
doubt other ways of analyzing this phenomenon than attributing it ta
ambiguity. E.g., say that it is "unspecific" as ta how many donkey-
buyings there are, and that the sentence really says that alI buyings -
including the case where there is but one - were shortly followed by a
selling. But we have a somewhat different feeling about the "ambiguity"
in that one reading clearly seems ta say that there was (at least) one
donkey-buying yesterday followed by a selling (of that donkey, shortly
afterward) whereas the other reading seems ta predicate a habit ar
tendency an Pedro's part yesterday to sell-donkeys-he-bought-shortly-
after-buying-them. And this reading does not require him ta sell all his
donkeys.
Depending on the sentence, one ar the other of the alternative
readings can be made more salient.
24. a. When John went ta bed, Mary was asleep

b. When John went ta bed, he wore a nightcap
c. When John was going ta bed, he was wearing a nightcap
d. When John was going ta bed, he wore a nightcap
e. When John goes ta bed, he wears a nightcap
It seems that (24a) is most naturally read as a simple report of what

happened an a specific occasion of John's bed-going, and that (24e) is
most naturally read as a statement of a habitual practice of John's. The
sentences (24b)-(24d) vary in their affinity ta one or the other of these
understandings. Perhaps they are most easily seen as reporting what
happened an a specific occasion of John's bed-going in the past, but
they also can be read as making a statement about what John's habits
were in the past. (24b), for example, can be seen as saying that, there
was a past habit of John's: when he went ta bed he wore a nightcap.
(24d) can also be seen in this way: while John went ta bed (but not
necessarily when he was in it), he habitually wore a nightcap. Once this
understanding has been pointed out, it is easy ta get it for progressives:
(24c) might state a relationship between two of John's past habits - his
habit of going ta bed and his habit of wearing a nightcap - and the
relationship might itself be a habitual one. (24a) now can be seen as

expressing also a habitual statement, although it is such a strange
regularity that we tend to overlook it: it can express that the habitual
regularity of Mary-being-asleep-when-John-went-to-bed used to happen
in the past. (It is perhaps easier to get this when an adverb is supplied:
When John went to bed, Mary was usually asleep.) (24e), on the other
hand, can be seen perhaps as making factual prediction (not a habitual
claim) in a "future present" tense (rather like the "historical present").
We do not insist on this, but it is clear at least in the past and
progressive that such sentences are ambiguous in just the way we
described. In any case, it is the type of understanding of the sentence
expressing a habitual or regular occurrence that we wish to point to,
and (as before in the discussion of donkey sentences) call a "generic
reading" of these (unmodified) when-sentences.
VII. A LITTLE FORMALISM,

AND "LEGITIMATE <P-ALTERNATIVES"
As might have been gathered from the discussion above about "truth in
English" and "truth in a context", there are times when we wish to be
able to consider all the possible new contexts that could be generated
from "processing" a sentence in an old context. This is particularly
salient when we wish to know alI the ways (3x: 'II x) x could have
been true, or all the times at which PAST(W) was true, in order to be
able to talk about each or several of those things or times. The current
formulation of "truth in a context" gives only one way: that way
specified by the context at hand. So, we would like to be able to talk
about other contexts, sometimes. The point of this section is to define
this notion.
A context space is a quadruple (D, W, C, O), where D is a domain
of individuals partitioned into objects 0, kinds K and time intervals 1,
W is a set of worlds, and C is a set of contexts, all of these being non-
empty. An element of C, a context, is a (total) function on the expres-
sions of the language 2. O is a (total) function O: F X 1 X W X C --
C, called a context transJormation Junction, mapping formulas, times,
worlds, and contexts into (new) contexts. When we talk about a
particular O, we will abbreviate 0(<1>, i, w, [ ]) as <l>i, w[ ]; this notation
means the context which results from having "proces sed" or "evaluated"
<1> in the context [ ], at time i and world w, when it is understood that
214 LENHART K . SCHUBERT AND FRANCIS JEFFRY PELLETIER
we are talking about a particular context transformation function. We

intend ~ and all [] E C to meet various conditions, which are spelled
out in the Appendix.
As can be seen from the above definition, the context transformation
function ~ is a function; and therefore, at a time and world, given a
preceding context and a formula it will always generate the same new
context, and thus it will not allow us to consider "other contexts which
could have resulted from processing <l>". When we do wish to consider
variant contexts we appeal to the notion of a legitimate <l>-alternative of
context [] , which amounts to considering a context that could have
been generated by processing <l> in context [] (at i, w) using some
context transformation function ~. Various restrictions apply here: we
want only to consider context transformation functions which render <l>
true in [ ] (at i, w) and we want them to be "legitimate", i.e., to obey the
strictures on a context transformation function (given in the Appendix).
More formally, we introduce a function fs dependent on the context
space S = (D, W, C, ~ >which applies to a formula, a time and a world
of evaluation, and a particular context, to yield a set of contexts that
arise from applying other context transformation functions to that
particular context (with the condition that the evaluation of the formula
had to be true).
fs(<l>, i, w, []) = {[ ]'Ithere is a context space S' = (D, W, C,

~' > such that [ ]' = ~'(<l>, i, w, []) and
[<l>F w = 1 with respect to S'}.
Here the phrase "with respect to S'" is crucial. This is because <l> may
be compound, and the value of one part of <l> may be "conditioned" by
the effect on the context of another part; thus the overall value may
depend on the context transformation function, and hence on the
context space.
Earlier, we outlined two types of context change: changes in variable
assignments due to existential quantification, and changes in the refer-
ence time sequence r due to episodic sentences. Henceforth <l>-
alternatives to a context [ ] may be thought of as differing from [ ] only
in these assignments. (However, a more fully developed theory would
also allow for context-dependence of indexical constants as well as of
predica te extensions and intensions.)
VIII. "REFERENCE ENSEMBLES",

FREQUENCY ADVERBS, AND GENERICS
We mentioned at the outset of this paper that we find one of the most
salient features of habituals and (some) generics ta be their (explicit ar
implicit) reliance an an "ensemble of cases" with respect ta which their
truth is evaluated. We call the "ensemble of cases" which is relevant ta
a particular sentence the reference ensemble far that sentence. As
remarked in the introduction, we think that there are two types of
reference ensembles: ensembles of situations and ensembles of objects.
Ta evaluate correctly and accurately the truth of (certain) generics and
habituals, one must determine what these reference ensembles are. It is
aur view that these reference ensembles - whether of situations ar of
objects - are determined in various ways. One way would be from a
previous-sentence context. For example
25. a. John rarely misses
might have as its reference ensemble a group of situations consisting of
events wherein John is firing a gun at something, and we might have
this made obvious ta us by the presence of a previous sentence like
25. b. John is an excellent marksman.
The sentence
26. a. Baboons farm a protective circle with males an the
outside.
might have a reference ensemble of situations consisting of events
where leopards are approaching monkeys, perhaps given by means of a
previous sentence like
26. b. Most monkeys flee when leopards approach.
Another way reference ensembles may be determined is by presupposi~
tion, either of the VP ar by stress, within the sentence under considera-
tion. For example
27. Cats usually land an their feet
suggests, by means of the presuppositions of land, that the reference
ensemble will be cases of a cat dropping ta the ground. And the
reference ensemble of situations relevant to (boldface is used to

indicate stress)
28. Leopards usually attack monkeys in trees
contains those situations in which leopards attack monkeys somewhere,
and the sentence then says that in most of these situations the attacks
occur in trees. A third way reference ensembles might be determined is
by "characterizing properties" of the subject. For example
29. Bullfighters are often hurt
seems to require as a reference ensemble, a group of situations in which
a bullfighter is participating in a bullfight. And then the sentence says
that many of these are situations in which the bullfighter gets hurt.
Some sentences are ambiguous between a generic (or habitual)
reading and a "capacity" reading. Consider for example
30. a. This car goes 200 kph
31. a. Kim reads German
32. a. Robin rides horses.
In the "capacity" reading, the sentences mean that the car is capable of
going 200 kph, that Kim is able to read German, and that Robin knows
how to ride a horse. This sort of reading involves no reference
ensemble. (Instead, we suppose that it involves an implicit "is-able-to"
operator which modifies the overt predicate.) In the genericlhabitual
reading the sentences mean that the car often goes 200 kph, that Kim
habitually reads German, and that Robin regularly rides a horse. In this
reading the presupposed reference ensembles are:
30. b. This car is being driven.
31. b. Kim is reading.
32. b. Robin is riding (an animal).
The examples considered here suggest that when we evaluate any
generic or gnomic sentence with respect to a given "index" we do so
with reference to an ensemble of situations or of things that may not be
extant at that index, but rather at other (especially temporally dls-
placed) indices. The exact characterization of this ensemble (which
perhaps also involves a probability distribution, as we shall later
speculate) is a problem for "pragmatics" in that this will involve both

linguistic and extralinguistic context. But the formulation of truth
conditions with respect to a given index given such an ensemble is a
problem for semantics. Moreover, it is a problem for semantics to show
how explicit "case adverbials" and restrictive clauses (as in (33)-(50)
below) provide an ensemble with respect to which the embedded
matrix clause is evaluated.
33. Usually, when a cat is dropped to the ground, it lands on its
feet.
34. Sometimes, when John and Marvin play a game of ping
pong, Marvin beats John.
35. Often, when bullfighters are engaged in a bullfight, they get
injured.
36. When a cat is dropped to the ground, it usually lands on its
feet.
37. When John and Marvin play a game of ping pong, John
sometimes beats Marvin.
38. When bullfighters are engaged in a bullfight, they often get
injured.
39. When cats are dropped to the ground, they usually land on
theirfeet
40. When a dog gives birth, it gives live birth
41. Small fish are widespread when big fish are rare
42. Lizards are pleased when the sun shines
43. Around the New Year in Edmonton, it usually snows
44. In emergency situations, flight attendants are usually
effective
45. Cats dropped to the ground always land on their feet
46. Rats crowded together in a small cage are usually aggressive
47. Dogs with/that have blue eyes are (usuClly) intelligent
48. Dogs are (usually) intelligent if/when they haveblue eyes
49. Dogs dislike cats when they (the cats) have blue eyes
50. When cats have blue eyes, they are intelligent.
Following up the suggestions in our earlier paper, we would like to

take as our paradigm cases of habitual/generic sentences ones like
(33)-(40) and (50); these contain alI of the information that appears
relevant to the specification of the reference ensemble in an explicit
if/when clause. So ... what were these suggestions? With regard to
"episodic" or "situation" examples (like (33)-(40)), we would in
essence like to be able to "quantify" over the situations in an ensemble;
that is, we want to be able to talk about what happens in many, Of
most, or alI, or some, of the situations in the ensemble. We furthermore
have in mind that this quantification is potentialIy intensional as well as
extensional especially when there is no overt adverbial, as in (40) and
(50) (i.e., is evaluated with reference to situations in other possible
worlds, as well as the actual one); and we wish to be able to make a
"referential connection" between the entities in the situations quantified
over and the entities referred to in the matrix formula. For example,
with regards to the when <1>, W pattern of sentences like (39) we would
like to represent it in some manner such as the following:
39 '. USUALLY[WHEN (3x: R(x,,u(cats))drop-to-the-ground(x),

land-on-feet(x)].
That is, we alIow the when clause to pick out the reference ensemble,
and then use the quantificational adverb usually that appeared in the
<1> clause to relate the reference ensemble with the main clause.
USUALLY can be thought of as a "quantificational conditional" which
takes two arguments. Intuitively, the first is the ensemble of situations,
and the second is the matrix formula to be evaluated (with respect to
that reference ensemble and in accordance with the evaluation rule for
USUALLY). Note, incidentalIy, that a frequency adverb can syntacti-
calIy occur explicitly at the sentence level over the entire conditional
sentence, as in (33)-(35), or it can occur at (apparently) the VP level
in the "matrix" sentence, as in (36)-(38). SemanticalIy speaking, we
will treat these two type of modification by a frequency adverb
identicalIy: we treat the frequency adverbs as applying to the entlre
"when-sentence", as manifested in (33)'--(35). (In the logical form we
eventualIy propose, we omit the one-place operator WHEN, since its
effect on its argument can be assimilated to the effect of the two-

place operator USUALLY on its first argument, and then we move
USUALLY from the above prefix position to infix position, rewriting it
as --+ u to emphasize its interpretation as a quantificational conditional.
But we retain it here for intuitive clarity.) Note however, the "dangling"
variable in the matrix formula (the land-on-feet(x) clause), which is not
in the scope of any quantifier. As with the donkey sentences, we would
like to think of this variable as implicitly bound by the "quantifier
phrase" (i.e., the USUALLY-WHEN combination). That is, we have in
mind that we will "iterate through" the situations satisfying the WHEN-
clause to find choices of x which render the restriction true, and we will
use these choices to evaluate the matrix formula. Just as in the donkey
sentences, we would wish to say that, although the scope of the (3x)
quantifier is only the WHEN clause, we can use information from its
evaluation to do further evaluation of the matrix clause. The details of
this sort of analysis, in Section IX, form the centre-piece of our paper.
We would also want to represent the reference ensemble of objects
in a similar manner. For example, (50) might be represented as (50)':
50'. G[WHEN3x: R(x, ,u(catshave-blue-eyes(x1
[intelligent(x )1.
Here we have used the WHEN clause to pick out groups of cats-that-
have-blue-eyes. The Ghere is a two-place quantifier (differing from
USUALL Y in that it has nomic import) meaning "generally" or
"characteristically" or "as a rule" which gathers together those cases
described by its first argument for use in evaluating the sentence given
as second argument. In this type of case, however, we are apparently
not looking at periods of time in which there is a cat-with-blue-eyes, but
rather at cats-with-blue-eyes themselves - that is, at the objects. We
will attempt, in Section X, to give an account that unifies these
"temporal" and "atemporal" uses of restrictive when clauses, but it is
highly speculative. In essence, it claims that despite appearances, the
quantification in sentences like (50) does make reference to times.
While it may ultimately prove best to falI back to a "schizoid" account
that distinguishes temporal and atemporal when clauses (with a
Carlson-like analysis of the atemporal examples), our further investiga-
tions in Sections XI and XII instead lead us in the direction of Lewis-
like uns elective quantifiers. However, the appro'ach we sketch avoids
radical non-compositionality, retains many of the features of our
analysis of episodic generics, and is coupled to a view that the correct

analysis of generics must appeal to structured meanings of some sort.
IX. PUTTING IT ALL TOGETHER: EPISODlC GENERICS
As suggested in the last section, we shall take, as our basic generics/

habituals, such sentences as
36. When a cat is dropped to the ground, it (usually) lands on its
feet.
39. When cats are dropped to the ground, they (usually) land on
their feet.
51. When a cat is eating, it is happy
52. When John is driving, he drives fast.
In the terminology of Krifka (1987) these are I-generics/habituals. Of
course our treatment of (Krifka's) D-generics will be different. Such
singular generics as
53. The dog is common
54. Water is scarce
55. !ce cream comes in many different flavours
(and others) will receive a representation wherein the predicate directly
applies to a kind. It should be noted that there are "mixtures" of the
two, such as (41) in the last section and the following. (Krifka's
taxonomie readings, mixed with I-genericity; the example is adapted
from one of Godehard Link's).
56. When a species is well-attuned, it usually survives.
Here we are making an I-generie statement, but the "instances" we are
talking about are themselves kinds. So, this is a generic statement
(hence it needn't be universally true, but rather expresses a "law" true
about instances of is a species); but the instances are kinds, and so what
is predicated of these instances is itself a "kind level" property. We
intend our theory to accommodate these types of generie statements
also. "-
As remarked in Section VI (on tense), the when <1> adverb clause,
when applied to a sentence '11 without explicit frequency adverb, is

ambiguous: on the one hand, it might pick out a particular time interval
at which <1> is true and claim that A T -THAT -TIME('I1) (or THERE-
UPON ('11)) is also true, or it might be used "generically" to pick out
(aH) times at which <1> is true and claim that at most of these times A T-
THAT-TIME('I1) or THEREUPON('I1) is also true. In our representa-
tion this ambiguity is manifested by whether there is or isn't supplied
the default frequency adverb G ("generally" or "characteristically").
We have already illustrated how we wish to treat the "simple report
of an event" reading of such sentences (above, Section VI, on tenses).
Here we concentrate on the other, generic, reading. Our proposal is to
translate sentences of the form
57. B, when <1>, '11
- where B is a frequency adverb, and when- is one of our when-
adverbials - as
58. [<1> -+ B '11].
That is, we introduce into our logical form language a series of
"quantificational conditionals" (one for each of our frequency adverbs,
supplemented by G as default). The rules of evaluation for such
formulas are (schematically - we associate always with "aH", usually
with "most", etc.):
59. [<1> -+B 'I1]i,w= 1 iffforali/most/many/some/no ...
[]' E fs(, i, w, []), ~'I1]'i, w= 1.
Here we see the use of our "legitimate -alternative" function fs. Its job
is to find alI the "alternative ways" (relative to the initial [] and the
given i, w) that the when clause might have been true by considering
other objects which satisfy the when clause and by considering other
times at which some other objects satisfy the when clause. If in B-many
of these legitimate aIternatives (whatever B happens to be), the '11-
clause is true also, then the whole sentence is true (in the initial world,
at the specified time interval of the original context). Again, one might
note the similarity to Heim (1982: 190f), aIthough we here always
translate a singular indefinite NP with an existential quantifier. Part of
the reason we can do this is that we are in effect treating quantificational
conditionals as adding another le vei of "iterating" through bindings of
indefinites an tap of the "iteration" (or "search" for a suitable binding)
carried out by the existential quantifier. Thus, there is no need to

suppress the existential import of indefinites within the restrictive clause
of a generic, as in other approaches. Another reason lies in the
mechanism we have developed for "carrying forward" both values of
indefinites and event times from one clause to the next. In conjunction
with our semantics of quantificational conditionals (59), this supplies
the required "referential connections" between antecedent and con-
sequent without converting indefinites to free variables shared by
antecedent and consequent.
The correctness of our method for simple I-generics is easily seen at
an intuitive level (we will also supply formal details for one example).
Consider, for instance
60. Usually, when a cat is eating, it is happy.
We as sume that the only form of the present tense possible in a clause
with a frequency adverb (operating on the S or VP) is gPRES. We also
as sume that the tense of a when clause and the main clause it modifies
must coincide. Then it follows that both clauses of (60) are in gPRES
tense.
So, the legitimate alternatives of the antecedent, when a cat is eating,
will contain every possible cat, and every possible time within the
extended present (i.e., within [present]) such that the cat is eating at that
time. This is because there are context transformation functions
(different from whichever one is "given") which will "select" any cat-
time combination that "shows" the antecedent to be true. Each alterna-
tive context thus generated will encode a choice of cat as a value of the
variable quantified in a cat, and a corresponding choice of eating-time
as value r, the focal reference time. (If the sentence were past tense, the
eating-time would be encoded as the reference time "to the right of" -
i.e., a step backward in time from - r; this is because the focus at the
end of processing a clause is always [now], and this lies "to the
immediate left of" a simple past reference time.)
The consequent would be translated with a gPRES operator as well
as a default A T -THAT -TIME adverbial, and the pronoun would be
disambiguated to the variable quantified by a cat, say x, in the
antecedent. When this is evaluated in one of the alternative contex;ts
generated by the antecedent, gPRES first "checks" whether x (as
determined in that context) is happy at some time within the extended
present, and for such times, AT-THAT-TIME "checks" whether they
are at-the-end-of, or coincident with, the time of r. If such a time can

usually be found, for the alternative contexts generated, the sentence is
true.
It might also be remarked here that, in the case of these sorts of
simple generics/habituals with no indefinites in the consequent, there is
no distinction between "truth in a context" and "truth in English".
Recall that the latter amounts to "truth in some (legitimate) context";
but in the case of generics that are true, we have found al! (legitimate)
contexts with respect to the when clause and evaluated the consequent
as some frequency adverb-quantification of them alI. So the generic
sentence is true in context [ ] because its consequent is true in certain
alternative contexts. But this result is independent of which context we
start with (so long as we start at the same time and world).
Our basic account satisfactorily handles many other examples of this
ilk, belonging to a certain group of (I-generic) generic sentences. For
example:
61. When a cat drops to the ground, it usually lands on its feet
62. When Tabby drops to the ground, she always lands on her
feet
63. When Fritz speaks, he never speaks German
64. Often, when a bullfighter participates in a bullfight, he is
injured
65. When Pedro owns a donkey, he usually beats it
and so ono The crucial points involved in determining which ones our
basic account handles are: (1) The sentence should have a When <1>, 1p
structure with or without an explicit frequency adverb, (2) There should
be a name or indefinite NP in the when clause but none in the
main clause (at least none that should be "iterated over" by the
quantifying adverb), and (3) The when clause should express an
"episodic" predication.
We will shortly come back to generics that do not exhibit one or
more of these three properties, but first we will work out in some detail
the semantic evaluation of one of the above serVences, (65), to show
how our formal mechanism works. This sentence is translated into our
logical form as (gPRES3x: Dx)O(p, x -+ U gPRES(AT-THAT-
TIME(B(p, X. We assume that B, the translation of beats, incor-

porates a dispositional or habitual operator, so that it can be read as "is
disposed to beat" or "beats habitually". (Obviously, we could have
avoided this minor complication by using, say likes instead of beats.)
Since this is non-episodic, the appropriate default adverb is A T -THAT-
TIME.
Now, [gPRES(3x: Dx)O(p, x ~ v gPRES(AT-THAT-TIME(B
(p, x)]i' W = 1
(1) iff for most [ ]' E fs(gPRES:lx: Dx)O(p, x, i, w, []),
[gPRES(AT-THAT-TIME(B(p, X) ]'i, w = 1; Le.,
(2) iff for most [ ]' such that for some j during [present],
[ ]' E fs:lx: Dx)O(p, x),j, w, [ ]) and
[(:Ix: Dx)O(p, xW' W = 1,
[gPRES(AT-THAT-TIME(B(p, X)]'i, w = 1; Le.,
(3) iff for most [ ]' such that for some d E D and
for some j during [presentl,
[]' E fs(Dx & O(p, x),j, w, [lx:d) and
[D(x) & O(p, x)];:~ = 1 and [(:Ix: Dx)O(p, x)]i' W = 1,
[gPRES(AT-THAT-TIME(B(p, X)]'i, w = 1; i.e.,
for some j during [present],
[ ]' E f~(Dx & O(p, x),j, w, [ ]x:d) and
[D(x) & O(p, x)];:~ = 1,
[gPRES(AT-THAT-TIME(B(p, X)]'i, w = 1; i.e.,
(5) ifffor most []' such that for some d E D and
[]' E f~( O(p, x),j, w, r ]x:d, r:j) and
[D(x) & O(p, x)];:~ = 1,
[gPRES(AT-THAT-TIME(B(p, X)J'i, w = 1; here the
reduction of the conjunction (D(x) & O(p, x to O(p, x),
in the first argument of fs, preserves the equivalence since
the last equality entails [D(x)];:~ = 1, as required by rules
(8b) and (6b) of the Appendix)
Now the remaining atomic clause O(p, x), like the clause D(x) just
eliminated from the first argument of fs, merely sets r to j in the

context [ ]x: d, r:j' But ris already j, so the equivalence (5) amounts to
[ ]' = [ ]x:d,r:j and [D(x) & O(p, x)]~:':t = 1,
[gPRES(AT -THAT -TIME(B (p, x)]' i, w = 1; Le"
(7) ifffor most []x:d,r:j such that d E D,j during

[present], and [D(x) & O(p, x)t:':t = 1,
[gPRES(AT-THAT-TIME(B(p, x)]~:~I, r:j = 1; Le"
(8) ifffor most []x:d,r:j such that d E D, j during
[present], and [D(x) & O(p, x)]~:':t = 1,
i = ~now] and [AT-THAT-TIME(B(p, x];':d, r:j = 1
for some k during [present]x:d, r:j (= [present1); Le"
(9) iff for most [ ]X:d,r.j such that d E D, j during [present],
and [D(x) & O(p, X)];:~I = 1,
i = [now] and [B(p, x)];':d,r:j = 1 and k = end Of[r]x:d,r:j
for some k during [present]; i.e.,
(10) ifffor most pairs (d, j) such that d E D,j during [present],
and [D(x) & O(p, x)]~;~ = 1,
i = [now] and [B(p, X)];':d,r:j = 1 for k = end ofj.
So this says that the original sentence is true just in case for most
individuals d and time intervals j within the extended present such that
d is a donkey that Pedro owns at time j, Pedro beats d habitually at
time k = end of j. (Note that if j is an instant, k = j.) This is just what
our previous intuitive explanation had told us we wished; and we take
this formal confirmation of our intuitions to be evidence that this way
of looking at "contexts" and "reference ensembles" satisfactorily captures
the truth conditions for generic sentences of this type. (We will have
something more to say about the exact meaning of "most" later on.)
There is one objection, however, that can be raised to these truth
conditions. They require Pedro to beat most of the donkeys he owns at
any given time, not only to beat some donkey at'most times at which he
owns any. While the former is a possible, and natural, reading of the
sentence, so is the latter. As in the case of non-generic sentences, this is

particularly clear if one replaces beats it in (65) by rides it to town, or if
one considers generic versions of "parking meter" sentences, such as II
I have a quarter, I always put it in the parking meter. We will eventually
suggest, in Section XII, that the way to obtain the two readings is to
associate with any quantificational conditional, such as --> u, an addi-
tional argument supplying the variables it "controls". If only time is
controlled by --> u, then only temporal alternatives are generated, and
only one donkey need be beaten. If both time and the donkey variable
are controlled, we obtain the reading at hand. However, we set aside
those tentative revisions of our theory for the time being, assuming
instead that alternatives are generated as described.
Finally, we should point out again that if (65) had lacked a frequency
adverb, we would have used --> G as "default" conditional. According to
clause (16) in the Appendix, --> G quantifies over (or statistically
samples) "nearby" worlds, whereas the other quantificational condi-
tionals quantify over (temporally dispersed) cases within the given
world only. (We do not want the rather approximate truth conditions
in clause (16) to be viewed as committing us to a non-intensional
semantics for quantificational conditionals other than --> G relative to
other conditionals, and should perhaps treat al! of them as intensional.)
We now turn our attention to episodic generics that involve a bare
plural as subject of the when-clause (so that they do not manifest the
second of the three characteristics we described as central to our
treatment). Our proposal on these is less than firm; stiH, we think that
there is considerable plausibility in our attempt to treat them. An
example of this type of generic sentence is
66. When cats drop to the ground, they usually land on their
feet.
(We will ignore certain questions involving plurality, such as how many
cats are involved in each situation, and why, in this case, we alI believe
that individual cats are dropping to the ground. See the Coda (on
plurals) to this paper for further remarks.) Our suggestion for these
generics that have an episodic when clause but a bare plural in the
antecedent follows our suggestions in the earlier paper (Schubert and
Pelletier, 1987). We assume that the "pragmatic component" which
does such things as "disambiguate pronouns", will also apply certain
meaning postulates relating episodic formulas about kinds to episodic
formulas about realizations of those kinds (i.e., objects); hose realiza-
tions will then be available as referents for anaphoric pronouns. So, for
example, we would propose a meaning postulate such as
(MPa). For P an episodic object-Ievel predicate and k a kind,
D(P(k) - (3x: R(x, kP(x.
Such a postulate is needed in any case to provide pronoun referents in
texts like
67. Cats dropped to the ground. They landed on their feet.
Just as in this text, the postulate can provide the existentially quantified
variable needed in (66) to allow interpretation of they in accordance
with our method for evaluating generics with an indefinite in the
antecedent. 23 Thus we would translate (66) as (with D for "drops to the
ground", and L for "lands on its feet" - ignoring problems with the
embedded pronoun):
66. a. gPRES(D(,u(plur(cat ..... u
gPRES(THEREUPON(L(they)
and then apply postulate (MPa) to yield (with anaphoric they"resolved"
to x)
66. b. gPRES3x: R(x, ,u (plur(catD (x ..... u
gPRES(THEREUPON(L(x)
which, modulo certain facts about the interpretation of the plural kind,
cats, appears to give just the correct truth conditions, when combined
with our semantic interpretation method. The sentence is true just in
case for most collections c of cats and times i within the extended
present such that c drops to the ground at time i, c land on their feet at
a time j within the extended present, where j is either immediately or
shortly after i.
X. THE PROBLEM OF NON-EPISODIC GENERICS
Recall that the third characteristic of those generic sentences which are
central to our treatment is their episodic nature. We now examine the
possibility of extending our treatment to cove non.:episodic generic
sentences. We will find that some extensions in this direction are
possible, by regarding certain non-episodic sentences as "implicitly"
episodic. However, we will ultimately express scepticism about. this
approach, after considering further problems encountered in the attempt

to develop a unified, compositional account of generic sentences with
and without restrictive clauses (see Section XI).
As in our treatment of episodic generics, let us begin with an
example involving a when clause with an explicit singular indefinite:
68. When a cat has blue eyes, it is usually intelligent.
We would like to treat these sentences in a manner that parallels the
treatment of episodic generics, in view of their parallel syntax and
analogous meaning (only, with ensembles of individuals taking the place
of ensembles of situations). For (68), this seems technically straight-
forward; we take the logical form to be
68 '. gPRES3x: cat(xHBE(x -+ u
gPRES(A T -THAT -TIME(intelligent( x)
(where HBE is an abbreviation of "has blue eyes" in which we have
suppressed any quantification over "eyes"). Much as before, the sen-
tence is true just in case for most cats x and times i within the extended
present such that x has blue eyes at time i, x is intelligent at i.
A reasonable question about this interpretation is whether it even
makes sense to talk about quantification over time intervals here. While
quantification over time intervals (or episodes) seems clearly appro-
priate in the episodic generics, shouldn't the quantification in (68) be
simply over blue-eyed cats? Or to put the question in a slightly different
form, is it even coherent to talk about "most" time intervals during
which a cat has blue eyes, in view of the fact that these time intervals
can overlap and include each other, and that there are inJinitely many
of them (assuming time to be non-discrete)?
In addressing this question, we should point out that the interpreta-
tion of "most time intervals" is potentially problematic even in clearly
episodic habitual sentences such as
69. WhenJohn has a headache, he is usually irritable.
Here too, there are infinitely many overlapping time intervals during
which John has a headache, even if (intuitively) there is just one
headache-episode. Indeed, it seems to us that the puzzle illustrated by
(68) carries over to aU the episodic examf'les we have considered: to be
able to evaluate the truth of such sentences, one must have some
systematic or random method of "enumerating" or "sampling" the cases
.
over which one is quantifying - one cannot merely "consider the

domain, simpliciter". This is especially salient when the quantification
is done with one of the "proportional" or "comparative" quantifiers
(usually, rarely, often, half the time, etc.). For example, one might
have in mind some systematic or random method of enumerating blue-
eyed cats kept as pets in North American homes (or encountered in
one's personal experience, depending on the context, one's degree of
"objectivity", etc.) and determining the proportion of intelligent cats
among those surveyed. Presumably the sampling would not be confined
to the moment of time at which (68) is uttered, but would also probe
past times (conceivably, according to a probability distribution that
attaches greater weight to the immediate past than to the remote past).
In addition, if one wishes to read (68) as nomic, the sampling would
have to extend to "nearby" possible worlds. An explicitly statistical
interpretation of quantificational adverbs has been outlined by qvist et
al. (1979); and Bacchus and Schubert (1987) have proposed a prob-
abilistic extension of first-order logic which also allows representation
of certain generics as explicit conditional probability statements. We
shall not explore these possibilities here, but will simply assume that
phrases such as "for most cats x and time intervals i such that x has
blue eyes at i", or "for most persons x and time intervals i such that x
has a headache at time i" can be explicated in a technically satisfactory
way. On that assumption, we claim to have shown that the truth
conditions of non-episodic generics like (68) can be treated as identical
to those episodic generics we have already considered.
If a temporal interpretation seemed at first inappropriate for (68), it
will seem even more inappropriate for a sentence such as
70. When a number is a prime power of two, less one, it is
usually squarefree.
Unlike any of our previous examples (and like some of Lewis's, 1975),
this sentence is about entirely atemporal entities, and so it would
appear that time should not enter at alI into its interpretation. But while
this may be the correct view for a mathematician to take, it is arguably
not the correct view for a linguist or philosopher of language to take.
For in the first place, the semantic evaluation of a sentence like (70) is,
presumably, tied to a particular time index (or several such indices, as
in many theories of tense). In other words, despite the atemporal
character of numbers, we need to be able to talk about the truth value
of particular predications about them at particular times. Secondly, (70)

cannot simply be a statement about the proportion of squarefree
numbers among the "Mersenne numbers", because porportions are not
well-defined for infinite sets. However, proportions (or at least their
limit sups and infs) are well-defined for enumerations of infinite sets in
some order, as weB as for probability measures over fields of (finite or
infinite) sets. So the interpretation of (70), just like that of (68) and
episodic sentences, requires some underlying notion of systematically
enumerating, or randomly sampling, the individual cases at issue. In
particular, (70) might be interpreted with respect to an enumeration of
Mersenne numbers in their natural order (though in that case the
limiting proportion of squarefree numbers may be O, making the
sentence false), or with respect to some probability measure that assigns
relatively high probability to relatively small numbers (perhaps making
the sentence true).
The probabilistic interpretation is not implausible, if one considers
that informal generalizations like (70) (or like any of our episodic
generics) are, after all, based on the individual or collective experiences
of a human community. Experiences pertaining to numbers, just like
experiences pertaining to people and cats, include many encounters
with particular instances of them, and these may form the basis for a
subjective probability distribution (or something idealizable as such)
over instances likely to be encountered "in practice".
That some notion of sampling over an extended stretch of time is
involved even in non-episodic generic sentences is also suggested by the
foBowing anomalous examples:
71. a. ?When a number lies between 114 and 126, it is never
prime.
b. ? When a cat is among those that John presently owns, it
is usually intelligent.
Here the difficulty in associating a temporally extended sampling
process with the quantifying adverb seems to hinder their inter-
pretation.
In view of these considerations, we propose to translate (70),
consistently with our treatment of episodic sentences, as
70'. gPRES3x: N(xP(x --. u gPRES(Q(x.
This says of the ensemble of instances/times-encountered-in-practice of

prime powers of two, less one, that they are usually squarefree.
Venturing further, we now turn to generic when sentences in which
the when clause both contains a bare plural and does not have an
episodic verb phrase, such as
72. When cats have blue eyes, they are usually intelligent.
At the end of the last section, we were able to extend our "basic"
account to deal with some generic when sentences containing bare
plurals. In doing so, we appealed to a meaning postulate (MPa),
applicable to episodic predications about kinds, to provide realizations
of those kinds as referents of unbound pronouns in the main clause.
This solution is not so readily available in the present case, since a non-
episodic predicate like have blue eyes does not easily induce an
existential reading on a bare plural subject. (This was part of "The
Fundamental Intuition" of Schubert and Pelletier, 1987). Thus (MPa)
cannot reasonably be invoked for (72). Can we perhaps invoke another
meaning postulate relating kinds to their realizations, which applies to
kinds within non-episodic constructions like the sentence complement-
ing when in (72)? A plausible candidate (borrowed, like MPa, from
Schubert and Pelletier 1987) is
(MPb). For k a kind and
P E {white, loyal, soft, blue-eyed, four-Iegged, ... }'
D(P(k) .... Dn(My: R(y, kP(y
where M encodes a notion roughly verbalizable as "most of the
relevant", and o n is a weak, necessity-like modal operator, roughly
verbalizable as "it is inherent in the nature of things that ...".24 Let us
apply (MPb) to our representation of sentence (72):
72'. gPRES(HBE(,u(plur(cat -+ u
gPRES(AT-THAT-TIME(I(they)
to yield
73. Dn(gPRES((Mx: R(x, ,u(plur(catHBE(x) -+ u
gPRES(AT-THAT-TIME(I(they).
Is this the logical translation we are seeking? Unfortunately, we must
discount it as such, for several reasons. First, in order to evaluate the
consequent of this conditional by our method, we need to know what

context-dependent "binding" x should receive as a result of "process-
ing" (Mx: <P)'P - but we have no good answer to this. Second, to use
our method, we need to know what alternatives are generated by o n <P
- but again, we have no good answer to this. Third, the M ("most")
quantifier in the antecedent might be seen as an embarrassment if
usually in (72) were replaced by, say, rarely. The new variant of (73)
would then say something like "When most cats are blue-eyed, they are
rarely intelligent." And fourth and most crucialIy, (MPb) is expressly
designed to ensure that a sentence like CaIs have blue eyes will have the
same truth conditions as CaIs generally have blue eyes, i.e., it expresses
a nomic, "quasi-universal" generalization. But then (73) must have the
same truth conditions as
73 '. When cats generally have blue eyes, they are usually
intelligent.
But - though this is a possible reading of (72) - it is surely not the
interpretation of (72) we are trying to capture. (Instead, we wish to talk
about individual cats, saying of each of them that if they have blue eyes
then they are intelligent; and then asserting that such a statement is
generalIy true). The problem here is how to introduce a variable on
which to "hang" realizations of a kind, so that we can use our method
of "carrying forward previous quantificational bindings".
One more approach, which alIows (MPa) to be brought to bear
indirectly, is worth mentioning before we confront the most serious
problem of alI for our theory, that of interpreting when generics with
indefinites in the main clause (Section XI). This approach is suggested
by the folIowing rather puzzling phenomenon. In many cases where no
episodic interpretation is available for a clause in isolation, the when
construction can induce an intuitively episodic reading:
74. al. People are overweight
a2. When people are overweight, they often diet or exercise
b 1. Dogs have fleas
b2. When dogs have fleas, they should wear flea collars
c 1. Wisdom teeth are a nuisance
c2. When wisdom teeth are 'a nuisance, it is best to have
them pulIed.
(As usual, if could readily be substituted for when in these examples).

From a compositional-semantics point of view, it should not be possible
for one and the same clause to acquire a meaning in one syntactic
context which it entirely lacks in another.
Though we will take a different tack on this apparent non-composi-
tionality in Section XII, we draw the "obvious" conclusion here: we are
not dealing with instances of one and the same clause in the (1) and (2)
examples of (74). In fact, this move is not implausible from a syntactic
point of view, if we assume the following sort of structure for when
adverbials:
75. a. ADVL[when] -. when S/ADVL[temp].

I.e., we take the embedded clause to have a time-adverbial gap, where
this gap originates in the verb phrase of the clause:
75. b. S/ADVL[temp] -. NP[subj] VP/ADVL[temp].

Indeed, the oddity of examples with extra time adverbials, such as
76. ? When dogs barked at that time, the burglars fled
may lend support to this analysis (though there may be a semantic

explanation).
Once we have "breached" the embedded clause, placing a gap in the
VP, we can easily ascribe a translation to it differing from that of the
gapless clause. In particular, we can arrange for the translation of the
sentence predicate within the translation of the when clause to be
modified by an operator, E, which is absent from the translation of the
corresponding gapless clause. For example, the translation of People
are overweight in (74a1) would be gPRES(O(,u(plur(person, while
in (74a2) it would be gPRES(E(O(,u(plur(person)))) (at least, after its
incorporation into the when adverbial; this may involve A-conversions
whose details need not detain us here).
How does the presence of E help us? Suppose that we can assign a
formal meaning to E which makes E (P) intuitively express something
like "is/are exhibiting (enjoying, manifesting) the property of being P".
This is not quite an identity operator - the idea is that this should
convert a non-episodic to an episodic predicate; i.., we want to regard
77. People are exhibiting the property of being overweight
as implicitly being about "sm people", rather than about "people in

general". Accordingly, the E-operator would allow us to apply (MPa) to
exactly the same effect as in episodic when sentences with a bare
plural in the when clause. AlI we need to stipulate is that (MPa) applies
not only to the sorts of episodic predicates already indicated, but also
to non-episodic predicates operated upon by E.
So, returning to (72), the translation based on a "gappy" when clause
would be
72". gPRES(E(HBE(,u(plur(cat))))) .-. u
gPRES(A T -THAT -TIME(I(they)))
and after application of (MPa) and disambiguation of they,
72 III. gPRES3x: R (x, ,u(plur(cat))))E (HBE(x)) .-. u
gPRES(AT-THAT-TIME(I (x))).
Assuming that E (P) has the same truth conditions when applied to an
object as P, this translation quantifies over blue-eyed cats (cats
exhibiting the property of having blue eyes) as desired.
In principle, this approach also covers bare-plural versions of
sentences like (70) about atemporal entities:
70". When numbers are prime powers of two, less one, they are
usually squarefree.
However, we would probably be pushing our "episodic" strategy too
far, if we were to claim that this sentence is about numbers "exhibiting
the property of being prime powers of two, less one (at particular
times)". We will indicate possible directions for a more satisfactory
analysis after considering further obstacles to a uniform, compositional
analysis.
XI. FURTHER PROBLEMS
A broad class of episodic and non-episodic generics we have so far

carefully excluded from formal analysis is illustrated by sentences
(78)-(81):
78. A cat usually lands on its feet when it drops to the ground
79. Cats usualIy land on their feet when they drop to the ground
80. A cat is usually intelligent if it has blue eyes

81. Cats are usually intelligent if they have blue eyes.
These (like many examples mentioned in Section VIII) will be recog-
nized as of the type considered in Carlson (1979), though Carlson
confined his analysis to atemporal iflwhen sentences like (81). In
contrast with alI of the sentences we have so far attempted to analyze,
these generics contain singular indefinites or bare plurals in the main
clause. It is the possible instances, or realizations of these main-clause
indefinites that are intuitively involved in the "cases" over which the
frequency adverb (or an implicit generally) quantifies. This remains true
if the ij/when clauses are fronted, or deleted.
There is a tantalizing similarity between pairs of sentences like (82a)
(which is (78) with the when clause fronted) and (82b) (previously
(36:
82. a. When it drops to the ground, a cat usually lands on its

feet
b. When a cat drops to the ground, it usually lands on its
feet.
Not only are they synonymous, but even the structural differences seem
intuitively minor. Yet the variable locus of the indefinites (or of the
corresponding pronouns) presents a larger obstacle to a unified analysis
of generic sentences than either the episodic/non-episodic (temporal!
atemporal) distinction or the singular indefinite/bare plural distinction.
Let us amplify.
We were able to suggest some fairly plausible ways of extending our
basic analysis, centered around episodic when sentences with singular
indefinites in the when clause, to cover non-episodic generics with
singular indefinites, episodic generics with bare plurals, and even non-
episodic generics with bare plurals. Conversely, it seems to us that
Carlson's analysis could have been extended to deal with episodic
sentences such as (79), and perhaps also their singular-indefinite
analogues, such as (78) or (82a). For (78), a Carlson-like analysis
would have formed a kind, cats that drop to the ground, and predicated
usually land on their Jeet of that kind. While there are some difficulties
concerning the formal interpretation of cats thal drop to the ground
(are the realizations of this kind objects or stages?), these do not seem
236 LENHART K. SCHUBERT AND FRANCIS JEFFRY PELL E TIER
insuperable. Similarly, a Carlson-like analysis of (78) would have

treated a cat as kind-denoting, and indeed Carlson (1977a) had a
specific (though in our opinion unsatisfactory) proposal for doing this
(Schubert and Pelletier 1987). The analysis of (79) would then have
carried over to (78).
But neither our approach nor Carlson's seems capable of bridging
the gap, in one direction or the other, between generics like (82a), with
a main-clause indefinite, and those like (82b) with a when-clause
indefinite. Our own procedure, applied to (78), (82a), or (79), requires
a "pragmatic component" to find a suitable referent for the pronoun in
the when clause. But what could it be? One thing is certain, it cannot be
simply a replacement by the NP of the main clause, for the sentences
78'. A cat usually lands on its feet when a cat drops to the
ground
79'. Cats usually land on their feet when cats drop to the ground
quite obviously say something different from (78)-(79). On the other
hand, our method of translating the pronoun as a variable and allowing
it to "pick up" its reference from previously quantified sentences does
not seem to work either. With regard to (78), were we to have
83. gPRES(D(x ~ u gPRES3x: cat(xL(x

then the x in the antecedent could take on any value that it may have
had in a previous sentence (perhaps constrained by the fact that the
initial pronoun was neuter). It instead we were to somehow use the
binding of x acquired in the main clause, then we would be considering
an x which is a cat that lands on its feet. But this is incorrect also: we
do not wish to consider cats that land on their feet. What we really
need is to somehow shift the existential constraint (just the NP) from
the main clause to the when clause.
This is not impossible, even within a compositional framework, given
the resources of -abstraction and conversion. For instance, suppose
that we particularize our earlier phrase structure rules (75a and 75b),
for the "gappy" analysis of when adverbials, so as to encode pronoun
position as a syntactic feature:
84. a. ADVL[when, pron] ~ when S[pron]/ADVL[temp],

when'(S')
84. b. S[pron]/ADVL[temp] ->

NP[pron, subj] VP/ADVL[temp], P(VP'(P)(NP')).
Alongside each rule, we have indicated the corresponding translation of
the mother in terms of the translations of the daughters in the usual
GPSG fashion (Gazdar et al. 1985). The translation for (84b) "trans-
mits" a -abstracted variable, associated with the ADVL gap, upward
from the VP to the S in the manner proposed by Pollard (1983). Let us
further state a sentence formation rule for sentences like (82a),
combining a when clause having feature [pron] with a sentence having
feature [-pron] and a when-adverbial gap:
85. S -> ADVL[when, pron] S[-pron]/ADVL[when],

S'(ADVL').
If we now take the logic al translation of when in (82a) to be
86. when' = WUxW( V(U(V(x))))

and assume that drops to the ground (the gappy VP in (82a)) is
translated as R (R (D)) (intuitively, a -abstracted adverbial operating
on the drops to the ground predicate), then (84b) yields translation
P(P(D) (it)) for it drops to the ground, and (84a) in turn yields
UxU(D(x)) (it) for when it drops to the ground. Next, assuming that
usually lands on its feet, as a VP with a when-adverbial gap, is
translated as QQ(Sy(S -> u L(y))), application of (84b) but with
[pron] changed to [-pron] yields P[P(SyS -> u L(y))) C<:I cat)] as
the translation of the main clause of (82a) (where (:1 cat) is the
unscoped translation of a cat - see the Coda on scope, or Schubert
and Pelletier 1982). Finally (85), whose semantic rule applies the
translation of the main clause to the translation of the when adverbial,
yields D:I cat) -> u L(it) as a translation of (82a); after scoping, this
becomes :Ix: cat(x))D(x)) -> u L(it). (J/e have neglected tense; the
lack of intension operators is in accord with our "inherent1y inten-
sional" semantics - see the Appendix.)
Thus we have successfully - and compositionally! - exchanged the
logical subjects of the when clause and the main clause, in principle
allowing the application of our theory of when generics without change.
But we have not done so because we think the approach is plausible; on
the contrary, we wished to indicate that it is' implausible, though
possible. For one thing, itis prima facie implausible that people could
actually learn lexical semantic rules like (86). More importantly, it is

hard to see how this sort of "A-magic" could be generalized to deal with
multiple indefinites in the main clause, as in
87. A man often pets a cat when he sees it.
88. People often pet cats when they see them.
89. A chivalrous man usually helps a woman change a flat tire, if
he sees that she is having trouble with it.
Moreover, these additional indefinites can be referred to by definite
noun phrases instead of pronouns:
90. A cat usually foils a dog by running up a tree, if the dog
chases it.
Without significant modification, our approach seems very ill-equipped
for dealing with such examples.
Carlson's approach, as one might expect, has the opposite problem.
Any attempt to generate the relevant kind for a sentence like (82b)
again requires a translation of when of such complexity that it is
doubtful that people could learn it; and again, the approach breaks
down when there is more than one indefinite in the when clause, as in
91. When a man sees a cat, he often pets it.

In fact, Carlson's approach does not extend readily to sentences with
multiple indefinites in the ma in clause, such as (87)-(90) or (for some
non-episodic examples)
92. A man is usually taller than a woman if he and she are

siblings.
93. People are afraid of dogs when they believe those dogs to be
vicious.
A general sort of remedy which these difficulties suggest is that both
the main clause indefinites and the when-clause indefinites (and
perhaps the definites as well) ought to be "raised" somehow to a higher
level, where they will combine with the entire when-clause plus main
clause complex. However, "raising" in any standard sense will give one
of the following results: (i) if the frequency adverb is regarded as
operating on a single (conditional) sentence, we could raise the indefi-
nites so that they immediately follow the adverb and immediately

precede the single (conditional) sentence; i.e., the order of operation is
that we combine the indefinites with the (conditional) sentence and then
apply the frequency adverb to the result. But it is hard to see how
application of the frequency adverb to a single entity could possibly
give the right truth conditions. (In fact, since we've re-combined the
indefinites with the conditional sentence before applying the frequency
adverb, we have gained nothing from the raising operation.) We might
also raise the indefinites to the very top level, but this would clearly just
give us irrelevant wide-scope readings, such as "Some cat is such that
when it drops to the ground, it usually lands on its feet". (ii) It the
frequency adverb is regarded as operating separately on the when-
clause and main clause translations, then "raising" of indefinites can
only give us the irrelevant wide-scope readings.
XII. CAN COMPOSITIONALITY BE SALVAGED?
It would appear, from these considerations, that we are forced to adopt

some sort of non-compositional approach, such as postprocessing of
logical translations to shuffle indefinites into the when clause, or
treating indefinites as supplying constraints separate from both the
when clause and the main clause (e.g., in the style of Farkas and
Sugioka).
Arguably, this radical conclusion was reached only because we
assumed that sentences denote truth values (in a world). Consider again
92. Arnan is usually taller than a woman if he and she are
siblings.
The trouble we have in trying to quantify over "reference cases" is that
part of the information about what these cases are is in the main clause
(i.e., a man and a woman) and part of it is in the it clause. So any
straightforward logical translation which leaves the indefinites in the
main clause loses their contribution to the "reference cases", given that
the denotation of the main clause is just a truth value (at any time and
world). We would not Iose this information, though, if the sentence
denotations had more structure - specifically, if the relational struc-
ture specified by a sentence were stiH discemible in its meaning. Then
the frequency adverb, in operating on the main clause meaning, could
stiH make separate use of the "entities" involved in that meaning, and
the relation or property that they partake of. In particular, if these

"entities" (corresponding to the NPs of the main clause) are semanti-
cally identifiable as "indefinite" - perhaps by being represented as
indeterminates in a situation-style semantics - then we could stiH use
them to restrict the "cases" over which we take the quantification to
run.
One conclusion we might draw, therefore, is that the case for a
situation-based semantics, or in any event one which somehow pre-
serves the relational structure of sentences in their meanings, has been
bolstered by the difficulties we have encountered. Though this avenue
seems to us very worthy of pursuit, we will not pursue it further here.
Instead, we will briefly consider a variant of the approach we have here
developed, which is in a sense analogous to a structured meaning
approach.
Our sketch of this variant begins with the observation that our
semantics for quantificational conditionals already "digs beneath the
surface meanings" of the clauses its combines. A strictly compositional
semantics for a quantified conditional (<1> .... u 'II) would require its
intension to be specified strictly as a function of the intensions of <1> and
'II (perhaps with the intension of 'II determined by a context altered by
prior "processing" of <1. But that's not what we have done. Instead,
we've availed ourselves of a class of interpretations generated by <1>
from the given interpretation. These interpretations include variable
assignments - which certainly are in no sense part of the intension of
<1>. What's interesting about these variable assignments is that they give
us a handle on the ''fine-grained'' structure of sentence meanings; i.e.,
they separate out the possible "instances" of indefinites which make the
predication as a whole true - something not given by the usual
"coarse-grained" intension of <1>. So, our use of -alternatives seems to
be a way of "getting a hold on" the indefinites in the when clause
separately from the intension of the when clause. In a roundabout way,
they provide what "situation types" (or perhaps other sorts of struc-
tured meanings) would provide more directly.
Now, the difficulties we encountered in the attempt to extend our
theory to when generics with main clause indefinites were due to our
inability to "separate out" the semantic content of these main clause
indefinites from the relational content of the main clause. From the
above perspective, this is not surprising: our semantics for (<1> .... U 'II)
only "digs beneath the surface meaning" of <1>, but not of 'II. A remedy
within our framework, therefore, would consist of some ~ay of gener-
ating possible sets of variable bindings for the main clause, in much the
same way as we generated variable bindings for the when clause.
However, these bindings should respect only the semantic content of
the main clause NPs (indefinites, and also definites and pronouns),
without regard to the truth of the predication made about them. For
example,in
94. When it is cornered by a dog y, a cat x sometimes scratches
thedogy
(where we have added variables for convenience) we need to know just
what "bindings" for x and y satisfy "a cat x" and "the dog y" respec-
tively in the main clause. We want to use those bindings to restrict the
alternative contexts (interpretations) that satisfy the antecedent "x is
cornered by a dog y", and only then do we want to check what propor-
tion of these restricted contexts satisfy the consequent predication, "a
cat x scratches the dog y". We shall use the term "legitimate anchor-
ings" for the main clause bindings to be used in restricting when clause
contexts. Legitimate anchorings are partial variable-assignment func-
tions whose domain is a nonempty sub set of the variables which are
existentially or definitely quantified in the main clause, and which assign
denotations that satisfy the NPs associated with those variables. 25
Quantificational conditionals will have a third argument (besides
antecedent and consequent), namely the list of those variables they
"control", i.e., over which they "iterate bindings" (ef. Heim 1982). In
part, this additional flexibility is needed because bindings may not be
(independently) iterated for alI main clause definites and indefinites
(e.g., consider a tree in A cat usually runs up a tree when a dog chases
it, or a tail in A dog (always) has ataii). Also, a definite or indefinite
may be within the scopes of several nested quantificational conditionals
or quantifiers, any of which may control it, so that a disambiguating
syntax is required (e.g., consider lf a person occasionally contributes to
a charity when he receives an appeal, he will usually receive Jurther
appeals; here a charity may be controlled by occasionally or by usually,
depending on whether Jurther appeals is understood as Jurther appeals
from charities or as Jurther appeals fram that charity).
So, neglecting time and tense, the translation of (94) would be
94 '. (The x: neut(x (3y: dog(ycornered-Jy(x, y) -+ S(x,y)
(3x: cat (x (The y: dog(y scratches (x, y)

(The pronoun it can no longer be translated simply as x, since the
generation of alternative bindings for x in the antecedent requires that

it appear definitely or indefinitely quantified there. Besides, the gender
restriction seems like a nice touch.)
We can generate legitimate anchorings in much the same way as we
generate -alternatives. Le., roughly speaking, when the embedding
conditional controls x, (3x: <IW extends by (x, d) the anchorings
generated by ( & 'II) with x fixed at d, for each d which satisfies 
(when used as value of x). So basically, the difference between the
generation of alternative contexts by (3x: <I 'II and the generation of
legitimate anchorings is just that we do not require the value of x to
satisfy 'II (as well as cI. (The x: <IW is treated identicalIy with
(3x: <IW. (We also need to as sume that in both cases, the value of x
does not change if it is already defined; i.e., both definites and indefi-
nites can behave referentially.) We then use the legitimate anchorings of
the consequent (of a quantificational conditional) to restrict the alterna-
tive contexts generated by the antecedent, and for the resultant contexts
we ask, what proportion of them also satisfy the consequent?
Let us state this new method of evaluating quantificational condi-
tionals a little more formally - though with a good deal taken for
granted: we presuppose a revised semantics in which contexts no longer
assign values to ali variables, and in general admit truth value gaps,
existential quantifiers no longer assign new values to variables that
already have values, the notion of alternative contexts induced by a
formula is relativized to a set of "controlled" variables x, and the
sketchy definition of "legitimate anchorings" above has been formalized.
Assuming that these revisions can be carried out successfully, we would
define the semantics of ..... B(x) (where B is A, U, 0, S, N, ... , and x is
the set of controlled variables) as folIows (for a given context space S):
[ ..... B(x) W]i- w = 1 iff for all/most/many/some/no ....
[]' E {U' I[]' = [ ]H U & for some[ r E fs(x, , i, w, [])
and some legitimate anchoring & of x relative to
'II, i, w, [ ] such that [x r = &(x) for alI x E x for
which [x] is defined l'
H
[W]'i. w = 1.
Note that the "alternative contexts" function fs now has x, the set of
controlled variables (allowed to as sume alternative values) as an
additional argument. We have extended 'the alternative contexts, using
the union operation, to include variables that may occur in the
consequent without coreferential occurrences in the antecedent. At the

same time we have eliminated contexts which do not agree with any
legitimate anchoring of the consequent, at alI x E x where they are
defined. Thus, definites or indefinites in the consequent can be
"iterated" over even if they do not occur coreferentially in the
antecedent (as in Cats are intelligent, or in Cats rarely flee from dogs it
the dogs are very small); whereas definites or indefinites in the
antecedent can be "iterated" over only if they occur coreferentially in
the consequent.
We have so far glossed over the times that may be implicitly
referenced by the tense and time adverbials of the main clause. One of
these, too, may be included in its legitimate anchorings, and in the case
of episodic clauses, must be included.
For example, sentence (94) is of a type for which we previously
suggested introduction of a default adverb THEREUPON in the main
clause. Now a cat THEREUPON scratches the dog contains an implicit
indefinite reference to the time of the scratching, as well as an implicit
definite reference to the time of the antecedent event. It is the latter
which is here the natural choice of time to be controlled by sometimes.
In a sentence like
95. The world always tums
where there is no explicit or implicit when clause, and hence no default
time adverbial in the main clause, it would be the implicit indefinite
time "generated" by gPRES which is controlled by always.
There is a slight notational problem for time, since there are no
explicit time variables which can be included among "controllees" of
the quantificational conditional. We will not try to solve this rather
minor problem here (but see our last example below).
Finally, it seems that bare plurals (kinds) can be handled in much the
same way as singular indefinites. Again we have a slight notational
problem, in trying to indicate which bare plural translations are
controlled by which quantificational conditionals, if any. The problem
could be solved by using term subscripts rather than variables to fix
coreference and control (as Heim and others do), but this would be less
elegant in other respects. We therefore introduce the somewhat ad hoc
device of optionally changing a term like i (cat) to the equivalent
"quantified" term (The x: (x = i (cat). Then in the definition of
legitimate anchorings, we would say that when the embedding condi-
tional controls X, an expression of the form (The x: <I'P, where is

kind-Ievel with respect to x while 'P is object-Ievel with respect to x,
extends by (x, d) the anchorings induced by ( & 'P) with x fixed at d,
for each d realizing the unique kind satisfying .
This more or less completes our sketch of how we might construct a
general, uniform theory of generics, which would be capable of
accounting for virtually all of the examples we have discussed (in
particular the "problematic" ones from (78) onward). For additional
clarification, let us write down the translations of a few more sentences
(again neglecting tense):
81. a. Cats are usually intelligent when they have blue eyes
b. (The x: plur(T)(xHBE(x) --+ U(x)
(The x: x = # (plur(cat)I (x)
93. a. People are afraid of dogs when they believe those dogs to
be vicious
b. (The x: plur(T)(x (Those y: plur(dog)(ybelieve(x,
vicious(y --+ G(x,y) (The x: x = #(plur(person) (The
y: y = # (plur(dog)afraid-of(x, y)
96. a. A cat is intelligent
b. (3x)T(x) --+G(x)(3x:cat(xI(x).
In alI three translations, T is the universally true predicate, AX(X = x).
Note that in (96) we have invoked an arbitrary universally true formula
as default antecedent (and abbreviated (3x: T(x as (3x). In this
particular example, it might have been more appropriate to use the
presumption that the cat is "extant", i.e., exists (in a realistic sense) at a
given time:
96. c. (3x)extant(x) --+ G(x) (3x: cat(xI(x).
The "extant" presumption could be used quite often as default, though
clearly not in examples like
97. Agreat composer is always remembered.
It is also worth coming back here to the following (commonly dis-
cussed) type of example,
98. a. * A dog is widespread

b. (3x)extant(x) --+ G(x) (3x: dog(x W(x)
whose anomaly, on this translation, stems from the application of a

kind-Ievel predicate W to an object-Ievel argument x.
In considering our translations of sentences like (96) lacking an overt
restrictive clause, the reader should keep in mind our view that such
generics are not "basic", and instead are to be evaluated with a
"missing" or "suppressed" when clause; and it is our view that this
missing clause is to be supplied somehow by a "pragmatic component"
which can look at the previous discourse,26 at characterizing properties
of the subject, at presuppositions associated with verbs and stress
patterns, etc. Once we have provided a position for such a missing
clause, we are obliged to fill it, even in cases where there is no clear
need for such a "phantom clause".
As a final example, we come back to donkey sentences:
99. a. When Pedro owns a donkey, he usually rides it to town.
bl. gPRES, 3x: D (xO(p, x - U(x,')
gPRES(AT-THAT-TIME,The x: neut(x (R(p, x)
b2. gPRES,3x: D(xO(p, x - U(')
gPRES (AT-THAT-TIME , The x: neut(x (R(p, x).
As shown, we now have two possible translations, corresponding to the
choice of "controlled" variables. (There is no third possibility with x
only controlled, given our stipulation that a time variable must be
controlled.) The first has the same truth conditions as our original
analysis (as given by analogy with sentence (65. But interestingly, the
second would be "true in English" even if Pedro always owns numerous
donkeys, only one of which he habitually rides to town. This is because
the donkey variable x is not controlled by - u, and so assumes an
arbitrary value for each alternative context generated by the antecedent.
So, if there are choices of these values which usually render the
consequent equal to 1, the conditional as a whole will be "true in
English." This brings the analysis of quantified conditionals in line with
that of ordinary conditionals and apparent1y resolves the difficulty
inherent in our original analysis of such sentences (that they miss such
readings).
XIII. IN CONCLUSION
Needless to say, our proposals in Sections IV-X, for certain kinds of

episodic generic statements, are no more than a preliminary sketch of
what a satisfactory formal account of such statements might look like.

We have left open not only many details of the formal semantics, but
virtually alI details of the mapping from English surface form to logical
form. Nevertheless, we hope to have shown that the ideas of (1)
indefinites and tenses "loading" reference values into a context, and (2)
generic sentences iterating over combinations of such values, are quite
promising. The interaction of adverbs, when-clauses, indefinites and
tenses is rather intricate in our theory, but no more so, we believe, than
the subtlety of these phenomena warrants - indeed, there is much
room for refinement and elaboration.
In Sections X and XI we considered possible extensions of our
theory to cover larger classes of generic sentences, but concluded that
generic sentences containing singular indefinites or bare plurals in the
main clause were ultimately beyond the reach of our theory, in its
present form. Some such sentences are amenable to Carlson's analysis,
which in a sense starts at "the opposite end" of the spectrum of generics
(emphasizing non-episodic sentences with bare plural subjects); but
Carlson's analysis is not extensible to multiple indefinites, or to the
types of episodic sentences central to our own analysis. At this point,
prospects for a uniform compositional theory seemed poor.
In Section XII, however, we suggested that the apparent non-com-
positionality stems from the use of unstructured meanings (intensions).
We then sketched what further directions one might take in an attempt
to get at the "finer-grained" structure of sentence meanings, and to use
these in the construction of reference ensembles that combine informa-
tion from restrictive clauses and main clauses of generic sentences. In
this sketch, we strove, as far as possible, to stay within the framework
we had already developed, and filled in enough detail to make the
feasibility of the enterprise plausible.
However, the proposed semantics of --+ B(x) was not strictly com-
positional, i.e., it was not defined exclusively in terms of the intensions
of its operands (though it was defined in terms of the intensions of the
parts of its operands). Consequently, we think that the correct way to
proceed, ultimately, will be one that makes direct use of structured
meanings, such as some form of situation semantics. Nevertheless, we
expect such a theory to have much in common with the one we have
sketched. .
The reader who has kept in mind Lewis's theory of frequency
adverbs, interpreting these as uns elective quantifiers (Lewis 1975), will
have noticed the drift of our theorizing, in Section XII, toward

something rather closely resembling Lewis's views. Existential and
definite quantifiers lost some more of their autonomy, yielding it up to
the "control" of quantifying adverbs. However, unlike some theorists
who have followed in Lewis's footsteps, we have continued to insist on
translating English definite and indefinite noun phrases uniformly, and
not dislocating these arbitrarily from the clauses in which they occur.
In this respect, we have tried to adhere to Carlson's compositional
methodology.
One slightly unappealing aspect of our theory we have noted is its
reliance on "phantom" restrictive clauses in the translation of generic
sentences lacking such clauses. Could these phantom clauses be elimi-
nated? We think that this may be possible within a situation-semantic
framework. To do so, we would "embed" a situation type or kind
(whose instances provide the required ensemble of situations or
objects) directly into contexts, presumably as values of certain indexical
constants; Le., WHEN(<ll) [], the new context generated by having
"processed" a WHEN-clause, would differ from [ ] in having its
"embedded" situation type restricted to <ll-situations or objects. For
generic sentences lacking restrictive clauses, the appropriate ensembles
would be constrained by context. Frequency adverbs, rather than
having syntactically explicit restrictive clauses as arguments, would
instead make reference in their truth conditions to the situation type or
kind embedded in their context of evaluation. This seems to us a
promising direction for further research. In any case, we hope that our
investigation will encourage others to work in this challenging area.
XIV. COD A ON PLURALITY AND SCOPE AMBIGUITY
The point of this brief Coda is to explain, quickly, how we view

plurality and issues involving relative scope of quantifiers (and other
items such as connectives and adverbs). A fuller discussion of plurality
can be found in Schubert (1982), and of scope in Schubert and Pelletier
(1982).
Throughout this paper we have talked about bare plurals. We
occasionally represented a bare plural such as cats as t (cat) - the
kind, cats - formed from the predicate cat (which is true of individual
cats) and our kind-forming operator, t. But since plurals may be used
to refer to group or collections of things, we have generally made use of
plur( cat), plur being an operator that takes the predicate that is true of
individuals and forming the new predicate which is true of arbitrary
groups or collections. These groups may have as few as one ~lement in
them, in light of such sentence as
100. The stars called "the morning star" and "the evening star"
are one and the same.
The sentence would be 10gical1y false if "the stars" had to be at least
two stars. (There may sometimes be an implicature to the effect that
there are, or may be, more than one object, when a plural is used; but
the truth conditions should allow for groups of size one.)
So a predicate like plur( cat) is true of anY arbitrary group of cats.
Numerical modifiers - e.g., three - are viewed also as predicate
modifiers, so that three(plur(cat)) is a predicate which is true of groups
of three cats. Since these are predicates, they should be able to be
operated upon by our kind-forming operator, t. But there seems to be
a difference in meaning when one attempts to form the kind from a
"numericalized" predicate. Certain predicates, it will be recalled, apply
directly to kinds - be extinct, be common, be rare, come in many
different flavours, be out of stock, and the like. Such predicates happily
combine with paradigm cases of "ordinary" kind-denoting terms:
101. a. Dinosaurs are extinct
b. Sparrows are common
c. Albino ravens are rare
d. Ice cream comes in many different flavours
e. Eggs are out of stock.
Yet when numeric al modifiers are applied, the result does not seem to
combine readily with such predicates, indicating to us that we cannot
apply t to the term to yield a term denoting a kind. For example,
102. a. * 12 eggs are out of stock

b. * A dozen eggs are out of stock.
And this is the case even if the store in question only receives eggs
packaged in groups of 12, or that there were no eggs in stock, or that
only the eggs packaged in groups of }2 were out of stock, etc. The
above contrast is of the same sort as that observed between
103. a. A cat is intelligent.

b. * A cat is common.
We therefore propose to treat them analogously. Since we have
proposed to translate singular indefinites like that in (103a) (previously
(96a existentially even under a generic reading, we will assume that
bare numerical plurals, as well, are always translated existentially. Thus
12 eggs, for example, would be uniformly translated as (3
twelve(plur(egg ), just as a cat is uniformly translated as (3 cat ). In
both cases, the specific/generic ambiguity is accounted for by the
absence/presence of an implicit .... G operator. Contrary to our initial
hypothesis, then, bare numeric al plurals do not admit a translation in
terms of l. In a formal grammar, this constraint can be implemented by
using separate rules for bare non-numerical NPs and bare numeric al
plurals (see (10Sb, c) below). Sentences like (104a) would be then
translated ambiguously as (b) or (c):
104. a. 12 eggs cost a dollar

b. (3x: (twelve(plur(egg) (x cost-one-dollar(x)
c. (3x:)T(x) .... G(x)
(3x: (twelve(plur(egg) (x cost-one-dollar(x)
ignoring the exact representation of "cost a dollar". Our predicates
would need to be further subdivided into ones which properly can
apply only to groups, in addition to the ones which apply to kinds and
the ones that apply to individuals. The exact details of this we leave
open here: one possible position is that if a predicate applies to an
object, it always applies as well to groups of objects, and if it applies to
a kind, it always applies to groups of kinds as well. This would
legitimize the application of cost-one-dollar to groups x in (b) and (c)
above. However, it leaves open the question of whether (b) is to be
interpreted distributively (a dollar per egg) or collectively (a dollar for
the lot). A possible answer is that when an individual-level predicate P
is applied to a group x (syntactically, to a plural NP), it can be
translated either without change (yielding a collective reading) or with
distributive operator Distr, where Distr(P) (x) is understood as saying
"each member of group x has property P".
In Schubert and Pelletier (1982) a mechanisfn for generating the
relative scopes of quantifiers, adverbs, and coordinators was given. The
picture was that the parsing rules, which in a GPSG-like manner

yielded a semantic representation, only yielded an "ambiguous" logical-
form representation. Once this ambiguous representation was gener-
ated, then certain rules for determining the possible relative scopes
were applied, to yield the various possible readings. For example, a
quantified NP such as every cat is analyzed and translated by rule
(1OSa)
105. a. NP -> DET[+quant] N, (DET' N')
b. NP -> N[-num], t(N')
c. NP -> N[+num, plur] , (3 N')

neglecting various fine points. The angle brackets in (a) and (c) indicate
that the "generalized quantifier" they enclose is ambiguously scoped.
The translation of a sentence like (106a) is as in (106b):
106. a. Some dog chases every cat
b. chase((V cat)) ((3 dog)).
The two quantifiers can now be "raised" to encompass any sentential
formula that embeds them, at the same time introducing variables, with
the possible results
107. a. (3x: dog(x)) (Vy: cat(y)) chase(y) (x)
b. (Vy: cat(y)) (3x: dog(x)) chase(y) (x).
It is this mechanism which allows us to retain the view that names
denote individuals, and also to dispense with dual version of predicates
such as love' and love* (or love+), which Montague introduced to
distinguish the direct translations of English transitive verbs (whose
objects are intensionalized property sets) from extensionalized versions
(whose objects are individuals).
The same sort of scoping mechanism is used for coordination,
adverbs, and other operators such as tense. Thus (108a) is translated as
(108b):
108. a. Kim williove Robin or hate Sandy
b. FUTR[( V 10ve(Robin) h~te(Sandy)) (Kim)].
The ambiguously scoped conjunction allows the minimal scoping in
(109a) or the wider scope in (b), or the still wider scope in (c) (with
conversion from prefix to infix form):
109. a. FUTR[(Ax) ~ove(Robin) (x) V
hate'(Sandy) (x)] (Kim)]
b. FUTR[love(Robin) (Kim) V hate(Sandy) (Kim)]
c. FUTR~ove(Robin) (Kim)] V
FUTR[hate(Sandy) (Kim)].
As it turns out, alI three "readings" are equivalent. But if FUTR were
certainly (for example), (c) would be distinct from (a) and (b). Other
examples will make (a) and (b) distinct.
In the "logical form" language described in the Appendix, we do not
make use of this feature of our grammar, but rather treat the logic al
forms as already having gone through this process of disambiguation.
Were we to give an explicit grammar for the phenomena discussed in
this article, though, we would wish to have our grammar generate this
ambiguous logical form, and have some postprocessing to convert it
into the format actually used in the Appendix.
APPENDIX: SYNTAX AND SEMANTlCS FOR A

"LOGICAL FORM" LANGUAGE
In this Appendix we try to be somewhat more formally accurate about

what our "logic al form" language is and how it is semantically to be
interpreted. As indicated in the Coda (on scope ambiguity), we have in
mind a "layered" approach. The rules for mapping from English surface
structure will yield an "ambiguous" logical form. This logical form must
first be disambiguated; for the phenomena under consideration here,
this will happen in two stages. First there will be a "scope disambigua-
tion" in which our "ambiguously scoped" expressions are given some
unambiguous representation, and second there will be replacement of
our "pronoun markers" by some constant or variable. These two
aspects of generating a representation of a text are not considered in
this Appendix; instead, we start with the "layer" which results from this
disambiguation.
Our logical form language $' has constants a, b, c, ... (perhaps with
subscripts); individual variables x, y, z, ... (perhaps with subscripts);
indexical constants r, now, present, yesterday, then, 1, you (and possibly
others ... we leave this matter open for further investigation); n-place
predicate symbols; the connectives -, &, V, and ..... ; equality, =;
restricted quantification of the form (Qa: ctlP where Q E {\f, 3, The,
Most, Few, No, . . '}' ct>, lP are (open or closed) formulas, and a is a
variable; certain operators ("tense" operators: PRES, gPRES, PAST,
PROG, PERF; "temporal adverbs": YEST, THEREUPON, AT-
THAT -TIME; and "quantificational conditionals": ..... A, ..... U, ..... G, ..... S,
..... N, ..... o - corresponding to always, usually, generally, sometimes,
never, often); and a kind-forming operator l which takes a one-place
predicate (this includes the case of having other positions of an n-place
predicate filled with a constant or bound with a quantifier) and forms a
term; a plurality-forming operator plur which takes a one-place predi-
cate (including, as before, the case of having other positions "filled")
and forms a new (plural) predicate; ,-abstraction (over individual
variables); and numerical adjectives (two, three, ...) which take a
(plural) predicate and form a predicate.
The constants and variables are sorted into object- and kind-
constants/variables, where that distinction may sometimes be empha-
sized by superscripting the former with "o" and the latter with "k". The
predicate symbols correspondingly have their argument positions sorted
as being object-level or kind-level (and this sorting may sometimes be
emphasized by sequences of "o" or "k" superscripts). The syntax is
quite ordinary, generally written G(F( a, b, ...)) which indicates that
the sentential operator G is applied to the formula F( a, b, ...), and the
latter means that the predicate F is applied to the constants/ variables a,
b, .... The results of applying term-forming operators are generally
written G(F), where G is the term forming operator and F is the
argument it applies to. We will often use English-like expressions rather
than simply abstract symbols for predicates and names, so that, for
example, we might write l(dog) for the result of applying the term-
forming operator l to some abstract predicate F and relying on a
"translation scheme" to go from dog to F. We also regard an expression
of the form F(x, y) as equivalent by definition to (F(x)) (y) or to
F(x) (y) - thus we may regard many-place predicates as function-
valued single-place functions that absorb one argument at a time.
(However, in the main text we often violate this convention for
purposes of keeping our discussion in line with usual first-order
notation. For example, Pedro owns Annabelle is usually translated as
O(p, a); our "categorial grammar" version would be O(a) (p) - that is,
"owns Annabelle" is true of Pedro - which by our convention becomes
o (a, p), just the reverse of the usual notation! In this formal Appendix
we are more careful to obey our conventions.)
For our semantics, we assume a domain D of individuals (sorted into
objects 0,27 kinds K, and a set of time intervals 1), and a set of possible
worlds W, alI of which are non-empty. Intuitively, O contains not only
(single) objects, but also groups or collections of such objects. We leave
open the exact structure of 1, but intend its elements to be thought of as
time intervals having such relations as "during", "before", "in the
vicinity of", etc., detined on them. Finite time intervals are assumed to
have a beginning and an end, thought of as "instants" of time -
intervals whose beginning and end coincide. For our present purposes,
we need not impose a further structure on K, although we have in mind
that it will form a "lattice of kinds" - more or less in the same way we
explained towards the end of Pelletier and Schubert (1985) with
respect to count terms formed from mass terms (e.g., a wine, viewed as
being formed from wine).
Instead of the usual notion of an interpretation and its extension to a
valuation function, we wilI detine the notion of a context, [ ], though for
us, that notion is inextricably bound up with the notion of context
change, resulting from the evaluation of (previous) formulas and their
parts with respect to a particular time and world. The following
detinition is intended to formalize these notions in a way which meets
our present needs.
A context is a (total) function on the expressions of .z, yielding
appropriate values in our previously-given domains along with O and 1,
and various functions on these domains. A context transJormation
Junction o is a (total) function on wffs of .z, times, worlds, and
contexts into contexts. It F is the set of wffs of .z, then o: F X 1 X
W XC ..... C. A context space is a quadruple (D, W, C, o) where D is
the set of individuals (sorted into objects, kinds, and time intervals), W
is the set of worlds, C is a non-empty set of contexts, o a context
transformation function, and o and each of the [] E C meet the
following conditions. (We abbreviate o(, i, w, [ ]), where is a wff,
iEI, w E W, and [ ]is a context, as i, w[ ]. Intuitively it designates
the (new) context which is generated from having already "processed"
 in the (old) context [] - it wilI always be dear which o is
intended.)28
1. ~ a] E 0, for a an object constant or variable.

[ a ]E K, for a a kind constant or variable.
2. [now] , [present], [yesterday], ... E 1, with the constraints

that [now] during [present], [yesterday] during [present] (i.e.,
present denotes an "extended present"), [yesterday] is the
day before [now], ... etc. We leave these constraints
unformalized. [you], [1] ... E O.
3. [r] is a sequence of one or more elements of 1, where one
element is distinguished as "being in focus". The element in
focus is designated by r, and [r] E 1. It is allowed to have
something "in focus" which is not in the sequence (and then
we say that the focus is undefined); this can happen in two
ways: by having it undefined to the right or to the left.
Undefined to the left is denoted by _r, undefined to the
right is denoted r_.
4. [F] E D, * D 2 * ... Dn * I * W, for F an n-place predicate
symbol, where Dm is O or K, if F is sorted as being object-
level or kind-Ievel in its mth argument. Here A * B denotes
the set of functions from A to B, A * B * e denotes the set
of functions from A to the set of functions from B to e, and
soon.
In some of the remaining clauses, we will use the notation [ ]a: d to

denote a context identical with [] except that the denotation of the
constant or variable a is d. When a is now, the meaning of this
notation is slightly enhanced: the indexicals r, present, yesterday, etc.,
are appropriately reset along with now, so that the first element of r
shares the denotation of now, so that present denotes a time encom-
passing the time denoted by now and extending sufficiently far into the
past, so that yesterday denotes the day before the time denoted by now,
etc. When we mention these variant contexts, we keep aH values of
every other feature of the context fixed. A notational convention used is
that [F]a indicates application of the function [F] to argument a. More
generally, [F]a. b, ... n indicates successive application of [F] to a, b . .. n.
Should there be sufficient superscripts to designate ali the possible
argument-positions of F, then the value will either be 1 or O. In all of
the constraints, whenever we say that the value of some (metalinguistic)
expres sion equals 1 if and only if some condition holds, we mean that
its value is 1 under that condition and is O otherwise. (No truth value
gaps). It might be noted that, as given in clause 6, this semantics is
"inherent1y intensional". Application of a predicate function to an

argument is not interpreted as mapping an extension into an extension
at each world, but rather as mapping an intension into an intens ion.
This has been accomplished by making the times and worlds be the last,
rather than first, arguments to which a context is applied. An advantage
of this formulation is that no analogue of Montague's intension
operator is needed, not even for intensional locutions. Quantifiers and
truth-functional connectives, however, are extensional, and this is
reflected in the constraints concerning them, which are assumed to hold
for aU i E I and w E W. Starting with the sixth condition, we
subdivide the conditions into a and b parts, where the a-part provides
a constraint on the value of a certain type of formula, while the b-part
provides a constraint on the context change induced by "processing"
this type of formula. A text SI . S2 . . . Sn is assumed to induce a total
change: S~ W(S~.'~\(. . .(S~ w[ ]) .. ) in the initial context [ ], relative to
time i and world w. We now return to the conditions that each context
and transformation function must meet.
5. p,a] = {(d, []a :d) 1d E O or d E K, depending on

whether a is an object variable or a kind variable}
6. If is a predicate, then
a. [a] = []a, where a = [a]. is object level if a is an
object term, and is kind level if a is a kind term.
b, a i, w[ ] = [ t :i' if is monadic and [a F w = 1 (true
atomic predications leave context unchanged
except that the time "in focus" is set to the
"episode time").
= [ ] if is n-adic for some n > 1.
7. a. [ -]i, w=liff[]i, w=O
b. There are (at least) two plausible alternative hypotheses
for context change due to processing a negated sentence.
We will not decide between them, as this is not crucial to
our concerns.
(i) - i, w[ ] = [ ],i.e., the context is "restored" to the
value it had before - was evaluated (even though
evaluation of the embedded (p may have involved
various context changes itself). On this view the
pronoun in Kim does not have a favourite professor;

he is tall cannot be interpreted as referring to Kim's
favourite professor, even if Kim has one: the refer-
ence that would have been induced for a favourite
professor by further processing of the first clause
(from 11 b, below) is unavailable outside the nega-
tive context.
(ii) - <1>i. w[] = <1>i. w[], Le., the effect is just as for the
unnegated sentence. On this view the pronoun he
(above) can be interpreted as referring to Kim's
favourite professor. This seems useful for sentences
like It's not that 1 don't have a reason; I'm just not
telling you what it is, where the it of the second
clause can be interpreted as referring to my (con-
cealed) reason, if 1 am speaking truthfully in the first
sentence. Option (i) blocks such reference into a
doubly negated context.
(iii) Further options have to do with sentences like I'm
not reading a nove!; it's a biography, in which
neither (i) nor (ii) seem to work. Here we would
want the context induced by processing the first
clause (assuming it's true) to provide something like
"the book 1 am in fact reading" as referent, so as to
be "picked up" by the it of the second clause.
8. a. [<1> & 'II]i- w = 1 iff[<1>]i. W = 1 and <1>i. w['II ]i. W = 1
b. (<1> & 'II)i, w[ ] = 'IIi, W(<1>i, w[ ])
9. a. [<1> V 'II ]i,w= 1 iff[<1>]i,w= 1 or<1>i,w['II]i,w= 1
b. (<1> V 'II)i, w[] = ei, w[] for some e E {<1>, 'II} such that
[e]i, w = 1, if there is such a e. (Thus for example, in
John has a dog or a cat, and it has fleas, the pronoun can
be interpreted as referring to one of John's dogs if he has
dogs but no cats, to one of John's cats if he has cats but
no dogs, and to any one of his dogs or cats if he has both.
If he has neither we have placed no constraints on
possible interpretations of it.)
10. b. (<1> --+ 'Py w[] = ( - <1> V 'P)i, W[]. Then we have the
options for negation as listed above. If we take option (i)
- negation does not change the context - then this
amounts to 'P i , w[ ]; if we take option (ii), then (using the
V context change rule) this will give us one of a set of
contexts, some like what we got with option (i) here and
some like what we get with option (ii) for negation -
depending on whether the antecedent of the conditional
is false or the consequent true (or both).
Note that in the rules for the binary connectives, the second formula is
evaluated with respect to the context as altered by the first. This is
critical to our handling of the "donkey sentences" and related problems.
11. a. [(3 a : <1'11] i, w =, 1 iff some d satisfying [<1>] ~ ~d = 1 also
satisfies <1> i, w['P] ~ wd= 1,29
It is worth noting that this is equivalent to: some d satisfies [<1> & 'P]~~d
= 1. Clauses similar to (lla) hold for V, Most, Few, No, etc.
[(The a: <I'PF w = 1 iff [<1> & W]io w = 1 or else there is
a unique d E D satisfying [<1>] ~ ~d = 1, and for that d,
<l>i, w['P]~~d = 1 (The first possibility corresponds to a
"referential" interpretation, the second to a "Russellian"
one.)
b. (3a: <I'P i, w[J = (<1> & 'PY' w[ L:d for some d E D such
that [<1> & 'P j ~~d = 1, if such a d exists. (Thus when the
existential assertion is true, some possible value of the
variable a satisfying <1> and 'II becomes the value of a in
the modified context. This value can then be "picked up"
by subsequent clauses.)
- for the quantifiers V, Most, Few, No, ... we again have the options
similar to those for negation and the conditional; and again we leave the
matter open.
(The a: <IW i, w[ ] = (<1> & 'PY' 1 ], if [<1> & W]i, w= 1;
= (<1> & 'P)i, w[ ]a:d, if d is the unique
element of D suchthat []~~d = 1, and
for that d, <1> i, W['fl~ ~d = 1.
It is probably worth pointing out that the abovequantificational

treatment of indefinites (Le., as 3, interpreted as in lla) is a unique

feature of the present approach to discourse, and one generally thought
to be incompatible with the ability of indefinites to provide referents for
pronouns (see Heim 1982: Ch. 1, Section 1).
Recall that r is our "time reference vector" and that r stands for the
element of r that is "in focus". RecaB also that we have an operation on
r which "shifts focus" one element to the right (Y) (backward in time) or
one element to the left (F) (foward in time) of r, and finaBy, recaB that
the element of r that is in focus might be undefined to the right (r_) or
undefined to the left (_r). Aiso refer to rule (6b), which is the basic
ruIe for assigning r a new event time (besides the ruIe for progressive,
13b). Here then are our clauses for issues having to do with time.
12. (Time adverbs)

a. [YEST(<I]i,w= 1 iff[]i,w= 1 and
i is during [yesterday]
[THEREUPON(<I]i,w = 1 iff []i,w = 1 and is
immediately or shortly after [r].
[AT-THAT -TIME( ]i, w = 1 and
i = end of[r]
b. (A (<I)i, w[ ] = i, w[ ] for A a time adverb,
if [A (<I]i- w = 1
13. (Present and progressive)

a. [PRES(<I]i, w = 1 iff i = [now] and []j, w = 1 for some
j in the vicinity of i
[gPRES(<I]i, w = 1 iff i = [now] and [ F w = 1 for some
j during [present]
IPROG()]i- w = 1 iff [F w = 1 for some j which
includes i (ignoring certain modal characteristics of the
progressive).
b. (PRES(<IY' w[] = j, w[] for some j in the vicinity of

[now] such that [ ]i- w = 1, if there is such a j and i =
[now]
J
(gPRES(<I)i, w[] = '; w[ for some j during [present]
such that []i- w = 1, if there is such aj and i = [now]
(PROG(<I)i, w[ ] = []r:i' if[PROG()]i- w = 1
14. (Past and perfective)

a. [PAST(<I]i.W = 1 iff i = [now] aud [];W = 1 for some
j before i
[PERF(<I]i.w = 1 iff i = [r] and [];W = 1 for somej
before i. (Note: [r] must be defined.)
b. (PAST(<IY' w[] = (j. w[ ]r-}r- for some j before i such
that ~] W= 1, ifthere is such aj and i = [now]
(PERF(<I)i. w[] = (j. w[ ],), for some j before i such
that [] = 1, ifthere is such aj aud i = [r].
Legitimate -alternatives are defined relative to a context space, a
formula, a time, a world, aud a context. Intuitively, they are alI the new
contexts that could have been generated from the formula <1> and that
context, had we been given a context transformation function different
from the one specified as part of the given context space. (We as sume
that the time and world remain constant, and that these new contexts
and context transformation functions obey the strictures given about
them.) This set is picked out by the function Is for a given context space
S = (D, W, C, O):
Is(, i, w, [J) = {[]' I there is a context space S' =
(D, W, C, O') such that []' = 0'(<1>, i, w, []) aud []i.W = 1
with respect to S'}
15. (Simple temporal whens, a predicate over times)
a. [when-, 'II] i. w = 1 iff [<1> & 'II] i. w = 1 where it is
assumed that 'II has been translated to incorporate a
default adverbial A T -THAT-TIME or THEREUPON
b. (when-, 'lIy. w[] = (<1> & 'lIy. w[ ] if [<1> & 'lI]i. W = 1,
where it is assumed that 'II has been translated to
incorporate default adverbial AT-THAT-TIME or
THEREUPON 30
16. (quantificational and generic conditionals: a schema, B E
{A, U, 0, S, F, Ni)
a. [<1> ...... 8 'lIJi.w = 1 iff for B-many of the [1' E
1\(<1>, i, w, [ ]), ['lI]'i. = 1
W
[<1> ...... c 'lI]i. = 1 iff for "most" (J;y', ~ nEW X C such

W
that []' E fs(, i, w', []), ['lI]'i. w' = 1, where "most" is

to be interpreted in terms of some probability distribu-
tion favouring worlds w' similar to w, with regard to the

"inherent" or "essential" nature of things. (More crudely,
we could just require the worlds w' quantified over to
satisfy some appropriate similarity relation relative to w).
b. (<l> --+ B 'P) i, w[] = [] (and likewise for --+ G), i.e., em-
bedded indefinites cannot be referred to outside the
conditional. (Of course, indefinites that are "raised" to
have wider scope than --+ B can be referred to later on, by
rule(llb).)
We leave open what the possible worlds "similar to w with respect to
the inherent nature of things" are, and exactly how many contexts
counts as B-many. Rowever, as indicated in the text, we think that
complete formalization of the latter notion requires appeal to some
method of systematically enumerating, or randomly sampling, the
alternative contexts allowed by the antecedent, where the precedence
(in an enumeration) or probability (in a sampling process) of a context
depends not only on the variable bindings determined by that context,
but also on the reference times determined by it (via the value of r).
We close this Appendix with a treatment of a somewhat more
complex example text, to show our conditions on tense and time
adverbials at work:
a. John entered the room.
b. Mary had taken down his paintings.
c. Then she had hung up some cheap posters.
d. Re groaned.
Assuming that these events did transpire as reported, let us verify that
(a)-(d) are "true in English", when translated as:
a'. PASTThe x: room(x))enter(x,john))
b'. PAST(PERF The y: plur (painting) (y) &
belong-to(john, y)) take-down(y, mary)))
c'. PAST(PERF (TREN 3z: plur(poster) (z) &
cheap(z))hang-up(z, mary))))
d '. PAST(THEREUPON (groan(john)))
(We neglect issues related to the relative scopes of quantifiers, adverbs,
tense and aspect operators, the relative scopes of adjectiyes and the
plur operator (which, recall, forms a predicate true of collections of one

or more individuals from a predicate true of individuals), and the
implicit reference to locations in take-down and hang-up.) In evaluating
(a') at i = [now] and w = the actual world, with r denoting (il (a
singleton sequence whose element is in focus), we will obtain value 1 iff
the formula embedded by PAST has value 1 at some time j before i
(i.e., before now), with the focus of r shifted to the right (into the past)
from i (leaving r = (i-l, with the focal element undefined). By clause
(lla), the embedded formula (beginning with the definite quantifier)
has value 1 iff (room(x) & enter(x, john does. This in turn has value 1
(at the appropriate time j) iff x already designates the room John
entered. But the definition of "truth in English" quantifier over all
possible denotations of x (since no variable assignments are ruled out
by conventional English meanings), and so the embedded sentence is
true at time j. Moreover, evaluation of enter(x, john) resets r from (i-l
to (ij 1(by clause (6b, and after recursive ascent back to the top level
of sentence (a'), r has value (ij) (by clause (14b, and of course, the
sentence is true.
In (b'), r initially denotes (ij). The shifts induced by PAST and
PERF change this to (ij-l for evaluation of the embedded quantified
sentence. The effect of the definite, "the paintings belonging to John"
(much as in the case of "the room") is to ensure selection of a binding
for y which is indeed the salient collection of paintings. The episodic
predication, when evaluated to 1 at some time k, will change r to (ijk 1,
and upon recursive ascent to the top level of sentence (b'), this
becomes (ijk 1.
Evaluation of sentence (c') again shifts the focus to k, i.e., r again
becomes (ijk 1as a result of the PAST, PERF operators. The embedded
episodic sentence headed by THEN will then cause the iteration over
contexts compatible with English to "find" the time k' at which Mary
hung up the cheap posters, where THEN ensures that k' is immediately
or shortly after k. Consequently the new value of r becomes (ijk' 1, or,
after ascent to the top, ( ijk' 1.
Finally, evaluation of sentence (d') first shifts the focus right one
step (due to the PAST operator) giving (ijk' 1 as value of r. The
sentence embedded by the PAST operator is therefore evaluated with
reference to time j, the time of John's entering (not with respect to the
times of the events reported in the past perfect} The effect will be to
replace the value of r by (ij' k' 1, where j' is the time of John's groaning
(constrained by THEREUPON to be immediately or shortly after j, the
time of John's entering). After as cent to the top of (d'), the final value
of ris <ii' k').
ACKNOWLEDGEMENTS
We wish to thank the members of the University of Alberta Logical

Grammar Study Group for comments and suggestions during the
writing of earlier drafts of the paper: Matthew Dryer (Linguistics),
Brendan Gillon and Bernard Linsky (philosophy), and Sven Hurum
(Computing Science). Thanks are also due to Gennaro Chierchia for his
careful reading and searching critique of an earlier version of this
paper. Many of the final touches were put on during the 1987 Summer
Linguistics Institute at Stanford where Pelletier conducted a Workshop
on Generics and Habituals. We thank Ivan Sag and his staff for
coordinating the Institute; and the NSF for a grant which allowed the
Workshop to take place. It was truly an intellectually stimulating
environment. In this regard, thanks are due to the members of the
Workshop both for their comments on this paper and for comments on
the entire topie of generics: Greg Carlson, Gennaro Chierchia, Manfred
Krifka, Godehard Link, and Alice ter Meulen. Finally, the authors also
gratefully acknowledge the aid of NSERC Grants A8818 (LKS) and
A5525 (FJP).
NOTES
1 Essentially, it attempted to find a middle ground between Carlson's and Chierchia's

theories which would permit the simplicity of Chierchia's logical-form representations
of generic (and related) sentences to be retained, while remedying the defects that seem
to result from his relinquishing Carlson's operator Gn. The idea in this "conservative"
proposal was to shift some of the burden of providing correct truth conditions from the
rules of translation (into "logical form") to the rules of semantic evaluation (of the
"logical form"). Intuitively, we "liberalized" the logical syntax of predication slightly at
the expense of a slight complication of the rules of evaluation, by allowing "indirect
evaluation". One could look at this proposal as if it removed "types" from the syntax of
logical form but recaptured ali the information thereby omitted within the semantic
evaluation of the logical form. Godehard Link informs us that a similar idea is taken by
Fine (1985) with respect to his "arbitrary objects": an "ordinary object level" predicat:
F applies to an "arbitrary object" a just in case F applies to every instance of the
arbitrary object. Therefore, natural numbers are even or odd is true because the
"arbitrary object" natural number(s) has this property; but natural numbers are odd
and natural numbers are even are both false.
2 "Situations" is being used here in an ordinary sense, with no necessary connection to

the Situation Semantics of Barwise and Perry (1983).
3 Some c10sely related views can be found in McCord (1981) and qvist et al. (1979).
In the former, reference situations or objects are encoded into restrictions on variables
bound by unselective quantifiers, while in the latter, an explicitly statistic al approach is
proposed, though without grammatical underpinnings.
4 For example, Kamp (1981), Heim (1982), and Groenendijk and Stokhof (1987). The
latter authors sharply distinguish their approach from the former two, and do not call
their approach "Discourse Representation Theory"; but it nonetheless seems c1ear that
the approaches are quite similar, as can be seen from comparing "Dynamic Predicate
Logic" with the "File Card Semantics" of Heim (1982: Chapt. III). We therefore treat
them ali as variants of one single idea.
5 Manfred Krifka informs us that Uwe Reilly and Werner Frey have developed a
(computerized) naturallanguage understanding system which treats "plural objects" in a
Discourse Representation Theory.
6 We follow the practice of Schubert and Pelletier (1987) in our symbolism. (1) We use
"restricted quantification" in an obvious way: (3x: cf>x)'P x is an abbreviation for the

unrestricted (3x)(cf>x & 'Px), (Vx: cf>x)Hx is an abbreviation for (Vx) (cf>x .... 'Px).
Sometimes we use 'M' for most. (2) Where cf> is a common noun, we use t(cf as the
name of the kind, cf>. (3) 'R' (which stands for "realization") is a relation between
objects and the kind which they instantiate: R(a, t(cf) says that a is an instance of the
kind, t(cf. Some Carlson-inspired theories also use this R to describe the relationship
between an object and a "momentary stage" or "time slice" of that object, and also for
the relationship between a kind and these "momentary stages" of objects. As we argued
in the earlier paper, since we have time indices as part of our semantic evaluation
mechanism, we have no need of these latter uses of 'R'. Our use of "realization" is
restricted to an object's instantiating a kind. (4) sm is used to indicate the unstressed
'some'. (5) Some formulas are written in a "categorial grammar style", as for example
loves(kim) (robin) means: the 2-place operator loves combines with kim to produce the
1-place loves(kim), which then combines with robin to produce the O-place robin loves
kim. Some further comments on our use of the "categorial grammar style" and its
relation to ordinary first-order notation can be found at the beginning of the Appendix.
7 We pick a non-quantified version of a "donkey sentence" for discussion. We think
that ali the points we make carry over to quantified versions, and that furthermore this
prevents us from getting involved in pseudo-issues of relative quantifier scope.
8 By "reading" here we intend the truth-conditions most naturally associated with the
following English-like sentences. The sentence (7) has been seen to "mean" each of the
following "readings".
9 Although Heim (1982) does not treat the if-then construction as a material
conditional, but rather (following Kratzer 1979) as a modal proposition in which the if
part restricts an "invisible" necessity operator, her analysis nevertheless has the univer-
sally quantified material conditional (Le., the "standard reading") as a consequence. See
her 1982: 186-188. Roberts (1987) presents a related theory which also follows
Kratzer 1979 in interpreting if-then constructions as modal propositions; and like
Heim she endorses the "standard reading" of sentences like- (7), at least when attention
is restricted to the extensional portion of the truth conditions. See her 1987: Chapter 1.
10 Other arguments against a quantifier raising analysis seem to turn on having
quantified subjects. It would be inappropriate, it is contended, to analyze Every man

who owns a donkey will ride it to town tomorrow with a wide scope a donkey, since
then every man would owe the same donkey. But even in these sentences there is the
problem of its being vacuous: if there were a donkey which was not owned by every
man, the sentence would be true.
II The term comes from Geach (1962) who views some pronouns as replacements for
a stretch of previous text which we are just too lazy to repeat. Differences from other
types of anaphora are most easily seen when the pronoun is a replacement for a stretch
of text which itself contains a pronoun, and this contained pronoun becomes "depen-
dent upon" different NPs. Consider Any person who eats his food with a fork is socially
more adept than one who eats it with his fingers (ef. Karttunnen 1969). The it here
must stand for his food, and this new his will "depend upon" a different NP from the
original his.
12 LePore and Garson (1983) erroneously attribute to Partee (1972) and Evans (1977)
the Non-Specific Existential Reading.
13 "Roughly, the pronoun denotes those objects which verify (or that object which
verifies) the sentence containing the ... antecedent" (p. 469).
14 "Scope widening" as opposed to "quantifier raising". A quantifier raising analysis
takes the existential quantifier found in the antecedent and "raises" it to have wide
scope, yielding the NonSpecific Existential Reading. The scope widening anlaysis tries
to preserve logical equivalence by changing the existential quantifier to a universal
quantifier with wider scope.
15 There is a reading of (12) similar to this, based on a strictly temporal interpretation
of usually (Le., most of the time, cats as a species have the property of landing on their
feet). However, even in this reading, a universal quantification of the cat-variable is too
strong, since generic sentences do not necessarily apply to ali instances of the generic
term, but only to "typical" ones.
16 This particular complaint, that in an "unselective quantifier" approach existentials
will sometimes be translated as ordinary bound variables and sometimes as unselective

quantifiers, does not directly apply to Heim (1982) who always translates indefinites as
free variables. However, it applies indirectly since she binds the appropriate variables
by unselective quantifiers (especially those in restrictive clauses) and then existentially
quantifies remaining unbound variables ("Existential Closure" of her Chapt. II). The
problem manifests itself in an additional way: she sometimes translates universal
quantifiers as "ordinary quantifiers" and sometimes as "unselective quantifiers". See for
example her discussion in 1982: p. 14lf. Matters are somewhat different in the "File
Change Semantics" of her Chapt. III. This approach is somewhat closer to ours than her
Chapt. II approach, at least for the limited class of sentences she concerns herself with
there, but we shall not pursue this similarity here. (JVe still find important differences
also, such as her "building existential quantification (of indefinites in the consequent)
into the interpretation of the quantificational conditional" - while we always treat
indefinites as explicit existential quantification, but again we will not pursue these
differences. )
17 The alternative approach, due to Carlson, of treating generic existentials as denoting
certain properties of kinds was also argued (in 'Our earlier paper) to be flawed and
would lead to a similar non-uniformity of translation as Lewis's approach.
18 This method is also pursued in Groenendijk and Stokhof (1987), although their
solution is somewhat different than ours in that it yields the Universal Reading of
conditionals. It is akin to ours in that it tries to preserve the compositionality of
generating logical form representations from the initial English.
19 Of course, had we translated is behatted as is wearing a hat, then the story would be
different. In such a case, there would be a further context change induced by the
indefinite a hat.
20 We are here interested in the Indefinite Existential Reading, and not in other
possible readings such as the Deictic Reading, the Generic Reading, or the Specific
Existential Reading (which, following Fodor and Sag (1982), we view as some sort of
"referential" reading - see also Section XII).
21 There are a variety of "pragmatic" reasons why we might rule out one of the two
men from consideration. We might, for some reason, explicitly agree that we are talking
only about, say, the c10sest one, or the one we know. More saliently, it is often the case
that we agree to talk about only the things we in fact actually do notice; and should one
of them be hidden, say, then implicitly we can "agree" that such a man is not relevant to
the discourse. Further discussion of this last option is in the next paragraph.
22 We have in mind here that for some things it is not c1ear which of these sets they fall
into. For example, people just in the process of putting on a hat. The idealized context
will make a decision on this; but is not of any interest which decision, so long as the
idealized context agrees with "reality" in the c1ear cases.
23 (MPa) ignores various fine points about plurals, such as how we know that the
collections of cats dropping to the ground here are "really" individual cats doing
individual droppings-to-the-ground. In general we think that the issue is more a matter
of "world knowledge" than semantics. For example, When men lift a piano, they ... can
be either about individual men lifting a piano, or about a collection of men lifting a
piano. The referent of they in the consequent c1ause must somehow be able to pick out
the relevant denotation. Some pertinent remarks can be found in the Coda (on
plurality).
24 We intend to use this sort of a modal operator to account for the nomic force of
typical generics and habituals.
25 This terminology is inspired by L0nning (1987), whose related work recently carne
to our attention.
26 Consider such a sentence in the context of some previous discourse. "At the
Universitt Tiibingen, they have their priorities straight. A student is usually intelligent."
Here the "missing reference ensemble" would be students enrolled at Universitt
Tiibingen.
27 We think a further sub-division of O would be desirable. Perhaps it should contain
T (things), A (attributes), E (episodes), and P (propositions), but for present purposes
we will not consider such a further subdivision.
28 Thus, for example, given a context [], evaluation of (the translation of) A man
yawned in that context relative to time i and world w yields the new context
(PAST(3x: man(x))yawn(x));' w[ ].
This new context will contain new values for x (namely, some man who yawned), and
for the second element of the reference time sequence ,r (namely, the time of his
yawning). ,
29 The ordering of operations in <1>;' wPPJ~::, r:j is: <1>;' W([ Ld, r:j){lIJ))j. w; i.e., unless
bracketing indicates otherwise, constant or variable "resetting" is done first, then the
context transformation (if any) is done, then the resulting context is applied to the
expression it encloses (here: qI), and finally the function applications indicated by the
superscripts is done.
30 We recognize that the proposed semantics for simple temporal whens is not entirely
satisfactory, in that the when-<lI-adverbial in such sentences seems to presuppose, rather
than assert, <li. A sentence like When John arrives, Mary will drive him home seems to
presuppose, but not assert, that John will arrive at some (particular, but unspecified)
time. In this it is like At (or shortly after) the time John arrives, Mary will drive him
home, whereas sentences like When John arrives, Mary drives him home is like our
quantificational condiional.
REFERENCES
qvist, L., Hoepelman, J., and Rohrer, c.: 1979, 'Adverbs of Frequency', in C. Rohrer
(ed.), Time, Tense and Quantifiers (Tiibingen: Max Niemeyer Verlag), pp. 1-17.
qvist, L. and Guenthner, F.: 1976, 'Quantification Theory for Natural Language
Based on Multi-Dimensional Modal Logics with Demonstratives', in F. Guenthner
and M. Guenthner-Reutter, Meaning and Translation (London: Duckworth), pp.
327-353.
Bacchus, F. and Schubert, L. K: 1987, 'Predicates with Probabilities', Univ. Alberta
Dept. Science Technical Report.
Barwise, J. and Perry, J.: 1983, Situations and Attitudes (Cambridge: Bradford-MIT).
Carlson, G.: 1977a, Reference to Kinds in English, Univ. Mass. PhD dissertation,
available from Indiana University Linguistics Club.
Carlson, G.: 1977b, 'A Unified Analysis of the English Bare Plural', Linguistics and
Philosophy 1,413-456.
Carlson, G.: 1979, 'Generics and Atemporal When', Linguistics and Philosophy 3, 49-
98.
Carlson, G.: 1982, 'Generic Terms and Generic Sentences', Journal of Philosophical
Logic 11, 145-181.
Carlson, G.: 1985, 'Exceptions to Generic Generalizations', ms. Linguistics, Univ. Iowa.
Chierchia, G.: 1982a, 'Nominalization and Montague Grammar: a Semantics without
Types for Natural Languages', Linguistics and Philosophy 5, 303-354.
Chierchia, G.: 1982b, 'On Plural and Mass Nominals', Proceedings of the West Coast
Conference on Formal Linguistics 1.
Cooper, R.: 1979, 'The Interpretation of Pronouns', in F. Heny and H. Schnelle (eds.),
Syntax and Semantics 10: Selections [rom the Third Groningen Round Table (N.Y.:
Academic Press), 61-92.
Dowty, D.: 1982, 'Tenses, Time Adverbials, and Compositional Semantic Theory',
Linguistics and Philosophy 5, 23-55.
Evans, G.: 1977, 'Pronouns, Quantifiers and Relative Clauses (1)' and 'Pronoun~,
Quantifiers and Relative Clauses (2): Appendix', Canadian Journal of Philosophy 7,
467-536 and 777-798.
Farkas, D. and Sugioka, Y.: 1983, 'Restrictive iflwhen Clauses', Linguistics and
Fine, K.: 1985, Reasoning With Arbitrary Objects (Oxford: Blackwell's).

Fodor, J. D. and Sag, I.: 1982, 'Referential and Quantificational Indefinites', Linguistics
and Philosophy 5, 355-398.
Gazdar, G., Klein, E., Pullum, G., and Sag, I.: 1985, Generalized Phrase Structure
Grammar (Oxford: Blackwell's).
Geach, P.: 1962, Reference and Generality (Ithaca: Cornell UP).
Groenendijk, J. and Stokhof, M.: 1987, 'Dynarnic Predicate Logic'. Paper presented at
ASL/LSA "Logic and Linguistics" meetings, Stanford.
Heim, I.: 1982, The Semantics of Definite and Indefinite Noun Phrases, PhD disserta-
tion, Univ. Mass.
Hintikka, J. and Carlson, L.: 1979, 'Conditionals, Generic Quantifiers, and Other
Applications of Subgames', in A Margalit (ed.), Meaning and Use (Dordrecht:
Reidel).
Hornstein, N.: 1984, Logic as Grammar (Cambridge: Bradford-MIT).
Kamp, H.: 1981, 'A Theory of Truth and Semantic Representation', in J. Groenendijk,
T. Janssen and M. Stokhof (eds.), Formal Methods in the Study of Language
(Amsterdam: Mathematical Centre-tracts, Univ. Amsterdam).
Karttunen, L.: 1969, 'Pronouns and Variables', Proceedings of CLS V, 108-116.
Kratzer, A: 1979, 'Conditional Necessity and Possibility', in R. Bauerle, U. Egli and A
von Stechow (eds.), Semantics [rom Different Points of View (N .Y.: Springer-Verlag).
Krifka, M.: 1987, 'A Typology of Generics'. Paper presented to Conference on
Linguistic and Philosophical Aspects of Generics, Stanford.
LePore, E. and Garson, J.: 1983, 'Pronouns and Quantifier-Scope in English', Journal
ofPhilosophicalLogic 12,327-358.
Lewis, D.: 1975, 'Adverbs of Quantification', in E. Keenan (ed.), Formal Semantics of
Natural Language (Cambridge: Cambridge UP).
Lemning, J. T.: 1987, 'Collective Readings of Definite and Indefinite Noun Phrases', ms.
Matematisk Institutt, Universitetet i Oslo, Norway.
McCord, M.: 1981, 'Focalizers, the Scoping Problem, and Semantic Interpretation
Rules in Logic Grammars', Technical Report, Univ. Kentucky.
Partee, B.: 1972, 'Opacity, Coreference, and Pronouns', in D. Davidson and G. Harman
(eds.), Semantics for Natural Language (Dordrecht: Reidel), pp. 415-441.
Partee, B.: 1973, 'Some Structural Analogies Between Tenses and Pronouns in English',
Journal of Philosophy 70, 601-609.
Partee, B.: 1978, 'Bound Variables and Other Anaphors', Proceedings of TINLAP-2,
pp.79-85.
Pelletier, F. J. and Schubert, L. K.: 1985, 'Mass Expressions', in D. Gabbay and F.
Guenthner (eds.), Handbook of Philosophical Logic, VoI. IV (Dordrecht: Reidel),
forthcoming.
Pollard, c.: 1983, 'On the Correct Use of Lambda Calculus in Generalized Phrase
Structure Grammars', ms. Stanford University.
Richards, B. and Heny, F.: 1982, 'Tense, Aspect, and Time Adverbials, Part 1',
Linguistics and Philosophy 5, 59-107.
Roberts, C.: 1987, Modal Subordination, Anaphora, and Distributivity. PhD disserta-
tion, Univ. Mass.
Schubert, L. K.: 1982, 'An Approach to the Syntax and Semantics of Affixes in
'Conventionalized' Phrase Structure Grammar', in Proc. of the Fourth CSCSIISCEIO

Conference (Univ. of Saskatchewan), pp. 189-195.
Schubert, L. K. and Pelletier, F. J.: 1982, 'From English to Logic: Context Free
Computation of 'Conventional' Logic Translations', American Journal of Computa-
tional Linguistics, 26-44. Reprinted, with corrections, in B. Grosz, K. Spark-Jones,
and B. Webber, Readings in Natural Language Processing (Los Altos, Cal.: Morgan
Kaufmann), 1986, pp. 293-312.
Schubert, L. K. and Pelletier, F. J.: 1987, 'Problems in the Interpretation of the Logical
Form of Generics, Bare Plurals, and Mass Terms', in E. LePore (ed.), New
Directions in Semantics (London: Academic Press), pp. 387-453.
HENK ZEEVAT
REALISM ANO OEFINITENESS
INTRODUCTION
This paper is concerned with the consequences of the familiarity theory

of definiteness for the foundations of natural language semantics. What
I will in particular explore is the question whether it is possible to
interpret the claims in the familiarity theory of definiteness within a
strictly realist framework or whether it forces a psychological inter-
pretation. My claim will be that it forces such an interpretation, though
I do not wish to claim that there are no possibilities that could be
fruitfully pursued within a more realist framework.
Section 1 develops a theory of the use of definites that though it is
not strictly speaking a familiarity theory of definiteness, assigns a major
rale to uses of definite descriptions based on some form of familiarity.
In section 2 tries to give a philosophical interpretation of the familiarity
concept. Section 3 raises some wider issues conceming the psycho-
logical interpretation arrived at in section 2. A sketch of a formalization
of the theory in section 1 in terms of Oiscourse Representation Theory,
and a similar sketch of the correspondence theories of section 3 follow
in an appendix. This paper elaborates on some ideas in Zeevat (1987).
For a fuller statement of the theory of belief occasionally alluded to in
the next pages I also refer to that paper.
1. DEFINITES AND FAMILIARITY
Irene Heim's dissertation revived the idea, originally due to structuralist

linguists like Christofferson in the Thirties, of associating the contrast
between definite expressions and indefinite ones with oppositions like
''familiar'' versus "unfamiliar" or "given" versus "new".
In this view, both definite and indefinite expressions refer to objects,
and the contrast is converted to different instructions to somebody
interpreting the expression. In the case of indefinites the interpreter has
to create a new object as the referent (with which he is thereafter
269
270 HENK ZEEVAT
familiar), while definites require him to find the referent among the
objects he is already familiar with.
For indefinite expressions, the familiarity theory has led to analyses
of phenomena that were not properly understood before (Heim 1982,
Kamp 1981).
For definites, the concept of familiarity sheds light on the anaphoric
uses of pronouns and definite descriptions. By itself however, it does
not seem to be sufficient to describe alI uses of definites. This section
is an attempt to give an overview of the various uses and to see to
what extent they can be understood as depending on the concept of
familiarity.
The folIowing list contains the concepts that may be involved in the
use of a definite description.
convention
definition
demonstration
experience
communication
relation.
Under convention come the uses of definites that are most like
proper names. Certain languages (colIoquial German, ancient Greek)
allow proper names to be preceded by a definite article. Generic uses of
definites may also be understood as standing for their referent (the sort,
or the prototype of the sort) solely on the basis of a linguistic
convention. Some examples:
the sun
derJohann
the tiger (is a fearsome animal).
In these cases the referent is identified directly by a convention
whose content is exhausted by the fact that the expres sion is referring
to the entity it stands for. Some of the other uses come close to this: a
description may start out as a different use, but then become the
conventional way of referring to that object among a number of users.
This is not the sort of convention that defines a language as a systeQl,
but rather a convention that governs the use of the languag,e between a
group of speakers for a certain period of trtne in a certain context.
Definition covers the case that conforms most to the Russell analysis:
REALISM AND DEFINITENESS 271
The king of France is bald.

The murderer of Smith must be insane.
The first man to be born in the 21 st century will be an
Australian.
In these cases - we are interested in the readings where the referent
(if there is any) is not known by other means than the description - it
is not possible to assume that there is a direct convention to the effect
that the referent (if there is any) is named by the expression: such a
convention would presuppose a previous acquaintance or identification
of the object. The function of the complex noun is to describe the
object in such a way, that it is the only object that fits the description.
The success of this use of definite descriptions depends on the
recognition on the part of the interpreter that the speaker has
succeeded in giving a description that indeed singles out the referent.
These uses can be discriminated from the others by this property of the
noun, and the fact that no other means of access to the referent is
available to the interpreter.
Demonstration is operating in those cases where a definite is used to
single out an object within the non-linguistic context of the utterance.
Generally, on such a use the definite article can be replaced by a
demonstrative article, or by a demonstrative pronoun.
Look, the/that paddle is drifting away.

This use is closely related to definites referring to objects in the
common experience of the two communication partners or the wider
non-linguistic context. Here it is possible to employ descriptions whose
nominal part falls short of singling out the referent.! Such use can be
found in:
Did you go to the baker today?

The director is ill today.
The computer is down.
Anaphoric use of definites can be seen as a special case of the
previous use. 2
The clearest examples of anaphoric use of descriptions is where they
stand in for normal pronouns. That is, they take their antecedent in the
last sentence, and we can slightly change that senfence so that anaphora
by a pronoun instead of a description is the most felicitous. U sing
272 HENK ZEEVAT
descriptions instead of pronouns is obligatory in situations where there

are two or more antecedents with an equal status in the previous
sentence or sentence pair that cannot be distinguished on the basis of
gender or number. Contrastive pairs and conjunctions are the typical
cases where this happens.
John is quite happy with his results, but his teacher thinks
otherwise. John maintains that the teacher just does not like
him.
John is quite happy with his results, but his teacher thinks
otherwise. *He maintains that he just does not like him.
1 met John and Bill. He* /The first was coughing.
Other cases of this use can be found when the anaphoric element has
to pick up the antecedent over a longer distance. Mostly, anaphoric
pronouns have their antecedent in the previous sentence. 3 When the
antecedent is too far away or otherwise unavailable as the antecedent of
a pronoun, it is necessary to use a proper name or a (generally short)
description.
A special case are the relational uses of descriptions. Here the
referent is not an already experienced or mentioned object, but
something that is related to an item that can be referred to by a
pronoun in the context. The noun in the description does not refer to a
one place property as in the standard use of common nouns but must
be reinterpreted as a relation, e.g. by taking into account a possessive
relation as in the example below. Compare:
John bought a new car. It was not expensive.

John bought a new car. He had to exchange the motor.
A typical property of this last use is that the referent need not be
uniquely determined after the identification of the high focus item. The
definite in those cases instructs the addressee to establish the link with
the high focus item, but does not at the same time establish an
identifying concept for the referent. The contrast with an indefinite in
these cases is that the indefinite does not require such a link.
The familiarity concept in the strict sense only shows up in the
convention, demonstrative, experience and communication cases. Here
it is possible to speak of an already established connection of the
interpreter with the referent. In both the definition and the (strict)
relation cases it is the communication itself that establishes the link
between the object and the interpreter: he obtains an identifying
concept of the referent that may depend on his identification of the
high focus item in the context.
In conclusion one can say that the functioning of definite descrip-
tions is governed by a number of closely related concepts. One thing
that they alI seem to have in common is the expectation on the part of
the speaker that the interpreter will manage to single out a particular
referent, on the basis of the material provided by the noun.
There is empirical evidence for this last claim. For example, Clark
and Marshall (1981) show that after full descriptions in conversation
(which tend to introduce referents new to the conversation, but mostly
refer to objects that the speaker and the addressee are both familar
with) the speaker asks for confirmation of the addressee's identification
of the item he refers to, and elaborates if the addressee indicates that he
has not been able to identify it.
It is interesting to notice that if we take the phenomenon of
identification and the strategies discussed above as a criterion for
definiteness, it becomes plain that much of natural language consists of
definites. First of alI there are names. As in the conventional use of
definites, their referents are determined by linguistic convention.
Another group is formed by demonstrative terms and referential
pronouns. If we treat proper names as definites then common nouns,
adjectives and verbs should be understood in the same way: their
contribution is to let the interpreter identify a property or relation.
Lastly, the past tense in English can be understood as involving the
identification of the occasion on which the reported event or state is to
have taken place. The typical identification strategies that are involved
here can be taken as anaphoric (states), relational (events) and perhaps
even definitional. So it seems proper to regard the tense morpheme as a
definite expression.
There are some counterexamples to the claim that identification is
characteristic of definites in general. In English (where expressions
starting with a possessive are generalIy considered to be definite) one
cansay:
Heismyson.
even when there are other sons. 4 As we have mentioned relational
274 HENK ZEEVAT
uses of definites allow imperfect identification as well. The following

conversation is quite plausible.
How is your car? Fine, only the rear tyre is punctured.
(assume that the car has two rear wheels). In French (and other
languages), the examples with possessives can be realised by a normal
definite.
Il posa la/*une/*sa main sur son genou.
It is good to point out that in both cases we are dealing with
relational uses of definites, and that such uses do not show up in either
conventional, definitional or anaphoric ones. It may be best to regard
them as idioms, historically derived from regular dependent definites.
The original content of definiteness - identification - which in these
cases depends on the identification of something else and the establish-
ment of a link between that thing and the referent has given way to the
identification and establishment of the link, that originally was part of
providing the identifying concept for the referent.
The picture that seems to emerge from the discussion above is the
following: definites require the interpreter to identify an element among
either the objects he is already familiar with orby defining a new
object, maybe on the basis of already familiar ones. The objects found
in this manner are the referents of the definites: they are what the
definite contributes to the meaning of the expressions it occurs in.
Naturally, if the definite attributes a predicate to this object, then this is
part of what the speaker claims to be the case, but not necessarily a
part of the message the speaker wants to communicate.
An account of this type treats definite descriptions on a par with
proper names and pronouns. Definite descriptions differ however in
allowing a series of strategies for identifying the referent, where
both names (convention), anaphoric pronouns (communication), and
demonstratives (demonstration) are specialized to a single strategy for
identifying the referent. Several strategies (relation, definition) are
exclusively available for definite descriptions.
2. INTERPRETING THE FAMILIARITY THEORY
The theory in the last section - it can doubtless be improved upon -

is an attempt to cover definite descriptions from a linguistic angle. It
tries to give an account of the various uses of definite descriptions in

terms of a number of identification strategies that employ different
aspects of the context of utterance. Though it is not possible to reduce
every single use to the concept of familiarity, most uses do appeal to a
psychological relation of the interpreter with the object he identifies or
creates as a result of his interpretation of a definite.
In this section the philosophical consequences of such a theory of
definiteness will be explored. These are brought out most clearly by
contrasting the present theory with the classical position on definites
that can be attributed to Frege and Strawson. Here definite descrip-
tions, like names, are referring expressions. They stand, in a context of
use, for an object that is given as the unique object that satisfies the
property to which the nominal part of the description refers. If there is
no such property, the definite does not refer. If the definite occurs in an
extensional position in the expression, the expression can therefore fail
to be true or false. Since the definite gives a way in which the referent
could have been given, such expressions continue to have a meaning.
Thereby an expres sion that contains a non referring definite in an
intensional context can be true or false. Definiteness is here explained
as reference to an object by means of a property that uniquely holds of
the object, perhaps in a contextually given subdomain.
The last section should have made it clear that the use of definites in
natural language does not bear out the second characteristic: there are
cases where a definite is employed in the definitional way but they are
the exception rather than the rule. Definites refer more often on the
basis of the noun together with the familiarity of both speaker and
addressee with the referents than solely on the basis of the conceptual
content of the noun used in the description.
First of alI, this makes it necessary to have a different idea of what
goes on when the definite does not refer: in that case we have a
malfunction of communication rather than a meaningful utterance that
could have been true or false if the world had just been different. We
can attribute meaning only if we have a criterion that would determine
the referent in a counterfactual situation. But there is no such criterion:
there are only strategies that we can use to determine the referent given
a certain psychological state. So we can understand how a sentence
containing a definite for which we cannot identify a referent could have
meaning, but we cannot determine what that meaning would be.
The second question is the analysis of the relation between the
276 HENK ZEEVAT
referent of a definite and the person who interprets the definite as

having that referent. On the classical theory, the relation consists in
obtaining a concept of ihe referent: if the definite refers, the interpreta-
tion process is obtaining a criterion that holds of the referent but of
nothing else. The relation becomes much more complicated when we
take the other uses of definites into account. In the case of anaphoric
uses, or reference to background knowledge, the use of the definite
presupposes that the interpreter already has a concept of the object. It
is by no means clear that this must be a concept that determines a
criterion that holds of the referent and nothing else. The information
that the interpreter has conceming the object may be insufficient for
singling out any object uniquely. Such situations are quite common. A
story could start as follows:
Last week 1 was in London. 1 was in a bar there to have a
drink. The bar was one of these places where ....
When the definite in the third sentence is uttered all we know about
the bar is that tlie speaker had a drink there last week and that it is in
London. London contains many bars, and the speaker probably visited
a number of them during his stay there. Still the interpreter is able to
"identify" the bar: it is the one mentioned in the second sentence.
What then is the relation between a person and the identified object
when the person is interpreting the definite? The familiarity theory
suggests that it should be familiarity, but it is not clear at all how one
can be familiar with an object for which one does not have the capacity
of recognition or the possibility of conceptually singling it out from the
other objects. What is unclear is with which object one is familiar.
There is nothing about any of the various objects that may fulfill ali
the information that the interpreter attributes to the referent of the
description that makes that object a familiar one for the interpreter.
This is perhaps the most worrying aspect about familiarity. The
classical theory suggests the following picture. If one has a definite and
one can prove that the noun it contains is a criterion that singles out an
object, the definite is like a name: it contributes its referent to the
reference of the expres sion. This suggests a realist picture of the
proposition expressed by a sentence containing a definite: it should
have a reference to the referent of the description, plus maybe the
attribution of the property expressed by the noun and the proposition
that says of that property that it is uniquely satisfied. If one tries to
interpret the familiarity theory as an amendment of the classical theory,
the realist interpretation collapses. We can omit the assumption that the
property is unique without any problem, but we do not know any more
to which object the definite refers and consequently we cannot con-
struct the proposition as referring to the referent.
Not only can the referent be underspecified, it can also be com-
pletely lacking or exist only in hardly robust ways. The first situation is
encountered when we as sume that the speaker in the little story is lying:
he was at home the whole of last week. Although there is no real
referent in this case, it does not seem that the story is more difficult to
interpret: as before the referent is the bar mentioned in the last
senteIlce. The second case is encountered when the referent is intro-
duced in a special context (intensional, negative or quantificational),
and subsequently referred to by a definite description, as in the
following examples:
It is false that Buganda has a king and that the king has a
beard.
John believed that a drunk had stolen his wallet and Mary
thought that she had seen the man when he carne into the
house.
Whenever John had a job in those days, he was unhappy
about the salary.
Therefore, it is impossible to have a realist account of the familiarity
theory that follows the pattern of the classical theory: reference plus the
addition of some extra predications to the property and the referent.
The second classical account makes use of Russells definiteness
operator. This operator turns a property into a quantifier. Since the
operator is based on the uniqueness analysis of definites it is by the
considerations in section 1 not adequate as an analysis of natural
language definites. It cannot be amended, it seems, to a "familiarity
operator" either, since familiarity is not a logical notion.
Realist theories that fare better in respect of the problems noted add
extra domains of objects: indeterminates, possible objects, arbitrary
objects and others. In such a theory, it is possible to find referents for
the harder cases noted above. It is still, 1 think, a problem to account
for the relation between an interpreter and the referent if the latter is an
abstract object. Some of the difficulties in this respect will be spelled
out in the next section.
The natural analysis of the familiarity theory is as a psychological
278 HENK ZEEVAT
theory. At any moment of time a language user has a certain number of

concepts of objects. These concepts are the "objects" that he is familiar
with. The referents that he reidentifies are the concepts. Some of the
concepts have a special relation with an external object: they contain
the information about the external object that was the content of the
experience of that object. Some concepts have a relation with a concept
of an object of somebody else: they are the result of interpreting
utterances of the other person that express those concepts. Since realist
interpretations of the content of expressions containing a definite
function only in special circumstances (if we do not take account of the
Meinongian versions of realism for the moment), it seems that we have
to accept the psychological theory for the moment. Unfortunately, the
psychological theory of interpretation does not immediately give us a
theory of the meaning of expressions: it merely teUs us that interpreta-
tion leads to a representation with a certain structure. Content is
something that can be shared by a sentence on different occasions of its
utterance or that people can share. A discussion of the issues raised by
a psychological theory occupies the next section.
3. MEANING AND THOUGHT
What I am proposing is to take the theory of definites at face value and

interpret it as a psychological theory. The entities that have to be
reidentified are the representations that have been formed on the basis
of experience or communication. In the case of definition or relational
use, a new representation is formed that has the definition, or the
relation with the other identified object as its content. "Familiar" has
the straightforward meaning that the item is present in the mental state
and can be accessed by memory on the basis of the information
supplied in the noun. The result of the interpretation process is a
complex representation, with a logical structure that reflects the syn-
tactic structure of the sentence or discourse that is interpreted, and
whose ultimate constituents are the representations that were reiden-
tified or formed in the course of the process.
It is one thing to postulate an interpretation process of this kind, it is
quite another to maintain that hereby one has carried out the program
of natural language semantics. No characterisation of the truth condi-
tions of natural language has been given, which is an explicit aim in
formal semantics. Another aim that is of real importance in Frege's
philosophy of language, and that is automatically answered by any

notion of meaning that is objectiveor intersubjective, is the notion of
agreement between two people. In that case, one can describe it as two
people grasping the same meaning and judging it to be true. When the
theory of meaning doubles as a theory of thought, one also has an
account of what it means for people speaking different languages, or for
a person and a machine or an animal to agree. On the view given here,
both the notion of truth and the notion of agreement need to be
addressed. The next pagescontain a sketch of some possible answers.
Before embarking on these, there is another question that needs to be
dealt with. It has been suggested by proponents of a realist theory of
thought that alI that we need to do is to add a theory of possible or
arbitrary objects to the realist theory of thoughts in order to deal with
the problems that may remain for a realist account. Since my answer to
this last question has direct bearing on the answer to the question about
informative content, it will come first.
a. Extra Objects
On the view that 1 am proposing the representation of an entity has, as
such, some formal properties: an internal identity and a content that can
be thought of as its links with other representations, in particular with
representations of properties and relations, that determine the informa-
tional content of the representation. As such it is formally a concept
that objects in reality may conform to. But it also has other properties
that derive from its history and context: it may be formed on the basis
of a particular experience ar in the interpretation of an utterance, or it
may be part of a belief, a desire or a fantasy. I will employ the word
intention ta stand for this richer notion. The historical properties alIow
representations ta be intentions of external objects: they may represent
an experienced object, ar an abject mentioned in an utterance by
somebody else. They do so because they have a direct causal link with
that experience or with the communication and thereby with the mind
of the speaker. Representations of this kind are similar to proper
names. As in the case of names, one can formulate a causal theory of
reference similar to that for proper names: an intention can be of a
given abject, since the intention was formed in response to a mention of
the object based an another intention of the object that was in turn
formed an the basis of communication, and so an until we reach direct
280 HENK ZEEVAT
experience. In a more psychological view, it is also unproblematic to

allow for chains of intentions that end in some intention that was not
caused, directly or indirectly, by a real object: for fictional objects for
example, or objects postulated in an explanation. Without historical
properties it is not possible to have intentions of external objects: on
the sole basis of the content there is a criterion for an external object
satisfying the representation, but there are no links of a referential kind,
that derive from a causal order. Thus, the common situation in both
communication and experience that an object is experienced or dis-
cussed, but not yet sufficiently described to allow for a conceptual hold
an its identity is beyond a theory that can only invoke conceptual
content.
This is the problem for a theory that tries to add one or more
categories of ideal objects. Such objects are abstract and ideal, and
thereby cannot have historical or contextual properties. The only way in
which they can have an exclusive relation with a single real object is by
having a content which is so determinate, that one single real object
fulfills it. Since this is not generally the case, ideal objects do not always
form a solution in the explanation of definites. Take a sentence like:
John found a penny and gave it to the beggar.
There is no problem with John, if he exists. But there is one with the
penny: we are not told which one it is. So we must interpret it as an
ideal object: some arbitrary penny. The pronoun "it" must refer to a
familiar arbitrary object. But arbitrary pennies are, as such, not distin-
guishable from each other. So what we must do here is appeal to the
history of the arbitrary object: the fact that it has been used earlier on
in the interpretation. But at that point we use a historical property of
the arbitrary object.
The point comes out even more strongly in a dialogue. Suppose
somebody A says:
There is a car outside.
And then somebody else B says:
Is it a green car?
Then it is not clear how A and B could agree on the arbitrary objeCt
to associate with "a car" and "it" respectively. Only from the outside a
third person, watching the exchange, could decide that A and B
associate the same arbitrary object with their utterance, appealing to

B's intention that "is of" whatever A intended with "a car".
A theory that appeals to arbitrary objects in a theory of interpreta-
tion is thereby very similar to one that uses intentions, and only
seemingly more abstract.
It is possible to understand Discourse Representation Theory as a
theory of arbitrary objects, in which they are combined with an account
of how they come into being, and a model theoretic account that
explains their semantical role in terms of real objects. The theory on
their use in interpretation links them to history and context and makes
it possible to regard a discourse representation and its reference
markers as the result of a given interpretation process. In this view, it is
quite natural to think of them as being caused by a real object via
perception, or in communication via a chain of references that origi-
nates in a real object.
One comes close here to the fun notion of intention. The markers
are provided in this way with a history and a location, that make it
possible to relate them in a direct way with real objects.
b. Truth
Truth for thoughts can be explained as correspondence between

thoughts and reality. A first step in the explanation of correspondence
has already been taken: for concrete intentions it can make sense to say
that they are intentions of a real object. In that case, they correspond
with this external object. It may be the case that an intention corre-
sponds with more than one object, because of a mistaken identification.
Similarly, there may be more than one intention that has the same
external object. But there is no reason to as sume that in this case they
do not correspond with the external object(s).
Not an intentions are of external objects in this sense. Intentions may
result from hypothesizing a real object, or be general in nature. In such
cases, their correspondence with real objects is different. Real objects
may fulfill the content of the intention. In this case, the intention
corresponds with the class of objects that fulfills its content. 5 Using
these basic correspondences we can map from intentions to corre-
sponding objects, universals and intentions. This is sufficient to map a
thought in which one intention is predicated of a 'sequence of others to
the fact that the relation corresponding to the first intention is instan-
282 HENK ZEEVAT
tiated by a sequence of entities that correspond to the intentions in the

sequence. And, if we allow ourselves an appropriate syntax for general
thoughts the correspondence can be extended to more complex
thoughts as well. As usual, one must assume that the structural relations
between simpler intentions that make up complex thoughts correspond
with logical operations. This gives the following definition:
A thought is true iff it can be mapped, preserving both the
basic correspondences and the correspondences between
structural operations on thoughts and logical operations to a
fact in reality.
Notice that under this conception it makes sense to as sume that the
correspondences are sometimes filled in by objects that are themselves
intentions: they can be both external objects of an intention and objects
that fulfill the content of a certain intention. This possibility allows for a
natural treatment of the attitudes and of fictional and other names that
lack a referent.
As an example consider the thought that John and Bill believe that
the same unicorn walks. This thought is true iff a particular thought
belongs to the beliefs of both John and Bill. This thought cannot intend
a real unicorn since there are none available. But it can intend the
intention of one that Bill and John have corresponding intentions of. In
this way, the thoughts of John and Bill correspond to the same thought
(possibly either John's or Bill's) which does not correspond to any fact,
because there could not be one.
So one can state when a thought is true. Having information can then
be defined as having a true thought. The information may be equated
with the intention, and its content with the fact (or class of facts) with
which the thought corresponds.
This notion of truth brings us close to the realist notion of a thought:
some structure made up from real entities combined in a structure that
characterises the logical relations of these entities. There is a difference:
not every thought corresponds to such a structure. On1y those do, that
are true and whose ultimate constituents correspond with a real
external object, or are fulfilled by a real entity. Others may refer to
intentional entities, or simply fail to correspond sufficiently with
anything. Another difference is that, even when a thought corresponds
with some fact, it cannot be equated with that fact or a concept of that
fact.
c. lnformation
The notion of information described above may seem to be restricted

to language users or beings endowed with an intelligence. This section
tries to make two points: first, that the notion of information described
in the previous section is only a special case of a more general notion
and, second, that it is not so much language use that is responsible for
attributing structure to information, but rather the concept of informa-
tion itself. Attributing structured thoughts and information to animals
or machines is thereby in principle possible, but involves the postula-
tion of a certain functional and causal principles that govern the acqui-
sition of information and the effect of information on the behaviour of
the animal or the machine. Such attributions and their evaluation
thereby depend crucially on the correctness of our assumptions about
their functioning and on the possibility of translating their internal
organisation into our own. In the case of language users it seems we
have a special kind of causality and functionality available that makes
the attribution of a certain structure to their information simpler.
Having a true thought is on1y a special form of having information
about something. The notion of having information can also apply to
books, machines, fossiles, stones and practically everything that can
enter into causal relationships with something else. What makes some-
thing have information, are particular causal 6 principles such as:
Normally, chairs are warm iff they have been sat on.
The warmth of the chair bears the information that it has been sat on
in virtue of the principle above. Unless something unexpected happens,
like somebody storing another warm object on it for some time, or the
installation of heating in the seat, the warmth is a sign that it has
recently been occupied. The informational value is absent in the
abnormal situation: it is strange to say that the chair contains the
information that somebody sat on it, when it has been heated by, say, a
hot water bottle.
This changes however in the following situation, where the abnormal
situation seems to tolerate a stronger description. Some microorganism
has a primitive method of movement. The behaviour of the organism is
such that it moves whenever it does not find sufficient nutrition in its
environment, obviously in the hope of changing,this situation. Now if
the organism moves, we can attribute to it the information that there is
284 HENK ZEEVAT
no food in its environment. Suppose now that by clever engineering we

can produce a chemical environment that is full of food but stiH has the
organism moving about. We can say in this case that it thinks that there
is no food. What we invoke here, it seems, is the purpose of the moving:
leaving a foodless situation for a better one. The moving is thought of
as a functional response to the information that no food is available.
Since chairs do not have purposes, it is strange to attribute information
to it on the basis of the causal principle in case this does not apply. In
the case of the microorganism, the causal principle is linked with
functional purpose of its movement: the organism behaves as if it were
causally affected in accordance with the principle.
There is another point here. Suppose that the animal is swimming in
a pint of beer. Beer does not contain the particular food that it needs,
so it moves about. It would again be improper to ascribe to it the
information that it is swimming in beer or in a pint of beer. This has to
do with the fact that swimming in beer is (ceteris paribus) a sufficient
condition for it to move, which in turn is our only reason to attribute
information to it, but not a (ceteris paribus) necessary condition. It is
the nature of the causal principle invoked in the ascription of informa-
tion that determines how the ascribed information must be categorised
and structured.
The notion of structure comes out better in a more complex
example. Suppose a cat sees a mouse disappear under the refrigerator.
It seems natural to attribute that information to it: its behaviour after
witnessing this scene indicates strongly that it has got the information.
The causal principles that must be invoked here are more numerous
and complex. They contain at least the ability of the cat to distinguish
mouses from other things and other animals and its ability to perceive
the motion of things from one place to another and calculate the
consequences for its own movements. Whatever the nature of the causal
trace that the event leaves in the cat, it must contain somehow the trace
of the recognition of a mouse, based on the distinction abilities, and a
trace of its movement based on the movement perception. The struc-
ture that we attribute to the cat's information is a reflection of the
perceptual and conceptual skiHs we attribute to it. It is not necessary to
attribute language to it in order to ascribe information to the cat that
has a considerable amount of structure.
It is the dependency on the particular causality by which a fact
causes information to arise in something that is affected by it that
imposes a structure on the information itself. There are only certain

aspects of a given factual constellation that are efficient in causing
information. In case the informative state is dependent on a series of
caus al principles, the analysis of the fact that would supply the series of
efficient causes is inherited by the resulting information state. The
complexity of thoughts, does not reflect some inherent organization of
the facts Of of the storage mechanism, but the complexity of the causal
link from fact to thought.
d. Communication
The criterion for successful communication is the preservation of infor-
mation through coding and decoding. Somebody has the information 1,
corresponding to a fact F. He expresses his complex intention 1 by a
sign S that is perceived by somebody else. That other person interprets
S and in the course of that constructs a complex intention that also
corresponds with F.
An efficient communication system must therefore be an expression
system for constructing signs on the basis of intentions and an inter-
pretation system for constructing intentions on the basis of signs such
that whenever an intention corresponds with a fact F, and can be
handled by the expres sion system, the interpretation system delivers an
intention that corresponds with the same fact F.
One of the basic functions of natural language use is to be an
efficient system of communication in this sense. It can be argued from
an evolutionary perspective that, together with the possibility of
reaching agreement (which is crucial for collaboration), this is what
makes having a language a biological advantage over not having a
language.
Naturallanguage use can be divided in a system of coding and one
for decoding signs. The first is the expres sion system, that regulates the
expres sion of thoughts in signs. The other system deals with interpreta-
tion: how to construct a thought on the basis of a sign. Both the
expression and the interpretation systems can be thought of as a system
of rules. This is a large and complex set that can be understood as a
series of causal principles that, under normal circumstances, make a
thought of one person the cause of a thought of another person, in
communication between the two. The causality,here seems to place a
number of constraints on both the cause of the exchange - the
286 HENK ZEEVAT
speaker's representation, and on the result - the representation of the

interpreter: they must have a certain structure and a certain complexity.
Even the sign, whose parts and structural relations correspond with
rules of the communication system must reflect this structure to a
certain degree. Therefore, the study of semantical regularities of the
language in question both from a linguistic and a psychological point of
view provides insights that are relevant for the structure of these
representations.
There is another consequence of assuming a more or less causal
relation between the thoughts of one person and another. It can be the
basis of a counterfactuallike the following:
It somebody expresses a thought T to somebody else, thus
causing a thought T' to arise, then, if T corresponded to a
fact F, T' would have corresponded too.
It the counterfactual stood alone, it would be impossibly vague.
However it rests on the regularities that constitute the communication
system and those are an object of research. What it provides is an
approximation of what it is for two people to agree: it is having two
thoughts that would correspond to the same possible facts. To give this
criterion some content one needs to analyse the notion of possible fact
and what it is for a thought to correspond to a possible fact.
In principle this is the correct strategy for explaining agreement. E.g.
one can identify possible facts with thoughts and define correspondence
between thoughts. But in defining such a correspondence relation, one
has in principle already done more: a notion of corresponding is itself
a theory of what it means for two persons to agree with regard to a
thought.
Suppose we have identified a thought. That makes it possible to
consider whether the thought has a realisation in reality: we can check
whether there is a fact that corresponds to it. It is a similar problem in
abstracto to find out whether there is a thought of somebody else that
corresponds to it. Again we must find corresponding counterparts in
the mind of that other person, and ensure that the mapping can be
extended to the structural operations. Assuming that the structural
relations that build the complex intention correspond when they code
the same logical relations, the problem is limited to stating the corre~
spondences for the primitive intentions. ,A condition for correspond-
ence that is necessary, is that the two intentions correspond with the
same external object, or class of objects, if they correspond with

anything. Here as before intentions themselves may figure among the
external objects.
But this condition is not always sufficient: under this criterion alone.
Two Babylonians, one thinking that Hesperus is far away the other that
Phosporus is far away, do agree, if both the Hesperus concept and the
Phosphorus concept are derived from experience of Venus.
What seems to be required is an extra condition on the content of
the intentions. It may be possible to state the extra condition in terms of
the quasi causal relation given by a communication system: both
thoughts should be able to cause the other by such a system. It they do
not participate in the same system by speaking different languages, the
hypothetical causal link may be made more complicated by making it
dependent on trauslation.
Again there is a temptation here to identify aU thoughts that bear
such a correspondence relation, aud to think of thoughts as the appro-
priate equivalence classes. When one leaves out the intentional external
objects, and limits oneself to the criterion of external correspondence,
one obtains a version of pure realist thoughts. When the intentional
objects are added to that system, a system is obtained that admits ideal
objects among the real objects. It is harder to state what one obtains
when extra conditions are put on the correspondence: it would seem
that copies are aUowed of both real aud ideal objects in the set that
generates the set of thoughts.
AlI these possible theories of thoughts share with their predecessor,
the possible worlds view, that they are abstractions from the psycho-
logical level on which thoughts live. The question that I have tried to
raise in this paper is whether they have a direct value when one tries to
understand the interaction of language and thought as in the explana-
tion of the communication process. They may seem harmless abstrac-
tions at first, but one is here at precisely the point where the problems
start in an account of propositional attitudes, and where the level of
abstraction becomes such that the familiarity theory can no longer be
stated in a natural way.
APPENDIX: A SKETCH OF A FORMALIZA TION
This paper is concerned with philosophical issues concerning meaning,

not with the question of what formalism is most suited for dealing with
288 HENK ZEEVAT
definites. This section should therefore be read as an attempt to make

the ideas on definites in the section 1 more precise, not as an
expres sion of a preference for a certain kind of formalism: as far as that
is concerned, it seems that many formalisms would do and that the
decision among those depends on other issues as welI. The second
caveat concerning this appendix is that the subject of interpretation for
definite descriptions is changing rapidly as it becomes better under-
stood. This means that the rules given here may well have to be
changed or replaced. 1 would maintain however that the general picture
is close to what actualIy goes on in the interpretation of definites and
can serve as a first step.
Section 1 ended with the remark that there are a number of
strategies for identifying the referent. These can be related to rules in a
development algorithm for DRSs, if the latter are slightly enriched. This
enrichment is necessary, since not alI the restrictions and conditions
needed in the strategy are expressible in terms of the discourse
referents and the conditions in the old representation.
1 will begin by giving the rules for interpreting definites, and only
later attempt to define the notions by means of which they are stated
more precisely. The interpretation of the DRS under construction is
here as a characterization of the mental state of some interpreter of
sentences rather than as a record of the meaning of some discourse.
This alIows one to have the background knowledge of the interpreter in
the DRS, next to the record of the ongoing discourse.
The general pic ture of the expansion rule for definites can be given
as follows:
at a certain point in the derivation we find an unexpanded

definite description "the D" in an unexpanded expression
"S(the D)" in a box B in the partial DRS A and we decide
to deal with it. We now expand the (complex) noun D to
obtain a DRS D', separately, but as if it were in box B, with
respect to B in A, identify a discourse marker x by means of
one of the strategies and deal according to the same strategy
with both x and D '. We then continue with Sex) in box B.
We have to distinguish five possibilities.

a. We can prove on the basis 'Of the information available in
box B that there is only one thing that satisfies D' even if
we do not have it available as an accessible discourse
marker. It a discourse marker x satisfying D' is available

we add D'(x) at the highest possible level, and proceed
with processing Sex) in B.
b. There is on1y one discourse marker x available in the
current topic or extra linguistic context that on the basis
of the information available in box B satisfies D'. It can
be justified that no pronoun is used. We proceed with
Sex) in box B.
c. x is the most recent discourse marker in the present
discourse part of the DRS that satisfies D'. We proceed
with Sex) in B.
d. The background part of A concerned with the current
context of discussion contains on1y one marker x that
satisfies D'. Alternatively, there is nothing in that back-
ground part that tells us that there could not be such an
x, we create one and add D'(x) at the highest possible
level seen from B. We proceed with S (x) in B.
e. The topic or the clause under development contains a
marker x of which we can prove that it has a D'.
Alternatively the topic or clause under development
contains an x of which we can as sume that it has a D'.
We create a marker y and add it, has (x, y) and D'(y) to
the highest possible level. We proceed with S(y) in B.
Here (a) corresponds with the definition and convention uses

distinguished above. The definition enforces that conceptual uniqueness
is already provable. It allows for accomodation in case the discourse
marker is not already familiar. (b) corresponds with the demonstrative
and anaphoric use of definites. Here uniqueness is not conceptual: the
noun may hold of any number of things. The crucial criterion is being
in the focus of attention and familiarity. Accomodation is therefore
impossible. (c) corresponds with another anaphoric use of definites, of
which more or less the same holds as for (b). The difference is that
there is no almost complementary distribution with pronouns. (d)
corresponds with what we earlier called experience and communica-
tion, be it that in case the experience or communication are not
sufficient, we may accomodate the use of the pronoun. (e) finally deals
with the relational use. Again there is the possibility of accomodation.
290 HENK ZEEVAT
In order to make this work, we need to impose some extra structure

on the set of discourse referents. In this way, we obtain a formal
counterparts of the notion of "topic", the non linguistic context and of
the present discourse. As usual a DRS has a set of discourse markers. It
further contains a set of conditions that are either atomic formulae or
implications, coded as ordered pairs of DRSs. A special DRS is the
absurdum. In any stage of the interpretation process, we are at a certain
level of the DRS, either at the top, or in some box that is part of an
implication. From this position there are five mutually disjoint sets of
accessible discourse markers: the background (BG), the present dis-
course (PD), the seen (S), the topie (T) and the minimal domain (MD).
S moreover is included in the discourse markers of the top DRS. The
other three sets can contain markers from the top level, but may also
contain elements from the current DRS or from any intervening one.
What is seen changes by adding new elements when they are seen
during the discourse, and by removing elements when they disappear.
Information conceming those objects may be available in the conditions
of the top box. The previous discourse contains a record of all markers
that are both accessible from the current position and have been
referred to in the course of the discourse. The minimal dom ain contains
the referents of the elements from the clause that is current1y being
developed.
An approximation of the notion of topic that we require can be given
as follows. A minimal domain is given by the set of discourse referents
introduced by the subject, the objects and maybe the PPs depending on
a verbal group. As is well known the on1y anaphora among these
elements that is possible uses reflexives and possessives. It is not clear
to what extent markers from PPs in the NPs should be included among
these. Certain1y, relative clauses have their own minimal domain. 1 use
topic here for the possible antecedents for personal pronouns. A
personal pronoun can be anaphoric with and on1y with one of the
elements of T. A reasonable approximation is to say that T consists of
the elements outside the current minimal domain that belong to either
the minimal domain of a preceding sentence (which one depends on the
structure of the text) or to preceding subordinate clause, or to a
superordinate clause. A small example:
John visited Mary. As he left~ he noted arnan at the other

si de of the road. The man, who carried a pair of binoculars,
had been there before he went in.
This small text contains the following six minimal clauses:

1. John visited Mary
2. as he left
3. he noted a man at the other side of the road
4. the man had been there
5. who carried a pair of binoculars
6. before he went in
MD 1 = (john, mary)
MD 2 = (john)
MD 3 = (john, man, (road1)
MD 4 = (man)
MD 5 = (man, pair)
MD 6 = (john)
T1 =0
T2 =MD]
T3 =MD] U MD 2
T4 =MD 3 U MD 5
T5 =MD 3 U MD 4
T6 =T5 U MD 5
Definitions for the other notions are given below. No such notion is
given for recency, because it complicates things, and further, no
treatment is attempted of contextual restrictions in the background
knowledge. The notion of a current DRS during the development
process is dealing with the information in box B. The notion of a
highest possible level for the addition of new material is formulated,
like the accomodation in terms of the possibility of adding it without
making the level contradictory.
If A is DRS and B is a box in the DRS, current(A, B) is the
DRS consisting of the discourse markers that are accessible
from B in A, and the conditions that are accessible from B
in A.
The information available in a subDRS B of A equals the
information contained in current (A, B).
A 1- B iff for every model M and truthful embedding. f of
292 HENK ZEEVAT
A on M there is a truthful embedding g extending f with

respect to the discourse markers of B that is a truthful
embedding of B.
If B is a box in A and current(A, B) proves C, the highest
possible level from B for C with respect to A is that box D of
A that is accessible from B such that current(A, D) proves
C and no box E accessible from D has the property that
current(A, E) proves C.
We can add certain information C to a box B in a DRS A iff
not current(A, B) f- ..., C.
CORRESPONDENCES
As in the first part of this appendix, the remarks in this section are
rather tentative. 1 will try to assume as much structural similarity
between the two kinds of structured entities that 1 am considering here:
minds and worlds. In the case of minds it is possible to as sume more or
less the model of a DRS adopted before, except for the special features
we had to add for the interpretation of definites. Of course, this model
is very simplistic.
A world can be reconstructed as a special kind of DRS. If one starts
from a model, one can first introduce parameters for each of the
objects in the domain and then consider the set {A (al' ... , an):
A(xl , ... , Xn) is a DRS with markers Xl' .. . , xn and {(Xl' al )' ... ,
(Xn, an)l is a truthful embedding for A 1as the set of facts, or shortly,
the world determined by the model. A crucial assumption in a notion of
fact of this kind, is that the facts are structured in the sense that the
objects they are about are syntactic constituents of the facts .
There is a strong assumption made on the relation of mind and the
world by the requirement that they employ the same syntax and the
same non logical constants. The content of the assumption is that
somehow the properties, individual constants and the logic al operations
already stand in a perfect correspondence. This is hard to maintain and
can in fact be got rid of in generalizations of the present approach.
The above gives us a set of unrelated systems. We will now try to
integrate alI these systems in what ont\ could call a DRT for many
minds and with respect to a world. In such a theory one should ideally
have rules that determine:
1. the changes made to a mind when it perceives something,

and the changes in the relation of the mind with the world.
2. the changes made to the mind and its relation to another
mind by the interpretation of an utterance by the person
who has the other mind.
It would be interesting to consider other types of change as well,
notably what happens in certain kinds of reflection, or in attempts to
integrate reliable information, that contradicts information that is
already there. These however bring complications that it seems good to
avoid here.
An example of the first kind of rule is what goes on when somebody
sees a yellow boat, and perceives it as a yellow boat. If the person does
not recognize the boat, he should add a new discourse marker x to his
mind, and at least the conditions: boat(x) and yellow(x). If he recog-
nizes it, this information, if not already available should be added as
properties of the discourse marker that contains his earlier experience
of it. In the first case the correspondence between the world and the
mind should be extended by (a, x) . In the second case, this corre-
spondence may already be present (it was a correct recognition).
Otherwise it should be added.
The second kind of rules are exemplified by the ones standardly
found in DRT. One special rule is the one that deals with indefinites.
This states that the correspondence between the mind of the speaker
and the mind of the interpreter should be extended by a new discourse
marker and by a condition on that marker determined by the content of
the common noun. The correspondence added here is between the
marker of the speaker, and the new marker. If definites are interpreted
by creating a new marker, the same addition must be carried out as in
the case of the indefinites, when they are interpreted as old markers,
the new correspondence should only be added if it was not there
already.
What we have therefore is a system
(w, MI,, M n , C)
where W is a set of facts, MI to M n are minds and C' is relation over
(Par U (n + 1 x Var)) X (n + 1 X Var).
In this system it is trivial to formalize the notions encountered in
294 HENK ZEEVAT
section 3. Moreover, they behave as they should within the Iimitations

to objects.
A thought is a subDRS of a mind.
An intention of Mi is a pair (i, x) where x is one of the
discourse markers in Mi.
A content is a DRS with parameters and intentions substi-
tuted for its discourse markers.
A correspondence preserving substitution of a thought or
content is a substitution of all its discourse markers on the
highest Ievel by parameters or other discourse markers such
that when a or y is substituted for x, (a, x) or (y, x) are in
the transitive and reflexive closure of C.
A true thought or content is one for which there is a
correspondence preserving substitution of parameters that
maps it into W.
A and B agree with respect to some content C iff there are
correspondence preserving substitutions for both A and B
that transform C in thoughts of A and B.
A believes some content C iff there is a correspondence
preserving substitution that maps the content into a thought
ofA's.
The system given alIows facts to contain parameters for intentions as
well as parameters for objects. There are a number of true statements
and therefore facts, that one can make with names that are not
referring. Witness
Santa Claus does not exist.
Mr Bovary died shortly after his second wife.
Some astronomers believed that Vulcan was a planet.
It does not follow however from the existence of such cases, that
intentions as such pIay a role in the facts that make these sentences
true: one may assume that there are other ways of constructing them as
facts that do not refer to intentions. The plausibility of intentions in
facts therefore depends completeIy on the plausibility of an intention
based account of these cases.
Let us start with the last example. The sentence is true iff there is an
object ar intention x that is named by Vulcan such that each of the
astronomers has a counterpart y of x that is also named by Vulcan and
mareover attribute the property of being a small planet to y. Since
Vulcan is not the name of any object it can only be an intention. Since
alI astronomers must have a counterpart, it may very well be the
original intention, that arose from the postulation of Vulcan by the first
one who did so.
The second example may be explained in a similar way. In a piece of
literary criticism the name Charles Bovary may refer either to an
intention in the pseudo beliefs of the author that are expressed in the
novel, or to any intention arising from the interpretation of the noveI.
My favourite explanation of the first case also runs via intentions.
One's interpretation of the name gives an intention in the tradition of
using the name Santa Claus. The property "exists" holds of an object
just in case it is a real object or the counterpart of one.
What a theory of this kind should avoid is to claim that also normal
simple discourse should be understood as somehow involving inten-
tions when it comes to defining truth: typicalIy, we are entering the area
where intentions start to play a role when we start discussing minds as
in the attitudes, or fiction as in the last two examples. The classical
distinction between extensional and non extensional contexts can deal
with the first and the last of the three examples: within a negation one
can alIow a name to have an intentional reference and similarly within a
belief complement. Of the second example, one can say that literally it
is false or rather meaningless, but taken as part of a kind of attitude
report it wilI become true again; in the sense that the condition that the
attitude imposes on the content is fulfilIed.
NOTES
1 It has been noted on various occasions (Nunberg 1978, Nunberg 1979, Donnellan
1966) that the noun need not even be true of the referent, provided the noun contains
sufficient material to enable the addressee to make the identification. Donnellan
discusses "the man with the Martini", who just has red lemonade in his glass, Nunberg
discusses cases like "the mushroom omelet", used in a restaurant to refer to a customer
who had a mushroom omelet.
2 Though being anaphoric means that they can be bound by a quantifier, which
demonstrative and experiental uses cannot. An example:
If a pupil meets a teacher, the boy is very polite.
296 HENK ZEEVAT
Binding by quantifiers is also possible in the definitional and relational uses, though
here the binding is not of the referent of the description, but of either an anaphor in the
complex noun, or the anaphoric element to which the referent of the relational
description is related.
3 Long distance anaphora with pronouns seems to be limited to those cases where the
pronoun occurs in the first sentence of a resumption of an interrupted topic. The
interruption can be either a real interruption, or an elaboration of a subtopic. See e.g.
Grosz and Sag 1981, Sidner 1981 or Scha and Polanyi 1984.
4 Sebastian Loebner (Loebner 1986) uses these examples as an argument against
theories in the Russell tradition, which assume that the function of the noun in the
description is to uniquely determine the referent as the only element of its extension.
His paper argues that functionality is what distinguishes definites from non-definites,
rather than familiarity. Here functionality is the idea that the definite markers change
the type of the noun from a (sortal) property into a functional concept that determines
an entity given a specification of an implicit or explicit argument.
Against the familiarity view, it is argued that there are a number of uses of definites
that cannot be understood as anaphoric, as the familiarity theory seems to imply. This
point must certain1y be granted. If one wants to use familiarity in an account of
definites, one is really stating a research aim: to understand the whole class of definites
from the viewpoint of anaphora, i.e. to see the other uses as analogous to the anaphoric
one.
Loebner's observation of a meaning shift in the noun when it is used in a definite is
also quite convincing. Both approaches share the idea that definites must not be
interpreted as quantifiers but as singular terms similar to proper names. The function of
the nouns in definite descriptions in both approaches is primarily to identify an object
for the interpreter of the utterance. This involves constraints on both the meaning of
the noun (it should be capable of identifying) and on the cognitive state of the
interpreter (he should have some form of access to the object he is to identify). Both
characterisations are therefore perfect1y compatible. What they have in common is
identification: the speaker assumes that he has made a reference to a singular thing and
expects the interpreter to single it out - perhaps only as far as is required in the
conversation - on the basis of a concept and the linguistic and non-linguistic context.
5 This distinction between two kinds of correspondence is similar to the central
distinction in (Landman 1986). Here "pegs" are stores for information that can get
more informational content as new information flows in. "Alecs" typically bear a given
amount of information and serve for storing general information about the class of
entities of which this information is true.
6 1 employ caus al in a naive sense that includes regularities that result from rule or
convention governed behaviour or the regularities that result from habits. It is not
assumed that every causal principle is based in naturallaw.
REFERENCES
Clark, H. H. and Marshall, C. R.: 1981, 'Definite Reference and Mutual Knowledge', in:
Joshi, A. K., Webber, B. L., and Sag, 1. (eds.), Elements of Discourse Understanding,
The University Press, Cambridge, pp. 10-63.
Donnellan, K S.: 1966, 'Reference and Definite Descriptions', Philosophical Review

75,281-304.
Grosz, B. and Sag, I.: 1981, 'Focusing and Description in Natural Language Dialogs', in:
Joshi, A. K, Webber, B. L., and Sag, I. (eds.), Elements of Discourse Understanding,
The University Press, Cambridge, pp. 84-105.
Heim, I.: 1982, 'The Semantics of Definite and Indefinite Noun Phrases', Ph.D. Thesis,
University of Massachusetts, Amherst.
Kamp, H.: 1981, 'A Theory of Truth and Semantic Representation', in: Groenendijk, J.
A. G., Janssen, T. M. V., and Stokhof, M. J. (eds.), Formal Methods in the Study of
Language, Mathematical Centre Tracts, Amsterdam, pp. 277-322.
Landman, F.: 1986, 'Pegs and Alecs', in: Towards a Theory of Information, Foris,
Dordrecht, pp. 97-155.
Loebner, S.: 1987, 'Definites', forthcoming in Journal of Semantics.
Nunberg, G.: 1978, The Pragmatics of Reference, Indiana University Linguistics Club,
Bloomington, Indiana.
Nunberg, G.: 1979, 'The Non-Uniqueness of Semantic Solutions: Polysemy', Linguistics
and Philosophy 3,143-184.
Polanyi, L. and Scha, R.: 1984, 'A Syntactic Approach to Discourse Semantics', in:
Proceedings of the IOth International Conference on Computational Linguistics and
22nd Annual Meeting of the Association for Computational Linguistics, Stanford
University, Stanford, California, pp. 413-419.
Sidner, C. L.: 1981, 'Focusing for Interpretation of Pronouns', American Journal of
Computational Linguistics 7, 217-231.
Zeevat, H.: 1984, 'Belief', in: Landman, F. and Ve1tman, F. (eds.), Varieties of Formal
Semantics, Foris, Dordrecht, pp. 405-425.
University of Stuttgart
INDEX OF NAMES
qvist, L., 229, 263, 266 Chomsky, N., 71, 84, 119, 121, 125,
Aczel, P., 138, 164 156, 165
Adamczyk, K., viii Christofferson, 269
Ades, A., 67 Clark, H.H., 273, 296
Akmajian, A., 125, 129 Clark, R., 164
Anderson, S., 104, 125 Cocchiarella, N., 80, 83, 125
Andrews, A., 69, 79,105,110,116, Cole, P., 126
125 Comrie, B., 90, 125
Aoun, J., 164 Conti, S., vii
Cooper, R., 112, 125, 199,266
Bacchus, F., 229, 266 Cresswell, M.J., 19,20, 131, 132, 133,
Bach, E., vii, 67, 68, 96, 105, 123, 125, 138
126,152,153,163,164,165,177 Croft, W., 190, 191
Barwise, J., 73,125,131,165,168, Cruse, D.A., 106, 125
179,190,191,263,266 Culicover, P., 105, 125, 129
Batterman, N., 118, 127
Bliuerle, R., 20, 68, 127, 165,267 Dahl, O., 168, 178, 185, 187, 191
Bealer,12 Davidson, D., 82, 84,125
Belletti, A., 148, 165 Davis, S., 126
Belnap, N., 31,41,43,44,45,67 Declerck, R., 187, 191
Bennett, M., 31, 43,44,45,67 Dixon, R.M.W., 124, 126
Benthem, J. van, vii, 5, 67 Donellan, K.S., 295, 297
Berman, S., 165 Dowty, D.R., 6-10, 67, 69, 88, 90, 95,
Berwick, R., 150, 165 96, 102, 104-106, III, 120-123,
Bolinger, D., 175, 191 126, 127, 139, 141, 163-165, 168,
Bresnan, J., 71, 90, 94, 104, 125, 127 172,191,266
Buszkowski, W., 67 Dryer, M., 262
Carlson, O., 11-14,20,71,78,84,85, Egli, U., 267

93-96, 116, 117, 119, 125, 167, 168, England, N.C., 124, 126
170,173,187,191,193,198,201, Evans, O., 199, 264, 266
235,236,238,246,262,263,264,
266,267 Faltz, L., 66, 68, 153, 165
Carnap, R., 9, 20, 131 Faraci, R., 164, 165
Chao, W.,164 Farkas, D., 165, 173, 174, 191, 193,
Chierchia, O., vii, 9, 10, Il, 13, 19, 20, 202,239,266
65,67,70,79,80,83,85,102,119, Feder, D., 126
120, 125, 131, 145, 163, 164, 165, Feferman, , vii
183,191,193,262,266 Fillmore, C" 69,94, 95, 97, 109, 110,
Choe, J., 165 126
299
300 INDEX OF NAMES
Fine, K., 262, 267 Joshi, A.K., 296, 297

Fodor, J., vii, 102, 199,267 Jubien, M., 12
Frege, G., 10, 18, 131,275
Frey, W., 263 Kamp, H., 15, 16, 18,20,123,127,
195,197,198,201,263,267,270,
Gabbay, D., 168, 191,267 297
Garfield, J., vii Kaplan, R.M., 71, 127
Garson, J., 198, 264, 267 Karttunen, L., 31,41,42,43,44,66,
Gazdar, G., 66, 67, 237, 267 68,127,264,267
Geach, P.T., 15, 18,264,267 Kay, M., 72, 127
Gillon, B., 262 Keenan, E., 66, 68, 110, 123, 127, 153,
Greenberg, J.H., 128 165
Grimshaw, J., 94, 104, 126 Keil, F., 118, 127
Groenendijk, J., 2-6, 20, 21, 66, 68, Kiefer, F., 68
127, 128, 166, 192, 195, 197, 198, Kimball, J., 192
201,263,264,267,297 Klein, E., vii, 267
Grosz, B., 268, 296, 297 Kratzer, A., vii, 178, In, 263, 267
Gruber, J., 69, 74, 105, 106, 123, 126, Krifka, M., 220, 262, 263, 267
139,165
Guenthner, F., 266, 267 Ladusaw, W., 104, 105, 120, 127, 164,
Guenthner-Reutter, M., 266 165
Gvozdev, A.N., 115, 126 Landman, F., vii, 67, 68, 128, 131, 165,
296,297
Hanns, R.T., 126 Larson, R., vii, 93, 127, 163
Hausser, R., 66, 68 LePore, E., 191, 198,264,267,268
Heim, 1., 15, 16, 18,20, 195, 197, 198, Levin, B., 127
201,202,221,241,258,263,264, Levin, L., 104, 122, 123, 127, 128
267,269,270,297 Lewis, D., 165,202,229,246,247,267
Henry, F., 267 Li, C.N., 127
Higgins, R., 100, 126 Link, G., vii, 85, 100, 127,220,262
Hinrichs, E., vii, 123, 127 Linsky, B., 262
Hintikka, J., 17, 198,267 Unning, J.T., 265, 267
Hoeksema, J., 86, 100, 127 Loebner, S., 296, 297
Hoepelman, J., 266
Homstein, N., 267 Marantz, A., 77, 97, 115, 127
Hurum, S., 262 Marciszewski, W., 67
Marshall, C.R., 273, 276
Jackendof, R., Il, 20, 69, 72, 73, 74, McConnell-Ginet, S., 154-156, 165
75,78,79,104, 105, 106, 108, 116, McCord, M., 202, 263, 267
127, 123, 139, 140, 143, 165 McDonough, J., 165
Jacobs, R., 125 Menzel, C., vii
Jacobson, P., 102, 125, 126, 145, 163, Meulen, A. ter, 116, 117, 127, 128,
165 183,191,262
Janssen, T., 20,127,128,166,297 Miller, G.A., 128
Jones, C., 151, 164, 165 Mi!sark, G., 170, 191
Jongh, D. de, 20, 68 Mithun, M., 126
INDEX OF NAMES 301
Moerdijk, 1.,67,68 Rohrer, C., 266
Monnich, U., vii Rooth, M., 20, 45, 55, 61, 62, 66,67,
Montague, R., 1, 19 68, 113, 136, 165
Moortgat, M., 67, 68 Rosen, C., 110, 128
Moravcsik, J., 168, 191 Rosenbaum, P., 125
Muskens, R., 67, 68 Russell, B., 10, 18, 131,277
Nerbonne, J., 124, 128 Saarinen, E., 67

Nikiforidou, V., 126 Saddock, J., 126
Nishiguachi, T., 79,104,128 Sag, 1., 102, 103, 128, 199,262,267,
Nunberg, O., 295, 297 296,297
Scha,R., 128,296,297
Oehrle, R., 67, 68, 165 Schmerling, S.F., 124, 128
Ostler, N.D.M., 100, 128 Schubert, L.K., 14, 16, 17, 19, 190,
191,193,229,231,237,247,249,
Palmieri, M.A., viii 263,266,267,268
Panini,69 Schwarze, C., 20,68,127,165
Parsons, T., 84, 91, 92, 128 Sells, P., 113, 128
Partee, B., vii, 16, 20, 24, 25, 26, 45, Shopen, T., 125, 128
55,61,62,66,67,68,136,163, Sidner, C.L., 296, 297
165,172,199,264,267 Slobin, D.I., 115, 128
Pelletier, F.J., 14, 16, 17, 19, 190, 191, Smith, F., 128, 191
193,231,237,247, Smith, N.V., 191
249,263,267,268 Spark-Jones, K., 268
Perlmutter, D.M., 128 Stalnaker, R., 67, 68, 73, 133, 166
Perry, J., 73, 125, 131, 165, 168, 179, Stechow, A. von, 20, 68,127,165,267
191,263,266 Steedman, M., 67
Pesetzky, D., 148, 165 Stokhof, M., 2-6, 20,21, 66, 68, 127,
Peters, S., 67, 88,126 128,166,195,197,198,201,263,
Planle, F., 124, 128 264,267,297
Plotkin, vii Strawson, O., 275
Polanyi, L., 296, 297 Stump, O., 173, 174, 179, 192
Pollard, C., vii, 267 Sugioka, Y., 173, 174, 191, 193,202,
Postal, P., 70, 128 239,266
Pullum, O., 125, 126, 165, 267
Talmy, L., 105, 123, 128
Quine, W.V.O., 168 Thomason, R., 131, 163, 166
Traugott, E., 191
Rappaport, M., 90, 98, 123, 128 Trechsel, F.R., 120, 124, 129
Reilly, U., 263 Tumer, R., 65, 67, 163, 165
Rescher, N., 125, 126
Richards, B., 267 Veltman, F., 128, 131, 166,297
Riemsdijk, H. van, 71, 77, 128 Vendler, Z., 106, 187, 192
Rizzi, L., 148, 164, 165 Verkuyl, H.J., 106, 129, 187, 192
Roberts, C., 86, 128, 178, 191, 198, "-
263,267 Wall, R., 88,126
302 INDEX OF NAMES
Wasow, T., 104, 125, 129 Zaenen, A., 126, 128

Webber, B.L., 268, 296, 297 Zeevat, R., 17, 18, 19,269,297
Wheeler, D., 67, 68, 165 Zwarts, F., 67, 68
Wilkins, W., 105, 125, 165 Zwicky, A., 127
Williams, E., 71, 77, 104, 128, 129,
148, 151, 163, 166, 192
INDEX OF SUBJECTS
<D, W, C, B>,213 characterizing property, 216

<Il alternative, legitimate 213, 214 coargumenthood, 148
coindexation, 151
aboutness, 18 communication, 285
Ad-verb, 155 completeness of thematic roles, 82
adverb of frequency, 212, 215, 246 compositiona1ity,239
time,208 structured meaning and, 132
adverbial generics, 173 conjunction, generalized, of interroga-
modification, 161 tives, 42, 52
theory, 154 constituent interrogative, 49
agent thematic role, 69 structure, 60
agent-patient opposition, 109 context, 203
agentivity, 6, 8 change, 203, 207
anaphora, 2 in light of quantifier processing,
intersentential, 196 205
intrasentential, 196 space,213
anaphoric reference, 1 transformation function, 214
description, 271 control, 142
argument analysis, 91 by implicit arguments, 148
control, 149 entailment in, 145
indexing, 107 in purpose clauses, 157, 158, 160
NP,184 local theory of, 102
raising,54 overt obligatory, 146
shift, 63 semantics of, 141, 144
structure, semantic difference, 89 purpose clauses and, 152
AT-THAT-TIME, 209, 210, 221, 222, strict obligatory, 143, 163
224,228 structured meaning, and thematic
role, 131
bare plural, 17, 184 syntax and, 150
belief de re, 134 convention in definite description, 270
rell\tion, 133 coordinated interrogative, 36
coordination, 5, 34
by :-hrase. 8. 97 generalized, 22
as adjunct of nominal, 98 types, 51
correspondence, formal theory, 292
capacity,216
case adverbials, 217 definiteness, familiarity and, 269
ensembles of, 194 familiarity theory, 19
relation see thematic role realism and, 269
categorial approach to question, 3 theory, phiIosophical consequences
303
304 INDEX OF SUBJECfS
of,275 theory, 274, 278

definites, 17 fine-grainedness, 131
classification of, 18 flexible analysis of interrogatives, 33,
formal theory, 288 45
definition in definite description, 270 flexible grarnmar, 61
deictic pronoun, 201 constraints on, 62
reading, 198 overgeneration by, 61-64
demonstration in defmite description, frequency adverb, 212, 215, 246
271
derived nominal, role types and, 111 gender, grarnmatical, 112
detransivization, 147 natural, 112
discourse analysis, 205 generic construction, compositional
representation, 14, 17 structure, 13
of generics, 193 predication, 11
theory, 195 property, 11
disjunction of interrogatives, 35, 52 relation is not predicational, 189
distinctness of thematic roles, 82 sentence, 14, 189,215, 218
ditransitive verb, 154 discourse representation of, 193
donkey anaphora, 15 episodic, 220
donkey sentence, 194, 197,206 intensional elements in, 189
dummy subject generics, 174 intensionality and, 179
meaning of, 189
embedded interrogatives, 53 non-episodic, 227
embedding types, 51 NPs in, 185
ensemble of cases, 194 quantifiers in, 180
entailment, 5, 22, 28, 34, 37 relational analysis of, 177
between interrogatives, 48 relations of, 179
control semantics and, 142 semantic composition of, 167
in control, 145 sentential operator analysis of, 175
lexical, of verb subject-predicate analysis of, 168
of thematic roles, 84 generic when sentences, 231
pattern, 141 genericity, 14
relations, 140 global shift, 63
enumeration of infinite sets, 230 Gn operator, 201
episodic generics, 220, 234 goal thematic role, 69
event as n place relation, 80 gPRES, 209, 222,224,228,230
as primitive, 7 grarnmar leaming, role types and, 114
proposition and, 8 grarnmatical gender, 112
-noun argument, linking in, 99 relations, oblique, 92
existential reading, specific, 198
experience in definite description, 271 habitual sentence, discourse representa-
experiencer thematic role, 69 tion of, 193
extra object, 279
1 generics, 222
familiarity, 276 iflwhen sentence, atemporal, 235
definiteness and, 269 in'clependence of thematic roles, 83, 84
INDEX OF SUBJECfS 305
indexical context, 203 local theory of control, 102
infinitival, implicit subject of, 2 locative generics, 170
infinival structures, 151 logical form language, 251-262
information, 283 ominiscience problem, 132
unit, proposition and, 137 space theory, 134
instantiation, 12
intensional object, property theory of, meaning as a fine-grained notion, 29
65 constraints on, 144
semantic analysis of, 64 structured, 30
intensionality, 2 modifier analysis, 91
and generics, 179 Montague grarnmar, classical, 1, 19
intention of, 17
interrogative, atomic, 43, 44, 45, 48 natural gender, 112
coordinated, 36 nominalization, 13
coordination and entailment in non-episodic generics, 227, 234
analysis of, 41 NP in generic sentences, 185
denotational, 32
disjunction of, 35 optional argument, 155
domains for, in naturallanguage, 55, ordered argument system, 71
56 theory, 73, 74
embedded, 53
entailment, 37 pair list, 34
flexible analysis, 33 parking meter sentence, 226
n-place relation, 49 partition, 50
relations, 27 partition, question as, 46, 47
semantic analysis of, 27 PAST, 209, 210
semantic type, 22 patient thematic role, 69
syntactic category and semantic type PERF,209
of,26 plurality,247
type shifting analysis of, 32 possible world, 46
type shifting analysis of, 59, 60 semantics, 5
types, 39 pragmatic component, 236
interrogatives, categorial approach, 23 pragmatics, 217
propositional approach, 23 predicate structure, ordered argument, 6
semantics, 21, 24 predicate-argument association, 88
intersentential anaphora, 196 relation, 71
ordered argument theory, 73
karakas see thematic role PRES, 209, 233,243
probability measure over fields of sets,
language acquisition, thematic roles 230
and,115 PROG,209
lazy reading, definite, 199,200 pronoun, anaphoric uses of, 270
indefinite, 201 property theory, 65
lexical entailment of verb, 75 and interrogatives, 31
linking (sensu Higgins), 100 proposition, event aneI, 8
in event-noun arguments, 99 informatiort unit and, 137
306 INDEX OF SUBJECfS
thematic role and, 9 evaluation, 229

propositional approach to question, 3, 4 type of interrogative, 26
function, 158 semantics for question, 5
theory of question, 5 of control, !O, 141, 144
prototype, thematic role type as, 118 of generics, 17
purpose clause, 151 of interrogatives, 24
control in, 157, 158, 160 categorial approach, 25
control properties of, 154 propositional approach, 25
denotation of, 156 of that clauses, 135
disjointness in, 159 structured meaning in, 131
semantic value of, 162 sentence embedding, 2
semantics of, 152 evaluation in context, 204
truth conditions, 161 sentential interrogative, 49
operator analysis of generics, 175
quantified NPs, 197 singular indefmite subject, 184
quantifier in generics, 180 source thematic role, 69
raising, 197, 202 structure building predic;ation, 138
question, 2, 49 structured meaning, 10/ 162
as partition, 46, 47 approach, 134
as unsaturated structure, 3 compositionality and, 132
propositional theory, 5 control, and thematic role, 131
semantic analysis, 2 subject selection principles, 109
semantic approach to, 3 subject-predicate analysis of generics,
question-answer relation, 24, 27 168
syntactic category of interrogative, 26
realism, defmiteness and, 269 syntax and control, 150
reference case, 239
ensemble, 14, 15, 16,215 temporal adverb, 208
referent underspecification, 277 tense, 208
relational analysis of generics, 177 that clause, 133
type for interrogative, 57 semantics of, 135
use of description, 272 thematic hierarchy constraint, 74
relations between interrogatives, 27 relation see thematic role
representation embedded in a model, 19 thematic role, 2, 6, 7, 9, 11
resultant state, 153 as aid to leaming grammar, 116
rleatedness of generics, 179 as cluster of entailments, 76
role bearer, uniqueness of, 100 as relation, 7
type, evidence for, 104 bearer, 103
indirect, 113 completeness, distictness, and
strong theory, 103 independence of, 82
control, and structured meaning, 131
scope ambiguity, 247 definition, 139
widening, 202 entailment, 84
semantic analysis of thematic role, 70 entailment cluster as, 140
category of argument, 78 trivial individual, 77
content of thematic role, 73 . expressive power of theories, 86
INDEX OF SUBJECTS 307
independence of, 84 as unification, 56

individual,76 in analysis of interrogatives, 32, 59,
L type, 77, 83, 87,105,107 60
completeness and distinctness of, rules,21
78-79 theory, 34
semantic content, 81 two-sorted, for interrogatives, 49
neo-Davidsonian system, 7, 82, 87, unification, type shifting as, 56
88,91 unique answer theory, 41
ordered argument theory, 74, 88,91 uniqueness of thematic role, 85
proposition and, 9 USUALLY, 218, 219
semantic content of, 69, 73
strong theory, 74 value assignment to variable, 16
strong type theory, 103 variant context, 203
syntax generalization and, 108 verb as unsaturated predicate, 73
system,71 detransitivization, 94, 147
definition, 72 ditransitive, 95
type,l14 extensional interrogative-embedding,
as prototype, 117 54
uniqueness, 85, 99 intensional interrogative-embedding,
theme thematic role, 69 54
THEREUPON, 209, 211, 221, 243 lexical entailment of, 75
theta role see thematic role subcategorization, thematic role
transitive generics, 172 account of, 93
verb phrase modifiers, 152 verbal passive and nominal, by phrases
true-in-a-context, 206 in,97
true-in-English, 206
truth,281 WHEN, 218, 219
truth-in-a-context, 223 when clause, 236, 240
truth-in-English,223 extensional and intensional, 183
type change, 2 temporal, 182
coordination and embedding, 51 when sentence, episodic, 235
forinterrogatives, 39,44,48
type shifting, 5, 22, 54, 136 YEST, 209, 210

Gennaro Chierchia - Semantics

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Gennaro Chierchia - Semantics

Diunggah oleh

Hak Cipta:

Format Tersedia

PROPERTIES, TYPES AND MEANING

KLUWER ACADEMIC PUB:tISHERS

Properties, types and meaning.

(Studies in 1inguistics and phi1osophy

Published by Kluwer Academic Publishers,

Kluwer Academic Publishers incorporates the publishing programmes of

Sold and distributed in the U.S.A. and Canada

In ali other countries, sold and distributed

AII Rights Reserved

TABLE OF CONTENTS OF VOLUME I vi

JEROEN GROENENDIJK and MARTIN STOCKHOF / Type-

INDEX OF NAMES 299

INDEX OF SUBJECTS 303

PETER ACZEL / Algebraic Semantics for Intensional Logics, 1

This collection of papers stems originally from a conference on Property

It is quite uncontroversial that a semantic theory for natural language

ence, or in the analysis of belief sentences and sentence embedding.

GROENENDIJK AND STOKHOF:

Groenendijk and Stokhof investigate aspects of the grammar of ques-

strategies that are currently being pursued in connection with the

(3) a. who went to the rally?

relevant phenomena. Groenendijk and Stokhof seek a way out of this

(5) a. Mary wonders whether second order logic is categorical

intensional notion of semantic content. The work of Groenendijk and

DOWTY: THEMA TIC ROLES AND EVENTS

a certain slot in a relation should be definable independently of that

would be that verbs correspond to predicates that are typed as to the

CHIERCHIA: THEMATIC ROLES AND PROPOSITIONS

Chierchia's starting point is the notion of "structured meaning". Struc-

that sentences can be interpreted at some level as structured objects,

different entailments that these verbs associate with their argument

CARLSON : GENERIC PROPERTIES AND GEN E RIC PREDICATION

Genericity constitutes a traditionally difficult area of analysis and has

instances via a predication relation, just as Carlson links kinds to their

(2) a. It rains 30 a year, here

b. It snows in Florida once every decade or so

Clearly, the sentences in (2) are generic. Yet their compositional

SCHUBERT AND PELLETIER:

Pelletier and Schubert develop a fairly detailed proposal concerning the

ensemble". Sometimes, such a reference ensemble is explicitly provided,

as existentially quantified. But they can be used to constrain the context

ZEEV AT: DEFINITES AND DISCOURSE REPRESENTA TION

aboutness of information bearing structures. Zeevat's paper addresses

possibilia, indeterminates, abstract objects and the like. In this connec-

Classical Montague semantics and the variety of related approaches

Carlson, G.: 1980, Reference to Kinds in English, Garland, New York.

TYPE-SHIFTING RULES AND THE

example. For reasons of simplicity and elegance, it is attractive to assign

sketch of the ideas underlying the categorial and propositional

1. CATERPILLARS AND BUTTERFLIES

1.2. The Categorial and the Propositional Approach

Taking the orthodox view, on which there is a fixed category-to-type

corresponds a unique semantic type, the two approaches are incom-

1.3. A Puzzling Situation

different interpretation, thus failing to account for their assumed

within the overall meaning of an interrogative, which in the proposi-

wor1ds to (sets of) propositions whether, taken as structured objects,

sentential interrogatives are not related by entailment to each other

different perspectives, taking different constructions as their starting

2. COORDINATION, ENTAILMENT AND TYPES

(4) on the relevant reading contains only one wh-term, it is answered is

gave him for Christmas. Obviously, it is a matter of the internal

(9) Peter knows whom John loves or whom Mary loves