Professional Documents
Culture Documents
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
PREFACE vil
INTRODUCTION 1
v
TABLE OF CONTENTS
(Volume 1)
VI
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
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
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
(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
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.
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
(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.
(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-
16 GENNARO CHIERCHIA
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
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
18 GENNARO CHIERCHIA
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
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
Gennaro Chierchia, Barbara H. Partee, and Raymond Turner (eds.), Properties, Types
and Meaning, Il. 21-68.
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
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.
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
28 JEROEN GROENENDIJK AND MARTIN STOKHOF
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
THE SEMANTICS OF INTERROGATIVES 29
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
THE SEMANTICS OF INTERROGA TIVES 35
(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.
THE SEMANTICS OF INTERROGA TIVES 37
2.2. Entailment
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.
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.
46 JEROEN GROENENDIJK AND MARTIN STOKHOF
3. A FLEXIBLE APPROACH
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
48 JEROEN GROENENDIJK AND MARTIN STOKHOF
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
THE SEMANTICS OF INTERROGA TIVES 49
walks, and those in which he doesn't. This means that (37) partitions
the set of worlds in two:
no onewalks
everyone walks
W
THE SEMANTlCS OF INTERROGATIVES 51
(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:
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 ):
THE SEMANTICS OF INTERROGA TIVES 55
II IV
LIFT
(s, t) .. " (5, t), t), t)
LOWER
I III
56 JEROEN GROENENDIJK AND MARTIN STOKHOF
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:
58 JEROEN GROENENDIJK AND MARTIN STOKHOF
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-
60 JEROEN GROENENDIJK AND MARTIN STOKHOF
(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.
64 JEROEN GROENENDIJK AND MARTIN STOKHOF
4. CONCLUSION
NOTES
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
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
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.
68 JEROEN GROENENDIJK AND MARTIN STOKHOF
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
1. INTRODUCTION
(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
SEMANTIC CONTENT OF THEMA TIC ROLES 73
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
(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),
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):
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.
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.
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:
SEMANTIC CONTENT OF THEMA TIC ROLES 95
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
SEMANTIC CONTENT OF THEMA TIC ROLES 101
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.
SEMANTIC CONTENT OF THEMA TIC ROLES 103
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.
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),
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.
... 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
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
NOTES
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
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):
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.
SEMANTIC CONTENT OF THEMA TIC ROLES 125
REFERENCES
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
Gennaro Chierchia, Barbara H. Partee, and Raymond Turner (eds.), Properties, Types
andMeaning, II. 131-166.
1989 by Kluwer Academic Publishers. AII rights reserved.
132 GENNARO CHIERCHIA
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
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
STRUCTURED MEANINGS 135
S, hug'(M, 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
136 GENNARO CHIERCHIA
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)).
STRUCTURED MEANINGS 137
The reading in (9c) is more coarsely grained than the one in (9b). To
illustrate our approach, let us show how (9b) comes about.
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
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).
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.
STRUCTURED MEANINGS 139
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
146 GENNARO CHIERCHIA
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.
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.
STRUCTURED MEANINGS 149
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
150 GENNARO CHIERCHIA
4. PURPOSE CLAUSES
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:
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
154 GENNARO CHIERCHIA
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.
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))])
158 GENNARO CHIERCHIA
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
160 GENNARO CHIERCHIA
5. CONCLUSIONS
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
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.
STRUCTURED MEANINGS 165
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.
166 GENNARO CHIERCHIA
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
1. GENERIC SENTENCES
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
ENGLISH GENERIC SENTENCES 171
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
3.2. Transitives
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
(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.
4. A SENTENTIAL-OPERATOR ANALYSIS
(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
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
(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.
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:
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
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
ENGLISH GENERIC SENTENCES 191
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
Philosophy 70, 513-26.
Kratzer, Angelika: 1977, 'What 'Must' and 'Can' Must and Can Mean', Linguistics and
Philosophy 1, 337-55.
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
1. INTRODUCTION
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
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.
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
200 LENHART K. SCHUBERT AND FRANCIS JEFFRY PELLETIER
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
208 LENHART K. SCHUBERT AND FRANCIS JEFFRY PELLETIER
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.
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
GENERICALL Y SPEAKING 209
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(<I are (temporal) sentences. The
"tenses" we consider are: PRES, gPRES, PAST, and aspects PROG,
PERF. If <1> 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
210 LENHART K. SCHUBERT AND FRANCIS JEFFRY PELLETIER
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.
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)<I> 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
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.)
GENERICALL Y SPEAKING 215
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
216 LENHART K. SCHUBERT AND FRANCIS JEFFRY PELLETIER
49. Dogs dislike cats when they (the cats) have blue eyes
50. When cats have blue eyes, they are intelligent.
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
GENERICALL Y SPEAKING 219
(6) iff for most [ ]' such that for some d E D and
for some j during [present],
[ ]' = [ ]x:d,r:j and [D(x) & O(p, x)]~:':t = 1,
[gPRES(AT -THAT -TIME(B (p, x)]' i, w = 1; Le"
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
226 LENHART K. SCHUBERT AND FRANCIS JEFFRY PELLETIER
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.
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
228 LENHART K. SCHUBERT AND FRANCIS JEFFRY PELLETIER
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
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)
[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
GENERICALL Y SPEAKING 243
XIII. IN CONCLUSION
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,
(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.
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
GENERICALL Y SPEAKING 253
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>, i, w, [ ]), where <I> is a wff,
iEI, w E W, and [ ]is a context, as <I>i, w[ ]. Intuitively it designates
the (new) context which is generated from having already "processed"
<I> in the (old) context [] - it wilI always be dear which o is
intended.)28
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 [<I>]~~d = 1, and
for that d, <1> i, W['fl~ ~d = 1.
time of John's entering). After as cent to the top of (d'), the final value
of ris <ii' k').
ACKNOWLEDGEMENTS
NOTES
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
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
266 LENHART K. SCHUBERT AND FRANCIS JEFFRY PELLETIER
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
Philosophy 6, 225-258.
GENERICALL Y SPEAKING 267
INTRODUCTION
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
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:
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
REALISM AND DEFINITENESS 273
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
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
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
b. Truth
c. lnformation
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
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:
REALISM AND DEFINITENESS 293
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.
REALISM AND DEFINITENESS 297
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