Compositional Semantics

Compositional Semantics
An Introduction

The moral rights of the author have been asserted

First Edition published in 2022
Impression: 1
 Contents

Sections and exercises marked with * are more advanced and/or less central (and
on some occasions intended as open-ended speculation). These may be skipped
with little or no consequence for later sections.

Acknowledgments xv
Foreword: On Using this Book xvii

Part I: Foundational Concepts: Building a Fragment

1. Introduction 3
1.1. Goals3
1.1.1. Compositional semantics and (some of) the goals of
semantic theory 4
1.1.2. Direct Compositionality—and its role in this text 8
1.2. A brief note on the history of semantics within modern linguistic
theory 10
1.3. The notion of a “fragment” and its use in this text 12
*1.4. An intriguing puzzle 13
1.5. Appendix: Sets and Functions 19
1.3.1. Sets, members, and subsets 19
1.3.2. Union, intersection, and complement 21
1.3.3. Ordered pairs, relations, equivalence relations, and
partitions 21
1.3.4. Functions 24

2. Semantic foundations 27
2.1. Model-theoretic semantics 27
2.2. Truth conditions 28
2.3. Possible worlds 31
2.3.1. Introducing the notion 31
2.3.2. Characteristic function of a set 35
CONTENTS
 viii
2.3.3. Notation and terminology 36
2.3.4. Talking about worlds 37
2.4. Times 41

3. Compositionality, Direct Compositionality, and the syntax/semantics

interface 42
3.1. Building a fragment: First steps 43
3.2. Implicatures vs truth conditions 46
3.3. Folding in worlds (and times) 48
3.4. Negation: A first pass 49

4. Expanding the fragment: Syntactic categories and semantic types 52

4.1. Noun phrases 52
4.2. Intransitive verbs/verb phrases 58
4.3. A brief look at lexical semantics 60
4.4. Back to the compositional semantics 62
4.5. Illustrating the syntactic/semantic composition 63

5. Transitive verbs: Resolving an apparent syntax/semantics mismatch

66 5.1. First pass: Straw Man’s solution 67
5.2. Abandoning Direct Compositionality 70
5.3. Hidden S conjunction? 73
5.4. Back to the drawing board: Curry’ed functions 76
5.1.1. Recasting the meaning of transitive verbs 76
*5.4.2. Currying more generally 79
5.4.3. Solving Straw Man’s problems 80
*5.5. A note on three-place verbs 84

6. Categorial Grammar 88
6.1. Basics 89
6.2. Syntax/semantics correspondence 92
6.3. Refinements 94
*6.4. Some further observations about syntax: Case marking and
agreement 96
6.5. Further generalizing the rules 98
6.6. and, or, negation, and case-marking prepositions 100
6.7. Summary: The fragment so far 104
 ix
CONTENTS

7. The autonomy of syntax? 107

7.1. Good syntax—bad semantics 108
7.2. Bad syntax—good semantics 111

8. Adjectives, nouns, determiners, and more 113

8.1. Predicative adjectives 113
8.2. Nouns (“common nouns”) 116
8.3. the118
8.4. Prenominal adjectives and unary rules 120
8.4.1. An additional binary rule? 120
8.4.2. Interlude: Unary rules (“type shift” rules; morphological
rules) 121
8.4.3. Back to adjectives 124
8.4.4. Are all noun modifier adjectives intersective? 126
8.5. Relational nouns 128
8.6. Summary: The fragment so far 131

9. Interlude: The semantics of variables and the lambda calculus 133

9.1. First-order predicate logic (modified) 134
9.1.1. Variables 134
9.1.2. Quantifiers 138
9.2. The lambda calculus144
9.2.1. Lambda abstraction 144
9.2.2. Lambda conversion 147
9.3. Naming model-theoretic objects 151

Part II: Enriching the Domain

10. Returning to English: Generalized quantifiers 157

10.1. “NPs” with quantificational determiners 157
10.1.1. First pass: Individuals? 158 10.1.2. Second pass: Sets
of individuals? 159
10.1.3. Third pass: English as first-order logic? Straw Man
makes a cameo reappearance 160
10.2. Generalized quantifiers 162
 x
10.2.1. The semantics 162
10.2.2. The syntax 165
10.2.3. Extending to the other quantified cases 166
CONTENTS

10.3. The determiners 167

10.4. Interlude: Scalar implicature 168
10.5. Generalized quantifiers and conjunction 173
10.5.1. VP conjunction 173
10.5.2. Generalized quantifier conjunction 175
*10.6. Domain restriction 178

11. Ordinary NPs and type lifting 181

11.1. Montague’s solution 182
11.2. Partee and Rooth’s solution: Lift as a unary rule 185
11.3. Generalizing the rule 186
11.4. The lowest types hypothesis 189
11.4.1. Some additional predictions 189
*11.4.2. The lowest types hypothesis: Processing vs
grammar 192
11.5. Resolving some syntactic loose ends 194
11.6. Summary: The fragment so far 196

12. Generalized conjunction 198

12.1. Generalizing and and or 198
12.2. Remarks on noun conjunction 201
12.3. Summary: The fragment so far 203

Part III: Relative Clauses, Scopes, and Binding: Some

Theoretical Controversies

13. Relative clauses: Sketching two accounts 207

13.1. Preliminaries 207
13.2. What do relative clauses modify? 209
13.2.1. The noun modification argument 209
*13.2.2. But—not quite so fast . . . 213
 xi
13.3. An account with movement and traces: Assignment functions in
grammar 214
13.4. A Direct Compositional (and variable-free) analysis 221
13.4.1. Preliminaries: Function composition 221
13.4.2. Interlude: Coordination 224
13.4.3. Returning to relative clauses 227
13.4.3.1. The basic account 227
13.4.3.2. Syntactic loose ends 231
CONTENTS

13.4.4. A refinement: The Geach rule 233

13.4.4.1. The rule and its application to
relative clauses (and
coordination) 233
*13.4.4.2. Generalized and as generalized g-
sl? 238
13.5. Appendix: Reassuring the Appalled Syntactician 240

13.6. Summary: The fragment so far 242

14. Generalized quantifiers in object position: Two

approaches 244

14.1. A tempting but inadequate solution 245

14.2. Quantifier (Lowering)/Raising 248

14.3. A Direct Compositional solution 255

14.4. Some comparisons 263

14.4.1. The generality of the solutions 263

14.4.2. Antecedent-Contained Ellipsis 265

14.4.2.1. The “textbook” argument 265

14.4.2.2. Antecedent-Contained Ellipsis as Transitive

Verb Phrase Ellipsis 270
14.5. Generalized quantifiers in embedded sentences 272
14.6. Summary: The fragment so far 273
 xii
15. The interpretation of pronouns: Two accounts 274
15.1. Introductory remarks 274
15.1.1. Variables or not? 274
15.1.2. Pronominal binding: The empirical
phenomenon 276
15.2. The LF (variable-ful) view 278
15.2.1. The basic account 278
15.2.2. A wrinkle: Weak Crossover effects 282
15.2.3. Free pronouns 285
15.3. A Direct Compositional and variable-free approach 286
15.3.1. Pronouns and “binding” 287
15.3.2. Free pronouns 293
*15.3.3. Generalizing the rules 294
15.3.4. Weak Crossover 298
15.3.5. What happens when there is more than one
pronoun? 301
CONTENTS

15.3.2.1. Multiple pronouns which

are “the same” 301

*15.3.5.2. And yet another possibility 303

15.3.5.3. More than one pronoun—

different binders 304

15.4. More on Weak Crossover 305
15.4.1. Definite NPs and WCO 305
15.4.1.1. An apparent problem (for both

theories) 305

15.4.1.2. Not a problem after all 306

15.4.1.3. Independent evidence: Ellipsis and

the notion “sloppy identity” 307

*15.4.1.4. A more formal account of the

strict reading 312

 xiii
*15.4.2. A note on three-place verbs and WCO interactions
315
15.5. More on free “variables” 317
15.5.1. Free “variables” in the variable-ful account 317
15.5.2. Free “variables” in the variable-free theory 319
15.6. The i-within-i effect revisited: A solution 320

Appendices to Parts I–III: The full fragment 323

Appendix 1: The basic fragment 323

Foundational concepts 323

The lexicon 324

Word order rules 326

Unary and combinatory rules 327

*Appendix 2: Possible extensions of the fragment for

three-place verbs using Wrap 329

The lexicon 329

Word order rule 330

Revision of the definition of functions

corresponding to categories of the form A/B 330

Appendix 3: Informally discussed addition for VP Ellipsis 330

CONTENTS

Part IV: Further Topics

16. Negative Polarity Items, semantic strength, and scalar

implicature revisited 336
16.1. The notion of Negative Polarity Items 337
16.1.1. any and ever 337
16.1.2. Minimizers 339
16.2. The distribution of NPIs 340
16.2.1. First pass 340
16.2.2. Second pass 342
 xiv
16.3. The first breakthrough 343
16.3.1. Initial characterization of the environments 343
16.3.2. A note on many and few 349
16.4. But why? And how? 352
16.4.1. The insight from semantic strength 354
16.4.2. The second breakthrough: Kadmon 356
and Landman on any 356
16.4.3. ever 359
16.5. Open questions and further discussion 361
16.6. Semantic strength and scalar implicatures 363
17. More binding phenomena 367
17.1. Functional questions 367
17.1.1. The phenomenon 367
17.1.2. An analysis with complex traces 369
17.1.3. A variable-free analysis 371
17.2. Binding and coordination (in variable-free semantics) 373
17.3. “Paycheck” pronouns (and “donkey” pronouns?) 374
17.3.1. Paychecks 374
17.3.2. A variable-free analysis of paycheck pronouns 377
17.3.3. Donkeys 379
17.4. “Hidden variables” 381
17.4.1. Hidden variables can always be complex 381
17.4.2. Translating into variable-free 385
18. Additional semantic dimensions: The semantics of focus 388
18.1. Computing alternatives compositionally 389
18.2. The meaning of only 392

CONTENTS

18.3. Additional focus-sensitive words 392

18.4. Further compositional issues 393
18.5. Focus and binding 395
18.6. More generally: Multidimensional semantics 397

19. Intensionality and the syntax/semantics interface 398

19.1. Intensionality-sensitive words 399
 xv
19.2. Revising the fragment: Type-sensitive rules 403

References 407
Index 419
 xvi
Acknowledgments

This book began life as a series of handouts for my Formal Semantics course in 1998.
The handouts grew and were revised over the years until I realized they were trying to
become a book. I owe my greatest debt to the students in these classes. I never would
have undertaken to write this book were it not for the enthusiasm and insight of my
students who have always made teaching Formal Semantics just plain fun. I am also
indebted to three scholars whose work has influenced my thinking in ways
sometimes obvious and sometimes subtle. The influence of Barbara Partee will be
obvious throughout this book. To David Dowty, I owe an understanding of how to think
“semantically,” and my thinking about Categorial Grammar and its elegance as a
theory of the syntax/semantics interface owes much to the work of Emmon Bach. The
way of thinking that I learned directly or indirectly from these three semanticists has
shaped in various ways the point of view taken in this book, and the way the material
is put together.
Oxford University Press provided me with three detailed and extraordinarily helpful
referee reports. Probably none of the referees will be satisfied with all of my decisions,
but hopefully each will find ways in which their thoughtful comments have greatly
improved the manuscript. I thank also Chris Kennedy both for very helpful comments
on the manuscript and for “dry-running” parts of it in his courses.
I owe a huge debt to Peter Klecha, who did a heroic job of going through the
manuscript in detail—working through all of the exercises to make sure they were
doable (indeed sometimes they weren’t), to make sure they were not presupposing
material not yet introduced (indeed, sometimes they were), and to make sure the
formalism was consistent (indeed, sometimes it wasn’t). I have been amazed at
Peet’s ability to see just what a student new to the material could or could not be
expected to do, and at his astute reading of the material which saved me from some
embarrassing mistakes. I am also extremely indebted to Jackson Golden, who has
done a fantastic
ACKNOWLEDGMENTS

job in helping to put together the bibliography, and to Christopher Mescher for help in
compiling the index. I thank NSF for funding on my grant BCS 0646081; although this
is a textbook it also contains much of my own research which was funded under the
grant.
 Finally, I thank Michael Rosen, for always encouraging this project, for putting up
with its writing over the last two years, and for everything else.
 xviii
Foreword: On Using this Book

Each instructor will no doubt find her or his own most useful path through this book
(or parts of it), but some suggestions might be helpful. The book obviously contains
more material than can realistically be covered in a single semester’s introductory
formal semantics course. In teaching this material myself in introductory formal
semantics courses (with a mix of graduate students and advanced undergraduates) I
have found that I can teach most of the material through Part III plus Chapter 16 and
a brief foray into one other of additional topics in Part IV, although I do not go through
the material in the full detail given here. (One reason for writing such a book is so that
students can get more detail on their own.) Hence, an instructor might choose to do
(most of) Chapters 1–15, and one or two additional chapters in Part IV. One plan for
this is to spend about three to four weeks on the material through Chapter 5, one week
on Chapters 6 and 7 combined, and then approximately one chapter a week for the
rest (Chapter 12 could take less than a week and Chapter 15 probably more).
Depending on the students’ background, another reasonable semester’s goal might
instead be to work in detail through Part II and then approach some of the material in
Part III. The material would also probably work as the full text for a twoquarter or one-
year semantics course (perhaps supplemented at the end with some readings from
original sources), and the later material can be the basis of a more advanced course,
supplemented with readings from the literature. I hope that the book might also be
used by students already familiar with basic formal semantics but not familiar with
the viewpoint of Direct Compositionality (or not familiar with Categorial Grammar);
such a student can systematically work through the material from what would be for
them a new point of view. This could easily be part of a second-semester or
secondyear semantics course.
There are some choice points in the order in which to read the material. Chapter
16 (on Negative Polarity Items) right after Chapter 10. This is because once students
have learned about generalized quantifiers, they have the tools to approach the
domain of Negative Polarity Items. I feel it is an especially satisfying reward—after
working through
FOREWORD: ON USING THIS BOOK

much technical apparatus—to apply the apparatus to a domain that is rich and so
beautifully illustrates the relevance of the formal tools to the distribution of items that
are entirely natural and conversational. So right after Chapter 10 there is a fork point:
one can either go on to Chapter 11 or digress and read Chapter 16. I have set it up so
that neither chapter (nor subsequent ones) presupposes the other except in
occasional minor ways. Chapter 16 could also be read between Parts II and III.
Moreover, the chapters in Part IV can be read in any order.
The decision to include two theoretical points of view—the Direct Compositional
view and a view that instead uses Logical Form—poses some difficulties. While I
would love to have written the book purely from the point of view of Direct
Compositionality, I feel it would be doing students a disservice to not be able to read
and appreciate the rich body of work that uses the notion of a Logical Form. My
solution, then, was to illustrate these two different viewpoints side by side with
respect to a key body of data (relative clauses, quantifiers in object positions, scopes,
and binding). But this could be confusing for a student new to this material, so I have
set it up in such a way that a student can—on a first pass—work through the relevant
chapters carefully from just one of the points of view, and later go back to work
through the other. In general, the material on the other point of view will need to be
skimmed as it will sometimes contain empirical data of relevance to the discussion,
but the technical details can be postponed.
Another difficult decision centered on the material on binding. As is pointed out in
the text, I give two versions of pronominal binding: a fairly standard one that makes
use of Logical Form and variables, and one making no use of variables in the
semantics set within the Direct Compositional framework. (This is based in part
although not exclusively on my own work on variable-free semantics.) But this makes
it appear as if Direct Compositionality is crucially tied in with a variable-free
semantics. It is not, and in an earlier draft (as well as in the way I have often taught
this material) I also exposit an intermediate position: one which is Direct
Compositional but does make use of variables. In the end, I decided that to develop
three approaches to a single domain was bound to be just too much for an
introductory book. With some regret, then, I decided it was prudent to remove the
discussion of the intermediate position. Still, I feel it is worthwhile to be able to access
this material—both for the sake of historical accuracy (it was a well-developed
position) and to understand that the issue of whether or not to have variables is to
some extent independent of
FOREWORD: ON USING THIS BOOK xix

the issue of Direct Compositionality. For those who are not fans of the variable-free
approach, I don’t want to leave the impression that if variablefree semantics is wrong,
 xx
so is Direct Compositionality: there is a very good Direct Compositional semantics of
pronouns using variables. So in order to allow access to a discussion of that
intermediate position, I have made it available on the textbook website, which is
<http://sites.clps.brown.edu/ compositional.sem>.
One further decision that some have found unusual is to delay the formal
introduction of the lambda calculus until Chapter 9 which is the end of Part I. There is
a reason for this. I want to stress that this is just a convenient notation for writing
model-theoretic objects, and not some level of representation that is a crucial part of
the theory. (For example, lambda conversion is not intended as a rule of grammar, but
just as a way to convert one way of writing a model-theoretic object into another way
of writing it.) To stress the difference between the notation and the actual meanings
of expressions, I have always found it helpful to try to use ordinary prose in naming
model-theoretic objects wherever possible, even when this prose leads to
cumbersome descriptions like “the set of all sets that contain the dog-set as a
subset.” Eventually, of course, English prose gets too cumbersome (and is not
sufficiently unambiguous) and we need a better notation (hence the lambda
calculus). But I find that insisting on prose whenever possible alongside with a clearer
notation is helpful—not only for understanding that notation is just that, but also to
get a better intuition about fancy objects like sets of sets or functions from sets of sets
to truth values. However, an instructor who prefers to make use of the lambda
calculus earlier can move Chapter 9; it could have been placed after Chapter 6.
For the student with no prior background in elementary set theory, I have included
an appendix to Chapter 1 that contains all of the basic notions of set theory, ordered
pairs, relations, and functions that are necessary for the later material. There is quite
a bit of material there, and so I would recommend that a student with no relevant
background just read the discussion of sets first, and then come back to subsequent
notions as they appear in the text. It is often easiest to absorb this material when one
has a reason to do so, and when it is in the context of the fuller semantic theory for
which it is being pressed into service.
A word about the sections and exercises marked with an asterisk. These are more
advanced and any of them truly can be skipped (especially on a first pass through the
material). They are intended to provide greater coverage and depth for the ambitious
and curious student, and the starred
FOREWORD: ON USING THIS BOOK

exercises can be quite challenging. Some—especially toward the end—are actually

open-ended; I hope that a few might even spawn some interesting research topics for
students. In any case, the starred exercises are there to be enjoyed and not to be
frustrating, so I would recommend picking and choosing wisely according to a
student’s level of interest and ambition.
I do hope that some of these will be fun.
Finally, as noted above, there is a website for this book at <http://sites.clps.
brown.edu/compositional.sem>. It contains the supplementary material on binding
and will be updated from time to time with other supplementary material, along with
errata as they are discovered.
xxii

PART I
Foundational Concepts :
Building a Fragment
 1
Introduction
1.1. Goals 3 1.3. The notion of a “fragment” and its
use in this text 12
1.2. A brief note on the history of
semantics within modern *1.4. An intriguing puzzle 13
linguistic theory 10
1.5. Appendix: Sets and Functions 19
1.1. Goals

This book stems from a belief that linguistic semantics is a beautiful field, that the
tools used to study formal semantics have yielded a rich body of results about
fascinating and subtle data, that the field continues to produce exciting new
insights at an impressive rate,1 and that there are simple and

1 Readers wishing a taste of many of the ongoing developments in formal semantics and
in the syntax/semantics interface might want to look at the journals Linguistics and
Philosophy (Springer), Natural Language Semantics (Springer),
Journal of Semantics (Oxford University Press), and Semantics and Pragmatics
(online journal, available at <http://semprag.org/>), among many other journals. Regular
conferences at which cutting-edge research is presented include the annual Semantics
and Linguistic Theory (SALT) conference, Sinn und Bedeutung (also annual), the biannual
Amsterdam Colloquium for Language, Logic, and Information, and Semantics of
Underrepresented Languages of the Americas, as well as most of the more general regular
linguistics conferences. Of course, most of the work in these venues will not be accessible
to a student just learning formal semantics, but it is hoped that this book will give a large
part of the necessary background for following at least some of this research. In any case,
a glance at the list of papers in
elegant tools to model how the syntax and semantics of a natural language work
together. We begin with a very elementary “fragment” of English and proceed to
expand it further and further—adding tools as needed but aiming to keep the basic
machinery relatively simple. The goal of proceeding in this way is to account for a
domain of data which is sufficiently rich as to show the excitement of studying
formal semantics and its interaction with syntax. We note one limitation from the
outset: this book concentrates entirely on the analysis of English. The project of
 4 1. INTRODUCTION

modeling the semantics and the syntax/semantics interaction of any single

language already provides such a rich set of results that one can hopefully find this
limitation justified for an introductory book like this. In fact, the results that have
been gleaned from a detailed modeling of one language have in recent years
allowed the field to expand so as to provide a wealth of analyses of other
languages.2 This book hopes to give the foundation to approach that literature.

1.1.1. Compositional semantics and (some of) the

goals of semantic theory

One of the most striking and fundamental properties of language—any language—

is that speakers have the ability to produce and understand an unlimited number
of expressions that they have never produced or heard before (indeed many of
these will have never before been uttered by anyone). This simple point is stressed
in just about every introductory linguistics textbook, often phrased this way: “A
speaker of a language is able to recognize as well-formed an unlimited number of
expressions that s/he has never heard before.” Examples that demonstrate this
are easy to construct.
One can, for instance, note the existence of expressions like those in (1): any of

these venues can give the reader a taste of the richness of the domain of inquiry within

linguistic semantics.

2
Much cross-linguistic semantic work can be found in the journals and conference
proceedings cited in footnote 1. An early edited volume on this is Bach et al.’s Quantification
in Natural Languages. There is now also an annual conference
Semantics of Underrepresented Languages of the Americas with published conference
proceedings. And many of the specialized conferences on individual languages and
language families regularly include work on semantics.
(1) a. the tallest linguistics major
b. the tallest linguistics major who is graduating in December
c. the tallest linguistics major who is graduating in December who is enrolledin
formal semantics
d. the tallest linguistics major who is graduating in December who is enrolledin
formal semantics who took phonology last semester . . .
 1.1. GOALS 5
One can keep forming longer and longer expressions like this by adding new
relative clauses (each of the phrases that begin with who here is what is
commonly known as a relative clause). But while this is often put in terms of a
speaker’s ability to recognize that these are well-formed, that is surely only part of
the story. Even more interesting (at least to a semanticist) is the fact that speakers
know how to interpret these expressions. The rule system that speakers have
unconsciously learned is hardly just a system to determine whether a given string
of words is an expression of the language in question (here English); language
would be quite useless if it were just a collection of meaningless strings.
And so, in modeling what a speaker of English “knows” (in an unconscious
sense, of course) about her/his language we want to predict how it is that s/he can
understand expressions like those in (1) no matter how many relative clauses they
contain. Thus speakers obviously have as part of their knowledge a finite set of
basic items—call these the words and call the collection of the basic items the
lexicon. (Here and for most of this text we ignore the distinction between words
and morphemes.) Since the lexicon is finite, the meanings of the basic items can
be learned on a case-by-case basis. But this obviously cannot be the case for the
larger expressions: there has to be some systematic set of principles that
speakers have that allows them to understand their meanings on the basis of the
meanings of the smaller parts (ultimately the words) that make them up. This is
the system which is called the compositional semantics—and one of the jobs of
a theory of the semantics (of any language) is to model the rules and/or principles
which allow speakers to understand an unlimited number of expressions. This
book is primarily about just this.
Let’s look a bit more at the expressions in (1). When a speaker utters any of
these expressions—perhaps as part of a fuller sentence like in (2)—the act of
uttering these expressions takes place in a fuller discourse context, and we
understand them relative to facts about that context:
(2) We need to make sure to order academic regalia which is long enough to fit the tallest
linguistics major (who is graduating in December (who . . . ))
The role of context will be discussed more formally at various points in the text,
but the informal notion of a speech or discourse context is clear enough. So
suppose we are using the expressions in (1) in a context in which it is obvious that
we are concerned with the students at Brown University. Given this (or any other
context), we can see that any speaker of English immediately knows some
interesting facts about these expressions—facts which our model of the
compositional semantics needs to account for. Take for instance (1a). It refers to
 6 1. INTRODUCTION

some unique individual.3 The hearer may well not know who exactly that is—in
fact the speaker might not either (as is clear in a context like (2)). But both parties
assume that there is a particular individual (and only one) referred to by each of
these expressions. And there are many other inferences that can be drawn from
these. For example, we immediately know that if the individual described by (1a)
is Nora, then either she’s also the individual described by (1b) or else she is not
graduating in December. Moreover, if Nora is not the person picked out by (1b)
then whoever that person is, s/he must be shorter than Nora. Similarly, with each
successively longer phrase we either refer to the same person, or to one who is
shorter. Suppose that Zorba is the person described by (1b). We know that he is
shorter than Nora, and also know that if he is not the person described by (1c)
then he is not enrolled in formal semantics. And whoever the (1c) person is—let’s
say Otto—Otto must be shorter than Zorba. The addition of each successive
relative clause either keeps the referent constant or allows shorter and shorter
people to “rise to the top.” This kind of knowledge is automatic and immediate,
and it is the job of a model of the compositional semantics to explicitly account
for inferences like this.
We won’t give a serious account of any of this at this point, but can hint at one
possible account. Suppose that an expression like linguistics major refers to some
set of individuals. (Readers not familiar with basic notions of set theory should
consult the Appendix to this chapter.) When this set is put together with the tallest
(pretend that the tallest is a single word here), the entire expression ends up
referring to the tallest member of that set. Nothing

3 In reality there could conceivably be two individuals of exactly the same height. But use

of the expressions in (1) does seem to assume that there is a unique referent for these. This
is sometimes called a presupposition; these are rather odd expressions if the speaker
knows that there are two individuals with exactly the same height (in that case the speaker
might have said the two tallest linguistics majors).
surprising so far. But what is more interesting is what happens with the addition of
further relative clauses. It seems plausible that something like who is graduating
in December also refers to a set (obviously, the set of December graduates). The
above facts will make sense if the compositional semantics first combines the
two sets (the set of linguistics majors and the set of December graduates) and
intersects them to give a new set. (The intersection of two sets is all things that
are in both sets; again see the Appendix.) So (1b) ends up picking out the tallest
member of that set. It is now possible to demonstrate that the system correctly
 1.1. GOALS 7
predicts that if the referent of (1b) is not Nora, it can only be because she is not
graduating in December. For if Nora is taller than anyone in the linguistics major
set (call that L) then she is taller than anyone in the intersection of L with the
December graduates (call that D). After all, everyone who is in that intersection of
L and D is also in L. So if Nora is not the referent of (1b) it can only be that she’s
not in the intersection of D and L, and since she’s in L (by assumption) it follows
that she can’t be in D. It also follows that if (1b) refers to Zorba, he must be shorter
than Nora. By the definition of intersection, if Zorba is in the intersection of D and
L he is in L, but we already know that Nora is taller than everyone else in L. All of
this is very simple logic that we—the linguists—can work out in the form of an
informal proof as above. It could also be worked more formally if one were so
inclined. Pedantic though it may seem, it shows that our compositional procedure
(which involves intersecting two sets) can be used to correctly model inferences
that speakers of English effortlessly make.
Moreover, the appeal here is that this is perfectly general and extends no matter
how many new relative clauses are added. Take (1c). The semantics set-up above
extends immediately to this. The new relative clause in (1c) is who is enrolled in
formal semantics. This picks out yet another set—and so this now intersects with
the set that we already formed for (1b). The fact that the referent of (1c) can either
be Zorba or someone shorter than Zorba follows by the same logic shown above;
the reader can work out the details. And the procedure can be repeated over and
over no matter how many relative clauses are introduced.
1.1. Because this example is just meant to illustrate the notion of a
compositional semantics, we have made some assumptions about the order
in which the semantics put things together without justifying them. Suppose
that rather than the way it was set up here, the meanings of the two relative
clauses (1c) first combined, and then that combined with linguistics major.
Would that make any difference to the basic semantic compositional picture
that we have set up here? Would the procedure extend correctly to (1d)?

1.1.2. Direct Compositionality—and its role in this text

This book has a rather ambitious set of goals. On the one hand, I intend this to be
a stand-alone text for anyone wishing to have an introduction to formal semantics,
compositional semantics, or what is commonly known as the syntax/semantics
interface. In other words, we will be asking (as in the above example) what a
 8 1. INTRODUCTION

compositional semantics might look like: how can we model the tools available
(again, of course, unconsciously) to speakers of a language that allow them to
compute meanings of larger expressions from the meanings of the smaller ones
that make them up. What are the formal ways in which meanings combine? And
what are the types of objects that we need in order to model that? (For example,
the discussion above shows that some simple tools of set theory can be useful.)
But while most semanticists agree that (in general) the meaning of a larger
expression is built in some systematic way from the meanings of the parts that
make it up, just exactly how the syntactic system of a language and the
compositional semantics work together is a matter of considerable controversy,
and is one of the central questions addressed in this book. And so this book takes
one particular point of view on this: the point of view known as Direct
Compositionality. This view was explored perhaps most notably in Montague
(1970) and was either generally accepted or at least taken as a serious
desideratum in much of the work in linguistic formal semantics throughout the
1970s and 1980s (particularly work in what was then known as the Montague
Grammar program). It was also taken as the foundation for semantics in syntactic
theories such as Generalized Phrase Structure Grammar (Gazdar, Klein, Pullum,
and Sag 1985), and is assumed in a large body within current grammatical theories
that go under the rubric of Categorial Grammar, Type-Logical Grammar, and other
related theories.
To elucidate, a fairly uncontroversial claim is that the grammar of any natural
language is a system of rules (or principles, if one prefers) that define the set of
well-formed expressions of the language (i.e., the syntax) and a set of rules (or
principles) pairing these with meanings (i.e., the semantics). The hypothesis of
Direct Compositionality is a simple one: the two systems work in tandem. Each
expression that is proven well-formed in the syntax is assigned a meaning by the
semantics, and the syntactic rules or principles which prove an expression as
well-formed are paired with the semantics which assign the expression a
meaning. (An interesting consequence of this view is that every well-formed
syntactic expression does have a meaning.1) It is not only the case that every well-
formed sentence has a meaning, but also each local expression (“constituent”)

1 Of course one of the earliest arguments in Generative Grammar for divorcing the
syntax from the semantics (and thus a putative argument against Direct Compositionality)
is based on the claim that there are well-formed expressions that don’t have any meaning
(Chomsky 1957). This is addressed in section 7.1.
 1.1. GOALS 9
within the sentence that the syntax defines as well-formed has a meaning. Of
course putting it this way is arguably not much more than just a slogan: the
empirical content of this depends in part on just how the syntax works and what
one takes to be a meaning. This will be filled in as we proceed. It might also seem
at first glance that the hypothesis of Direct Compositionality is a fairly trivial one.
But in fact it is not always immediately obvious how to give a Direct Compositional
analysis. Even the example in 1.1.1 is a case in point. If the syntax and semantics
work together, then the analysis given above leads to the conclusion that in the
syntax a relative clause like who is graduating in December combines with
linguistics major rather than with the tallest linguistics major. But this very
question regarding the syntax of relative clauses has been debated in the literature
since the 1960s, and many researchers have claimed that the syntactic
constituent structure of the tallest linguistics major who is graduating next year is
not the structure that was used above for the semantic analysis. We will actually
revisit this particular question in later chapters (see, e.g., section 13.2). So one of
the goals of this book will be to see what it takes to give Direct Compositional
analyses of a variety of constructions.
While the material in this book is generally exposited from the Direct
Compositional point of view (along with discussion of the challenges to this
hypothesis), the book is also intended to be a perfectly reasonable
 10 1. INTRODUCTION

stand-alone textbook for any formal semantics course. Thus it is suitable for any
linguistics student or linguist wanting a ground-up introduction to formal semantics,
and for a philosophy or logic student wanting a background in formal semantics within
linguistics. In the service of being a stand-alone text in modern formal semantic
theory, the book will, where relevant, also develop the mechanics of at least one fairly
standard nonDirect Compositional theory of the syntax/semantics interface. This is
done especially in Part III (Chapters 13–15) where some phenomena are discussed
from both direct and non-Direct Compositional points of view. There are several
reasons for expositing parallel Direct and non-Direct Compositional accounts of
some domains. One is to enable readers to approach the wide range of literature
written from either point of view. Second, this allows for a serious comparison of two
different approaches. Third, learning more than one set of details for the analysis of
any construction allows for a deeper understanding of the basic generalizations and
results— generalizations which often transcend the particulars of one theoretical
implementation. Finally, a student who has already learned formal semantics from a
non-Direct Compositional point of view can hopefully also profit from this book by
seeing an interesting fragment of English explicitly analyzed from the Direct
Compositional point of view.

1.2. A brief note on the history of semantics within modern

linguistic theory

The subfield of semantics as a core field in modern linguistic theory is relatively recent
and is one of the fastest growing subfields. 5 Early work within the general enterprise
of generative grammar had little to say about semantics. To be sure, by the end of
about the 1960s and the early 1970s there was considerable discussion as to how
the syntax and the semantics interacted; such discussion was mostly framed in
terms of a debate between Generative Semantics (see, e.g., McCawley 1971; Lakoff
1971) and Interpretive Semantics (see Chomsky 1970; Jackendoff 1972). We will not
discuss the content of that debate here, but much of the work framed within these

5
A much more extensive and authoritative history of the development of formal semantics
within modern linguistic theory can be found in Partee (forthcoming).
 11
1.2. THE HISTORY OF SEMANTICS

two competing points of view did not incorporate a systematic view of the semantics
itself. Of course, the linguistics literature during that period contained many seminal
observations about semantic notions such as scope, negation, and “binding,” but
these were generally not embedded within a full-blown theory of semantics designed
to capture semantic notions like entailment and truth conditions (see Chapter 2),
although they easily could have been embedded into such a theory. The fact that
semantics was not taken a subfield in and of itself during this period comes—at least
in part— from Noam Chomsky’s emphasis on syntax during the early development of
generative grammar. Chomsky (1957) explicitly rejects the notion that semantics is
relevant in the construction of grammars, and this notion persisted for quite some
time.
It is probably fair to say that modern formal semantics as a subfield within linguistic
theory began in the early to mid-1970s with the cross-fertilization of linguistic theory
and philosophy of language (including semantics) sparked by Barbara Partee,
Richmond Thomason, David Lewis, and others. Partee’s work was a particularly
influential bridge between linguistics and philosophy as she had originally been a
student of Chomsky’s at MIT and always had a strong interest in the connections
between language and logic, and hence in topics like quantifiers, negation, etc. As an
assistant professor at UCLA, she became acquainted with the seminal work of
Richard Montague, a philosopher and logician who (among his many other
contributions within philosophy and logic) had a major interest in modeling the
semantics (and the syntax/semantics interaction) of natural language (although
Montague himself dealt only with English). In fact, the program of Direct
Compositionality is advocated in his work (see especially Montague 1970).
We will have more to say about his specific contributions as this book proceeds; for
now, we note that one of the appeals of his work from the point of view of a linguist
was his notion that the semantic composition of natural language reflects and
respects its syntax. Partee saw the relevance of Montague’s work to linguistic theory
and wrote a series of papers aimed at synthesizing some of the insights from
Montague’s work with results within Transformational Grammar (see, for example,
Partee 1973). At the same time, the appearance of Lewis (1970), Stalnaker and
Thomason (1973), and other work in the philosophy of language also helped launch
modern formal semantics and cement its connection to linguistic theory. Such work
within philosophy as well as Partee’s early group of students (both at UCLA and later
at the University of Massachusetts) continued the tradition, broadening the domain
 12 1. INTRODUCTION

of inquiry and results. From there was born the enterprise known as Montague
grammar6 which eventually gave rise to the more general subfield of formal
semantics. Montague himself died in 1971,7 and the field of formal semantics evolved
in many ways quite different from the original work in Montague grammar.
Nonetheless, many of the basic tools of linguistic formal semantics as it is developed
to this day stem from some of this early work cited above. Since the late 1970s the
field has blossomed, and is now within linguistics generally considered as one of the
core areas along with at least phonology and syntax.

1.3. The notion of a “fragment” and its use in this text

Inspired by the work of Montague in papers such as Montague (1973), much work in
formal semantics within the 1970s and 1980s took it as axiomatic that a goal was to
formulate fully explicit grammars (in both syntactic and semantic detail) of the
fragment of the language one is concerned with (English in most such work). The term
“fragment” got extended to mean not only the portion of the language being modeled,
but also the portion of the grammar being proposed as an explicit account of the facts.
The strategy of writing fragments (of grammars) has the advantage of giving an explicit
theory which makes testable predictions, and of making theory and/or proposal
comparison easier.
Unfortunately, the goal of formulating fully explicit fragments went out of style
during the last two decades or so. This is in part due to the fact that linguistic theories
often promised that many of the particular details did not need to be stated as they
would fall out from very general principles. It is certainly reasonable to hope that this
is ultimately true, but the relevant principles often go unstated or are stated only
rather vaguely, making it extremely difficult to really compare proposals and/or
evaluate theories and theoretical claims. Having rules and principles be as general as
possible is, of course, highly desirable. But this does not mean that they should not
be

6
An excellent introduction to the general program of Montague semantics and an
explication especially of Montague (1973) can be found in Dowty, Wall, and Peters (1981).
7
Montague was murdered on March 7, 1971. No arrest was ever made in conjunction with
the murder.
 *1.4. AN INTRIGUING PUZZLE 13

formulated explicitly—only that more mileage will be gotten out of explicit

formulations.
The present text is therefore committed to trying to revive the notion of explicit
fragment construction. We cannot promise to give every detail of the domain of
English syntax and semantics we are trying to model. Some parts will be left
tentative, some stated informally, and some simply omitted. Nonetheless, the
goal is to give a reasonable amount of an explicit fragment. We will therefore
periodically take stock by summarizing the fragment constructed so far, and a full
summary is provided at the end of Part III.

*1.4. An intriguing puzzle

This introductory chapter concludes with an illustration of a puzzle, a solution to

which is proposed in section 15.6. However, the goal here is not to champion any
one particular solution, and readers may safely skip this section and return to the
data only in section 15.6. But we include this in the introductory remarks for the
reader who wants a preview of just what kinds of complex and subtle data a
theory of syntax and semantics ultimately hopes to account for. To fully
appreciate the particular puzzle here, one should keep the following in mind. The
contrasts are quite real; the judgments have been checked with many speakers
over the years by myself and many others. Yet—like other subtle facts in syntax,
phonology, and semantics—these are not generalizations which we have ever
been consciously taught nor even generalizations that most of us are even aware
of until we see them in a linguistics course (or book). What, then, is there about
our unconscious knowledge of the grammatical system that predicts these
judgments? This is the sort of puzzle that theories of semantics and its interaction
with syntax ultimately seek to solve.
So, consider what we will call the A-B-C party scenario. I go to a small party
consisting of only myself and three married couples: Alice and Abe, Betty and
Bert, and Cathy and Carl. I learn that Alice and Abe met each other only a few
years ago, and similarly for Cathy and Carl. But interestingly, I also find out that
Betty and Bert have been sweethearts since childhood. I like Betty a lot, and spend
a good part of the evening talking to her. The next day, you ask me how I enjoyed
the party and if there was anyone that I especially enjoyed meeting. I certainly can
answer with (3):
(3) Oh yes, I especially enjoyed talking to Betty.
 14 1. INTRODUCTION

But now suppose that I can’t remember Betty’s name, although I do remember
that her husband’s name is Bert. I can answer with either (4a) or (4b):
(4) a. Oh yes, I especially enjoyed talking to—oh, I can’t remember her name— you
know, the woman who is married to Bert.
b. Oh yes, I especially enjoyed talking to—oh, I can’t remember her name— you know,
the wife of Bert.

((4b) would sound more natural if we substituted Bert’s wife for the wife of Bert;
this will not impact on the ultimate point and the exposition is simplified using
(4b).)
Now, let us tweak the scenario slightly and assume that I am one of those
people who just doesn’t remember names very well. As a result, I remember
neither Betty’s name nor Bert’s name, although I do remember the interesting fact
that they are the only couple at the party who have been sweethearts since
childhood. As an answer to your question, (5) would be quite natural:
(5) Oh yes, I especially enjoyed talking to—oh, I can’t remember her name—you know,
the woman who is married to her childhood sweetheart.

But what is striking is that I can’t answer with (6):

(6) *Oh yes, I especially enjoyed talking to—oh I can’t remember her name—you know,
the wife of her childhood sweetheart.

(We are taking liberties with the * notation here. This is generally used in works in
syntax to indicate something that is ill-formed. (6) in fact is fine, just not on the
intended reading; and we will continue to notate a sentence with an asterisk in
front of it when we mean “bad on a particular reading” provided that it is clear what
the intended reading is.) It should be noted that some speakers find the contrast
rather subtle but there is general agreement that (6) is stranger than (5).
All of these examples contain various extra material (the parentheticals, etc.)
which are there to make them sound natural and conversational. But as we
proceed it will be convenient to strip away the parts that are irrelevant to figuring
out the semantics, so we can recast (6) as (7)—also impossible as an answer to
the question and in this context:
(7) *I especially enjoyed talking to the wife of her childhood sweetheart.
Of course, (7) is a perfectly good sentence, but it cannot be used in our party
scenario as a way to identify Betty.
Since some readers do find the contrast subtle, two points are worth noting.
First, one should resist the temptation to recast (7) in one’s mind as I especially
 *1.4. AN INTRIGUING PUZZLE 15

enjoyed talking to the one who’s the wife of her childhood sweetheart or I
especially enjoyed talking to the woman who’s the wife of her childhood
sweetheart. That would be cheating; the point is not to find a closely related way
to say the same thing but to notice that the actual way in (7) contrasts with I
especially enjoyed talking to the woman who is married to her childhood
sweetheart (and contrasts with the above variants too). As to why these variants
are good, we return to that shortly. Moreover, while the contrasts above may be
subtle for some speakers, there is a related mystery where the facts dramatically
pop out. Thus take (8) in the same scenario, where the only people at issue are
Alice, Betty, and Cathy:
(8) Betty is the only woman who is married to her childhood sweetheart.

This can be making two different claims. The obvious one in this scenario is that
Cathy is not married to Cathy’s childhood sweetheart, and Alice is not married to
Alice’s childhood sweetheart. The other is the “non-polygamous” reading: it
asserts that Bert (or whoever Betty’s husband might be) has only one wife. Since
we (generally) assume that people have just one wife, this reading (given standard
assumptions) is not the first one that someone would think of since it is less likely
to be conveying any interesting information. But despite the fact that the non-
polygamous reading is the less obvious one for (8), it is the only reading (or at least
the one that pops out first) for (9):
(9) Betty is the only wife of her childhood sweetheart.

Why should that be? We’ll put (8) and (9) aside for the moment, and return to the
simpler case of (7).
So the mystery is why (7) is bad as a way to identify Betty. This is especially
puzzling in that both (4b) and (5) are perfectly good—or, to give their stripped-
down versions, (10) and (11) are both fine. Each one differs minimally from our
bad case, yet neither of these two has any problem.
(10) I especially enjoyed talking to the woman who is married to her childhood
sweetheart.

(11) I especially enjoyed talking to the wife of Bert.

So surely there is nothing incoherent or wrong with the meaning that (7) is trying
convey, for (10) is just a slightly different form and conveys exactly this meaning.
Hence the puzzle has something to do with the mapping between syntax and
semantics: why one is a good way to package the relevant information while the
other is not.
 16 1. INTRODUCTION

We can informally recast the puzzle in the following way. Compare the two
expressions the woman who is married to Bert and the wife of Bert.
(Following a long tradition within linguistics, we will refer to these as NPs, which
comes from “noun phrases.” They are also in much modern literature referred to
as DPs, for “determiner phrases,” but we stick to the more traditional terminology
in this text.)8 Both of these can correspond to meanings that we can (roughly and
informally) represent as (12):
(12) the x: x is a woman and x is married to Bert

But while the object NP in (10) can be represented as in (11), the object NP in (7)
cannot:
(13) the x: x is a woman and x is married to x’s childhood sweetheart

The basic phenomenon here was discussed in, among others, Jacobson (1977)
(where it was called Langendoen’s constraint), Chomsky (1981 under the rubric
of i-within-i condition), and many since. As there seems to be nothing wrong with
the meaning, we can assume that the phenomenon in question has something to
do with the way the syntax and semantics interact.
Notice that we have given a kind of formula (and one that uses a “variable” x) to
represent the meanings in question, but for now we should think of these simply
as placeholders to bring out the intuition. After all, recasting her in the above
examples as x doesn’t really immediately give us the tools for computing the
meanings of the expression: we have traded a

8
In theories which use the term DP, the NP is used instead to refer to material after the
Determiner; e.g., mother of Romeo in an expression like the mother of Romeo. Here we will
be calling this simply N (i.e., a “noun”) and allowing terms like N to refer both to simple
material consisting of just one word and to complex material. This is discussed further in
Chapter 6. We are aware that this will initially cause some confusion to a reader who is
used to using “NP” to mean a noun and its complement, but it is well worth becoming fluent
in both sets of terminologies. The terminology here is the traditional one found in large
amounts of literature not only in syntax and semantics but in neighboring fields like
psycholinguistics, philosophy, cognitive science, etc.
pronoun her for a variable x. But this accomplishes little until we have a way to
think about what a variable like x means. (Indeed this is explored in detail in
Chapters 9 and 15, including developing an alternative view that does not make
use of variables in the semantics.) We thus caution that formulas like (12) and
(13) are best seen simply as informal and helpful ways to bring out the intended
 *1.4. AN INTRIGUING PUZZLE 17

meanings. Similarly, one often sees indices used in the literature as a way to bring
this out; one will find discussions using the notation in (14) and (15) to make the
point, where the indexation in (14) indicates a good possible reading for the NP
while (15) cannot be understood in the intended way:

(14) the womani whoi is married to heri childhood sweetheart

(15) *the wifei of heri childhood sweetheart

Much work in grammatical theory actually assumes that NPs and pronouns come
with (obviously silent) indices in the syntax; here we will be using indices from
time to time simply as a way to notate intended readings without any commitment
to their being actual pieces of grammatical machinery.
Before leaving this (for now), there’s one other interesting point to notice.
However we ultimately state the principle, the claim is that an NP like the wife of
her childhood sweetheart cannot correspond to the meaning shown earlier in
(13):
(13) the x: x is a woman and x is married to x’s childhood sweetheart

But one might think that this is not really correct, since it is in fact just fine to use
(16) as a way to identify Betty:
(16) I especially enjoyed meeting the woman who is the wife of her childhood
sweetheart.

This point was made earlier; many speakers on reading (7) tend to recast it in their
minds as (16). Similarly, (17) is impeccable on the understanding where her is
Betty:
(17) Betty is the wife of her childhood sweetheart.

But a close reflection reveals that this does not threaten the generalization.
Again, using indices or variables simply as a convenient way to elucidate the point,
it is easy to see in (17) that her need not be “coindexed with” (or “correspond to
the same variable as”) wife but rather it just refers directly to Betty. That is, we can
represent it as in (18a) using indices, or as in (18b) using the more spelled-out
formula (though still quite informal).
(18) a. Bettyj is [NP the wifei of herj childhood sweetheart].
b. Betty, y [y = the x: x is a woman and x is married to y’s childhood sweetheart]

Since we are asserting identity between Betty and the person married to Betty’s
childhood sweetheart, it of course follows that Betty is married to Betty’s
 18 1. INTRODUCTION

childhood sweetheart and so the full sentence (16) will end up with the relevant
meaning.9 But the claim that the object NP itself (the wife of her childhood
sweetheart) does have the meaning represented in (12) is not threatened. The
same point holds for (17), whose meaning can be represented as (19a) or (19b).

(19) a. the womanj whoj is the wifei of herj childhood sweetheart

b. the y: y is a woman and y = the x: x is a woman and x is married to y’s childhood
sweetheart

Is there a way to confirm that this is the right sort of explanation for these
apparent counterexamples? Indeed there is, and it centers on the contrast
between (8) and (9) which was discussed earlier. We leave it to the interested
reader in the exercise to play with this and get a sense of why (8) is ambiguous and
(9) is not. Having completed that, one should be able to see how it is that this gives
support for the explanation offered above as to why (17) does not threaten the
claim that the wife of her childhood sweetheart cannot correspond to the meaning
shown informally in (12).

*1.2. Work out—using the informal representations either with indices or the
representations with variables—why it is that (7) is ambiguous and (8) is not.
Of course you will need to think a bit about how to treat only, but nothing very
complex is required. You can be perfectly informal in your treatment of only, but
you should be able to get a feel for why these two differ.

9
This general observation—although for a slightly different case—was made in
Postal (1970) who distinguished between “presuppposed” coreference and “asserted”
coreference. Here the fact that Betty and the wife of her childhood sweetheart end up
“referring” to the same individual is exactly what the sentence is asserting.
 1.5. APPENDIX: SETS AND FUNCTIONS 19

As noted at the outset of this section, the goal here is just to provide a mystery
to whet the reader’s appetite; the tools needed to provide a hypothesis as to the
explanation of the mystery will be developed later.

1.5. Appendix: Sets and Functions

1.5.1. Sets, members, and subsets

Since the notions of sets and of functions are crucial throughout this book, some
formal definitions and discussion are provided here for readers not entirely
familiar with these notions. We begin with the notion of a set. A set is simply any
collection of objects (it can have a finite number of objects, an infinite number, or
none at all). For example, we can talk about the set of positive integers less than
10; sets can be notated by listing the members and enclosing the list in curly
brackets: {1,2,3,4,5,6,7,8,9}. The order in which they are listed makes no
difference; a set is just a collection of things without any order. So if we were to
write {2,5,3,4,9,7,8,1,6}, this names the same set. Each integer in this set is called
a member or an element of the set. If we were to name this set A, then the notation
4 ∈ A means that 4 is a member (or element) of A. Something either is or is not in
a set; it makes no sense to say it occurs twice (or more) in the set. Note also that
a set can have a single member; this is called a singleton set. Thus {4} is the set
with only one member; this set is distinct from 4 itself. (4 is a member of {4}.)
A set can have an infinite number of members; the set of positive integers for
example is infinite. Obviously this can’t be named by listing the members. One can
in this case specify the set by a recursive procedure. Call the set I, then one can
specify I by two statements: (a) (what is known as the base step): 1 ∈ I, and (b)
(the recursion step) if n ∈ I then n+1 ∈ I. (It is understood when one lists things this
way that nothing else is in I.) One will also often see a notation which describes
rather than lists the members. For example, we can write the following set, call it
B: {x|x is a New England state}. This names a finite set, and so we could also give
B in list form as follows: {Maine, New Hampshire, Vermont, Massachusetts,
Rhode Island, Connecticut}. These two are just different notations for naming the
same set. This can also be used, of course, for infinite sets. Take, for example, the
set {x|x is an integer and x > 9}. This names the set of integers greater than 9. And,
a set can have no members. There is only one such set; its name is the null set or
the empty set, and is generally written as . Of course, there are other ways one can
 20 1. INTRODUCTION

describe the null set. For example, the set of integers each of which is greater than
9 and less than 10 is the empty set. The cardinality of some set refers to the
number of elements in that set; the notation |B| means the cardinality of B. Hence,
given our set B above, |B| is six.
Take some set A. Then a subset of A is any set all of whose members are also
in A. Suppose, for example, we begin with a set C which is {1,2,3}. Then {1,2} is a
subset of C, as is {1,3} and so forth. The notation for the subset relation is . The full
definition of subset is as follows: B A if and only if every member of B is a member
of A. From this it follows that every set is a subset of itself (so for the set C above,
one of its subsets is the set {1,2,3}). It is, however, sometimes convenient to refer
to those subsets distinct from the original set; in that case we can talk about a
proper subset of some set. The symbol for this is , so B ` if and only if B A and B
6¼ A. Since the definition of subset says that B is a subset of A if and only if
everything that is in B is also in A, it follows that if nothing is in B then B is a subset
of A. thus the null set is a subset of every other set. Sets themselves can have sets
as members, and so one can talk about the set of all subsets of a set A. This is
called the power set of A, written as P (A). For example, given the set C above, P
(A) = {, {1}, {2}, {3}, {1,2}, {1,3}, {2,3}, {1,2,3}}.

*1.3. If a set A has n members, then the number of subsets of A is 2 n. Try to see
why this is true. Hint: for every member x of some set A, then for each subset B
of A, x is either in B or is not in B.
1.4. How many members does the following set have: {}?
*1.5. What is P ()?

We will also have occasion to talk about the reverse of the subset relation—i.e.,
the superset relation. A is a superset of B if and only if B is a subset of A. The
notation for this is A B. Once again this is defined in such a way that every set is a
superset of itself; a superset of B which is not identical to B is called a proper
superset, and the notation for this is .
1.5.2. Union, intersection, and complement

Take any two sets A and B. Then there is a set C which consists of everything that
is in A and everything that is in B. This is called the union of A and B, and is written
A [ B. For example, if A is {1,2,3} and B is {2,4,6} then A [ B is {1,2,3,4,6}. Moreover,
 1.5. APPENDIX: SETS AND FUNCTIONS 21

for any two sets A and B the intersection of A and B is the set of all things that are
in both A and B. This is written A \ B. So, for example, in the case directly above,
the intersection of A and B is {2}. Or, if we were to intersect the set of integers
which can be evenly divided by 2 (the set of even integers) with the set of integers
which can be evenly divided by 3, we end up with the set of integers that can be
evenly divided by 6.

1.6. a. For any two sets A and B such that A B, what set is A [ B?

b. For any two sets A and B such that A B, what set is A \ B?

One final useful notion here is the complement of a set. The complement of
some set A is the set of all things which are not in A (this is sometimes notated as
A0). Usually one talks about this notion with respect to some larger domain.
Strictly speaking, the complement of {1,2,3} would include not only all integers
greater than 3 but also all sorts of other numbers (like 1/3), the sun, my dog Kiana,
and the kitchen sink. Rarely are we interested in that sort of set; so in practice
when one talks about “the complement of some set A” this is generally with
respect to some larger set B of which A is a subset. Then the complement of A
refers to all things in B that are not in A (this is notated as B-A). For example, when
restricting the discussion to the set of positive integers, the complement of {1,2,3}
is the set of all integers greater than 3.

1.5.3. Ordered pairs, relations, equivalence relations, and

partitions

Sets are unordered collections of objects. But it is quite useful (as will become
very apparent as this book proceeds) to be able to talk about pairs of objects that
are ordered in some way. An ordered pair is just that: it is two objects with some
ordering between them. If the two objects are a and b, then (a,b) is an ordered pair;
(b,a) is a different ordered pair. An ordered pair need not contain distinct items:
(a,a) is an ordered pair. In applying this to actual natural relations that exist in the
world we are generally interested in sets of ordered pairs. (One can generalize this
notion to ordered triples and so forth; an ordered n-tuple means an ordered list of
n items.)
This notion is easiest to grasp with some concrete examples. Take again the set
{1,2,3}, and take the relation “is greater than.” Then this can be seen as a set of
ordered pairs; if we are restricting this to items from our little 1-2-3 set, this would
 22 1. INTRODUCTION

be the set {(2,1), (3,1), (3,2)}. Now suppose we instead take the following set of
ordered pairs: {(2,1), (3,1), (3,2), (1,1), (2,2), (3,3)}. Then (restricting this again to
our 1-2-3 set) we have now actually listed the relation “is greater than or equal to.”
Or, take the set {(1,1), (2,2), (3,3)}.
That is the relation “is equal to” (defined for the set of integers {1,2,3}).
In other words, what we are calling a relation is just some set of ordered pairs.
In the example above, both the first and second member of each ordered pair was
drawn from the same set (the set {1,2,3}). But this is not necessary; we can have
a set of ordered pairs each of whose first member is drawn from some set A and
the second member from some set B where A and B are different (they can, but
need not, have some of the same members). For example, the relation “is the
capital of” is a relation between cities and states; it can be expressed as a set of
ordered pairs of the general form {(Providence, Rhode Island), (Boston,
Massachusetts), (Springfield, Illinois), (Pierre, South Dakota), . . . } (the . . . here is
a shorthand for the remaining 46 pairs).
Take two sets A and B (they could be the same set or different). Then A x B refers
to the set of all ordered pairs whose first member is in A and whose second
member is in B. (This is also called the Cartesian product of A and B.)
As in the case above, it is helpful to give the intuition of this by coming up with
some concrete example. Suppose we take as our set A some group of
professors—say, Professor Magoo, Professor Carberry, and Professor Glazie. Call
that set P (for shorthand, let’s call its members m, c, and g, so P is the set {m,c,g}).
Now suppose we have a set S which consists of three students who we will just
indicate as x, y, and z (so S = {x,y,z}). Then P x S = {(m,x}, (m,y), (m,z), (c,x), (c,y),
(c,z), (g,x), (g,y), (g,z)}. Suppose that Magoo wrote a letter of recommendation for
all three students, Carberry wrote one for only y, and Glazie wrote one for y and z.
Then the relation “wrote a recommendation for” is a subset of P x S, and is the set
of ordered pairs
{(m,x}, (m,y), (m,z), (c,y), (g,y), (g,z)}.
More generally, we define a relation (between members of A and B) as any
subset of A x B. There are some special and interesting properties that can be
defined when the two sets are the same. That is, we are now looking at subsets of
A x A. Consider a relation R (some subset of A x A) which is such that for all x in A,
(x,x) is in R. Such a relation is called a reflexive relation. (These need not be the
only kinds of pairs to be in R for R to be reflexive; other pairs can be in there too.)
For example, if talking about the set of integers again, the relation “is greater than
or equal to” is reflexive; for all numbers n, (n,n) is in the set of ordered pairs
 1.5. APPENDIX: SETS AND FUNCTIONS 23

described by that relation. A relation R is called irreflexive if for all x in A, (x,x) is

not in R. Further, consider any two members x and y (both members of A). Then if
it’s the case that for all x and y if (x,y) is in R then (y,x) is also in R, the relation is
called symmetric. Imagine, for example, a lovely world with no unrequited love.
Then is in love with is symmetric in that world. If our set were {m, c, g, and p}, then
if the pair (m,c) were in our relation R (i.e., “is in love with”) the fact that R is
symmetric means that (c,m) is also in R. (Notice that our definition neither
requires (c,c) to be in R nor excludes that; either is possible.) Or, to look at a
relation which is symmetric by definition: consider the relation is a sibling of.
(While is a sibling of is symmetric, is a sister of is not. Why not?) One final useful
definition is a transitive relation. A transitive relation R is one for which for every x,
y, and z, if (x,y) is in R and (y,z) is in R, then (x,z) is in R. (The relation “is greater
than” is transitive, as is the relation “is greater than or equal to”).
Any relation R which is reflexive, transitive, and symmetric is called an
equivalence relation. As an example of such a relation, consider the set of
students (call it S) at an elementary school that services grades 1 through 6. Then
“is in the same grade as” is an equivalence relation in S x S. (While it is unusual to
use the phrase “in the same grade as” when referring to the same person it seems
false to say Johnny is not in the same grade as himself so we can see that this
relation is reflexive.) It is also obvious that it is symmetric and transitive. Note that
this—and any other equivalence relation—divides up the original set (here, S) into
a group of non-overlapping subsets. The set of these subsets is called a partition.
Thus, a partition of any set S is a set of subsets of S such that for each distinct
subset A and B, A \ B = , and the union of all the subsets is S. To show that any
equivalence relation induces such a partition, take any x in S and define S x as
{y|(y,x) is in R}. Since R is reflexive, we know that x is in Sx (and hence we know that
Sx is guaranteed not to be empty). Moreover, the fact that R is reflexive means that
each member of S is guaranteed to be in at least one such subset, so we know
that the union of all of these is S. We can further show that for any two such
subsets Sa and Sb, they either have no members in common (i.e., they have a null
intersection) or they are the same. Thus, take any c which is in both S a and Sb. By
definition, this means that (c,a) is in R and (c,b) is in R. By the fact that R is
transitive and symmetric, it follows that (a,b) and (b,a) are in R (the reader can
work through the necessary steps). But then, for all x such that (x,a) is in R, (x,b) is
also in R. To show this note again that R is transitive. If (x,a) is in R and (a,b) is in R
then (x,b) is also in R. Hence given the initial premise that there is a non-empty
intersection between Sa and Sb, it follows that everything in Sa is in Sb. That
 24 1. INTRODUCTION

everything in Sb is also in Sa follows in the same way, and so the two are the same
set. Each subset in a partition is called a cell in that partition.
In the example above, the cells correspond to the different grades. (There don’t
have to be six cells—it could be that one of the grades has no student in it. But
there can be no more than six; recall that by definition a cell can’t be empty.) Just
as any equivalence relation induces a partition, given any partition one can give an
equivalence relation that corresponds to any partition; this is the relation that
holds between any two a and b in S such that a and b are in the same cell in the
partition.

1.5.4. Functions

A function takes every member of some set A and assigns it a value from a set B
(B could be the same set as A, but need not be). This can also be formalized using
the notion of a set of ordered pairs. Thus, consider two sets A and B (which again
could be the same but need not be). Then, a (total) function from A to B is any set
of ordered pairs (i.e., any subset of A x B) such that for each a in A, there is one and
only one ordered pair with a as first member. Thus if we think of the function f as
assigning to each a in A some values in B, note that the criterion above ensures
that each member of A is indeed assigned a value, and is assigned a unique value.
A is referred to as the domain of the function, and B is referred to as the co-
domain. For any function f and any a in the domain of f, we write f(a) to indicate
the value that f assigns to a. (To use other common terminology, f(a) means the
result that one gets by applying the function f to a.) There is no restriction that each
member of B must appear as second member of some ordered pair; the term
range of the function f is the set of all b in B such that there is some a such that
f(a) = b. Note that these definitions are such that the range of a function is a subset
of the co-domain. In practice (at least in works within linguistics) the terms
“range” and “co-domain” are often not distinguished.
As noted above, there is no restriction that each member of B appear as second
member of an ordered pair. Nor is there a restriction that it appear only once. If
each member of B is used as a value only once (that is, for each b in B, there is a
unique a such that f(a) = b) then B obviously can be no smaller than a. It can have
more members, or it can be the same size. If the latter is the case, then it also
follows that for every b in B, there is some a such that f(a) = b. When both
conditions above hold (i.e., for each b in B, there is one and only one a such that
 1.5. APPENDIX: SETS AND FUNCTIONS 25

f(a) = b, we say that there is a one-toone correspondence between A and B. Note

that for any function f which is a one-to-one correspondence, there is a
corresponding function f-1 which is just the reverse: it is a function mapping each
member of B to a member of A such that for all a in A and b in B, if f(a) = b then f -
1
(b) = a.10
We will have some occasion to talk about the notion of a partial function.
A partial function is one where not every member of A is actually assigned a value
by f; f is undefined for some subset of A. (Of course any partial function f is also a
total function with a smaller domain.) We can illustrate this by returning to our
earlier example of ordered pairs of US cities and states, where the first member
of each ordered pair is the capital of the second. This is a partial function from the
set of US cities to states (not every
US city is a capital). We can reverse it, and have each state as the first member of
the ordered pair and the second as its capital (this function could be expressed in
prose as has as its capital). This is now a total function from

10 Incidentally, the notion of the availability of a one-to-one correspondence can be used

to define what it means for two sets to have the same cardinality. Obviously for two finite
sets it is clear what it means to have the same cardinality, since we can count the
members. But consider the case of infinite sets. Take the following two sets: A = the set of
positive integers {1,2,3, . . . } and B = the set of positive even integers {2,4,6, . . . }. Both are
infinite. Surprisingly (when one first hears this) they are also of the same cardinality,
because one can establish a one-to-one correspondence between them (each member of
A is paired with a member of B by multiplying by 2: we will never run out of members in B).
the set of states (every state does have a capital) to the set of US cities. But it is
not a one-to-one correspondence for the same reason that our original relation is
not a total function; there are many cities without the honor of being a capital.
Occasionally in this text it will be useful to list out some actual functions—that
is, to name every member in the domain and name what the function at issue
maps that member to. There are a variety of ways one could do this. To illustrate,
take a domain of four children {Zacky, Yonnie, Shelley, and Baba} (call that set C)
and four men {Abe, Bert, Carl, David} (call that set M). Suppose there is a function
f from C to M which maps each child to their father. Assume that Abe is the father
of Zacky and Yonnie, Bert is the father of Shelley, and David is the father of Baba.
Then one can write this information out in various ways. One would be to simply
give the set of ordered pairs: {(Zacky, Abe), (Yonnie, Abe), (Shelley, Bert), (Baba,
David)}. Usually this notation, however, is not terribly easy to read. We could also
write this out in either of the ways shown in (20):
 26 1. INTRODUCTION

(20) a. f(Zacky) = Abe b. Zacky ! Abe

f(Yonnie) = Abe Yonnie ! Abe

f(Shelley) = Bert Shelley ! Bert

f(Baba) = David Baba ! David

Or, sometimes it is more convenient to list out the domain on the left and the co-
domain on the right and connect them with arrows as in (21):
Zacky Abe

Yonnie Bert

Shelley Carl

Baba David

Which notation is chosen makes no difference; the choice should be dictated by

clarity.
 2
Semantic foundations
2.1. Model-theoretic semantics 27 2.3. Possible worlds 31
2.2. Truth conditions 28 2.4. Times 41
2.1. Model-theoretic semantics

The primary focus of this book is the syntax/semantics interface—that is, how the
syntax and semantics work so that sentences (and other wellformed linguistic
expressions) are paired with a meaning. But of course this task is impossible without
some idea of what meaning is. Can we talk about meaning without relegating it to the
realm of the mysterious, or leaving it solely to folks who work on cognition to deal
with? The answer (of course) is yes—there is a rich tradition within linguistics and the
philosophy of language for modeling linguistic meaning.
In some early work within linguistic theory—especially in the 1960s— meaning
was taken to be just a symbolic representation (call it a Logical Form, or LF). While it
remains an open question as to whether such representations play an actual role in
the way the grammar pairs expressions with meanings, this book (along with much
other modern work in formal semantics) assumes that meaning is not just some string
of symbols, but rather some actual object out there in the world. Call this a model-
theoretic object. (More precisely, we are taking meaning to be an object which forms
part of a model which is an abstract representation of the world: hence the term
model theory.) Of course, we need some way to name these objects, and so
throughout we will use strings of symbols as ways to name them. But the point is that
the grammar maps each linguistic expression into something beyond just a symbolic
representation. Otherwise—as so aptly pointed out by David Lewis (1970)—we are
simply mapping one language (say, English) into another (what Lewis termed
“Markerese”). Yet language is used to convey facts about the world; we draw
inferences about the world from what we hear and we gain information about what is
true and what is not. So semantics must be a system mapping a linguistic expression
to something in the world.
But what exactly is meant by model-theoretic objects? These can in fact be quite
abstract. Still, they are the “stuff” that is out there in the universe— something
 28 2. SEMANTIC FOUNDATIONS
constructed out of actual bits of the universe (or, at least, the ontology of the universe
as given by language). This would include things like individuals, times, possibilities,
and perhaps others; just what are the basic objects that we need is an open question
and is part of what semantic theory addresses. The strategy here will be to use a fairly
sparse set of primitive objects, and construct more complex objects out of these. Let
us, then, begin by setting up two basic building blocks which are foundational in much
of the work in linguistic formal semantics.

2.2. Truth conditions

A common adage in semantics is: “To know the meaning of a (declarative) sentence
is to know what it would take to make it true.” We can use this adage as a first step in
constructing the building blocks for meanings: a fundamental fact about declarative
sentences is that they are either true or false 1 (and since we use language to
communicate information about the world, a listener will in general assume that a
sentence they have just heard

1
Henceforth we use the term “sentence” to mean a declarative sentence. There is actually
no reason to consider questions to be of the same category as declarative sentences even
though they also are traditionally referred to as “sentences.” Questions have a different external
distribution from declarative sentences (for example, wonder can occur only with a question
as its complement, not an ordinary sentence, while the reverse is true for believe); they have a
different kind of meaning, and they have a different internal structure. Whether imperatives and
declaratives should be considered the same category is a bit less clear (they are more similar),
but we will not deal with those here either.
2.2. TRUTH CONDITIONS

is true, and uses that fact to enrich their knowledge of the world). Thus (1) is true and
(2) is false:
(1) Barack Obama moved into the White House on Jan. 20, 2009.

(2) John McCain moved into the White House on Jan. 20, 2009.

Hence, one basic notion used for the construction of meanings is a truth value—for
now assume that there are just two such values: true and false. (More on this directly.)
The claim that truth values are a fundamental part of meaning is also motivated by
 29
noting that—as shown by the examples above—speakers have intuitions about truth,
given certain facts about the world, just like they do about acceptability. And these
judgments can be used to test the adequacy of particular theories of meaning.
Following standard practice, we use 1 for true and 0 for false. Thus the set of truth
values {1,0} and we will also refer to this set as t. Let us use [[Æ]] to mean the
semantic value (i.e., the meaning) of a linguistic expression Æ. Then (temporarily) we
can say that [[Barack Obama moved into the White House on Jan. 20, 2009]] = 1.
Some worries should immediately spring to mind. The most obvious is that
something seems amiss in calling the meaning of (1) “true” even if we are willing to
accept the fact that it is true. We will enrich the toolbox directly to take care of that.
But there are other objections: does it really make sense to say that all declarative
sentences are true or false? Clearly not—for some sentences the truth value depends
on who is speaking (and on when the sentence is spoken). Take (3):
(3) I am President of the United States.

This is true if spoken by Barack Obama in 2011, but not if spoken by John McCain and
not true if spoken by Barack Obama in 2006. So this has no truth value in and of itself.
Nonetheless once certain parameters are fixed (time of utterance and speaker) it is
either true or false. So we might want to think of the meaning of (3) as a function into
{1,0}—it does yield a truth value but only once we fix certain parameters. But it seems
inescapable that a declarative sentence is telling us something about the world, and
so truth values are certainly one fundamental piece.
In fact, there are many parameters that need to be set in order to derive a truth
value. Certain words like I, you, here, now, etc. quite obviously have the property that
their value depends on when, where, and by whom these are spoken (these are called
indexicals). There are also more subtle cases—such as (4) and (5)—where truth
seems to depend on what is at issue in the discourse context in which these are
uttered:
(4) Magic Johnson is tall.

(5) Every student got an A on their formal semantics midterm.

(4) might be true if we are comparing Magic Johnson to the general population (Magic
Johnson is a former basketball player for the Los Angeles Lakers) but perhaps not true
if we are comparing him to the set of basketball players. The context of utterance
usually makes it clear what is the relevant comparison class. (5) may be true if we are
restricting the interpretation of students to those students in my formal semantics
class in 2011, but not if we mean every student in the world, or even every student at
 30 2. SEMANTIC FOUNDATIONS
Brown or every student who has ever taken formal semantics from me. We return to
cases like these in (among other places) sections 8.4, 10.6, and 17.4.

2.1. In what ways do the following sentences also show a dependency on context
of utterance? (Some may show this in more than one way.)

(i) The dog did not get a walk today.

(ii) Samantha hasn’t eaten yet.
(iii) All of the ex-presidents of the US gathered at the State of the Union address.

So far, we have two reasons that it is oversimplified to say that the meaning of a
sentence is true or false: (a) even once we do determine the truth value of a sentence
we surely don’t want to call that “meaning,” and (b) often the truth value can’t be
determined until we know the context of utterance. Two further worries have to do
with the fact that (c) there are vague sentences where some have the intuition that
the truth value is something in between 1 and 0, and (d) some sentences (even once
we fix the context of utterance) seem to be neither true nor false (nor anything in
between).
As to (d), these are cases which have generally gone under the rubric of
presupposition failure. The classic example from Russell (1905) is (6); another
example would be (7) said of someone who never went to an aerobics class:
 2.3. POSSIBLE WORLDS 31

(6) The present King of France is bald.

(7) He stopped going to aerobics class.

Surely (6) is not true, but do we want to say it is false?2 For the most part, we will
make the expository simplification of assuming all sentences are true or false,
although we will return to the issue of presupposition from time to time.
This still leaves the issue of vagueness. For example, even once we do fix the
comparison class at issue in a sentence like (4) there remains some vagueness:
is a 6 foot 9 inch basketball player tall for a basketball player? If so, exactly where
does one draw the line between tall and not tall? There is a rich literature on the
problem of modeling vagueness and some have attempted to model it using
intermediate truth values (where a sentence can have any value between 1 and
0). Like many other domains, we can set this aside in this introductory text (but for
a nice discussion aimed at introductory students, see Chierchia and McConnell-
Ginet 1990: chapter 8). For the present purposes, take all declarative sentences
(once fixed to a context) to have just one of two truth values: 1 or 0.

2.3. Possible worlds

2.3.1. Introducing the notion

The one important observation that we will not set aside is the first problem noted
above: it hardly makes sense to think of the meaning of a sentence like (1) as true.
There are plenty of reasons to find this absurd. First, we can know the meaning of
a sentence without having the slightest idea of whether it is true or false: I know
the meaning of (8) but have no idea of its truth value:
(8) The tallest man alive anywhere on January 1, 2010 had pomegranate juice for
breakfast.

2
Russell (1905) answered this question “yes”; under his analysis (6) is false.
Strawson (1950) argued “no” and introduced the notion of presupposition by which a
sentence like (6) is neither true nor false.
And surely we do not want all true sentences to have the same meaning (ditto for
all false sentences). In fact if this were all there is to meaning we would be able to
 32 2. SEMANTIC FOUNDATIONS

express only two things: “1” or “0.” Obviously a rather strange notion of what we
do with language! But let us return to the adage: “To know the meaning of a
sentence is to know what it would take to make the sentence true.” Thus enters
the notion of a set of possibilities, otherwise known as a set of possible worlds. 3
This gives us just the tool we need: to know the meaning of a sentence is to know
what the world would have to be like to make it true: i.e. the meaning of a
declarative sentence, then, is a function from ways the world could be (possible
worlds) to {1,0}. (Readers not familiar with the notion of a function should consult
the appendix to Chapter 1.)
This makes good sense. We can know the meaning of a sentence without
knowing its truth value not because of our lack of linguistic knowledge but
precisely because we don’t know what is the actual world in which we live. But if
we did—if we were omniscient—then for every sentence we would of course
know whether it is true or false. Moreover, think of what it means for one’s
information to grow as we communicate. As we hear a sentence (and if we also
assume that the speaker is speaking truthfully) we narrow down the space of
possible worlds that we could be in; all worlds mapped to 0 by the relevant
sentence are eliminated. This also provides a nice way to capture the notion of an
entailment. (9), for example, entails (10) (Mitka is the name of the author’s former
dog):
(9) Mitka killed the bird that had been trapped on the porch.

(10) The bird that had been trapped on the porch died.

Hence, we can say that a sentence S1 entails S2 if and only if every world mapped

to true by S1 is mapped to true by S2. Put differently (but equivalently), the set of

worlds in which (9) is true is a subset of worlds in which 3

The terminology of “possible worlds” often causes people to cringe—it brings with it
some unfortunate terminological baggage. While there may be deep philosophical
questions that ultimately need to be addressed here, one can go quite far without worrying
too much about the real status of “alternate possible worlds.” Surely we do talk about other
ways the world might be (see the discussion below). And surely there are many facts about
the actual world that we do not know, hence think of a possible world as a way the world
might be compatible with our knowledge state.
 2.3. POSSIBLE WORLDS 33

(10) is true. (The reverse of course does not hold; (10) can be true because a
different dog killed the unfortunate bird, or because it died of a broken
heart.) Similarly, we can call two sentences synonymous if and only if they
entail each other—i.e., they have exactly the same truth conditions.4
As soon as one begins to talk about possible worlds and truth conditions, an
interesting question arises. Consider a sentence that is necessarily true— that is,
one which seems to be true no matter how the world happens to be and hence
true in all possible worlds. If the meaning of a sentence is a function from worlds
to values, then all necessarily true sentences have the same meaning. But do we
want to say that all necessarily true sentences really have the same meaning?
(Similarly for all sentences that are necessarily false.) These are thorny issues. But
some of the discomfort posed by modeling the meaning of a sentence as a
function from possible worlds to the set t can perhaps be allayed by taking a rather
charitable view of what it means to be a possible world. Consider the (often cited)
case of mathematical facts. Such facts are what they are—they are not really
contingent facts about the world (the world could not change in such a way as to
make them false). Yet there is a strong intuition that (11a) and (11b) do not have
the same meaning:

(11) a. The largest prime number less than 1000 is 997.

b. The largest prime number less than 2000 is 1999.

Many find this an unsatisfying definition of “synonymy” since two sentences can have
4

the same truth conditions but a variety of subtle differences such as differences in when
they are appropriately used, what they draw attention to, etc. Consider active–passive
pairs:
(i) a. Mitka killed the bird that had been trapped on the porch.
b. The bird that had been trapped on the porch was killed by Mitka.
One can imagine a variety of situations in which it would be appropriate to utter (ia). For
example, if I come home at the end of the day and ask my husband what Mitka did during
the day, (ia) would be a reasonable answer. It is also a reasonable answer to the question
of why the porch appears to be full of feathers. But while (ib) is reasonable in this second
scenario, it is an odd answer to the question of what Mitka did during the day. These kinds
of differences often go under the rubric of information structure. Ultimately, then, one might
well want a much finer-grained notion of synonymy that includes not only the truth
conditions of a sentence but other factors as well. But this is simply a matter of
terminology: here we elect to use the term synonymy in this rather coarse-grained way
(referring to having the same truth conditions). One could substitute in “mutual entailment”
if one prefers.
 34 2. SEMANTIC FOUNDATIONS

Arguably, though, this is because the set of possible worlds includes more than
just the mathematically possible worlds; rather, it includes all worlds which we
can conceive of. Until writing out (11a) and consulting a number theorist I had no
idea whether it was true or not: surely the space of epistemically possible worlds
(worlds compatible with my knowledge state) includes worlds in which (11a) is
true and worlds in which it is false. Similarly for (11b). Indeed, the truths of
mathematical statements are subject to discovery (and ditto for truths that hold
via physical laws of the universe) and so there’s no real reason to restrict the
notion of possible worlds to those in which those laws hold. (If one prefers, we
should perhaps rename these as imaginable worlds—but we stick here to the
term possible worlds out of convention.) Thus at least some of these concerns
can be allayed in this way.
This still leaves the case of sentences known as tautologies (true by logic
and/or by definition) such as the pair in (12). No amount of imagination would
make these false (someone who thinks they are false presumably just does not
know the meaning of or) and so the theory says they have the same meaning:
(12) a. Henry had breakfast yesterday or Henry didn’t have breakfast yesterday.
b. Sally climbed Mt Washington or Sally didn’t climb Mt Washington.

We will simply agree to live with this problem (if it is really is one). Perhaps literally
these two do have the same meaning. As will be discussed at points throughout
this text, there is a distinction between the literal meaning of sentences and other
things that they might be used to convey (the latter being the domain of
pragmatics). It is not unreasonable to think that the pair in (12) have the same
meaning but could be used for different purposes. A full-blown theory of
pragmatics should be able to predict that while both sentences are always true,
they can indirectly convey different information and hence be used in different
circumstances. So the notion that the meaning of a sentence is a function from
possible worlds (broadly construed) to truth values does not seem fundamentally
threatened by these cases.
Incidentally given the meaning of a sentence as a function from worlds to the
set {1,0}, the notion of a partial function can be used to model presupposition
failure. Hence a common account of the fact that The present King of France is
bald seems somehow just nonsensical in our world is to suppose that it assigns
neither true nor false to this world. The meaning of this sentence is thus a function
from worlds (and times) to truth values, but a partial one—undefined for this
world. Whether this is the right way to think about presupposition is a topic that
 2.3. POSSIBLE WORLDS 35

we leave open in this work, although we will be confronting the issue(s) from time
to time.

2.3.2. Characteristic function of a set

Take any function f from some domain D into the set t (i.e., the set {1,0}). Then
consider that subset of D consisting of all and only the members of D that are
mapped to 1 by f; call it D'. Then we can say that D' is the set characterized by f.
(Note that D' can be .) Similarly, if we have a set D' which is a subset of D, then we
can construct the function f that maps every member of D' to 1 and every member
of D not in D' to 0. f is called the characteristic function of the set D'. Note that for
any function f into {1,0} we can recover the set D'. Note moreover that for any D' a
subset of D and any D, we can uniquely recover the total function f that maps all
members of D' to 1 and everything else to 0. In this way we can go back and forth
between talking about certain functions (those whose domain is {1,0}) and the
sets that they characterize.
This is useful in thinking about the meaning of sentences. We have taken the
meaning of a sentence to be a function from worlds to {1,0} but one can
equivalently take it to be a set of possible worlds (namely, all worlds mapped to 1
by the relevant function). It is useful to be able to go back and forth between these
two conceptions. Note that the function is uniquely recoverable from the set only
under the assumption that every sentence is indeed assigned 1 or 0 in every world.
Thus if we treat presupposition failure as best modeled by allowing partial
functions—a sentence which is true in some worlds, false in others, and assigned
no value in yet other worlds—then there is no way to uniquely recover the relevant
function from the set of worlds mapped to true. (Those not in the set might have
been mapped to 0 or might have been given no value.) However, as long as the
function is total (thus all worlds are assigned 1 or 0 by the relevant function) it can
be uniquely recovered. The reverse holds in either case: one can always construct
the set characterized by a function f whose co-domain is {1,0}: that set is the set
assigned 1.
2.2. Consider the function f whose domain is the set of New England states and
co-domain is {1,0} which is as follows: f(Maine) = 0, f(New Hampshire) = 1,
f(Vermont) = 1, f(Massachusetts) = 1, f(Rhode Island) = 0, f(Connecticut) =1.
Show the set it characterizes.
 36 2. SEMANTIC FOUNDATIONS

2.3. Now consider the case of partial functions from the set of New England
states to {1,0}. That is, we are considering functions which map each state into
either 1, 0, or where the function can be undefined for some states (some
states will be mapped into neither value). Consider the set {Maine, Vermont,
Rhode Island}. List out all of the possible functions (partial or total) which
characterize this set. (For readers unfamiliar with American geography, the
remaining New England states are as given above—they are New Hampshire,
Massachusetts, and Connecticut.) See Appendix to Chapter 1 for ways to write
out functions.

2.3.3. Notation and terminology

Recall that t is the set {1,0} (i.e., the set of truth values). We will use w to mean the
set of worlds. We will use the notation <a,b> to mean the set of all functions
whose domain is the set a and co-domain is the set b.5 Thus <w,t> is the set of
functions from worlds to truth values. We will speak of the semantic type of a
syntactic category such as a “sentence” (S) as being <w,t>; this means that every
expression of this category has as its meaning some member of <w,t>. (Thus
<a,b> itself means the set of all functions from a to b; the terminology “of type
<a,b>” means that an expression has as its meaning a member of <a,b>.)
Take any expression whose meaning is a function from possible worlds to
something else. (So far we have seen this just for the case of sentences.) We call
that function the intension of the expression. Thus the intension of an expression
is its actual meaning, not its value in some particular world.

5
This is very slightly different from the way this notation is used in some other works in
semantics (in particular this differs slightly from its use in Montague 1973 which is how it
was introduced into linguistic semantics). However, the differences are inconsequential
for the purposes here.
When talking about the value of some expression Æ at some particular world, we
will use the term extension of a at a world w. Most often we are concerned with
its value in the actual world, and in such cases we will simply refer to the
extension of the expression.
 2.3. POSSIBLE WORLDS 37

2.3.4. Talking about worlds

Possible worlds have been added to the basic toolbox as a way to capture the fact
that the meaning of a sentence is not simply 1 or 0, but rather can be thought of
as a function from worlds to truth values. But there is a second reason for
incorporating possible worlds into the theory, which is that natural language
contains all sorts of expressions and constructions which actually make
reference to worlds other than the actual one. Put differently, there are words
whose meanings crucially combine with the intensions of sentences (or other
expressions).
A rather striking example of how language allows us to talk about possible
worlds is the case of counterfactual conditionals. These are sentences of the form
If S1 then S2, where there are special forms of the verbs (or auxiliaries) in these
sentences (in particular, the then- clause often contains the auxiliary would).
Consider (13):
(13) If the US Supreme Court had not stopped the Florida recount in 2000, nuclear power
plants would no longer be in use in the US.

Some relevant background: the outcome of the 2000 US Presidential election

(between Al Gore and George W. Bush) hinged on a few hundred votes in Florida.
Although Bush was initially declared the winner of Florida (and hence of the US
election), Gore asked for a recount of the votes in key precincts. There was a
complex series of court decisions, leading up to the Florida Supreme Court ruling
in favor of Gore’s request. Bush then took this to the US Supreme Court. While the
recount was in process (and hence no one knows for sure what the outcome
would have been), the US Supreme Court halted the recount, and subsequently
ruled that it should not proceed. Bush therefore remained the victor in Florida and
hence won the US presidency. An additional bit of relevant background is that Al
Gore is very opposed to the use of nuclear power.
Given this, it is easy to understand (13), and to understand the reasoning
leading a speaker to say this. What is interesting about such sentences is that we
can argue about whether they are true or false but we can never really know, as
we are arguing about a chain of events in another possible world. A reasonable
approximation of the truth conditions of a counterfactual conditional sentence of
the form If S1 then S2 .is that it is true if in the closest world to ours where S1 is true
S2 is also true. “Closest” means the world just like ours where we make minimal
changes. In other words, we look at a world in which the if-clause (known as the
 38 2. SEMANTIC FOUNDATIONS

antecedent) is true and where nothing else changes except those facts that follow
from the truth of the antecedent. (Usually these are uttered in a situation where
the antecedent clause is not true in the real world and hence the closest world is
not our world, but this is not necessary; the closest world in question could be the
actual one.6) Of course in those cases where the relevant other world is not ours,
we are not omniscient; we can’t just jump over to another possible world. And this
is exactly why we argue about the truth of counterfactual conditionals: we don’t
really know what the facts are in our relevant alternate reality. In fact, (13) is
interesting because we don’t even know for sure whether Gore would have won
the election had the Supreme Court not stopped the recount. People argue about
this (different newspapers tried to do counts, and came up with different results),
so we actually don’t even know the truth value of a simpler sentence like If the
Supreme Court had not stopped the Florida recount, Gore would have become
President. But we do know what these sentences all mean: we know what it would
take to make them true.
Notice, incidentally, that we have restricted the semantics to looking only at the
closest world in which the antecedent is true (the world most like ours where we
change as little as possible). It would be incorrect to posit, for example, that the
consequent must be true in all worlds in which the

6 An illustration of the fact that the construction does not require the antecedent to be
false in the actual world is the following “detective” scenario (modified from discussion in
Anderson 1951): A great detective is analyzing the scene of a murder in which there are
bloody footprints of a size 12 shoe, an empty vial of an extremely rare poison available only
in Saskatchewan, and a note clearly scribbled with a magenta felt-tip marker. Our detective
reports his thinking as follows. “Let’s see now. If Mr Magoo had committed this murder he
would have left bloody footprints from a size 12 shoe. If he had committed the murder he
would have purchased his poison in Saskatchewan. And if Mr Magoo had committed the
murder he would have left a note with his favorite magenta felt-tip marker. I therefore
conclude that
Mr Magoo is indeed our murderer.”
antecedent is true. For surely there is a possible world in which the recount
proceeded, Gore did get more votes, was assassinated his second day in office,
and where his successor Lieberman had no qualms about nuclear power. Or for
that matter, one cannot doubt the existence of a possible world (quite remote) in
which the majority of votes in Florida had actually been cast for Patrick Buchanan
or even Mickey Mouse.
 2.3. POSSIBLE WORLDS 39

The interesting point here, then, is that while we cannot really determine the
truth value of (13), it does indeed have truth conditions (otherwise there would be
nothing to argue about), and we can rather precisely state the truth conditions—if
not the truth value—using the notion of possible worlds. Our ignorance, as always,
comes from the fact that we are not omniscient: we do not have access to the full
set of facts in our world nor to the full structure of other possible worlds. Thus in
order to give a semantics for if . . . then, the theory needs access to possible
worlds.
In fact, one does not even need to consider constructions with such exotic
syntax to see that we reason and talk about alternative possibilities in everyday
language. Even the semantics of the verb form called the progressive (as in I am
writing a book) arguably requires access to possible worlds.
(The discussion below—although not the particular example—is based on the
analysis in Dowty 1979.)
Consider the following (silly but wonderfully illustrative) two scenarios.

Scenario 1: Mr Peepers lives alone and spends a lot of time watching the activity
in his street out his window. A turtle who he has named Myrtle lives in a nearby
pond, but every morning Ms Goodheart who lives across the street from Mr
Peepers places food on the curb (across the street from Mr Peepers) for Myrtle.
And so every morning at the same time, Myrtle slowly ambles across the street,
gets to the curb on the other side, eats her food, and slowly ambles back. One
fateful day, Myrtle gets exactly a third of the way across the street, and a large
truck comes and hits her and Myrtle dies.
Mr Peepers is reporting this, and can quite naturally say:
(14) Poor Myrtle the Turtle was crossing the street when she got hit by a truck and died.

Scenario 2: Exactly the same, except that in this case Ms Goodheart always puts
the food right in the middle of the street. Of course this means that every day,
Myrtle goes just halfway across the street, has her breakfast, and then returns.
But the rest is the same, and so on the fateful day, Myrtle gets exactly a third of the
way across the street, and a large truck comes and hits her and Myrtle dies.
Can Mr Peepers report Scenario 2 with (14)? The judgment here is quite robust:
no. Yet—in terms of the event that would be described by (14)—the two scenarios
contain exactly identical events! What is the difference? It is clear that in Scenario
1 we expect that Myrtle would have crossed the street had she not been hit by the
truck. (Notice the use of the counterfactual conditional in reporting this.) In other
 40 2. SEMANTIC FOUNDATIONS

words, the semantics of the progressive is appealing to another possible world in

which the truck had not killed Myrtle. (Dowty calls the relevant other possible
world the “inertial” continuation of our world: the world that would have
materialized had nothing extraordinary intervened.) So even ordinary everyday
constructions like the progressive require reference to possible worlds to fully
flesh out their truth conditions.7

7
One might dispute the claim above that exactly the same events happened in the two
scenarios: in Scenario 1 Myrtle had the intention to go to the other side of the street and in
Scenario 2 she did not. So one might argue that the difference does not come from the
different possible worlds which are the inertial continuations, but rather is located in the
intentions of the subject of a progressive verb.
However, one can repeat the scenario (albeit not as colorfully) with subjects that denote
non-sentient beings. Scenario 1: A bowler named Jack always throws his bowling ball
exactly the same way, and it always lands in the gutter. On the day in question, I watch him
play a game—he throws his ball in the usual way, it goes on the usual path. But just as it is
five inches from the gutter, a very heavy overhead light falls from the ceiling of the bowling
alley and smashes the ball. I can report this as (i):
(i) Jack’s ball was rolling into the gutter when an overhead light fell on it and destroyed it.
Now we modify the scenario much as we did with Myrtle. Scenario 2: Again, Jack always
throws the ball in exactly the same way, and he throws it in exactly the same way as in the
first scenario. However, the bowling alley is also home to a small genie who magically
appears every time that the ball is three inches from the gutter, corrects its course, and Jack
gets a strike every time. (I have resorted to the genie here to make sure that the ball itself is
thrown in exactly the same way in the two scenarios. There is no hidden torque on it in
Scenario 2; the change in outcome is due neither to the ball nor to Jack, and until the genie
appears the events in the two scenarios are exactly identical.) Once again the light falls on
the ball when it is five inches from the gutter. Here again (i) is not a good report of Scenario
2.
 41
2.4. TIMES

2.4. Times

It is not enough to think of the meaning of sentences as functions from ways the world
might be to truth values. Their truth is also time-dependent. As noted earlier, (3) is
true when uttered by Barack Obama in 2010 but not in 2006. Similarly, consider an
ordinary past-tense sentence:
(15) George W. Bush was president of the US for 8 years.

This is true now, but when uttered in 1998 it was not true. So the value of a sentence
is not only a function from worlds but also times to truth values.
The relation between worlds and times is a tricky one. For the present purposes it
will work well to think of the set of times and the set of worlds as completely separate
sets. Notice that we can easily talk about “now” in other possible worlds:
(16) If I weren’t here now, I’d be outside enjoying the beautiful weather.

It seems that here we are importing the “now” time to another possible world—which
is most easily made sense of if we treat the worlds and the times as independent sets.
Thus the full meaning or intension of a sentence is a function from worlds and times
to truth values, and since the two are independent we can think of it as a function from
world–time pairs to truth values. (And its extension is the value of this function taken
at some particular world and some particular time.) We will use i to represent the set
of times. Whenever it is convenient to ignore the worlds but not the times, we will
speak as if the value of a sentence is something in the set <i,t>. When it is convenient
to ignore the times but not the worlds, we will speak of it as having a meaning of type
<w,t>. In reality, though, it is a function from both: and the standard terminology is to
use s to name the set of world–time pairs; thus the semantic type of a sentence is
<s,t>. As is the case with worlds, there are many linguistic expressions whose
meanings crucially combine with time-dependent objects (the past-tense morpheme
is one). We explore this briefly in the final chapter. The strategy in this book will be, for
the most part, to give an extensional semantics (ignoring both worlds and times) as
one can get quite far in considering the syntax/semantics interface by looking only at
constructions in which intensions don’t interact with the syntax. However, the final
chapter is concerned with the fact that some expressions are crucially sensitive to
intensions, and considers the implications of this to further issues concerning the
syntax/semantics interface.
 3 Compositionality, Direct
Compositionality , and the
syntax/semantics interface
3.1. Building a fragment: First steps 43 3.3. Folding in worlds (and times) 48

3.2. Implicatures vs truth conditions 46 3.4. Negation: A first pass 49

As noted in the preceding chapter, the program of semantics often takes as its point
of departure the adage “To know the meaning of a (declarative) sentence is to know
what it would take to make it true.” There is a second foundational adage: “The
meaning of the whole is computed in some predictable way from the meaning of its
parts.” This is also often known as Frege’s principle or the principle of
compositionality. Taken in its broadest sense, this is fairly weak: obviously speakers
of a language know the basic units (words, or more accurately, morphemes) and know
some system to combine these into larger meaningful expressions. Surely at least
most of language has to be compositional, or there would be no way to systematically
combine units into larger meaningful expressions which have never been uttered (or
heard) before.2 In other words, we can see the grammar of a natural language as a
specification of the set of well-formed expressions of the language, combined with
rules which map each well-formed expression
3.1. BUILDING A FRAGMENT: FIRST STEPS

into a meaning (that is, some object built out of our model-theoretic toolbox, which
so far consists only of truth values, worlds, and times).
But while it is generally accepted that the meanings of larger expressions are
computed in some predictable way from the meanings of smaller ones, what is not

2 But there is a question about whether everything is compositional. Idioms are a classic
case of items that at least appear to be non-compositional.
 43
so clear is just what that predictable way is and how complex it is. A very simple and
elegant view on this is as follows. The syntax consists of statements that predict the
existence of well-formed expressions on the basis of the existence of other well-
formed expressions. At the base of this is the lexicon; the list of basic units (for our
purposes, words). The semantics works in tandem with the syntax: each syntactic
rule which predicts the existence of some well-formed expression (as output) is
paired with a semantic rule which gives the meaning of the output expression in terms
of the meaning(s) of the input expressions. This is what we mean by Direct
Compositionality, and is the point of view that will mainly be pursued in this book.

3.1. Building a fragment: First steps

Every expression of a language, including the basic expressions (the words), can be
seen as a triple of <[sound], syntactic category, [[meaning]]>. Borrowing notation
from phonology, we enclose the sound part in square brackets (although we simply
use English orthography rather than phonetic transcription), and the meaning part is
enclosed in double square brackets [[ . . . ]]. A rule is thus something which takes one
or more triples as input and yields a triple as output. A reader who is used to thinking
of the syntax as consisting not of a number of very specific “rules” (such as phrase
structure rules) but a set of much more general rules (or “principles”) such as the X-
bar schemata need not worry: Chapter 6 will give the rules in far more general form. 2
Since we will often be adopting rules which are temporary and will later be revised,
simplified, or generalized, any rule not part of the 2
A certain amount of modern linguistics has eschewed the term “rule” on the grounds that
the grammar hopefully does not consist of a list of a large number of “rules” rather than having
the apparatus stated in much more general format. Nonetheless, most agree that the grammar
contains statements that predict the set of well-formed expressions and assign them a
meaning, and here the term “rule” is being used in the most general sense here to simply mean
these statements. We will
44 3. COMPOSITIONALITY & DIRECT COMPOSITIONALITY

final fragment will be notated as TR (followed by a number). (Rules which do

remain in the ultimate fragment are simply labeled with R.)
Assume that the set of syntactic categories includes S which we use to mean a
sentence.3 We continue here to assume that all Ss have as their truth value 1 or 0
(i.e., none are undefined in some worlds). Moreover, for now let us go back to
ignoring the world (and time) parameters, and give a purely extensional
semantics. Then we begin with the following:
TR-1. If Æ is an expression of the form <[Æ], S, [[Æ]]> and is an expression of the form <[],
S, [[]]>, then there is an expression ª of the form <[ Æ-and-], S, 1 if and only if [[Æ]]
= 1 and [[]] = 1>. (Hereafter we use the standard iff to abbreviate “if and only if.”)

There are a variety of other ways this rule could be written. For example, we
could write the syntax as a (context-free) phrase structure rule and pair this with
a semantic part. This notation was used in much of the literature within
Generalized Phrase Structure Grammar (GPSG) (Gazdar, Klein, Pullum, and Sag
1985):

TR-10. S1 à S2 and S3; [[S1]] = 1 iff [[S2]] = 1 and [[S3]] = 1.

A phrase structure rule by itself, then, is a rule which gives the phonological and
syntactic part of a rule such as TR-1 but not the semantics. When supplemented
with the semantics—as in TR-10—this is just TR-1 in a different format. Note also
that TR-10 uses subscripts in the syntactic part of the rule, but these are not
actually meant as a part of the syntactic category. 3 These are needed just to state
the semantics; this is avoided in the notational system in TR-1.

indeed be moving towards a view of grammar which contains just a few rather general
statements to this effect, but there is no reason not to call these rules.
3
In some current theories of syntax (Minimalism and some of its predecessors) this
category is instead called TP (Tense Phrase) and in slightly earlier versions of that theory it
is called IP (Inflectional Phrase). As with the label NP, we are using the terminology
standard in many other theories and in many other related disciplines; readers familiar with
the terms TP or IP instead can make the appropriate translation.

3 A context-free phrase structure rule specifies the well-formedness of an expression of

some grammatical category X in terms of concatenations of strings of other expressions
in the language. These other expressions are described either listed on a case-by-case
basis (like and above) or are described by their grammatical category. The set of
grammatical categories is finite, and it does not include things like S 1 etc.
 45
3.1. BUILDING A FRAGMENT: FIRST STEPS

To forestall one objection: one should not be misled by the fact that the word
“and” is used in the semantic part to give a meaning for English sentences
containing and. The word “and” in the first part of the triple in TR-1 is the English
word [ænd] where English is the object language—the language whose syntax and
semantics is being modeled. But we need a language in which to describe the
grammar, and we have chosen English for that purpose as well. So here English is
also being used as the metalanguage. The confusion that arises by using English
as the metalanguage to model English as the object language would go away if we
were using English to, e.g., model the grammar of Swahili—or if we wrote the
semantics for English using some sort of purely symbolic notation (whose
meaning had been agreed on). There are other ways to alleviate the potential
confusion; we could just as well have written the semantic part of the rule by
means of what is known as a truth table. This is illustrated below where we rewrite
the semantic part of TR-1 by listing the possible combinations of values for each
of the Æ and followed by the value that this gives for the string [Æ-and-]:
(1)[[Æ]] [[]] [[Æ-and-]]
1 1 1
1 0 0
0 1 0
0 0 0

From now on we use the format in TR-1 as our official notation, although at
times will use phrase structure rule format or other notational systems as
convenient. Incidentally, TR-1 introduces and directly as part of the rule rather
than treating it as an item in the lexicon with a grammatical category. This is known
as treating it syncategorematically. This decision to treat and not as a lexical item
but rather as only something introduced by a rule is not essential (and is almost
certainly incorrect), and will be revised later.
The rule for or is analogous (here we will highlight the lack of need to use the
English word “or” in stating the semantics).

(Of course we could posit such a category, as long as there are a finite number of them, but
that is not the intent in the rule above.) In the rule above, the idea is that the syntactic rule
is just S à S and S. The reason that the subscripts are there in the rule is because the
semantics needs to refer to them. Note that this is just an artifact of this particular
notational system; the rule in TR-1 doesn’t use any subscripts in the syntax.
46 3. COMPOSITIONALITY & DIRECT COMPOSITIONALITY

TR-2. If Æ is an expression of the form <[Æ], S, [[Æ]]> and is an expression of the form <[],
S, [[]]>, then there is an expression ª of the form <[ Æ-or-], S, 0 iff [[Æ]] = 0 and [[]]
= 0 >.

3.2. Implicatures vs truth conditions

Before continuing with the fragment, a digression is in order about the meaning of
or. (More extensive discussion is found in section 10.4.) According to the
semantics in TR-2, the sentence in (2) is true if Carol skydives and doesn’t bungee
jump, if she bungee jumps and doesn’t skydive, and also if she does both.
(2) Carol will go sky dive on her vacation, or she will go bungee jumping on her vacation.

This is what is known as inclusive or: both sentences can be true. But is that result
correct? One might at first think not—there surely is a strong temptation to read
(1) as saying she will do one or the other but not both. On the basis of this it is
tempting to think that English or actually has as its meaning what is known as
exclusive or—either conjunct could be true but not both. If that were correct then
of course TR-2 needs revision.
But we can show that English or is actually inclusive. Although we often take
sentences like (2) to convey that both events will not happen, this is not a part of
its actual truth conditions. Note first that if I say (2) to you and it turns out that
Carol goes completely wild and does both, you can hardly accuse me of lying. And
it is easy to construct cases where there is no exclusive or suggestion and where
things would not make sense if the meaning of or were “one or the other but not
both.” So, take a linguistics department which is very strict about enforcing
prerequisites. There is a course called Phonetics and Phonology II (Phon/Phon II).
The course brochure contains the language in (3):
(3) Students may not enroll in any course unless they have had the prerequisite(s). You
will be allowed to enroll in Phon/Phon II if and only if you have taken Phonology I or
you have taken Phonetics I.

Suppose you are a student in love with all of phonetics and phonology, so of
course you have taken both Phonetics I and Phonology I. You now go to enroll in
Phon/Phon II. Surely you would be shocked if the Registrar told
 47
3.2. IMPLICATURES VS TRUTH CONDITIONS

you could not enroll in this course. Anyone reading this course brochure would
certainly come to the conclusion that someone who had both courses is eligible
for the more advanced one. But if or meant “one or the other but not both” then
that is exactly what (3) would be saying.5
Of course there is still one other possibility: that or is ambiguous and that (2)
uses “exclusive or” and (3) uses “inclusive or.” But if it were ambiguous, then each
of these sentences should be too—it should be possible to read (3) in such a way
that the Registrar would be entitled to block you from enrolling. But (3) has no such
understanding; the Registrar simply could not legitimately come back to you and
say “Oh, sorry you misunderstood. You’ve taken both Phonetics I and Phonology I;
so you do not fit the eligibility requirement for enrollment in Phon/Phon II.” 4

5
This point is amusingly illustrated by the following true story. An Ivy League institution
which will remain nameless has a teaching center as a resource for graduate students to
hone their teaching skills, and gives a certificate to any student who completes a program
that includes a series of workshops and other activities (including a “micro-teaching”
session). In February of 2011, students in the program received the following e-mail:
Attendance in Workshop #5 is MANDATORY if you belong in any of the following categories:
a. You haven’t completed workshop #5, or
b. You haven’t completed the Micro-Teaching requirement, or
c. You haven’t completed workshop #5 AND you haven’t completed the MicroTeaching
requirement.
Of course anyone reading this would be completely amused and wonder why on earth (c)
was included here. Surely, no one who had both not completed the workshop and not
completed the Micro-Teaching requirement would really think that they were exempt. The

4 The argument is not quite complete, because one might still think that or is
ambiguous, but that the reason the exclusive interpretation does not emerge is simply
because it would be strange in this situation. Our knowledge of how prerequisites work
makes this a highly unlikely interpretation in this scenario. But we can tweak the scenario
to make the “exclusive” interpretation one that might make sense, but is still absent.
Suppose that the linguistics department is also very unhappy with students taking courses
that overlap too much with what they already know. And suppose that Phon/Phon II
overlaps in some ways with Phonetics I and in some ways with Phonology II. In other
words, the department is concerned not only about underqualified students, but also
overqualified ones. Yet even in this situation, you would be genuinely outraged if the
Registrar blocked your enrolment in the course.
48 3. COMPOSITIONALITY & DIRECT COMPOSITIONALITY

fact that we react to this e-mail as being a silly instance of bureaucratic language confirms
the point.
Thus while an example like (2) often carries with it the suggestion that only one
of the two conjuncts is true, this is just a suggestion and is what is known as an
implicature rather than being part of the truth conditions. Sometimes the
exclusive implicature is quite natural, as in (2), and sometimes it disappears, as in
(3). In fact, when embedded under if (as in (3)) the exclusive suggestion
associated with or systematically disappears. Ultimately one would hope for a
single, stable meaning, where other principles can predict when the exclusive or
implicature arises and when it doesn’t. Indeed, by assigning it the inclusive
semantics in TR-2 combined with a theory of language use (pragmatics), this can
be done. (Many readers will be familiar with this under the rubric of scalar
implicature.) This is the subject of sections 10.4 and 16.6.

3.3. Folding in worlds (and times)

TR-1 and TR-2 are obviously inadequate in that they are given purely extensionally:
they both ignore the fact that the value of a sentence is not 1 or 0 but a function of
type <w,t> (or really <s,t>). For convenience, this book generally gives the rules in
purely extensional fashion, but this is always just for expository simplification. We
can easily write these rules in a more adequate fashion. Thus let us rewrite TR-1
as what we will call INT-TR-1 (INT here for intensional) as follows (actually this
includes just the world argument and not the time; the full intensional version
would replace “w” with “s” (a world–time pair):
INT-TR-1. If Æ is an expression of the form <[Æ], S, [[Æ]]> and is an expression of the form
<[], S, [[]]>, then there is an expression ª of the form <[Æ-and-], S, for any
world w, [[ª]](w) = 1 iff [[Æ]](w) = 1 and [[]](w) = 1>.

Very often the world argument is given as a superscript, so one can rewrite the
semantic part of this rule as given below:
[[ª]]w = 1 iff [[Æ]]w = 1 and [[]]w = 1

Indeed, if the intent were to keep out both the underqualified and the overqualified student,
the brochure would have to be revised to say “if you have taken Phonology I or you have
taken Phonetics I and you have not taken both.” So even in a scenario designed to bring out
the “exclusive or” interpretation, it is not there.
 49
3.4. NEGATION: A FIRST PASS

One will commonly find this notation in the literature (and this will be used from
time to time in this book). Notice that [[Æ]] w means exactly the same thing as
[[Æ]](w) –the superscript notation has an implicit “for all worlds w” at the
beginning of the whole rule. One should keep in mind that nothing (other than taste
and expository ease) hinges on the choice of notation. The grammar is a system
that proves expressions well-formed and assigns each expression a model-
theoretic interpretation; the notation that one chooses to express the model-
theoretic object being assigned is of no theoretical consequence.

3.1. Rewrite the intensional version of TR-2 (i.e., INT-TR-2) using both of the
notations for dealing with the world argument shown above.
3.2. For any function f that characterizes a set, let us use the notation fS to mean
the set characterized by f. Now recall that we can think of the value of a
sentence as a function of type <w,t> (a function from worlds to truth values)
or—equivalently—as a set of worlds; the set mapped to 1 by the function.
Thinking in set terms, what does and do? (That is, what operation on sets does
it correspond to?) What about or? Using the notation suggested just above,
write out the semantics for INT-TR-1 in set terms. Do the same for INT-TR-2.

3.4. Negation: A first pass

Readers familiar with first-order logic (or simple propositional logic) will recognize
that the syntax and semantics of English sketched so far looks a bit like the syntax
of simple logical languages in which two propositions p and q may be connected
with ^ (“and”) and v (“or”). A third operator that one might be used to seeing from
elementary logic is a negation operator which is prefixed in front of a proposition.
(Thus, in the development of an elementary propositional logic, there is a rule
saying that if p is a well-formed formula then so is ~p. The truth value of ~p is the
reverse of that of p.)
If we try to model English directly using the tools from first-order logic in this
way, we immediately reach a challenge for the Direct Compositional program. For
50 3. COMPOSITIONALITY & DIRECT COMPOSITIONALITY

a sentence like (4b) is the negation of (4a): if the first is true the second is false
and vice versa:
(4) a. Olympia Snowe voted for Bill 1838 (the only time it came up for a vote).
b. Olympia Snowe didn’t vote for Bill 1838 (the only time it came up for a vote).

Yet the syntax of English is completely different from the syntax of propositional
logic. Not is not prefixed at the beginning of a sentence. It is not even prefixed at
the beginning of a verb phrase. It generally shows up as a clitic on an auxiliary (as
in didn’t above, where it is a suffix attached to did.) Semantically, though, it seems
reasonable to think of it as negating the entire sentence in which it sits.
There are two choices for resolving this apparent mismatch between the syntax
and the semantics. One is to reject the Direct Compositional hypothesis. A
common solution here is to imagine that actual English syntax is mapped into
another level (called Logical Form, or LF) and that the grammar proceeds
containing a rule or series of rules that map (4) into something like (5) (or perhaps
something even more abstract), and that the compositional semantics interprets
(5):
(5) not [Olympia Snowe voted for Bill 1838 (the only time it came up for a vote)]

Then not here can be given the interpretation exactly like the symbol ~ is in logic.
The second possibility is to conclude that it is a mistake to try to copy logic in giving
a meaning for English n’t—its meaning might have something to do with “~” but
perhaps it is a more complex packaging of negation. This is the strategy that we
will ultimately aim for; it allows for a much simpler conception of the grammar.
Temporarily, though, we take a third tack: which is to postpone incorporating
into our fragment the natural way to do negation in English, and instead pretend
that English has a single word it-is-not-the-case-that which can be prefixed to a
sentence to negate it. Hence we enrich our fragment with TR-3 and its intensional
counterpart INT-TR-3:
TR-3. If Æ is an expression of the form <[Æ], S, [[Æ]]> then there is an expression of the
form <[it-is-not-the-case-that-Æ], S, 1 iff [[Æ]] = 0>.
3.4. NEGATION: A FIRST PASS

INT-TR-3. If Æ is an expression of the form <[Æ], S, [[Æ]]> then there is an expression

of the form <[it-is-not-the-case-that-Æ], S, [[]] is a function from worlds to
truth values such that for any w, [[]](w) = 1 iff [[Æ]](w) = 0>.
(or, using the shorthand, the semantic part would read [[]] w = 1 iff [[Æ]]w = 0).
 51
For truth in advertising, while the final fragment will be closer to actual English, it
will still not be a complete account of English n’t, simply because space
precludes an account of the English auxiliary system. But TR-3 will be improved
upon.

3.3. As in Exercise 3.2, rewrite this in set terms. What is the operation on sets
that corresponds to negation?
 4 Expanding the fragment:
Syntactic categories and semantic
types
4.1. Noun phrases 52 4.4. Back to the compositional 62
semantics
4.2. Intransitive verbs/verb phrases
58
63
4.3. A brief look at lexical semantics 4.5. Illustrating the syntactic/
semantic composition 60

4.1. Noun phrases

English has a variety of expressions like the woman who is jogging down
Blackstone Boulevard, the most disobedient puppy in Alexandra’s puppy class,
the house on the southeast of Chace Ave. and Hope St., etc. We will take it as
given that these are all of the same grammatical category and will refer to them as
NPs (see Chapter 1, n. 8). Let us assume that all expressions of the same
syntactic category correspond to the same type (very broadly speaking) of
semantic object. This assumption was already made for the case of S; we took
each S to have a meaning of type <s,t>, i.e., a function from world–time pairs to
truth values. Recall that our basic semantic toolbox so far consists of the set of
truth values, worlds, and times. But none of these are obvious candidates for NP
meanings. What, then, is the type of thing that NPs have as their semantic values?
For the NPs illustrated above the answer seems simple: add to the toolbox a
set of individuals, and assume that each NP picks out an individual. The notation
e will denote the set of individuals; hence the value of an NP is some member of
e. Of course, one immediately observes that this has a problem analogous to the
problem of saying that the value of a sentence is a truth value. After all, the
individual picked out by the most disobedient puppy in Alexandra’s puppy class is
entirely contingent on how the world happens to be: it could have been the puppy
Mitka but it could instead have been the puppy Brunhilde. Besides, this can also
depend on when this is uttered—the referent of the puppy-NP will change from
year to year.
 4.1. NOUN PHRASES 53
Moreover, one can know the meaning of an NP without knowing who it picks out: I
know what it means to say the tallest man in the world on Jan. 1, 2011 but I have
no idea who this is. So the solution is also analogous to the situation with
sentences: the intension (the actual meaning) of an NP is a function from worlds
and times to individuals—that is, some function in <s,e>. Note that this says that
if one knows the meaning of an NP and knows everything there is to know about
the actual world and time, one can indeed pick out the relevant individual. But we
are not omniscient, so we know the meaning (the intension) without necessarily
knowing the extension. Our toolbox for constructing meanings now consists of the
set of worlds (w), the set of times (i) (we continue to use s for the set of world–
time pairs), the set of truth values (t), and the set of individuals (e). Incidentally,
one might wonder whether the individuals are separate from the worlds: does it
make sense to talk about an individual independent of the world in which it is
located? As for the case of times, we will here treat the set of individuals as
separate from the set of worlds and the evidence for this is quite similar (we can
march an individual from world to world using a counterfactual conditional). We
return to this shortly.
The claim that NPs pick out (in any given world and time) an individual has some
interesting consequences. First, this means we need to think of “individual” rather
broadly; the domain of individuals needs to include quite abstract things like those
things picked out by the NPs the idea that the earth is flat, the desire to leave, etc.
Moreover, in addition to our normal individuals like you, me, and my dog we have
plural individuals (the dogs). We will not deal with these here, save to say that they
can be treated as a special kind of individual (Link 1983). Finally, there is a class
of expressions traditionally called NPs which really do seem to resist being treated
as individuals: these are “NPs” with quantificational determiners like every dog,
no dog, few dogs, and so forth. These will simply be excluded from the fragment
for now. They form the basis of some of the most interesting and rich work in
semantics, and will be the subject of Chapter 10.
Consider now the case of names like Mitka, Barack Obama, etc. Traditional
grammar tells us that these are “nouns” (they are considered “proper nouns” as
opposed to “common nouns” like dog, table, stream, but both are called nouns).
But in fact names (in English) have nothing in common syntactically,
morphologically, or semantically with common nouns. Unlike ordinary common
nouns, they do not take plural morphology. They do not occur with determiners
like the.1 And, as we will see below, while NPs denote individuals, common nouns
surely do not (the hungry dog picks out an individual, but dog does not). On the
other hand, names have exactly the same syntactic distribution as do ordinary
54 4. SYNTACTIC CATEGORIES AND SEMANTIC TYPES

NPs; the name Mitka can occur in all the same syntactic environments as an NP
like the disobedient husky.
Moreover, names and NPs are similar semantically: clearly a name picks out an
individual, just like more a complex NP containing the. Given that names have the
same syntactic distribution and the same semantic type as NPs in general, the
conclusion is obvious: they are NPs. Our fragment will treat these as listed in the
lexicon as NPs. Incidentally, we will also for now take complex NPs such as the
disobedient husky, the first man on the moon, and so forth to just be listed in the
fragment as single items; the internal structure of these is the subject of Chapter
8.
Note that this departs from a traditional view that a category label like NP (noun
phrase) is an appropriate label only for expressions containing more than one
word. Along with this traditional view goes the idea that single words must have
as their category one of what is traditionally called the “basic parts of speech”—
noun, verb, adjective, preposition, determiner (or article), adverb, and perhaps a
few others. (These are the “parts of

1 There are examples where proper nouns occur with determiners, as in (i) and (ii):
(i) The Berrymans are coming over to dinner tonight.
(ii) Every Dmitri in the semantics class might be a bit annoyed at the use of thename Mitka
in so many example sentences.
But we can analyze these as cases where a proper noun shifts its class and is used as an
ordinary common noun. English is very free with shifting items from one category to another
(along with a meaning change). It should, however, be noted that a number of semanticists
have explored a different view of proper names according to which Mitka
instead means something like “the one who is named Mitka.”
speech” that still form the subject of colorful posters decorating elementary
school walls, along with such odd definitions as “a verb is an action word” and “an
adjective describes a pronoun.”) The traditional parts of speech are not entirely
fiction—the notion that at least some of these categories are appropriate labels
for a class of words is based on morphological commonalities. “Nouns,” for
example, can (generally) combine with the suffix /z/ to form plurals; “verbs”
combine with (among others) the suffix -ing, “adjectives” can (generally) combine
with -er to form comparatives, etc. Thus the grammar may well need access to
the notion of a class of “nouns,”
“verbs,” and “adjectives” for the purposes of the morphology. But even granted
that, we have already seen that proper nouns do not show noun-like morphology.
There is no reason then not to call them NPs. The category name is arbitrary; it
could as well be 342, D, or Butternutsquash; a category is just the name for a
 4.1. NOUN PHRASES 55
group of expressions with the same syntactic distribution and semantic type.
(Under the Categorial Grammar syntax developed in Chapter 6 it turns out that
many category labels are not actually arbitrary since the names themselves
encode distributional information. But there will be a set of basic category names,
including NP, which remain arbitrary.) Incidentally, there are plurals (like dogs) and
mass nouns (like water), which can function both as nouns (they can occur with
the determiner the to give an NP) or as NPs as in Dogs make good pets, Water is
plentiful on this island.
When they are NPs they have some similarity to proper names: they name “kinds”
of objects. (For much more detailed discussion, see Carlson 1980 among others.)
Still, there does seem to be one difference between a name such as Barack
Obama and an NP like the president of the US in 2011. As discussed above, the
referent of the latter depends on how the world happens to be. One can imagine
worlds in which things had worked out differently and this NP picked out John
McCain, or more distant worlds in which this picked out Dennis Kucinich. One can
even imagine yet more distant worlds in which it picked out the lead dog of the
winning Iditarod team (just imagine a massive change in the political system, a
revision in the duties of the president, and a bit of dog worship). But the referent
of a proper name like Barack Obama is stable; change the world all you want, let
in a bit of dog worship and a massive change in the political system, and the
referent of this NP remains the same. One can even change some rather
fundamental properties— assume that Barack Obama not only is not president
but happens to have been born in Kenya, moved to France, and became a master
sommelier. The referent remains the same—just many things about him are
changed.2
The account of this proposed by Kripke (1980) is to treat names as rigid
designators. This simply means that a name does indeed denote a function from
worlds to individuals, but the relevant function maps every world to the same
individual. Any function which assigns the same value to everything in its domain
is called a constant function. Note that this hinges on the assumption made
earlier: that the domain of individuals is not worlddependent—the set of
individuals is independent of the worlds. 3 Odd though this may seem at first
glance, we can use the tool of counterfactual conditionals to give plausibility both
to the claim that the same individual can march around from world to world, and
for the claim that proper names are rigid designators. Consider the following
sentences:
(1) If Barack Obama had been born in Kenya, he would not have been eligible to be
president of the US.
56 4. SYNTACTIC CATEGORIES AND SEMANTIC TYPES

(2) If Barack Obama had grown up in France, he would have become a master sommelier.

The semantics in Chapter 2 for counterfactual conditionals leads to the

conclusion that there is some individual—call that individual o—whose

properties can vary from world to world but is nonetheless a single individual.

Moreover, the name Barack Obama picks out o in our world, where he is President

of the US in 2013 and was born in Hawaii, and continues to pick 2

Notice that we can change large facts about a person—even their birthplace—as we
move them from world to world while still considering them the same person. Lest one
think that a person’s birthplace is a necessary fact about them—rather than a contingent
fact—consider the claims of the “Birthers.” These are a group of people who have
continually insisted that Barack Obama was actually born in Kenya (and hence ineligible to
be president of the US). Notice that while this group insists on a world in which the Barack
Obama individual was born somewhere else, they never once disputed that Barack Obama
is Barack Obama. The dispute kept the individual constant, and simply centered on a world
(not the actual one) in which that same individual had a different birthplace from his actual
one (which is Hawaii). 3
One might wonder how could this be, since it is a contingent fact as to who is born,
and therefore who is “actual” in any given world. We can, however, assume that there is an
equivocation here about the word “existence.” Let us imagine some very abstract notion of
“existence” which holds for all individuals who are present in any world. And of course one
can also wonder about the proper account of fictional individuals like Santa Claus. These
worries are all beyond the present scope.
out o in the world in (1) in which o was born in Kenya, and in the world in (2) in
which o grew up in France.
While difficult, one can even change more basic facts about the individuals. My
dog Mitka happens to be a Siberian husky, but I could meaningfully assert (3a) and
even (3b):
(3) a. If Mitka had been a Labrador Retriever, I would have been able to let him off
leash.
b. If Mitka had been a wolf, his fear of thunder would not have let him survive.

Again these sentences make sense if we can take the individual rigidly referred to
by Mitka, locate him in a different world, and change his properties, while still
tracking that it is the same individual. We will thus adopt Kripke’s semantics for
these. This allows us to account for the observations above, while keeping their
semantic type the same as that of other NPs: functions of type <s,e>. They just
happen to be constant functions.
 4.1. NOUN PHRASES 57
*4.1. The verb think takes as its object (or complement) a sentence as in Chris
thought that the earth was flat. This verb is a good example of something which
crucially combines with intensions. The object of
Chris’s thoughts is not just a truth value 1 or 0, but a set of worlds (all worlds
compatible with what he believes). If one wishes to simplify, one can think of
this as simply a single world—for the purposes of this question it makes no
difference.
Notice that there is an interesting ambiguity in a sentence like the following:

(i) Bill thinks that the mayor of Providence will win the marinara sauce
competition.

See if you can detect what the ambiguity is and describe—perfectly

informally—how one might account for the ambiguity. (Do note that this
exercise is not looking for a formal account, as the tools for that are developed
only in Chapter 19. The hope is that the informal intuition might be accessible.)
Keep in mind that phrases like the mayor of Providence are functions from
worlds to individuals; this is crucial. Assume further that in putting the whole
semantics together, some actual world is “plugged in” as the argument of the
function so that this NP ends up denoting an individual, but the individual can
vary from world to world.
Incidentally, this phrase is also time-dependent (it actually denotes a

(cont.)

function of type <s,e>) but this fact can and should be ignored here; keep the
time constant.
Now notice that (ii) does not display the same type of ambiguity:

(ii) Bill thought that Buddy Cianci would win the

marinara sauce competition.

Comment on why this is not surprising. (A historical note: Buddy Cianci is not
the mayor of Providence at the time of writing this book. But he was, and indeed
he did make award-winning marinara sauce.)

Before leaving this, we can also note that some NPs fail to refer to anyone in the
actual world, giving rise to the classic notion of presupposition failure. The time-
honored example (from Russell 1905) is the (present) King of France. As
mentioned earlier, this book contains no in-depth discussion of presupposition,
58 4. SYNTACTIC CATEGORIES AND SEMANTIC TYPES

but we can note that one account is to treat the meaning of such NPs as partial
functions from worlds to individuals, undefined for this world.

4.2. Intransitive verbs/verb phrases

The toolbox so far contains as basic building blocks a set of worlds, a set of times,
a set of truth values, and a set of individuals. While there is controversy as to
whether other primitives (e.g., events) are necessary, one can get surprisingly far
using only these sets and constructing other semantic objects using sets and
functions built out of these. As a first step, consider the case of sentences with
intransitive verbs. First a small bit of housekeeping. Sentences with simple
intransitive verbs in English can contain the verbs in the past tense, as in:
(4) The president coughed.

(5) Mitka howled.

But we are ignoring the semantics of the past tense (as we wish to ignore times),
and so in using examples of this type it should be kept in mind that the semantics
is incomplete. Note, incidentally, that we could not give a
 59
4.2. INTRANSITIVE VERBS/VERB PHRASES

fuller account by recasting into the syntactic form which is called the simple
present, as in (6) and (7), for the semantics of these is even more complex:
(6) The president coughs.

(7) Mitka howls.

Sentences with “eventive” verbs (coughs, runs, howls) in the tense which is called
simple present in English actually do not make a statement about an event holding
true at the moment of utterance. For example, (6) does not entail that the
president is coughing at this moment, but that this happens from time to time with
some regularity. These are called habitual statements. So we will refrain from
using these sentences, and stick to past-tense cases.
We begin with the syntax. We will refer to the category of intransitive verbs as
VI, and will at this point assume that English contains the two phrase structure
rules in (8):

(8) a. S à NP VP b. VP à VI

Of course, a different possibility would be to adopt only the rule in (8a) and simply
think of cough, howl, and so forth as lexical VPs. After all, it was shown above that
single words could be lexical NPs, and it would make sense to take the same
solution here. But in the present case the situation is more complex, for here
transitive verbs (like, kill, etc.), ditransitive verbs (give, put, etc.), and intransitive
verbs (and many other things that are traditionally call ed “verbs”) do have
common properties. They are the class that takes past-tense morphology, the -
ing suffix, etc. In other words, the morphology treats them as a single class, giving
some basis for the traditional notion that these form a single “part of speech.” For
now, then, let us stick with tradition and call these all verbs (using a subscript
feature as above in order to distinguish their syntactic distribution—and, shortly,
their semantic type) which leads us to adopt the two rules in (8). All of this will be
modified in Chapter 6 when we turn to Categorial Grammar.
The simplest semantics to associate with the rule (8b) is one which says that
meaning of the VP is the same as the meaning of the V1. Put in the official notation
then, (8b) can be recast as follows:
TR-4. If Æ is an expression of the form <[Æ], VI, [[Æ]]>, then there is an expression of the
form <[Æ], VP, [[Æ]]>.
 60 4. SYNTACTIC CATEGORIES AND SEMANTIC TYPES

The remaining challenge is to determine the type of meaning for the intransitive
verb (and hence the VP) itself. This can easily be done without introducing any new
primitives into the basic set: let these pick out (in a world and at a time) a set of
individuals.
At first, this seems odd. But first, note that—as with NPs and Ss—there is a
distinction between the intension of an intransitive verb and its extension.
The intension is a function from worlds (and times) to sets; it is only the extension
that is a set. Indeed, if you knew everything about the world (and time) that you
were living in, and you knew the meaning of the word cough, you could determine
the set of coughers. So intransitive verbs have meanings of type <s,<e,t>>. To
make this a bit more intuitive, we digress briefly to talk about word meaning.
Modeling the meaning of intransitive verbs as functions into sets of individuals
gives some nice tools for talking about the relationship between the meanings of
different words that clearly have certain semantic properties in common.

4.3. A brief look at lexical semantics

Thus while this book is primarily about compositional semantics—how the

meanings of smaller expressions (including individual words) combine to give the
meanings of larger expressions—that project itself is obviously intimately tied up
with lexical semantics, i.e., the project of determining the meanings of the basic
units (the words). And the tools developed so far are quite useful for lexical
semantics as well as compositional semantics. Consider, for example, the fact
that (9) entails (10) and that this is clearly a fact about the meanings of dance and
move:
(9) Sabrina danced.

(10) Sabrina moved.

If we stick to the idea that the extension of any intransitive verb is a set of
individuals, there is a simple way to express the relation between these two verbs.
One part of the meaning of [[dance]] is that for every individual in the set it picks
out, that individual is also in the [[move]] set. Put differently, the extension of
[[dance]] at any world is a subset of [[move]] (in that world). (Of course [[dance]]
is richer than just this, so the full lexical entry will say more.)
 61
4.3. A BRIEF LOOK AT LEXICAL SEMANTICS

This also provides tools for the meanings of certain words which are often
thought of as being decomposable into other bits of meanings. To illustrate, we
turn from intransitive verbs to nouns. As will be discussed more fully in Chapter 8,
nouns also plausibly have as their meanings world- and timedependent sets of
individuals. Some noun meanings lend themselves to a combination of other
concepts. A famous example (see, e.g., Katz and Fodor 1963) is bachelor, which
in early work on lexical semantics within generative grammar was treated as being
composed of a few primitive concepts. Oversimplifying somewhat, we can
reconstruct this idea by saying that [[bachelor]] is a function from worlds and
times to sets of individuals such that for any world w and time i [[bachelor]] at w
and i is the set: {x|x ∈ [[male]] (at i and in w) and x ∈ [[adult]] (at i and w) and x 2=
[[married]] (at i and w)}. This recapitulates—in model-theoretic terms—the early
“lexical decomposition” analysis of this in terms of primitive concepts. Of course
the above needs refining, but we have the tools to make the necessary
refinements. For a bachelor is not any unmarried male adult, but one who has
never been married. Thus we can revise the above by requiring that for each x in
the [[bachelor]] set at world w and time i, there is no time i' prior to and including i
at which x ∈ [[married]].

4.2. The situation is even more interesting: possible worlds also appear to be
necessary to give a full treatment of even a simple word like bachelor which is
often trotted out in lexical semantics courses to illustrate how words
decompose into other concepts. Thus most speakers have the intuition that a
Catholic priest is not felicitously referred to as a bachelor, nor is a gay man in a
state which does not recognize same-sex marriage.

(i) How can the use of possible worlds provide a way to capture these
intuitions?
(ii) Having done (i), does this also us to eliminate any other part of the above
definition? Note too that we did not put [[human]] in the definition above
although this might have struck the reader as a missing ingredient. But
given (i), do we need it?

Moral: A speaker’s ability to use perfectly ordinary words like bachelor involves
counterfactual reasoning.
 62 4. SYNTACTIC CATEGORIES AND SEMANTIC TYPES

4.4. Back to the compositional semantics

With this much, we can now formulate the semantics for the rule in (8a).
Recasting into the official notation yields TR-5 (given extensionally):
TR-5. If Æ is an expression of the form <[Æ], NP, [[Æ]]> and is an expression of the form
<[], VP, [[]]>, then there is an expression ª of the form <[Æ-], S, 1 if [[Æ]] [[]], and
0 otherwise>.

Notice that TR-5 as formulated above leaves no room for treating certain kinds of
presupposition failure as a case of a sentence being neither true nor false. In other
words, for the sake of discussion take stop-smoking as a single intransitive verb.
It is sometimes assumed that when predicated of someone who never smoked
the result is neither 1 nor 0 but is simply undefined. The semantics above provides
no way to accommodate for this view; every sentence of the form NP stopped
smoking will be either true or false. We remedy this directly below.

4.3. Give the intensional version of TR-5.

TR-5 is based on treating the meaning of an intransitive verb as (relative to a

world and time) a set of individuals. But as shown in Chapter 2, any set can be
recast as the characteristic function of that set. Thus an intransitive verb like
cough or howl can instead be taken as picking out (in any world and time) a
function from individuals to truth values. Using e to mean the set of all individuals,
this then picks out a function of type <e,t> (so its full meaning is a function of type
<s,<e,t>>). Using functions rather than sets has two potential advantages. First, it
provides a more general tool for modeling meanings, as will be clear later. Second,
it is a richer notion than the set version because potentially there could be partial
functions (in this case, a function defined only for a subset of the domain of
individuals). If these do exist, one can always recover a set of individuals from a
(possibly partial) function of type <e,t>: the relevant set is the set mapped to true
by the function. But one cannot recover the function from the set (if something is
not in the set, it could be that it is assigned 0 by the function, or it could be that it
is simply not in the domain of the function). For the most part this book will make
little use of partial functions, but will leave open the possibility of using it for
presupposition failure. Under this technique
 63
4.5. THE SYNTACTIC/SEMANTIC COMPOSITION

[[stop-smoking]] maps an individual to true (at a time i) if that individual smoked

at a time earlier than i and does not smoke at i; false (at time i) if that individual
smoked at a time earlier than i and also smokes at i, and undefined otherwise.
Having made this revision on the semantic types, we revise TR-5 as follows
(given extensionally here):
TR-50. If there is an expression Æ of the form <[Æ], NP, [[Æ]]> and an expression of the
form <[], VP, [[]]>, then there is an expression of the form <[Æ-], S, [[]]
([[Æ]])>.

If there are partial functions and if [[Æ]] is not in the domain of [[]] then we have
undefinedness (and hence the result is undefined at the world and time in
question); otherwise the VP-function applies to the individual in question and
yields either 1 or 0.

4.4. Give TR-50 in its intensional form.

4.5. Illustrating the syntactic/semantic composition

Let us illustrate the workings of the fragment that has been developed so far (along
with a suitable lexicon). Consider (11).
(11) The disobedient husky escaped and the obedient husky slept.

We assume the lexical items escaped, slept, etc. of category V1 with meanings of
type <s,<e,t>>; and the-disobedient-husky and the-obedient-husky of category
NP with meanings of type <s,e>. Extensionally, a V1 denotes a function in <e,t> and
an NP denotes some member of e. If we actually knew all the facts about the
world in question, we could show out the actual extensions of each of these at
that world. By way of illustration, imagine a “universe” with four individuals
{a,b,c,d}. Further, imagine a world w 1, whereby [[the-disobedient-husky]](w1) = a
and [[the-obedient-husky]] (w1) = c. Assume further that [[escaped]](w1) is the
function mapping a and c to 1 and b and d to 0 (thus it characterizes the set {a,c},
while [[slept]](w1) maps b to 1 and a, c, and d to 0 (thus it characterizes the set
 64 4. SYNTACTIC CATEGORIES AND SEMANTIC TYPES

{b}). We can now use the notation in (12) as one way to illustrate how it is that the
grammar proves (11) as well-formed and puts together its meaning.
In the semantic part, we are showing only the extension of each expression at w 1.
We give each expression as a triple of phonology, category, meaning (actually,
extension). For visual ease we will omit the brackets < . . . > around each triple,
and we will omit the brackets [ . . . ] and [[ . . . ]] around the sound and meaning
parts respectively.
(12) the-disobedient-husky; NP; a
escaped; VI; a à 1, b à 0, c à 1, d à 0
these two lines just show the information that is in the lexicon
escaped; VP; a à 1, b à 0, c à 1, d à 0 (by TR-4) the-disobedient-husky
escaped; S; 1 (by TR-50) the-obedient-husky; NP; c
slept; VI; a à 0, b à 1, c à 0, d à 0 (these two lines again given by the lexicon)
slept; VP; a à 0, b à 1, c à 0, d à 0 (by TR-4)
the-obedient-husky slept; S; 0 (by TR-50)
the-disobedient-husky escaped and the-obedient-husky slept; S; 0 (by TR-1)

There are many other ways one could illustrate this. Since we generally do not
have all the facts of some actual world at our disposal, we could write the
semantic composition more generally. For example, the semantic part of the
fourth line can be rewritten as follows: [[escape]] w([[the-disobedienthusky]]w).
Remember that formulas like this contain a hidden clause at the beginning: so this
says that for all worlds w, the value of this expression is the value [[escaped]] at
w applied to the value of [[the-disobedient-husky]] at w. Thus one can show part
of the composition as in (13):
(13) the-disobedient-husky; NP; [[the-disobedient-husky]] w
escaped; VI; [[escaped]]w escaped; VP; [[escaped]] w (by TR-
4)
the-disobedient-husky escaped; S; [[escaped]] w ([[the-disobedient-husky]]w (by
INT-TR-50)

In fact, in general we will illustrate the semantics this way (since we could not list
the actual function which corresponds to, e.g. [[escaped]]). Among other
difficulties in doing that, we would have to list this out for every possible world.
But it is sometimes useful to pick a world and demonstrate the composition
extensionally as in (12) as a reminder that the semantics is not computing
formulas, but actual model-theoretic objects. Formulas—as in (13)—are just
 65
what we humans need in order to write out those objects. It is also generally not
necessary to put in each rule that licenses the next line;
4.5. THE SYNTACTIC/SEMANTIC COMPOSITION

this is useful only when it serves some clarificational purpose. The literature
contains other methods for visually displaying the syntax and semantics at work.
Often this is shown as an annotated tree. A traditional tree shows just the
phonological part—what precedes what—and the categories of each expression,
but not the meanings). One could then “decorate” each node label with the
semantics.

4.5. Consider the following model of a tiny universe. We will be concerned with
just one world and one time, so we can consider everything extensionally. Our
universe has four individuals {a,b,c,d}. At the world in question, [[howl]] maps
a and c to 1, b and d to 0. [[cough]] maps b to 1 and all other individuals to 0.
[[Balto]] picks out a; [[the hungry dog]] picks out d. Show the full syntactic and
semantic composition for

(i) Balto howled and the hungry dog coughed.

In this case, the task is to show the actual extension of each expression at the
world in question; thus use (12) as the model for showing this.
4.6. To make things interesting, and to give a flavor for intensionality, let us add
in another world. World 1 (which we will label as w l) will be exactly the world
set up in Exercise 4.5. World 2 (labeled w 2) is as follows: [[Balto]] still picks out
a (of course, since it is a rigid designator); [[the hungry dog]] picks out c; [[is-
howling]] maps all four individuals to 0; and [[is-coughing]] maps a, c, and d to
1 and b to 0. Now show the syntactic and semantic composition where at each
step you should show the full function from each of the two worlds to its value,
and show how this is computed step by step.
 5
Transitive verbs : Resolving an
apparent syntax / semantics
mismatch

5.1. First pass: Straw Man’s solution 67 5.4. Back to the drawing board:
Curry’ed functions 76
5.2. Abandoning Direct
Compositionality 70 *5.5. A note on three-place verbs 84 5.3. Hidden S

conjunction? 73

This chapter is somewhat didactic—for a reason. The immediate goal is to

discover the semantic type of transitive verbs, which could be done in a few pages.
But we will instead go down a lengthy and winding path (which will turn out to be
fruitless). This involves setting up a somewhat naive “Straw Man,” whose
hypothesis requires giving up Direct Compositionality and then leads to greater
and greater complications. After Straw Man has taken enough twists and turns,
we abandon that path, and turn to a simple Direct Compositional solution. In one
sense the lengths to which Straw Man goes is a bit silly, for in reality most work in
modern semantics does not take the Straw Man position. The ultimate solution
arrived at in this chapter is fairly standard (even in non-
Direct Compositional theories).5 But this is precisely what makes this domain an
important one. For here the Direct Compositional solution turns out to be

5 An exception to this is work done under the rubric of Neo-Davidsonian event

semantics; see, e.g., Parsons (1990).
 67
5.1. FIRST PASS: STRAW MAN’S SOLUTION

reasonably obvious with some fairly simple tools that allow the semantics to be
set up in such a way as to respect the syntax. Perhaps, then, this basic strategy
extends to other domains; perhaps slightly more subtle tools can be pressed into
service in other cases that appear to challenge Direct Compositionality.

5.1. First pass: Straw Man’s solution

Consider sentences with transitive verbs like those in (1):

(1) a. Juliet loves Romeo.
b. The pig chased the wolf.
c. Lee kissed Sandy.

(Once again t eventive verbs like chase and kiss are put into the past tense to avoid
the habitual reading that emerges with present-tense morphology.) What kind of
meaning should be assigned to kiss, which is often thought of as a relation
between individuals? Straw Man consults an elementary logic book, in which
sentences like (1c) are translated into a formula of first-order predicate logic and
are represented as in (2):2

(2) K(l,s)

Recall that an intransitive verb has as its value a (world/time-dependent) function

of type <e,t>. If asked how to construe (2) as a way to name a model-theoretic
object, a reasonable answer is to take l and s in the formula above to denote
individuals, and to take [[K]] to be a function whose domain is the set of all ordered
pairs of individuals and whose co-domain is the set of truth values. In other words,
(l,s) picks out an ordered pair of two individuals, and [[K]] is a function mapping
such a pair to true or false. In set terms, [[K]] characterizes a set of ordered pairs
of individuals.
Straw Man applies this to the semantic composition of English, and concludes
that English [[kiss]] is like [[K]]: a function from the set of ordered pairs of
individuals to t. (Recall that the set of all ordered pairs whose first member is in
set A and second is in set B is represented as A x B.) Thus the

2
68 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?
While we will show that this is not a good way to represent the semantic composition
of English, the point here is not to criticize the use of such formulas in a logic textbook: most
work using formulae like this does not claim to be modeling the grammar of natural
language. The goal of the logician is quite different (to provide a way to represent valid
reasoning systems).
type of meaning for a transitive verb would be <e x e,t> (or, more accurately, <s,<e
x e,t>>). Incidentally, one need not have taken a logic class to come up with the
hypothesis that transitive verbs take ordered pairs as their arguments. We often
speak of kiss as denoting a (two-place) relation between two individuals. But
recall that the formal notion of a (two-place) relation is a set of ordered pairs (and
of course any set can be recast as the characteristic function of that set, which
gives the type above).3 It is also common parlance to say that the verb applies to
a pair of a “kisser” and “kissee,” or to an “agent” and a “theme,” and this also leads
naturally to the hypothesis that the type for transitive verbs is as above. There is,
in fact, nothing wrong with this informal terminology; the central question of this
chapter is whether the type above is the right way to formally model the informal
intuition.
Let us call the syntactic category of a transitive verb V2. Then the semantic type
assumed above leads us to posit FR-1. (Fictional rules that will quickly be
discarded will be notated as FR.) To minimize notational collision, an ordered pair
whose first member is a and second is b will (non-standardly) be notated as {a,b}
(with boldface curly brackets to distinguish from ordinary curly brackets which are
used for sets). This notation is not ideal, but unfortunately all of the other types of
bracket are used elsewhere in our formalism (the standard representation of an
ordered pair is (a,b), but parentheses are also used to indicate the argument of a
function). FR-1 is given extensionally here; the interested reader can supply the
intensional version:
FR-1 If there is an expression Æ of the form <[Æ], NP, [[Æ]]>, an expression of the form <[],
NP, [[]]> and an expression ª of the form <[ª], V T, [[ª]]> then there is an expression
of the form <[Æ-ª-], S, [[ª]]({[[Æ]], [[]]})>.

Since the meaning of the transitive verb combines with the meanings of its subject
and object simultaneously in the semantics (i.e., by applying to an ordered pair)
the syntax also directly combines the two NPs with the transitive verb.
But wait, says the syntactician. We learn from our introductory syntax texts that
this cannot be the right syntax. The phrase structure rule embodied in FR-1 is S à
NP V2 NP, but we know (or at least we think) that loves Juliet in (1a) is a constituent
and similarly for kissed Sandy in (1c). Moreover, we know (or at least we think) that
these expressions are of exactly
 69
3
A two-place relation means a set of ordered pairs. There are also three-place relations
which are sets of ordered triples, and more generally one can speak of ordered n-tuples.
5.1. FIRST PASS: STRAW MAN’S SOLUTION

the same category as expressions consisting of only single intransitive verbs— i.e.,
they are VPs. It is actually worth occasionally reminding oneself of the arguments
that lead to this conclusion; we mention just one here. Expressions like kissed
Sandy, watered the plants, or fed the pig can conjoin with each other, and can also
conjoin with simple intransitive verbs:4

(3) a. Lee watered the plants and fed the pig.

b. Lee watered the plants and meditated.
c. Lee meditated and fed the pig.

If transitive verbs were introduced in the syntax in the flat way given in FR-1, we
would need three additional syntactic rules for these:

(4) a. S à NP VT NP and VT NP
b. S à NP VT NP and VI
c. S à NP VI and VT NP

But this would still not be enough for these conjunctions can be iterated:
(5) a. Lee fed the pig and chased the wolf and meditated.
b. Lee fed the pig and meditated and watered the plants and studied.

Additional rules would thus be needed, but since one can iterate these
indefinitely, no amount of new rules will be enough. Positing a category VP solves
all of this: simply add a syntactic rule of the form in (6), and assume the two VP
rules in (7):
(6) VP à VP and VP

(7) a. VP à VI
b. VP à VT NP

Note that the remarks above concerning and extend also to or as in (8), and so we
also add the rule in (9):
(8) a. Lee milked the cow or fed the pig.
b. Lee milked the cow or meditated.

(9) VP à VP or VP 4
70 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?
There are actually many versions of generative grammar—including earliest
Transformational Grammar—in which it is assumed that all verbs must have a subject (at
least at some level) in the syntax. This would invalidate the argument below. For fed the pig
in (3a) for example would at some level have a deleted or silent subject and be a full
sentence; and all conjunction would actually involve S conjunction. We will not consider
these here, but see the discussion in section 5.3.
Ultimately, there are many refinements one could make to (6) and (9), but these
formulations are fine for now as we turn back to the semantics.
All is well as far as the syntax is concerned, but since the goal is a theory of
syntax and semantics it is striking to notice that the rules above give only the
syntactic side. What would be the semantics that goes along with (7b)? Well, if
we continue with the Straw Man hypothesis (i.e., [[kiss]] is a function that takes as
its argument some ordered pair) there can be none. The reason is simple: V 2 is not
of the right type to combine with just the one individual denoted by the object NP.
And since there is no semantics for (7b), it also does not make sense to ask about
the semantics associated with (6) and (9); there is none.

5.2. Abandoning Direct Compositionality

Straw Man is stubborn and does not wish to give up on the idea that [[kiss]] is a
function whose domain is e x e. But he has learned his syntax well and is happy
enough with the syntactic rules above. Thus he decides to abandon the
hypothesis of Direct Compositionality. He will leave the syntax of (7) alone, and
adopt a different model of grammar from the one we have been assuming so far.
Before developing this, a word about the role of trees. Probably a bit surprisingly
for a linguist, there has so far not been a single tree drawn in this book. This is
because we have simply stated the rules, and the rules so far make no reference
to trees. Consider a complex sentence like, for example, Mitka howled and Porky
grunted. (Porky made his debut in semantics in Lewis 1970.) One can of course
draw a tree for this, but the tree is just a representation of the steps in a proof of
the well-formedness of the sentence. Suppose that the rules combining
expressions do nothing more than put such expressions next to each other: this is
known as concatenation (this assumption will be subject to greater scrutiny in
5.5). Put differently, the syntactic portion of the rules are just those that could be
written as a context-free phrase structure grammar (see Chapter 3, n. 4). 5 Under
this

5
 71
It is well known that ultimately a theory of grammar will need to include rules beyond
just context-free phrase structure rules (which only concatenate expressions). While the
jury may remain out for the case of English, considerable debate about this point during the
1980s (especially within the theory of Generalized Phrase Structure Grammar) showed
that additional devices beyond context-free phrase structure grammars will be needed for
other languages. But just how much “extra” is needed remains open. See section 5.5 for
discussion of these points.
5.2. ABANDONING DIRECT COMPOSITIONALITY

view, the grammar never needs to refer to trees as it never refers to the internal
structure of some expression. A tree is always just a way to represent the proof of
the well-formedness of some expression, and given the Direct Compositional
hypothesis, it is also a representation of the steps in the semantic composition.
Under Direct Compositionality, then, combined with a rather impoverished view
of the set of syntactic operations, a tree is a convenient representation for the
linguist but it is not something that the grammar itself “sees.”
But of course much work in syntax and semantics assumes that the grammar
does have access to the trees, and the view of semantics to which we are about
to turn also makes this assumption. Thus in order for Straw Man to both adopt (7b)
in the syntax and at the same time maintain that [[kiss]] is a function of type
<exe,t> (or the intensional counterpart), the syntax and semantics must be pulled
apart. The syntax not only proves strings well-formed but keeps track of the steps
used so that in the end it is a proof of the well-formedness of a tree. Moreover, it
“feeds” into the semantics as a separate system: a system which provides
interpretations for trees. Hence we adopt the syntactic rules in (7), and add a
semantic rule interpreting a tree as in FR-2. (The notation here is inspired by the
notation developed in Heim and Kratzer 1998, although they do not use it for this
case as they adopt the same solution for transitive verbs as is developed later in
this chapter.)
FR-2 S

NP 1 VP

VT NP 2

= [[VT]] ( { [[NP1]], [[NP2]] } )

(This is given extensionally; it is easy enough to fold in intensionality.) If this move

is made, the entire fragment would be revised; each syntactic/semantic rule pair
72 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?
developed up to this point will be stripped of its semantic part. However, in all of
the cases we have discussed so far the necessary revisions are straightforward;
the revisions are left to the reader in the following exercise.
5.1. Take the rules in the fragment so far (this means only those rules labeled
TR). For each one, assume that they give just the syntactic and phonological
part (they could be stated as phrase structure rules or they could be stated in
the notation used here). Assume further that for each such rule there is a
separate semantic rule whose input is a tree. Then state each of these
semantic rules, using FR-2 as a model for the statement of the semantic rules.
You should note that FR-2 differs from the other rules you will be constructing
in that all of those rules have as their input a local tree—that is, only a mother
node and one or more daughter nodes—while FR-2 (by necessity) refers to a
bigger chunk of tree— mother, daughter, and granddaughters.

The Direct Compositional strategy vs “syntax feeds semantics” strategy is

discussed in Chapter 7, but for now we can note that the syntax feeds semantics
strategy for this particular case comes at a cost. In the case of a sentence like Lee
kissed Sandy, the fact that kissed Sandy has no meaning was remedied by the
adoption of FR-2. But then, how does the semantics work for the case of
conjoined VPs (i.e., cases where VPs are connected by and or or)? Recall that
Straw Man posits the syntactic conjunction rule in
(7). Now consider the syntactic structure in (10):
(10)
S

NP VP

Lee VP and VP

VT NP V

kissed Sandy sang I

Do the rules in Straw Man’s grammar provide an interpretation for this sentence?
The answer is no: there is no rule to interpret either the topmost VP here nor the
entire S. FR-2 is not relevant; it interprets only Ss of the form [ S NP [VP VT NP]], and
none of the rules that you will have formulated in your answer to Exercise 5.1 will
help either. So there would need to be an additional rule to interpret this tree.
 73
5.3. HIDDEN S CONJUNCTION?

5.2. Formulate the relevant rule (remembering to keep in mind the types that
Straw Man is assuming).

But of course even adding the rule asked for in Exercise 5.2 does not complete the
task. For the additional rules for (10) will do nothing for cases like those in (5),
which have (as one possibility) the shape in (11):
(11)
S

NP VP

VP and VP

VP and VP

Due to the recursive nature of the and-rule, Straw Man could never finish stating
the semantic rules. This is completely parallel to the reasoning as to why the
syntax needs a VP constituent: the same reasoning leads to the conclusion that
we need to find a meaning for each VP, which will be done in section 5.4.

5.3. Hidden S conjunction?

But first, Straw Man makes one final attempt to save his hypothesis. And this is to
posit that the semantics actually does not interpret structures with VP
conjunction. Rather, at the level at which these are interpreted, these are actually
the conjunction of two sentences. This is actually not a straw man; this view has
been taken in one form or another in many theories of grammar, and has been
implemented in various ways. For example, in early Transformational Grammar it
was assumed that there were no phrase structure rules like that in (6) and (9).
Sentences like (3) were actually at an underlying level (deep structure) instances
of S conjunction, with a transformational rule called Conjunction Reduction which
mapped trees of the form in (12) to those in (13), provided that NP1 was identical
to NP2 (see, e.g., Jacobs and Rosenbaum 1968):
(12) S ⟹ (13) S
74 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?

S and S 1 VP
NP
NP1 VP1 NP2 VP2 VP1 and VP2

(The rule can be generalized to extend to the case of or.) Additionally, it was
assumed that the input to the semantic interpretation was the deep structure, and
so the semantics would be assigning a meaning to structures of the sort shown in
(12) and not to (13). Since we already have semantics for S conjunction, there
would be no need to worry about interpretation for the case of VP conjunction.
Hence for a sentence like Romeo kissed Lee and watered the plants we don’t
need to worry about giving a meaning for kissed Lee. The semantics would
separately interpret the two sentences Romeo kissed Lee and Romeo watered
the plants (along with the rule interpreting the conjunction of the two). FR-2 will
do the job. (One might notice that under this view of conjoined sentences, the
argument given earlier for a VP constituent has been undermined, but we will not
pursue the consequences of this.) A variant on this strategy is to view the picture
in reverse. That is, assume that the phrase structure rules do give something like
(13) directly, but this is mapped to (12) (as its “Logical Form”). The semantics
operates so as to compositionally interpret that level. Again FR-2 (combined with
the other rules suitably reformulated in tree terms, as in Exercise 5.1) will now be
adequate.
But the hidden S conjunction strategy also comes at a cost. First and foremost
the syntax has again been complicated, no matter which variant is chosen. Either
variant posits an additional level of representation and hence needs a set of
explicit rules to map one level to the other. There is also an empirical difficulty
centering on sentences whose subjects contain quantifiers. Consider the pairs in
(14) and (15). In each pair the first involves conjunction or disjunction of just bare
VPs, and the second is the corresponding conjoined Ss with (syntactically)
identical subjects:
(14) a. Some pig chased the wolf and grunted.
b. Some pig chased the wolf and some pig grunted.

(15) a. Every pig chased the wolf or grunted.

b. Every pig chased the wolf or every pig grunted.

The meaning of (14b) is weaker than (14a); in the former the chaser and grunter
must be the same. In (14b) the chaser and grunter could happen to
 75
5.3. HIDDEN S CONJUNCTION?

be the same, but they need not be.6 Hence (14a) entails (14b) but not vice versa.
The reverse situation holds in the every case. Here (15a) is weaker— the most
salient reading for (15a) is each pig did one or the other of these things, but they
need not have all been in agreement as to which act to perform. (15b) has the
stronger requirement that they all did the same thing. Incidentally, it is true that
(15a) also (at least marginally) has a second reading which is the same as (15b);
this is best brought out by the continuation in (16):
(16) Every pig chased the wolf or grunted—but I can’t remember which.

Nonetheless, the most salient reading of (15a) is clearly the weaker one and is
true in situations in which (15b) is not. (The second reading is accounted for in
Chapter 11; its existence does not affect the point at hand here.)
Consider a theory in which the Logical Form in (12) is related to the surface
form in (13) and where the requirement for this is that the two NP subjects in the
two clauses in (12) must be identical. Then, of course, we get the wrong
prediction. In the cases above, the (a) sentences should have the same meaning
as the (b) sentences. As noted above, there is a secondary reading of (15a) which
is synonymous with (15b) but this is of little consolation since the correct theory
needs to also pair (15a) with its primary meaning. To be sure, there are ways to
solve these problems with more complex Logical Forms and additional conditions
on the rules effecting the mapping to (or from) Logical Forms.7 But the methods
involved require yet further rules. Perhaps it is time to abandon Straw Man’s path.

6 It might appear at first glance that (14b) requires the wolf-chasing pig and the grunting
pig to be different pigs. But closer inspection reveals that this is just an implicature (see
section 3.2) and not part of the truth conditions. If a speaker knew that a single pig did
both, they would more likely have used (14a)—hence the conclusion on hearing (14b)
that different pigs are involved. But this can’t really be part of the truth conditions. For if
I hear a grunting pig and have no idea who it is and at the same time have some other
indirect evidence about there being a wolfchasing pig, I can certainly say (14b) with no
commitment one way or the other as to whether it is the same pig. (And if it turns out to
be the same pig, I have not said anything false.)
7 Readers familiar with the “Quantifier Raising” solution to quantified NPs (see section
14.2) might think that one automatically gets the right meanings of the (b) sentences once
one combines Quantifier Raising with the rule that produces the twosentence LF. But it is
not automatic, for a principle is needed to ensure that the two processes happen in the
right order. We leave it to the interested reader to verify this.
76 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?

5.4. Back to the drawing board: Curry’ed functions

5.4.1. Recasting the meaning of transitive verbs

The complications above are all the result of Straw Man’s insistence that the
meaning of a transitive verb is a function from ordered pairs to truth values. The
problem is that this makes semantic composition look very different from the
syntax, for the semantics treats the composition as “flat” while the syntax does
not. But there is a simple way to recast things so that the semantics respects the
syntax. The technique to be used here relies on the observation that any function
whose domain is a set of ordered n-tuples (in this case, ordered pairs) can be
mapped to a function that takes the arguments one at a time. The observation that
this can be done for any such function is due originally to Schönfinkel (1924) and
later H. B. Curry (see, e.g., Curry and Feys 1958). Although Schönfinkel’s work is
earlier, this is generally known as Currying a function (although see Heim and
Kratzer 1998 who do use the historically more accurate term Schönfinkelization).
In the case at hand, we begin with the original [[K]] (a function from e x e to t)
and recast this as a function which takes one individual and returns a function of
type <e,t>. This can be done without loss of information—this is just repackaging
the meaning. Thus [[kiss]] is a function of type <e,<e,t>>. Note that there are
actually two different ways that one can recast the initial logician’s [[K]] (which
maps an ordered pair {x,y} to true just in case in the world in question x does kiss
y). One way is to create a new function such that for any x and y such that
[[K]]({x,y}) is true, [[kissed]](x)(y) is true. In other words, the “kisser” is the first
argument of the new function, and the “kissee” is the second. The other mapping
does just the reverse: for any x and y such that [[K]]({x,y}) is true, then
[[kissed]](y)(x) is true.
It turns out that the second strategy is the correct one, given the assumption
that meanings are packaged in such a way as to reflect the syntactic composition.
Our new function combines with the meaning of the syntactic direct object first,
because this is what happens in the syntax.
It combines next with the subject. This is often confusing when first encountered
because of the left-to-right order in which we write things: in a sentence like Lee
kissed Sandy, the semantic composition will be [[kissed]]
([[Sandy]]) ([[Lee]]). (The other meaning can be thought of as the meaning of was-
kissed-by).
 5.4. CURRY’ED FUNCTIONS 77

To illustrate in greater detail, suppose that the universe consists of three

individuals, a, b, and c. We will, moreover, be considering just one world w 1. To use
English to describe the facts of the world, let it be the case that a kisses b and c, b
kisses c, and c kisses a and c. Then the function [[kiss]] in w1 is as follows:
(17) a→0a→b→0c
→1

a→1b
→b→0c→0

a→1c
→b→1c→1

In order to get any mileage out of this, we also need to make explicit what was
implicit in the discussion above—i.e., the syntactic and semantic combinatory
rule for combining transitive verbs with their objects:
TR-6. If Æ is an expression of the form <[Æ], VT, [[Æ]]> and is an expression of the form <[],
NP, [[]]> then there is an expression ª of the form <[Æ-], VP,
[[Æ]]([[]])>.

Hence consider the composition of Lee kissed Sandy. Assume that [[Lee]] (in all
worlds) is a and [[Sandy]] is b. Then kissed Sandy has a meaning; it is the function
shown in (17) applied to the argument b, so it is that function of type <e,t> which
maps a to 1 and b and c to 0. In set terms, kissed Sandy characterizes the set of
Sandy-kissers (which in this case is the singleton set {a}). Then this VP combines
with its subject by the syntactic and semantic rule TR-50, and the resulting
semantics is 1, as we would expect.

5.3. The picture above in (17) is the extension of [[kiss]] in a world that we have
labeled w1. Now take a different world w2 in which the facts are that a kisses a,
b, and c, b kisses no one, and c kisses only c. (Ignore the obvious lack of
pragmatic plausibility here!)

(cont.)

First, show the extension of [[kiss]] at w 2 where this again will be shown as a
Curry’ed function of the same basic structure as (17). Then show the full
syntactic and semantic composition of Sandy kissed Lee, where at each step
 78 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?

you will be showing the meaning of each expression as a function from worlds
to some other object. (You already have [[kiss]]; it is a function mapping w 1 to
the object shown above in (17) and mapping w2 to the new object that you will
have constructed.) The two NPs are listed in the lexicon and each is a constant
function from worlds to individuals as given above. So your task is to show how
the syntax and the semantics put this all together step by step. Do this using
(12) in Chapter 4 as the model.

Does this strategy introduce undue complexity into the modeling of meanings?
Not at all. We have simply taken [[kiss]] to be a function whose co-domain itself
is a function. The general approach of the kind of model-theoretic semantics being
assumed here is to have a small set of primitive model-theoretic objects (worlds,
times, truth values, and individuals) and take other kinds of meanings to be
constructed from these using the notion of functions. 8 There is no reason to be
surprised to find that certain expressions denote functions which themselves
have functions as their co-domain (or, for that matter, as their domain; see
Chapter 10). In fact, such objects are already in the system. In the fully intensional
semantics, the meaning of an intransitive verb, for example, is a function of type
<s,<e,t>>.
Incidentally, one finds different notations for the semantic composition of a
sentence like Lee kissed Sandy, even among researchers who agree that [[kiss]] is
the Curry’ed function of type <e,<e,t>> (combining first with the individual denoted
by the object). Given the world above, the composition is most accurately
represented as [[kiss]](b)(a). However, one sometimes finds more reader-friendly
notation when the actual step-by-step composition doesn’t matter, and so one
might see something like [[kiss]](a,b), which looks like it is going back to the
hypothesis that [[kiss]] takes

8
There are debates as to whether others are needed—see especially the work in event
semantics (e.g., Parsons 1990) which takes events as primitives. But these will suffice for
what we will be considering in this book.
an ordered pair as argument. But generally this is intended simply as a way to make
things easier to read; this is called the relational notation. It’s ultimately helpful to
get used to going back and forth between the two notations.
 5.4. CURRY’ED FUNCTIONS 79

*5.4.2. Currying more generally

The above illustrates the technique of Currying to a function that has as its domain
a set of ordered pairs, but this is quite general. For any function f which takes an
ordered n-tuple as its argument to give some result x, there is a corresponding
function f 0 which takes a single argument, and returns as value a function g from
an n-1 tuple to x. This can apply recursively, until we have a function that takes its
arguments one at a time.
Generalizing the remarks above, concerning the fact that there are different
ways to Curry a function like [[K]], there are many ways to effect the mapping from
f to the one-place function f ', depending on which position of the n-tuple is peeled
off (A one-place function is one which takes just a single argument, not an ordered
tuple). Call f 0-i the function that peels off the ith position of the n-tuple. Then there
is a unique way to map f to f 0-i (and vice versa). Notice that the relationship
between the naive meaning of kiss (a function taking ordered pairs of kissers and
kissees) and our hypothesized actual meaning of English kiss is such that [[kiss]]
is [[K-2]] since we took the second member of the ordered pair to be the first
argument of [[kiss]]. This is crucial: this was done to model the way in which the
syntax works, as the syntax first combines kiss with its object and then with its
subject (and, simply because it is traditional, we took “naive-kiss” to map a pair
{a,b} to true just in case a kisses b).

*5.4. Take a set A with two members {a,b}, a set B with two members {c,d}, a
set C with two members {e,f}, and a set D with three members {1,2,3}. Consider
a function f whose domain is the set of ordered triples in A x B x C and whose
co-domain is D. (In other words, it takes each triple that can be composed by
taking a member of A, a member of B, and a member of C and maps each such
triple into some member of D.) The actual function is as follows:

(cont.)
 80 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?

{a,c,e} à 1
{a,c,f} à 1
{a,d,e} à 3
{a,d,f} à 2
{b,c,e} à 1
{b,c,f} à 3
{b,d,e} à 2
{b,d,f} à 3

Convert this to the corresponding Curry’ed function of type <C,<B, <A,D>>>.

5.4.3. Solving Straw Man’s problems

Let us consider whether this strategy solves the problems encountered by

Straw Man. First, as seen above, a VP such as kissed Sandy now has a meaning.
And, having Curry’ed the transitive verb, this VP has a meaning of exactly the same
type as the meaning of a VP with only an intransitive verb. Both are world/time-
dependent functions of type <e,t> (put differently, they characterize sets of
individuals). Because of this, we also now have a simple way to give the semantics
for conjoined VPs without having to assume that these really involve hidden S
conjunction. For expository simplicity, we ignore the possibility of allowing the
denotation of a VP to be a partial function from the domain of individuals, and so
assume that each VP maps each individual into either true or false. 9 Consider the
phrase structure rule in (6) which introduces VPs conjoined with and. (Keep in
mind that this will eventually be generalized to allow coordination of all
categories.) Up until now we have in our fragment only a semantics for and that
conjoins sentences (TR-1), but the discussion in 5.1 led to the conclusion that
there is VP conjunction as well. Direct Compositionality requires a semantics to
go with (6), and we are now in a position to see what this

9
As noted several times, a possible way to model presupposition failure is to use partial
functions. To use an example we have seen before, we might ultimately want to treat the
presupposition in a predicate stopped smoking by saying that it is undefined (at a
world/time pair) for all individuals who never smoked. If this is correct, the rule for
conjunction given below will need refinement.
semantics is. Thinking in set terms, the contribution of and is simply intersection.
(Ultimately one would hope for single and for sentences, VPs, and other
 5.4. CURRY’ED FUNCTIONS 81

expressions; this is accomplished in Chapter 12.) In other words, consider a

sentence like a slight variant of (3c) Lee meditated and fed Porky.
The intuition is easiest to get in set (rather than function) terms. Since [[feed]] is of
type <e,<e,t>>, it combines with [[Porky]] (some individual) to give the
characteristic function of the set of individuals who fed Porky. [[meditated]] also
characterizes some set. These two sets can intersect; and the result is the set of
individuals in both the pig-feeding set and the meditating set. This is exactly right;
when this combines with Lee, the sentence says that Lee is in this set—i.e., the
set which is the intersection of the meditators and Porky-feeders. By the definition
of intersection, it follows that Lee is in each set. Notice, then, that with ordinary
individual-denoting subjects (like Lee) the truth conditions for the case of
conjoined
VPs are the same as the truth conditions for the case of conjoined sentences, as
in Lee fed Porky and Lee meditated. But this follows from the semantics that
intersects the two VP meanings, and so the correct truth conditions are arrived at
without having to pretend that there are really two sentences.
To formalize the rule for conjoined VPs, note first that since we are not
considering the case of partial functions here we can safely go back and forth from
talking about sets of individuals to the corresponding characteristic function of
this set and vice versa. Given a function f of type <x,t> (for any set x) we have
introduced the notation fS to indicate that subset of x mapped to 1 by f. Now take
any set z which is a subset of some set x. We now introduce the notation z F to
indicate the function of type <x,t> that characterizes z (i.e., the function mapping
all members of z to 1 and all other members of x to 0). Using these notational
devices, we can state the VP conjunction rule (extensionally):
TR-7. If Æ is an expression of the form <[Æ], VP, [[Æ]]> and is an expression of the form
<[], VP, [[]]>, then there is an expression ª of the form <[Æ-and-], VP, ([[Æ]]S \
[[]]S)F>.

Thus the meaning of the conjoined VP is the function characterizing the

intersection of the two sets characterized by the daughter VPs. The rule
introducing or is similar, where here the relevant semantics is set union:
TR-8. If Æ is an expression of the form <[Æ], VP, [[Æ]]> and is an expression of the form
<[], VP, [[]]>, then there is an expression ª of the form <[Æ-or-], VP, ([[Æ]]S [ [[]]S)F>.
 82 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?

5.5. Take a universe with three individuals {a,b,c}. Take a world w 1 such that
[[kiss]](w1) is as shown in (17) above. Assume further the following:
[[Juliet]](w1) = a and [[Romeo]](w1) = c. Moreover [[sleepwalk]](w1) is the
following function:
a à 0, b à 1, c à 1.

Take the sentences in (i) and (ii):

(i) Juliet sleepwalked and kissed Romeo.

(ii) Romeo kissed Juliet or sleepwalked.

Show the full syntactic and semantic composition of this sentence

extensionally with respect to w 1. In other words, show the value of each
expression at w1 which the syntax proves well-formed. (Use the model in (12)
in Chapter 4 for how to display this.) For all of the VPs, show these both as sets
and as the functions which characterize those sets. What is crucial is to see
why the sentence doesn’t need to be converted into two separate sentences in
order to arrive at the correct semantics for the whole; showing out the
extension of each of the well-formed VPs in set terms helps to elucidate that
point.

This account will eventually be refined in several ways. First, and and or will
ultimately be listed in the lexicon (so they themselves have a meaning) and will be
members of a single syntactic category. Second, there appears to be evidence that
both and and or actually take their arguments one at a time such that the structure
of, for example, walks and chews gum is actually
[walks[and[chews gum]]] (see, e.g., Munn 1993). Note that this means that [[and]]
is (thinking in set terms) the Curry’ed version of the intersection operator. (The
interested reader can use prose to state just exactly what this function is.) Both of
these desiderata are accomplished in the revision in the next chapter. And finally,
as noted above, conjunctions like and and or are cross-categorial: for any two
expressions [Æ] and [] of the same category X, there is also an expression of the
form [Æ-and-]. The complex expression itself is of category X; we know this
because it has the same distribution as any other expression of category X.
Ultimately then a single and and a single or should be listed in the lexicon which
serves to conjoin expressions of any category. A full generalization is given in
Chapter 12 (and section 13.4.4.2 contains yet another way to accomplish the
generalization). But for the moment we will not collapse S conjunction (TR-1) with
VP conjunction (TR-7). Note, though, that while the semantics for VP conjunction
 5.4. CURRY’ED FUNCTIONS 83

and for S conjunction look at first blush to be completely different, they have
something in common. Recall that the value of an S is a function characterizing a
set of worlds; the fully intensional version of the and-rule in the case of S
conjunction also intersects these sets.

5.6. In light of the remarks above, it is tempting to try to collapse the two
conjunction rules as follows:
FR-3 If there is an expression Æ of the form <[Æ], X, [[Æ]]> and an expression of the
form <[], X, [[]]> then there is an expression ª of the form <[Æand-], X, ([[Æ]]S \
[[]]S)F>, for X ranging over S and NP>.

Aside from the fact that ultimately we want to allow in conjunction of other
categories besides just VP and S, there is a “cheat” embodied in FR-3. Why is it
cheating to try to collapse the case of VP conjunction and S conjunction in the
way done in FR-3?

Straw Man’s final problem was that his conjecture that VP conjunction was
secretly S conjunction gave (without additional complexities) the wrong meanings
for sentences like (14a) and (15a) where the subjects are quantified:
(14) a. Some pig chased the wolf and grunted.

(15) a. Every pig chased the wolf or grunted.

Under the two-sentence analysis of these, these should have the same meaning
as the corresponding sentences where the subject is repeated, and yet they do
not. But we cannot truly boast that the Curry’ed transitive verb solution solves all
of Straw Man’s problems until it is shown that this problem is avoided here. In fact,
we are not yet in a position to do that: but this is only because the meaning of NPs
with quantifiers has not yet been discussed. We postpone this until Chapter 10. It
turns out that armed with a meaning for expressions like every man, etc., the
meaning (or at least preferred meaning) of (14a) and of (15a) is exactly as
predicted. (It will also be shown that the secondary meanings can be derived.) At
this point we can only say “stay tuned.”
*5.5. A note on three-place verbs

English has many verbs that occur with two constituents within the verb phrase,
as in gave the new toy to Mitka, which consist of the sequence V-NP-PP (we will
 84 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?

call the verb V3 here). There are also cases of verbs which take sequences of two
NPs such as another version of give: gave Mitka the new toy. (The relationship
between the two versions of give is much discussed in the generative grammar
literature; here we focus only on the version that takes an NP and PP, but see also
Exercise 9.10.) And some verbs take NPs and sentential-like expressions
(which we will call CP) as in told Chicken Little that the sky is falling and so forth.
Here we focus on the V-NP-PP verbs such as give. Before going further we need
to say something about the meaning of to Mitka. Assume that to functions much
like a case marker and contributes nothing to the meaning, so that to Mitka means
the same as Mitka. Returning to the full VP gave the new toy to Mitka, there are four
logical possibilities for how this is put together in the syntax (indeed, all four
possibilities have been suggested):

1. The VP is “flat”: in phrase structure rule terms there is a rule of the form:
VP à V3 NP PP (see, e.g., Gazdar, Klein, Pullum, and Sag 1985).
2. The NP and PP actually form some sort of constituent; call it Z and there are
two rules:
VP à V3 Z and Z à NP PP (see Kayne 1984).
3. give and the new toy form a constituent which then combines with a PP (as
does the simple verb dash). The new rules would thus be something like:
VP à V4 PP and V4 à V3 NP.
This means that give and the new toy combine to give a complex verb of the
same category as things like dash; gave the new toy then combines with the PP
to Mitka and the result is a VP (see, e.g., Dowty 1988).
4. The last possibility is one which requires additional apparatus in the syntax, as
the syntax involves more than just concatenating expressions. This would be
to assume that give actually first combines with to Mitka and the resulting
expression then combines with the new toy. Versions of this (either for this
case or for related cases) are proposed in Chomsky (1957) (for a VP like
consider George a fool), and explored in much
 85
*5.5. A NOTE ON THREE-PLACE VERBS

subsequent Categorial Grammar literature (Bach 1979, 1980, Dowty 1982, and
others), within GPSG by Jacobson (1987), and then again in Government and
Binding theory in Larson (1988).

The basic idea of the proposals in 4 above is that gave to Mitka in the expression
give the toy to Mitka is a discontinuous constituent. There are two broad classes of
ways to implement this intuition. One makes use of two levels of representation in the
grammar, and posits a movement rule to break up the constituent. This in turn has two
possible implementations. (i) The basic structure is as in (18i) and to Mitka moves to
the right as shown there, or (ii) the basic structure is as in (18ii) and give preposes (we
suppress details of the actual proposals, and we arbitrarily adopt the label C for the
expression that is the discontinuous constituent).

i. VP ii. VP

C NP PP V NP C

V PP the new toy the toy V PP

gave to Mitka gave to Mitka

Chomsky (1957) proposed a version of (i) (although for a different set of cases), and
Larson (1988) proposes a version of (ii).
A rather different way to implement the basic intuition is put forth in Bach (1979,
1980). This relies on the observation that the existence of discontinuous constituents
does not inevitably lead to the conclusion that there is more than one level of
representation. Rather, the syntactic rules combining two expressions might perform
operations other than just putting one (or more) expressions next to each other.
Perhaps there are also cases where one expression is infixed into another. 10 (Note that
10 A common misconception is that the existence of “discontinuous constituents” requires
more than one level of representation because a sentence must be represented as a tree, and
trees do not admit of discontinuous constituents. If the grammar contained an infixation
operation, the representation of the sentence would involve an object with crossing branches—
not a tree. But the underlying assumption—that trees are sacred objects—makes little sense;
a tree may or may not morphological processes include infixation.) Bach dubbed the
 86 5. TRANSITIVE VERBS: A SYNTAX/SEMANTICS MISMATCH?

process in syntax as Wrap, and proposed that give first combines with to Mitka to form
a complex transitive verb give to Mitka. Further, any complex transitive verb combining
with its object (here, the new toy) does so by “wrapping” around its object. Bach’s
particular formulation of the operation referred to word boundaries: Wrap places one
expression after the first word of the other. Wrap operations have been formalized in
other ways as well. For example, Pollard (1984) suggests that the grammar refers not
just to unanalyzed strings but headed strings. Each string has a distinguished head,
and Wrap places the infix after the head of the other expression (or, in some
languages, before the head of the other expression). If there are Wrap operations, the
remarks made earlier regarding the role of structure in grammar are not quite correct:
the grammar must keep track of just enough structure within an expression as to
specify the infixation point. But none of the Wrap proposals requires the grammar to
refer to structures as rich as full trees. We will formalize a variant of Pollard’s proposal
in Chapter 6.
This fourth solution is the most complex in terms of the syntax: it requires a Wrap
operation (or a movement rule). Nonetheless, there seems to be some evidence for it
(see section 15.4.2). To the extent that we deal in this book with three-place verbs,
the Wrap solution is what we will (tentatively) adopt. (Note that this is compatible
with the general architecture assumed here while the movement solution is not.) And
even though the evidence for Wrap in the case of English is not entirely conclusive,
there is independent evidence that grammars containing only the equivalent of
context-free phrase structure rules will not be enough for other languages (see n. 5
and, e.g., Shieber 1985). Wrap operations have

be the appropriate representation as that depends entirely on what we discover is the correct
rule system. (A tree is a sensible representation of how the rules work in a context-free phrase
structure grammar; it is not a sensible representation in other systems.) Notice that if there are
multiple levels of representation then it is also the case that a tree is not the appropriate
representation for a sentence. Rather, a sequence of trees is. There is no particular reason to
prefer a sequence of trees to some other kind of object. The important point is that
representations represent the rule system(s); the task is to discover the kinds of rules
employed by natural language.
 87
*5.5. A NOTE ON THREE-PLACE VERBS

been shown to be useful in the analysis of some of the key cases that lead to this
conclusion.11
In any case, the main interest here is in how the semantics works. Given the
hypothesis of Direct Compositionality, each of the four logically possible analyses of
three-place verbs sketched earlier leads to a slightly different way to flesh out the
semantics associated with the syntax. Here we consider only the possibilities labeled
(3) and (4) above. While we have seen that (4) is syntactically more complex, (3) and
(4) are equally straightforward in terms of the semantics. Either way, the idea is that—
once again—the verb in the syntax is combining with its arguments one at a time. And
since both of these are individuals, the resulting verb is of type <e,<e,<e,t>>>. The
meanings would be different depending on whether we adopted (3) or (4). If (3) is
correct, [[give]] as a function maps an individual which serves as the gift to a function
taking the recipient to a function from givers to 1 or 0. If (4) is correct, the recipient is
the first argument, the gift is the second argument, and the giver is the last argument
in.

*5.7. (i) If (1) were correct, what might be its associated semantics?
(ii) One way to give a semantics that goes along with (2) would be similar
to the way you (hopefully) constructed for (i). Make this explicit.

11 One such example is

the analysis of “cross-serial dependencies” in Dutch given in Pollard
(1984). The key phenomenon was discussed in Huybregts (1976). To illustrate the basic
phenomenon, we use subordinate clauses (because Dutch, like most other Germanic
languages, always has the main verb in second position in main clauses, which complicates
the situation). Thus we find embedded sentences in Dutch like the following:

(i ) . . . dat Jan Piet Marie Cecilia zag helpen laten

zwemmen that Jan Piet Marie Cecilia saw help
make Swim
“that Jan saw Piet help Marie make Cecilia swim”

This consists of a series of NPs and a series of verbs, and the “subject” of the nth verb in the
verb chain is the nth NP in the NP chain. (This is not an isolated example but a systematic
phenomenon.) While a set of context-free phrase structure rules could literally give the relevant
string set it cannot give it in a way which—assuming a Direct Compositional view—will give a
sensible semantics. Pollard (1984) showed that supplementing concatenation rules with Wrap
can account for this.
 6
Categorial Grammar
6.1. Basics 89 6.5. Further generalizing the rules 98

6.2. Syntax/semantics correspondence 92 6.6. and, or, negation, and case-marking

prepositions 100
6.3. Refinements 94
6.7. Summary: The fragment so far 104
*6.4. Some further observations about
syntax: Case marking
and agreement 96

While the goal of this book is not to develop a detailed theory of syntax, a work on
the syntax/semantics interface obviously needs to be rooted in some theory of
syntax, even if only for expository purposes. We thus now develop (one version of)
the theory of Categorial Grammar, albeit in a somewhat rudimentary fashion. 1 The
reason for adopting Categorial Grammar is that it provides an elegant account of
how the syntax and semantics work together. However, one should bear in mind
that most of the remarks in the rest of this book on the syntax/semantics interface
could be imported

1
Categorial Grammar actually predates other versions of generative grammar, and was
first developed in Ajdukiewicz (1935). Montague (1973) adopted a version of Categorial
Grammar as his syntax for much the same reason as we are using it here: it provides a
transparent fit between the syntax and the semantics. There are a number of different
versions of Categorial Grammar. See, for example, Steedman (1987, 1996) which has often
gone under the name Combinatory Categorial Grammar (the basic version adopted in this
text is quite close to although not identical with this). Much recent work in the general
framework has gone under the rubric of “Type Logical Grammar”; see Carpenter (1998) for
an introduction.
 89
6.1. BASICS

into other theories as well, with of course some modification. Indeed some of the
proposals in this book are based on ones originally developed within the theory of
Generalized Phrase Structure Grammar (Gazdar, Klein, Pullum, and Sag 1985)
(which adopted the semantics of Montague 1973). Head-Driven Phrase Structure
Grammar (HPSG) (see, e.g., Pollard and Sag 1994) incorporates a rather different
theory of semantics, but nonetheless maintains essentially the same architecture
as the one here for the syntax/ semantics interface. And while Government–
Binding Theory and Minimalism both have very different views of the organization
of the grammar, many of the basic insights to be discussed later regarding the
semantics carry over to or even were developed within these theories. As we do
not wish the main body of results reported in this book to be locked into a single
theory of syntax, the reader familiar with other theories is strongly encouraged to
translate the results into the terminology of those theories from time to time
where possible.

6.1. Basics

A fundamental premise of Categorial Grammar (hereafter, CG) is that a syntactic

category label is an expression of its distribution. The grammar contains a small
set of primitive categories and defines other categories from these. Let us take the
primitives to be S, NP, N, PP, and CP.8 The fact that some of these (NP, PP) are of
the form “XP” (standardly used to mean X-phrase) while others (S, N) are not is
arbitrary and is just a bow to tradition: there is no distinction between categories
which typically contain only one lexical item and those which typically contain
more. Indeed this was noted already in section 4.1. There, complex expressions
like the disobedient husky were treated as being of the same category (NP) as
simple names like Mitka.
In addition to the primitive categories, we recursively define additional
categories as follows:

8 CP is the name for the category of expressions like that Mitka howled in a fuller
sentence like Sabrina thinks that Mitka howled. Thus this is distinct here from S (it is an S
combined with that), although we will assume that that Mitka howled has a meaning of
type <s,t>, just as does S.
90 6. CATEGORIAL GRAMMAR
(1) If A is a category and B is a category, then A/B is a category (this category is pronounced
as “A slash B”).
The intuition behind labeling some expression with the category A/B (for some A
and some B) is that such an expression combines with an expression of category
B to give an expression of category A. In other words, the name of a category
encodes what it syntactically combines with (as a sister) and the category of the
resulting (mother) expression.
Obviously missing from the above list of primitives is the category “VP.” This is
because any expression of this category can be recast as something which
combines with an NP to give an S;3 the traditional “VP” is thus renamed as S/NP.
We will, however, often use the label VP as a convenient abbreviation for S/NP.
Consider then the case of transitive verbs such as kiss, love, chase, etc. Each of
these combines with an NP to give a VP, and so each is of category (S/NP)/NP. A
verb like dash, on the other hand, which takes a PP complement within the VP
(The frightened fox dashed into a hole) is of category (S/NP)/PP. In other words,
the complement that a verb takes within the VP is encoded into its category. What
about intransitive verbs such as sing, walk, etc.? We now officially adopt the view
that the grammar does not distinguish between complex phrasal expressions and
expressions that have the same distribution but happen to consist of just a single
lexical item.4 Since walk has exactly the same distribution as the complex
expression chew gum and the same type of meaning (as was demonstrated at
great length in the preceding chapter) a simple intransitive verb like walk is also of
category S/NP. The astute reader will notice that we have suddenly glossed

3
In fact a variety of other constituents besides NPs can serve as subjects in
English. This includes “CPs” (sentences introduced by that) as in That Sarah was the
nominee was extremely surprising, and infinitive VPs as in To ski while being pulled by a dog
is tremendous fun. Arguably verbs do not impose any idiosyncratic restrictions on the
category of their subject in English: unlike in object position the category with which they
occur seems to be predictable purely semantically. (See Grimshaw 1982 for discussion.)
It may well be then that a “VP” should really be defined as S/X for X a variable over a set of
categories, but we will not pursue this here. 4
There is one situation in which English grammar seems to be sensitive to whether an
expression has more than one word in it. Simple adjectives (like red) occur prenominally,
but complex adjectives (happy that Bill left) occur only postnominally when they modify
nouns. Of course this means that the two do not have exactly the same distribution so they
could be given a different category. But in all other respects they are the same, so we will
not take that route. We leave it open as to how exactly the grammar refers to this difference;
see section 8.4 for a treatment of adjectival nominal modifiers.
 91
6.1. BASICS

over an observation made earlier: the traditional notion of “verb” is motivated by

morphological considerations. As things are set up so far, kiss and walk have
nothing in common: one is of category S/NP and the other of category (S/NP)/NP.
But these items are treated alike by the morphology: they all take past-tense
morphology, they can take the suffix -ing, and so forth. This will be addressed
shortly.
Thus the intuition is that (like in many other theories), the combinatory rules can
be stated in a very general form. Most of the information about what kinds of
complements an item can combine with has to be listed in any case as part of the
lexical entry. Given that, the rest should follow from some very general rule
schemata. This is much like the idea in, for example, classic X-bar theory but the
particular rule schemata will be quite different from those adopted in X-bar theory.
(Readers unfamiliar with X-bar theory can consult an introductory syntax text for
discussion, but it will play no role in this text.)
As noted above, the basic idea is that if something is of category A/B, it can
combine with an expression of category B to give an A. This however is not quite
enough: the category so far does not specify whether the B expression goes to the
left or to the right of the A/B expression (or, in light of the remarks in section 5.5,
possibly as an infix). We thus elaborate the categories in such a way that they
encode that as well. Using a notation not entirely standard in Categorial Grammar,
an expression of category A/RB is something which combines with a B to its right
to give an expression of category A, and A/ LB takes the B to its left.5 (Cases
involving Wrap—if they exist—are dealt with in the fragment summary at the end
of this chapter.) The recursive definition of categories is then expanded as follows:

5
There are two other common notational systems for this which unfortunately collide.
In one system (cf., Lambek 1958 and much work in Type Logical Grammar) A/B means
something which takes a B to its right to give an A, while B\A is something which takes a B
to its left to give an A. In Steedman (1987, 1996) and related work A/B continues to mean
something that takes a B to its right to give an A, while A\B is something which takes a B to
its left to give A. Thus the two systems collide on the meaning of X\Y: in one the result is X
and in the other the result is Y. It is because of this potential confusion that we adopt the
subscripts. Subscripts are also convenient if the grammar turns out to include a Wrap
operation; a “Wrapper” can be notated as A/WB. (This is what is known in morphology as a
circumfix, as in the German morpheme ge...en indicating past participle. Something which
is an infix would be an A/IB.)
92 6. CATEGORIAL GRAMMAR
(2) If A is a category and B is a category, then A/RB is a category and A/LB is a category.

The intuition behind these categories has already been discussed, but
technically all we have in our grammar so far is a list of lexical items and
categories. Thus the grammar also includes two general rule schemas, which
replace rules TR-50 and TR-6 in the earlier fragment (for now, the semantics is not
spelled out: “ . . . ” is used as a placeholder). We will label these rules as R-1a and
R-1b. Lest it appear that we are cheating by counting this as one rule when there
really are two, later discussion (section 6.5) will collapse them into a single rule.
R-1a (partial): If Æ is an expression of the form <[Æ], A/RB, [[Æ]]> and is an expression
of the form <[], B, [[]]> then there is an expression ª of the form <[Æ-], A,
. . . >.
R-1b (partial): If Æ is an expression of the form <[Æ], A/LB, [[Æ]]> and is an expression
of the form <[], B, [[]]> then there is an expression ª of the form <[-Æ], A,
. . . >.
At this point these two rules replace only two rules in the fragment constructed
earlier, and so there is no sense in which these rules are more general. But that is
obviously only because the fragment so far is rather small: there will be many
other applications of the rules R-1a and R-1b.
Speaking informally for the moment, there is a sense in which an expression
with a category of the form A/LB is a function from strings to strings. (This will be
formalized in 6.5.) Thus sings maps Madonna to Madonna sings. Similar remarks
hold for right-slash categories. Hence, when an A/B combines with a B we refer to
the former as the function and the latter as the argument. Unfortunately the
syntactic notation reverses the domain and codomain from the semantic
notation; A/B takes an expression of category B and returns one of category A,
while <a,b> is a function taking something from domain a and returning a value in
co-domain b. This confusion is an unfortunate accident of history which simply
takes some getting used to.

6.2. Syntax/semantics correspondence

What about filling in the semantics? For the two instances of the schemas that we
know of (combining VPs with subjects and transitive verbs with objects) the
semantics involves applying a function to an argument. In order
 93
6.2. SYNTAX/SEMANTICS CORRESPONDENCE

to state this generally, the correspondence between syntactic categories and

semantic types needs to be made explicit. Each primitive syntactic category
corresponds to some semantic type, i.e., all expressions of some primitive
syntactic category have as their meanings some member of the same set. The
extensional type of S is t—in other words, the value of any S is a member of {1,0}.
Its intensional type is <s,t>. The extensional type of NP is e (its intensional type
<s,e>). We also see that S/NP is of type <e,t> (intension: <s,<e,t>>) and (S/NP)/NP
of type <e,<e,t>> (intensional type: <s,<e,<e,t>>>). Assume then the following
correspondence between syntactic category and semantic type: for an
expression of category A/B, its extension is a function from the extensional type
of B to the extensional type of A. (There is a caveat: some expressions are
functions from intensions not extensions; this is the subject of Chapter 19.) This
correspondence holds regardless of the directional feature on the slash. Note
incidentally that while this assumes that all expressions of a given syntactic
category have meanings of the same types, the reverse does not hold. Thus A/ LB
and A/RB are syntactically different categories, but have the same type of
meaning. Put differently, word order has no reflex in the semantics and is purely a
syntactic phenomenon. And Chapter 8 discusses two other syntactic categories
besides S/NP whose meaning is of type <e,t>. The set of syntactic categories,
then, is richer than the set of semantic types.
Having made explicit the correspondence between syntactic categories and
semantic types, it is now obvious how to fill in the “ . . . ” part in the above rules.
We give the full rules extensionally, and we will refer to R-1a as r-app and R-2a as
l-app:
R-1a: If Æ is an expression of the form <[Æ], A/RB, [[Æ]]> and is an expression of the form
<[], B, [[]]> then there is an expression ª of the form <[Æ-],
A, [[Æ]]([[]])>.

R-1b: If Æ is an expression of the form <[Æ], A/LB, [[Æ]]> and is an expression of the form
<[], B, [[]]> then there is an expression ª of the form <[-Æ],
A, [[Æ]]([[]])>.

6.1. Give INT-R-1a, using the model for stating this shown in
section 3.3.
94 6. CATEGORIAL GRAMMAR

6.3. Refinements

There are some aspects of this system that might strike a reader familiar with
syntactic theory as objectionable, but these objections are easily addressed.
First, we don’t want to conclude that every intransitive verb is listed in the lexicon
of English as having category S/LNP, where the directional feature (L) is stipulated
on a case-by-case basis. For it is a perfectly general fact about English syntax that
subjects (which can be defined here as the last argument in forming an S) go to
the left. Similarly for the right slash: all transitive verbs take their objects to the
right and this should not be listed lexical item by lexical item. And the
generalization is even broader: verbs in English that occur with a single
complement within the VP take it to the right regardless of the category of that
complement. In fact, most instances of expressions of category A/B—provided
that A is not just the simple category S—combine with things to the right.
Prepositions, for example, are of category PP/NP and they take the NP to the right;
the is listed as NP/N and it takes the noun to the right, etc. In many syntactic
theories, all of these facts are subsumed under the observation that English is a
head-first language. (Verbs are thus seen as the head of the VP; determiners as
head of the NP (often called DP for just that reason), prepositions as the head of
the PP, etc.)
While the version of CG developed here has no formal notion of head,6 the basic
generalizations that go under the rubric of “head first” are easily accounted for.
Note that there is a rough correspondence between the standard notion of head
and the notion here of a function. Thus an expression Æ which is of category A/B
can (roughly) be thought of as the head of an expression formed by combining it
with a of category B. To the extent that in general heads precede their
complements, the generalization in CG can be recast as the observation that
slashes in English are generally marked with a R-feature. An exception, of course,
is the slash in S/NP (or S/X for other categories which can appear as subject). 7 6
Such a notion could be added in if need be, though it will play no role in this text except
possibly for the formulation of Wrap operations (see section 5.5).
7
In standard X-Bar theory, the fact that subjects precede VPs is accounted for by not
treating the VP as a “head” but rather treating the subject with a separate label: “Specifier.”
There is no corresponding notion of a Specifier in (at least this version of) CG.
 95
6.3. REFINEMENTS

Assume, then, that a lexical item is listed in the lexicon with an underspecified
category, and that there are general rules adding further information to the
category. We will use the term lexeme to refer to that which is listed in the lexicon.
Thus the directional feature is not listed on a caseby-case basis on individual
lexemes: kiss is just listed in the lexicon (S/NP)/NP. But only a more fully specified
version (with a slash directional feature) can actually combine with another
expression. Hence the two rules in (3) apply in English, where X and Y are variables
over any category.
These are displayed informally but the intent should be clear (the notation
. . . (S/X) . . . means any occurrence of this material; this might be contained within
a more complex category):
(3) a. ...(S/X)... b. ...(X/Y)....

L R

If (b) applies after (a) the effect will be that a right slash is the default. (Thus a rule
like this applies only to slashes not already marked with a directional feature.)
We add one additional directional rule. While modifiers have fairly free
placement in English, a rough rule of thumb is that they occur to the right of what
they modify. Modifier can be given a technical definition in CG: anything of
category X/X is a modifier. Hence to the rules above can be added (c), which also
applies before the (default) right-slash rule above (the order of application of (a)
and (c) is not relevant):

c. ...(X/X)...

A second point to address is the fact that—despite the earlier observations

about the morphological commonality of “verbs”—we seem to have no such
notion at our disposal. Similar remarks would ultimately hold for nouns and
adjectives. While ordinary intransitive nouns are of category N, nouns that take PP
complements (including relational nouns like aunt which will be discussed in
Chapter 8) are presumably of category N/PP. While beyond the scope of the
semantic treatment here, there are also nouns that take CP complements (belief
that the world is round) and hence presumably of category N/CP. Yet nouns like
96 6. CATEGORIAL GRAMMAR
this all have a morphological commonality (they all can have plurals). Indeed—as
noted in Chapter 4—the traditional parts of speech are motivated by
morphological considerations. But there is no problem here: it is perfectly easy to
recover the traditional notions in terms of the ultimate result category. A verb is
something whose ultimate result category is S (this will be slightly refined in
section 8.1). A noun is something whose ultimate result category is N.
Morphological processes are presumably sensitive to classes defined in this way.

*6.4. Some further observations about syntax: Case

marking and agreement

Once one starts thinking of a syntactic category A/B as the information that “my
sister will be a B and my mother an A,” there are other interesting consequences
for syntax. As in most other theories of syntax, we can refine the basic categories
given above to actually be feature bundles (let S, for example, be one feature in
the bundle). Gender, number, case, and so forth are additional features on these
labels. In CG, phenomena like case marking, agreement, and so forth are naturally
handled as selection for elaborated categories (rather than as a process
transferring features onto actual linguistic material). For example, walks is
presumably S/NP[SING] while walk is S/NP[PLUR].8 Presumably there is a single
entry in the lexicon which is not specified for singular or plural, and morphological
rules apply to the lexeme WALK of category S/NP and map this to [walk] of
category S/NP[PLUR] and [walks] of category S/NP[SING]. There is also a
semantic effect of this process. It was briefly noted in Chapter 4 that plurals can
be thought of as individuals of a special type, and so the meaning of WALK is a
function defined for the full set of individuals (both singular and plural). The
mapping of this to the more fully specified version S/NP[SING] would remove the
plural individuals from the domain of this function (and the reverse for the plural
version).
Similarly, case marking is really case selection. Take a language with robust
case marking but where the case marking is quite predictable—

8
One might well want a more sophisticated theory of features where a feature is a name
and a value, and so there could be a single feature named NUM whose possible values in
English are SING/PLUR. In languages with a singular/dual/plural distinction, the NUM
 97
feature has three possible values. See Gazdar, Klein, Pullum, and Sag (1985) for a well-
worked-out theory of features along these lines.
*6.4. CASE MARKING AND AGREEMENT

suppose for example all subjects are nominative and all objects accusative. Then
all the transitive verbs in the language are actually of category (S/NP
[NOM])/NP[ACC]. As for the case of English word order, one would not want to
conclude that this information is listed in the lexicon on a case-bycase basis.
Hence, let the verbs be listed in underspecified form, with the following two rules
to add the case features to the NP slots:
(4)a. ...(S/NP)... b. ...(S/NP)/NP

NOM ACC

Even more interesting is the situation in a language like German in which the
bulk of the verbs assign a predictable case to their objects (hence this is called
syntactic case) while a small number assign a different case. For example, the
verb helfen “help” assigns dative case, and it is generally assumed that this is a
lexically idiosyncratic fact which must be learned by German speakers on a case-
by-case basis (this is thus often called lexical case). On the other hand, speakers
do not need to learn a list of verbs assigning accusative case, for that is the default.
Here we assume that German contains the two rules in (4), but that helfen is listed
in the lexicon as (S/NP)/NP[DAT]. As was the situation with the directional
features, case features are assigned only to slots not already specified for a case
feature. Hence they can add but never change the case feature on a slot. Thus
phenomena like predictable (syntactic) case vs idiosyncratic (lexical) case are
easy to handle.

*6.2. The remarks above show how one might ensure that NPs with the
appropriate case markings appear in the correct position, but there remains the
question of ensuring that the correct internal structure of an NP with a certain
case marking. In some languages the actual case marking shows up in the
morphology of the noun, in some it shows up in the determiner, and in some
both.
Consider the case of an imaginary language in which the noun shows the
case marking. We will call this language English’, and imagine then that an NP
98 6. CATEGORIAL GRAMMAR
in subject position which translates as English the dog is, in English’, the dog-
us. In object position the NP would be the dog-um.
Assume (following, e.g., Gazdar, Klein, Pullum, and Sag 1985) a system as
noted above where Case features can be analyzed as a feature name

(cont.)

and a feature value: the name is CASE and the possible values are NOM and
ACC. Then we take dog-us to be N[CASE:NOM] and dog-um to be
N[CASE:ACC].
Give the category for the in English’. You may need to make up some
notational conventions, or you may be able to borrow one from phonology. The
interested reader might also formulate the category of the suffixes -um and -
us.
*6.3. Consider the case of a language (like German) where the case marking
shows up in the particular form of the Determiner. Thus (mixing up German and
English) let our imaginary language be English’’ which has NPs of the form dem
dog in subject position and den dog in object position. Give the categories for
dem and dem.

6.5. Further generalizing the rules

At the moment, R-1a and R-1b are—despite the numbering—two separate rules.
But their semantics is exactly the same, suggesting that there should be a way to
capture that generalization. Of course, there are only two rules at issue, and so it
is not quite so obvious on the basis of just these two that any major generalization
is being missed, but if there are similar additional rules such the Wrap process
briefly discussed in section 5.5 there would be yet another case of an A/B
combining with a B with the same semantics.
There are two conceivable resolutions (which are actually not mutually
exclusive). The first is what is known as type-driven interpretation, which was put
forth in, among others, Klein and Sag (1985), who were also working within a
Direct Compositional framework. While their theory of syntax was different than
that taken here, their basic solution carries over directly. Thus they proposed that
each individual rule (schema) is not listed in the grammar with the semantic part,
 99
but that the semantic part can be filled in with the following principle (their
terminology is slightly different):
(5) For any rule (schema) which allows an expression Æ whose meaning is a function in
<a,b> to occur with an expression whose meaning is a member of set a, the
semantics is [[Æ]]([[]]).

The intuition, then, is that there are general principles in the grammar putting
things together in the way suggested by the types, and so the
6.5. FURTHER GENERALIZING THE RULES

semantics could be pulled out of what is listed as part of the particular rules. (The
type-driven idea has also been borrowed—with modification—into a large body
of work in a non-Direct Compositional framework in which the trees are the input
to the semantics. The idea under this view is that semantic composition of each
node in the tree is predictable from the type of its daughters. See Heim and Kratzer
1998 for one implementation.)
But it is not clear that type-driven principles (of the sort envisaged by Klein and
Sag) are ever needed given the generality of rules in Categorial Grammar. The other
option here is to collapse R-1a and R-1b (and if there is a Wrap rule it would
collapse with these as well). This is the strategy that will be adopted here, which
is why the two rules were numbered as 1a and 1b.
In order to collapse these, we first formalize the notion introduced informally
earlier: that syntactic categories correspond to functions from strings to strings.
While the necessary formalization might at this point seem like a lot of trouble to
go to just to collapse these two rules, it actually has significant other benefits
which will emerge in later chapters. Hence, each category of the form A/ XB (where
X is either L or R) corresponds to a function. The category itself is the name of a
set of expressions, but it corresponds to some function which takes two strings
(one at a time) and returns a third. We will notate the corresponding function as
FA/B, and define them as follows:

(6) a. FA/R B ([Æ])([]) = [Æ-] b. FA/ LB ([Æ])([]) = [-Æ]

We can now state a single R-1 (call it app):

R-1. If Æ is an expression of the form <[Æ], A/B, [[Æ]]> and is an expression of the form
<[], B, [[]]> then there is an expression ª of the form <FCAT(Æ)([Æ])([]); A; [[Æ]]([[]])>
(where CAT(Æ) means the category of Æ).
100 6. CATEGORIAL GRAMMAR
For example, take the expression sings which is of category S/ LNP. The function
which corresponds to this category is one which applies to the string sings and
then yields a function which applies to Madonna to give Madonna sings. So the
fact that Madonna goes to the left of sings when they combine is a result of the
fact that R-1 says that the string resulting from the combination of Madonna with
sings is the value of the function FA/ B when applied to the string sings and then to
L

Madonna. For expository convenience, we will often continue to refer to R-1a and
R-1b separately.
 101
6.6. and, or, negation, and case-marking prepositions

Before summarizing the new fragment, we are ready to make some modifications
regarding and and or. In the fragment developed earlier, these words were not
listed in the lexicon but were introduced syncategorematically. Moreover, both
and and or were introduced by rules which were limited to just VP and S
conjunction, but it is well known that any two expressions of the same category
can combine (with either and or or) to give an expression of that category. That is,
the resulting expression has exactly the same syntactic distribution as does each
of the daughters.9 Each of the italicized expressions below illustrates this point:
(7) a. The candidates spoke to the moderator and to the audience.
b. Sam cooked and ate lentils for supper.
c. A friend and colleague of mine was just named Dean of the Faculty.

So conjunction (and disjunction) should be

generalized to be crosscategorial.
Moreover, in order to list and and or in the lexicon and to do so without
introducing any new rules, we can assume that they combine with their arguments
one at a time. In other words, both are of category (X/X)/X, for

One might ask whether coordination involves only the combining of two expressions of
9

the same category or whether there are other cases of coordination. There is a large
literature on this question, and many have concluded that there are other kinds of
coordination. For one thing, there appear to be cases where categories apparently of
different types conjoin, as in (i). Second, there are cases of so-called Non-Constituent
Coordination (including what is sometimes known as Right-Node Raising) as in (ii):

(i) Linda is a bank teller and proud of it.

(ii) a. Mary loves and John hates model-theoretic semantics.
b. Captain Jack served bananafish yesterday and lobster today.

But it will be shown in section 13.4.2 that within Categorial Grammar cases like (ii) of so-
called “non-constituent conjunction” reduce to ordinary constituent conjunction (Dowty
1988). The cases in (i) are also not threatening to the generalization that all and only like
constituents may conjoin with a more sophisticated notion of categories. This will not be
pursued in this book, but the interested reader can consult Bayer (1996) for a CG treatment
of these. A quick preview: this incorporates the notion of “union” categories and
“intersection” categories. If categories name sets of strings, this becomes quite natural.
See also Morrill (1994).
6.6. AND, OR, NEGATION, AND CASE-MARKING PREPOSITIONS
102 6. CATEGORIAL GRAMMAR

X a variable over categories. The claim that expressions like Mitka and Kolya do
combine with their arguments one at a time actually has some motivation; see,
e.g., Munn (1993).

6.4. Given the word order rules spelled out in (3a–c), what would be the
directional features on each of these slashes? Show exactly how the grammar
(in the syntax) thus derives expressions like fed the pig and meditated (ignoring,
of course, the presence of the past-tense morphology). It is worth noting that—
given certain assumptions shared in both Government–Binding theory (and its
descendants) and in some versions of CG—there actually is evidence that the
slash directionality features given here are correct. See Munn (1993) for the
point phased in GB terms, and Jacobson (1999) for a similar observation within
the framework here.

We are, however, not yet in the position to give a fully general meaning for and
and for or (partly for the reason hopefully discovered in Exercise 5.6, and also
because there are cases where and and or connect expressions that don’t denote
sets, such as transitive verbs). But we do know that whatever the general meaning
of and is, it has the effect of intersection for the case where it conjoins VPs. In this
case, then, its meaning is the Curry’ed version of the intersection operator. So
assume that the general meaning is such that whenever two set-denoting
expressions are conjoined, the result is (Curry’ed) intersection. For now, then, we
just write a complicated meaning for the lexical entry which says this. This
meaning is spelled out in the fragment at the end of this chapter, and all of this is
cleaned up in Chapter 12.
Note too that so far all of the other lexical items have as their meanings
functions from worlds and times to extensions. Items like and and or seem
different: these do not appear to have values that are in any way world- (or time-
)dependent. Given two VPs combined with and, whatever world we are in is such
that the result (in set terms) is the intersection of the extension of those VPs. That
the value of [[and]] is not world-dependent can be demonstrated as follows.
Consider the sentence The senior senator from Nebraska coughed and sneezed.
Imagine two worlds w1 and w2 such that this sentence is true at w1 and false at w2.
Then either the senior senator from Nebraska has a different referent in those two
worlds, the set of coughers is different, or the set of sneezers is different. If
Nebraska’s politics remain unchanged and the coughers and sneezers do too, this
 103
must have the same value at both worlds. This would not follow if [[and]] was
world(/time)-dependent.
But this is no different than the situation that we saw for proper names which
(at least arguably) have a constant extension regardless of the world. Thus one
way to keep the semantic combinatorics uniform, is to assume [[and]] and [[or]]
are indeed functions from worlds to (roughly) Curry’ed intersection, but are
constant functions. In fact, this is true in general for what are traditionally thought
of as closed-class items.9 As we will see, prepositions, the, be, etc. also do not
vary in their extensions from world to world. We will thus treat all of these as
constant functions from worlds. Since these do not vary, we will from now on
simply write their meaning extensionally. (The correlation between closed- vs
open-class items and constant vs non-constant functions is not perfect, since the
values of names are also (arguably) constant functions, but these are surely
openclass items.) To a reader who feels that there is something rather clumsy and
artificial about giving prepositions, and, etc. intensional meanings whose
extensions are always constant across worlds, a possible modification of this will
be very briefly explored in section 19.2.
We can also now give a somewhat more adequate treatment of English
negation. Recall that the only kind of negation in the earlier fragment is introduced
by TR-3, which appends it-is-not-the-case-that in front of a sentence and
reverses its truth value. In reality, English negation generally involves either the
word not after the first auxiliary, or n’t as a clitic on the first auxiliary. But since a
treatment of the English auxiliary system is beyond the scope of this book, we
cannot here provide a fully accurate account of English negation. Still, we can do
quite a bit better by pretending
6.6. AND, OR, NEGATION, AND CASE-MARKING PREPOSITIONS

9 For readers not familiar with this terminology: closed-class items are those whose
syntactic category contains a limited number of items and does not easily admit of new
such items (e.g., prepositions, determiners, conjunctions), while openclass items are
those in the traditional categories like noun, verb, adjective, to which new items can be
added with no difficulty (new nouns, verbs, and adjectives are coined daily). The former
category is also sometimes called “function words.” We make no commitment here to this
distinction having any real theoretical significance since the difference in terms of ability
for new items to be added is really a matter of degree and not of kind. Still, it is interesting
to note that this traditional distinction does correlate roughly with the semantic
distinction of having constant vs nonconstant values across worlds (and/or times).
104 6. CATEGORIAL GRAMMAR

that didn’t/doesn’t are unanalyzed units which combine with VPs. 11 This word,
then, is of category VP/RVP12 and its meaning—in set terms—is set
complementation. Hence [[didn’t dance]] is the function characterizing the set of
individuals in the complement of the set characterized by [[dance]]. In function
terms, [[didn’t]] takes some function [[VP]] of type <e,t>, and maps to 1 everything
that [[VP]] maps to 0 and vice versa.13
One final housekeeping detail concerns the case of so-called case-marking
prepositions, as in give the toy to Mitka. We are leaving three-place verbs
somewhat in limbo in the official fragment. But regardless of how these are
treated, it is clear that to functions much like a case marker. That is, one cannot
list the verb give just as selecting for an NP and a PP within the VP; it specifically
selects for a PP with to as the preposition. A verb like talk also selects for a PP
with to, as in talk to the principal. There are a few other

11
There are cases where negation occurs with other kinds of expressions, as in Not
everyone left. But unlike the case of and and or, not is not fully cross-categorial.
Thus (1a) is not good, although we do get things like not Mitka in (ib):
(i) a. *Not Mitka ate the rawhide.
b. Kolya but not Mitka ate the rawhide.
The full range of English negation is rather complex, and so we confine ourselves to VP
negation and introduce it by the single word didn’t.
12
For convenience, we have built the feature R into the lexical specification here to make
sure that this overrides the general rule in (3c) according to which items of category X/X
(i.e., modifiers) take their arguments to the left. In fact all auxiliaries might appear at first
glance to need this kind of lexical stipulation, since they are all of category VP/VP but all
take their VP complements to the right. But actually, the fact that auxiliaries (including
didn’t) take their complements to the right is unsurprising, for strictly speaking they are
actually not of the simple category VP/VP. Each auxiliary specifies a particular
morphological form for the item it combines with. be, for example, is something roughly like
VP/VP[PROG] and so forth. Thus the actual directional feature on didn’t will be specified by
the general word order rules in the predictable way. 13
As usual, we ignore the possibility that some VPs denote partial functions on the
domain of individuals; the usual assumption is that if some individual is not in the domain
of [[VP]] then it is also not in the domain of [[didn’t VP]]. In other words, if an individual is
not assigned a truth value by [[VP]] then [[didn’t VP]] also assigns no value to that
individual. In fact, a standard test for presupposition is that it “survives” negation, meaning
exactly what is said above: if [[Bill]] is assigned no truth value by [[stop smoking]] because
Bill never smoked, then [[Bill]] is also assigned no value by [[didn’t stop smoking]].
 105
prepositions in English which appear in places where they are not interchangeable
with any other preposition. This includes of in mother of Romeo (see section 8.5),
on in cases like rely on the weather, and by in (was) chased by the wild boar.
Syntactically, then, we need to mark the PP as a particular kind of PP. Borrowing
a convention from Gazdar, Klein, Pullum, and Sag (1985) we will add a feature on
such PPs, and—since there is nothing interesting about the choice of name of this
feature—will make it mnemonic by using the name of the preposition as the
feature. (Thus, give selects an NP and also a PP[TO].) The preposition to has the
category PP[TO]/NP. What about the meaning of to, of, etc.? (We exclude the
actual directional and/or locative uses of to.) In section 5.5 it was speculated that
to Mitka means the same as Mitka. This means that the semantic type of the
relevant PPs is type e.14 One might then determine that to is literally meaningless.
But if so, we would need to add a new combinatory rule to the fragment to account
for the semantic part of combining to with an NP. Thus instead of adding a new
rule, it is perfectly straightforward to give to a real meaning: let it denote the
identity function on individuals. Its semantic type is thus <e,e> (exactly as one
would expect, given its syntactic category), and it takes an individual as argument
and returns the same individual. That is the extensional meaning of to (and the
other case-marking preposition); in light of the remarks above its full meaning is a
constant function from worlds to the identity function.

6.7. Summary: The fragment so far

Here and henceforth we list the open-class and the closed-class items

separately. Because we are not dealing with lexical semantics, there is in general

14

The semantics of PPs is one place where one might question the wisdom of assuming
that all items of the same syntactic category have the same type of meaning. Cases like the
ones at hand where the preposition is a case marker appear to be of type e. But PPs also
occur both as noun and as VP modifiers, and even can occur with a few verbs that explicitly
subcategorize for “adverbs” as in examples like word the letter carefully/word the letter in
a careful manner (Bresnan and Grimshaw
106 6. CATEGORIAL GRAMMAR

1978). On the other hand, it may simply be a mistake to think that all of what we call “PPs”
are of the same syntactic category: their distribution does wildly differ.
6.7. SUMMARY: THE FRAGMENT SO FAR

little to say about the meaning of the open-class items. For these, then, we list
only their syntactic category and not their meaning.

Lexicon:
Open-class items:
NPs include: Mitka, Balto, Barack Obama, the-disobedient-husky, the-pig, . . .
S/NP: walk, talk, . . .
(S/NP)/NP: kiss, likes, feeds, . . .
(S/NP)/PP[TO]: talk, speak, . . .
(S/NP)/CP; believe, think, . . .

Closed-class items:
(X/X)/X: and, or
[[and]]
if X is of type <a,t> (for any type a) [[and]] = Curry’ed intersection if X = S,
[[and]] is the following function:
1 1 1
00

0 1

0 0
0

[[or]]: is left as an exercise

VP/RVP: didn’t; [[didn’t]] takes a function that characterizes a set S, and returns the
function characterizing the complement of S
PP[TO]/NP: to; [[to]] = identity function on individuals CP/S; that;
[[that]] = identity function on propositions

Word order rules:

i. ...(S/X)... ii. ...(A/A).... iii. ...(A/B)....

L L R where (i) and (ii) must apply before (iii)

Combinatory rule:
Definition of functions corresponding to categories of the form A/B:

a. FA/RB ([Æ])([]) = [Æ-] b. FA/LB ([Æ])([]) = [-Æ]

 107
R-1 (app):
If Æ is an expression of the form <[Æ], A/B, [[Æ]]> and is an expression of the form
<[], B, [[]]> then there is an expression ª of the form <A/B F([Æ])([]), A,
[[Æ]]([[]])>.
For expository convenience, we often break this into the two separate rules below:
R-1a: If Æ is an expression of the form <[Æ], A/RB, [[Æ]]> and is an expression of the form
<[], B, [[]]> then there is an expression ª of the form
<[Æ-], A, [[Æ]]([[]]>. (call this r-app)
R-1b: If Æ is an expression of the form <[Æ], A/LB, [[Æ]]> and is an expression of the form
<[], B, [[]]> then there is an expression ª of the form
<[-Æ], A, [[Æ]]([[]]>. (call this l-app)

*Possible extension for three-place verbs:

A. If the Wrap analysis is incorrect, simply add to the lexicon items of category
((S/NP)/PP[TO])/NP: give, donate, . . . B. If Wrap is correct: Lexicon:
((S/NP)/NP)/PP[TO]: give, donate, . . . Word
order rules: Add: ((S/NP)/NP)/X

Note: X here ranges over any category. One could also allow it to range over any category
including nothing; then ordinary transitive verbs would also be marked with a wrap slash.
Presumably, though, the effect of wrap in that case would just be vacuous.
Formalizing the Wrap operation:

a. Assume that each expression of the language is a string with an infixation point.
We write such a string as [x|y].
b. Assume that each lexical item in English is a string of the form [x|]. (This isactually
needed only for a small set of items, but we will give it in full generality here.)
c. Adopt a convention by which the infixation point in a new expression is inherited
from the function. This can be done by refining the above definitions of the
functions corresponding to syntactic categories as follows:

Definition of functions corresponding to categories of the form A/B:

Let Æ be some expression whose string is [x|y]. Then:

a. FA/RB ([Æ])([]) = [x|y-] b. FA/LB ([Æ])([]) = [-x|y]

d. Define a function appropriate for A/WB as follows. Let Æ be some expression whose

string is [x|y]. Then: FA/WB ([Æ])([]) = [x|y]

 108 7. THE AUTONOMY OF SYNTAX?

7 The autonomy of syntax?

7.1. Good syntax—bad semantics 108 7.2. Bad syntax—good semantics 111

As noted in the Introduction, this book will at times consider alternatives to the
Direct Compositional architecture both to elucidate comparison and to
familiarize the reader with alternative theories. Much of that discussion is featured
in Part III. Here, however, we pursue a narrower goal: to defend Direct
Compositionality against some challenges that one finds in the literature. Before
proceeding, it is important to tease apart two questions which are sometimes
conflated. The first (which is the subject of this chapter) is the question of whether
the syntax and semantics proceed “in tandem” (as assumed in this text), or
whether instead the syntax computes (i.e., proves well-formed) representations
where these representations are the input to the compositional semantics. Those
representations themselves could just be direct representations of the actual
pronounced sentences, or could be more abstract representations such as
Logical Forms. While the term autonomous syntax has been used in a variety of
ways throughout the history of modern linguistic theory, we use it here only to refer
to the view that the syntax computes representations which are “sent” to the
semantics. The second question is whether what is interpreted—either by a direct
compositional interpretation or via the autonomous syntax approach—is the
actual pronounced sentence or a more abstract level of LF.
All four combinations of answers to these two questions are possible. The
approach in this text combines Direct Compositional interpretation with the claim
that the syntax directly builds—and the semantics interprets in tandem—the
actual pronounced sentence. At the other end of the spectrum is what has come
to be a standard view both within Government–Binding (GB) theory and some
subsequent work in Minimalism. This is that the syntax computes representations
via a sufficiently rich procedure that there is one representation for the
pronounced sentence (i.e., “surface structure”), and a distinct LF which inputs the
compositional semantics. In Government–Binding theory, the hypothesis is that
surface structures are mapped into LFs by a series of rules. Variants of this have
been explored especially within recent work under the rubric of Minimalism, but
these differences are not relevant for discussions in this book. In either tradition
(GB or Minimalism) it is a Logical Form which is assigned the model-theoretic
interpretation, and in both views, the syntactic system which proves well-formed
sentences and pairs them with LFs does not work in tandem with the
modeltheoretic semantic interpretation. So these views maintain LF and also
maintain autonomous syntax (as defined here).
But there are other logically possible combinations regarding the organization
of the grammar. For example, one can imagine a Direct Compositional syntax and
semantics which builds LFs and simultaneously interprets them, followed by
additional syntactic rules mapping those LFs to surface forms. This is similar to
(although not exactly identical to) the architecture assumed under the theory of
Generative Semantics which was explored in the 1960s and 1970s. Thus this view
has in common with the GB/Minimalism view that there is a distinct level of LF,
but it does not entirely divorce the syntactic computation (of LF) from the
semantics, and so does not fit the definition of an autonomous syntax at issue
here. The question of whether LF is needed is addressed throughout Part III of this
book; the goal here is just to answer some classic arguments for the autonomy of
syntax.

7.1. Good syntax—bad semantics

The most famous argument for the autonomy of syntax comes from the earliest
days of Generative Grammar, when Chomsky (1957) constructed the by now
legendary sentence Colorless green ideas sleep furiously as evidence that there
can be syntactically well-formed but meaningless strings. But there is no clear
sense in which the syntax-autonomous model(s) being discussed here and the
Direct Compositional view actually handle such cases in any significantly
different fashion. In the first place, it is probably
7.1. GOOD SYNTAX—BAD SEMANTICS 109

not true that a sentence like this really has no meaning—it just has a rather silly
meaning. The reality of this kind of example is that each of the pieces fits together
in such a way that the combination either involves a presupposition violation
and/or results in something which is massively uninformative. We can illustrate
 110 7. THE AUTONOMY OF SYNTAX?
with just the combination of ideas (which is a plural noun but here functions as an
NP) and sleep, so we can simplify the point with respect to the sentence Ideas
sleep. Given the lexical meaning of sleep, it either is always necessarily false to
say that [[ideas]] sleep, or else that [[sleep]] is undefined for [[ideas]]. So, for
example, it could be [[sleep]] is a partial function of type <e,t>, where [[ideas]] is
not in its domain in any world. Then the value of [[ideas sleep]] is undefined for
every world w. Or, it could be that [[sleep]] is a total function, always mapping
[[ideas]] to 0. Either way, the sentence is uninformative. Thus having the syntax
and semantics work together has no problem accounting for the fact that the
syntax allows these two words to combine, but the semantics predicts that they
will feel like nonsense.
It is especially striking that the autonomous syntax view would say nothing
different. For the words can of course be put together in the syntax, and if syntactic
representations are sent to the semantics, the same basic story would have to be
told. Each word has a meaning, and the grammar has to contain rules (possibly
quite general ones) whose input includes structures of the form [ S NP VP]. These
are needed for things like Sheep graze. And so Ideas sleep fits that description;
each word has a meaning, and so divorcing the syntactic and the semantic
computation leads to no different result from the result given above. Either
[[sleep]] is undefined in all worlds for [[ideas]] or it yields 0 in all worlds. Hence
there is no clear sense in which autonomous syntax has any advantage over Direct
Compositionality for cases like this; indeed both succeed in predicting that the
syntax is fine but that the meaning could not be informative.
A related but slightly different case centers on “Jabberwocky” examples: cases
of material put together from words which obey English phonotactics, look like
they have an obvious syntactic category (due both to their distribution and their
morphology), but also “have no meaning.” The most famous Jabberwocky
example comes from Lewis Carroll’s Jabberwocky poem itself, which begins
’Twas brillig, and the slithy toves did gyre and gimble in the wabe and was
designed to be nonsense. Yet clearly it is syntactically well-formed (did is
capitalized in the original because of the line breaks). For the sake of argument,
suppose that brillig, slithy, toves, etc. are truly meaningless; i.e., that a listener
would have absolutely no idea of what these mean, and therefore no idea of what
each sentence in the entire poem means. Nothing about Direct Compositionality
precludes the existence of a word with phonology, syntactic category, and no
meaning: we simply would need to add in a principle that when such a word
combines with additional stuff the result also has no meaning. Note again that this
is no different under autonomous syntax. These are made-up words, but words
they are (otherwise there is no way that the syntax can put them together). And
since any theory takes words to usually be triples of sound, category, and
meaning, these have to be words with only the first two parts. Hence again, there
is no substantive difference between the theories that leads to any conclusion
concerning the organization of the grammar.
In fact, though, it is not at all clear that this is the right way (under either theory)
to think about such examples. For language users are sufficiently hungry to view
language as an object with meaning that listeners make guesses about the
meanings of these novel words, and come up with some (at least vague and
general) meaning for the whole sentence. They do have some idea of what this
sentence is about. And, strikingly, a speaker of
English would certainly say that the above sentence is not synonymous with
’Twas not brillig, and the slithy toves did not gyre nor gimble in the wabe. But this
judgment would be a complete category error if the sentence were literally
meaningless. (Incidentally, one can construct the same argument for the
colorless green ideas case.) Speakers can also conclude that the Jabberwocky
sentence entails There were slithy toves that gyred and gimbled in the wabe.
Again, this robust judgment about entailment would make no sense if the
Jabberwocky sentence literally had no meaning. For an English speaker who
knows absolutely no Mandarin, a sentence of Mandarin is truly meaningless (and
is not part of English) and the speaker will have no judgments about synonymy
and entailments. Presumably, then, even though slithy is not part of a speakers’
lexicon, s/he will add in this word (perhaps temporarily), making an extension of
English, and will make a guess at its meaning. Thus unlike the Mandarin sentence,
a Jabberwocky sentence is English (or English-extended) and has some sort of
meaning. What examples like this do show is that we can sometimes use
distributional and morphological cues to guess the syntactic category of an item
more easily than its meaning (different listeners will come up with different
meanings for toves but everyone agrees that this is a plural noun), but this does
not bear on the question of the Direct Compositional architecture.
7.2. BAD SYNTAX—GOOD SEMANTICS 111

7.2. Bad syntax—good semantics

The flip side of the argument above is that there are cases which are syntactically
ill-formed but perfectly interpretable. Such examples are easy to construct: they
 112 7. THE AUTONOMY OF SYNTAX?
can involve agreement mismatches as in (1a), the wrong case marking as in (1b),
or even category selection violations as found in (1c). (The interest in (1c) is that
dislike—unlike like—has the surprising property that it cannot occur with bare
infinitive VP complements but only with NPs, including gerundive NPs. Note then
that (1c) becomes fine if we substitute like for dislike.1)

(1) a. *The men is singing. / *The man are singing.

b. *Him is singing.
c. *I dislike to lose at poker. (vs I dislike losing at poker.)

Yet there is no problem understanding these sentences. If the semantics is

divorced from the syntax and very general semantic principles are stated to
interpret syntactic representations, then these principles could conceivably apply
even in cases where the syntactic representation is ill-formed. This, for example,
could be accomplished by translating the type-driven idea stated in Chapter 6 in
the terms of Klein and Sag (1985) into principles for the interpretation of trees; see
Heim and Kratzer (1998) for just such an implementation.
But a closer reflection of the situation shows that both Direct Compositionality
and autonomous syntax once again have essentially the same account of these.
Consider how these are interpreted under the autonomous syntax view. The
hypothesis put forward in the above paragraph is that the rules can apply even
where the syntactic representation is ill-formed. But what syntactic
representation? In order for the interpretive procedure to work under autonomous
syntax, there has to be some syntactic representation to input the semantics—
even if it is a defective one. So an autonomous syntactician must conclude that
speakers have access to a kind of secondary grammar which contains (fairly
minimal) deviances from the actual rules of English. Otherwise we would have no
idea what defective representation is inputting the semantics, and no way to
account for the fact that (1c), for example, is not understood as “I am hungry.” In
some cases the secondary

1
A related set of examples (although not this case) is discussed in Grimshaw
(1982).
grammar would involve slightly altering a lexical entry: in the “Englishish” grammar
that assigns a representation to (1c) the subcategorization restrictions of dislike
are slightly altered. In (1a) the morphology is slightly different; is is the form of be
taking plural subjects. Such representations are then allowed by an Englishish
grammar and can input the interpretive rules since the latter are stated quite
generally.
But once one allows for the possibility that listeners can understand (1)
because they have access to a grammar for Englishish, then the autonomous
syntax view has no advantage over the Direct Compositional view: the exact same
story holds under Direct Compositionality. Once again, (1a) is constructed by
slipping dislike from its English category that allows only for an NP complement in
the VP to its Englishish category that allows an infinitive
VP as complement. 10 Arguably the two items have the same meaning (so no
meaning revision is required). But even if they don’t, it is not particularly difficult
to imagine how a listener might slightly alter the meaning of dislike (analogous to
like). Either theory (autonomous syntax or Direct Compositionality) would need to
say exactly the same thing about the meaning of the Englishish dislike. Under
Direct Compositionality, Englishish is would once again be the form of be that
wants a plural subject. In sum, to account for the interpretation of ill-formed
material both theories need to endow the listener with the ability to slightly tweak
English into Englishish—a skill which speakers surely do have.

7.1. Sometimes the term “autonomous syntax” is used to mean that not
everything about the syntax can be inferred just from the semantics. This claim
is compatible with and indeed endorsed in the program discussed in this text;
syntax does not reduce to semantics. Discuss the ways in which the Categorial
Grammar and more generally the Direct Compositional view elucidated so far
endorses this claim.

8 Adjectives, nouns, determiners,

and more
8.1. Predicative adjectives 113 8.4. Prenominal adjectives and
unary rules
8.2. Nouns (“common nouns”)
116
120
8.3. the 8.5. Relational nouns
118 128

10 In many theories complements like to lose at poker are not VPs but sentences with
hidden silent subjects. But this debate does not bear on the question at hand.
 114 7. THE AUTONOMY OF SYNTAX?
8.6. Summary: The fragment so far 131

8.1. Predicative adjectives

The fragment so far does not include predicative adjectives such as those in (1)
(we return in 8.4 to prenominal adjectives as in the disobedient husky):
(1) a. Mitka is disobedient.
b. Kolya is affectionate.

Of course is disobedient and is affectionate are VPs and so presumably have

(extensional) meanings of type <e,t>. But what is the semantic type of the
adjective itself, and what is the contribution of is? Note too that expressions that
combine with is to form VPs can be complex adjectives, as in fond of
thunderstorms, eager to wash the dishes, etc. For the moment we confine the
discussion to single-word adjectives. A reasonable assumption is that
disobedient, affectionate, etc. themselves denote sets and so—like intransitive
verbs—are of type <e,t> (their intensional type of course is <s,<e,t>>).
This means that simple adjectives and intransitive verbs have the same
semantic type, but have a different syntax (adjectives generally occur with be, do
not show verbal morphology, and similarly do not themselves encode
 115
tense). But while intransitive verbs and adjectives have the same basic type of
meaning, it does not follow that what is lexicalized as an adjective and what is
lexicalized as an intransitive verb is totally arbitrary; there are some rather broad
tendencies (holding for English at least) differentiating the two classes. The most
obvious is that many adjectives express what are known as gradable properties:
one can be happy, disobedient, affectionate, and even fond of thunderstorms to
different degrees. This is an aspect of adjective meaning that we will not discuss
here; see, among many others, Kennedy (2007). Another difference (keeping in
mind that these are only tendencies) is that adjectives often denote sets whose
membership is more stable over time than run-of-the-mill intransitive verbs. In
other words, they tend to denote predicates which are true of an individual for a
relatively long length of time; this is tantamount to saying that the set of individuals
in, e.g., [[disobedient]] will be more stable from moment to moment than the set
of individuals in, e.g., [[howl]].
What about the syntax of adjectives? Because these are of type <e,t> one might
expect them to occur with NPs to give something of type t. Indeed this expectation
is borne out: they occur with subjects in certain environments as, for example, the
complement of prefer:
(2) I would prefer Mitka (a bit more) obedient.

Phrases like Mitka (a bit more) obedient also combine with with to give sentential
adverbial clauses:
(3) With Mitka (a bit more) obedient, he might manage to graduate puppy class.

Such cases have gone under the rubric small clause (Stowell 1981, 1983). The
intuition behind this terminology is that expressions like Mitka (a bit more)
obedient are somewhat like sentences. They are, however, not quite the same as
normal Ss for they never occur as root clauses which means they do not stand
alone. In fact, they occur in a rather limited set of environments. But while they do
have a different distribution from ordinary Ss, they also have something in
common with Ss. For just as VPs take the NP with which they combine to their
left, so do adjectives when giving a small clause. This suggests that perhaps what
we are calling small clauses can be thought of as a feature bundle which has
something in common with S but also has some difference. Let us add in two
features [A] or [V] on S categories. Ordinary “VPs” are thus S[V]/NP while phrases
like obedient or proud of Mary are S[A]/NP. For convenience, we will often refer to
an expression of
116 8. ADJECTIVES, NOUNS, DETERMINERS, AND MORE

8.1. PREDICATIVE ADJECTIVES

category S[A]/NP as an AP, which is exactly analogous to our shorthand use of the
term VP. It should be noted that items of category S[V] generally contain additional
features to account for their distribution. For example, an intransitive verb like ran
(or runs) is not just an S[V]/NP but must contain an additional feature; assume it
is actually S[V,Tense]/NP (for V and Tense two separate features). This is needed
because only Ss of the form S[V,Tense] occur as root clauses. In fact, verbs occur
with a variety of morphological features and VPs such as chose the best
candidate, choosing the best candidate, and chosen the best candidate all occur
in different environments. In particular, each occurs with a different auxiliary. All
of these distributional facts can be accounted for with suitable use of features on
the result S category.
We can now refine the definition of a verb which was given earlier and was
needed for morphological purposes. Presumably a basic lexeme like HOWL is
listed in the lexicon in its bare and underspecified form S[V]/NP, and
morphological rules will map this to more fully specified forms (howls, howling,
howled, etc.). Adjectives undergo their own morphological processes (e.g.,
addition of the comparative suffix -er). So a “verb”—for the purposes of the
morphology—is anything whose ultimate result category is S[V] and an
“adjective” is anything whose category is S[A]. We will henceforth omit the V and
A features whenever not crucial for the discussion. But it is still the case that no
matter how many and which complements a verb takes in order to give a VP, one
can give a common definition of this class in terms of the ultimate result. The
same is true of adjectives; fond is of category S[A]/PP while tall is just S[A], but
they have the same result category and thus have morphological commonalities.
In other words, the traditional parts of speech are reconstructed in terms of the
ultimate result category of each lexeme.
The next task is to discover the category of be and its meaning. The first is
simple: it is (S/NP)/AP, or—in more detail—it is (S[V]/NP)/AP. (Be does have a
broader distribution than just this, but this will do for present purposes.) As to its
meaning, we give it extensionally since (like the case of case-marking
prepositions) its extension does not vary with worlds. Ignoring the contribution of
tense or other contributions of morphology in the actual inflected forms, be
contributes nothing—obedient and is obedient are the same set. This is quite
similar to the situation with case-marking prepositions. Hence [[be]] is the identity
 117
function on properties (i.e., functions of type <e,t>); it takes any such property P
and returns P.
8.1. Take a “universe” with the three individuals {a,b,c,d} and a world w 1 in which
[[Kolya]] at w1 (and at all worlds) is b, and [[disobedient]] S at w1 is {a,c}. Show
the syntactic and semantic composition for:

i. Kolya is disobedient.

8.2. We have given a general syntactic category for and which allows us to
compute things like Kolya is obedient and affectionate. However, in terms of its
meaning, we have seen instances only of what it means when it combines two
VPs or two Ss. Continuing to postpone the question of just how to state the
meaning of and in a fully general way, what is its semantics when two
adjectives are combined as in the case above? 8.3. What type of meaning does
fond have, and what is its category?

This leaves us the question of how to treat adjectives when they occur
prenominally, as in the disobedient husky. But first we turn to nouns and the.

8.2. Nouns (“common nouns”)

What about the type of meaning (and category) of ordinary common nouns such
as dog, cat, table, etc.? While NPs (the dog, Balto) denote individuals, obviously
this is not the right type of meaning for nouns. In fact—as was the case for VPs
and APs—it makes sense to think of nouns as denoting (world/ time-dependent)
sets of individuals. So, for example, [[dog]] is (at a given world and time) some set
of individuals. A noun like dog denotes a relatively stable set across times
(individuals who are dogs rarely jump out of that set into the enchanted frog set).
The extension of other nouns such as [[student]] is more time-dependent. And
interestingly what are sometimes called the basic-level nouns tend to denote
mutually exclusive sets: a dog is not a cat, and a chair is not a table. But this of
course is not true for all nouns, certainly [[professor]] S and [[assassin]]S can have
a non-empty intersection.
There is an interesting and somewhat surprising fact about ordinary nouns. This
is that—unlike VPs and APs—nouns never occur with subjects, even in the
118 8. ADJECTIVES, NOUNS, DETERMINERS, AND MORE

environments that support small clauses. Thus we do not get things like (4) and
(5), all of which have been constructed to be
8.2. NOUNS (“COMMON NOUNS”)

pragmatically plausible. To bring this out, (4a) and (5a) have been surrounded with
some context; the contexts for the (b) sentences seem rather obvious: 1

(4) a. Context: Jim thinks that linguists are people who know lots of languages, and so
when he finds out that his daughter Nora is majoring in linguistics he says:
Oh good. *With Nora linguist, she’ll be able to travel everywhere and speak the
language.
b. *With his new puppy husky, he’s going to have to build a 6-foot fence around his
yard.

(5) a. Context: Same as in (4a), but Jim is disappointed as Nora switches her major to
geology, and says:
*I would really prefer Nora linguist. (or *I would really prefer Nora linguistics major.)
b. *I would really prefer his new puppy labrador retriever.

What this suggests, then, is that—in terms of the syntax—there is just a primitive
category N which includes dog, cat, table, linguist, etc. Thus, despite the fact that
these denote functions of type <e,t>, their corresponding syntax is not something
of category S/NP as we might expect. Nothing in the basic framework that has
been outlined in Chapter 6 requires this; the only requirement is that every
expression of category A/B has as its extension a function from (the extension of)
B-type meanings to (the extension of) A-type meanings. The reverse is not
required (although is admittedly somewhat surprising to find).

1
There is a class of apparent exceptions to this claim; there are nouns such as
president, mayor, queen, etc. which do appear here:
(i) a. With Michelle president of the US, the EPA will quickly be dismantled.
b. He would prefer Hillary president.
But these are actually only apparent counterexamples: it is well known that this class of
“nouns” can also serve as NPs without determiners. They also occur as predicate NPs with
be:
(ii) Sabrina is president of Atlantis/mayor of Providence/queen of Middle Earth.
This shift from nouns to full NPs (without determiners) is restricted to a small set of nouns;
they must denote sets which are (possibly because of cultural and/or political conventions)
singletons. We will not give an analysis of these bare singular noun NPs here; the interested
reader should be able to supply such an analysis by the end of this chapter.
 119
8.3. the

The fragment at this point will include only the single determiner the. We leave
quantificational determiners such as every, some, most, etc. until Chapter 10;
indeed the analysis of expressions like every cat and most students is one of the
most striking achievements in Montague (1973) and will be taken up in due
course. For now, the goal is just to give a semantics for the.
Clearly the is of category NP/N and so we can infer that [[the]] maps a set to an
individual. Thus the dog denotes some individual but dog denotes (the
characteristic function of) a set; and so the (extensional) type of the must be
<<e,t>,e>. But what is its exact meaning: how does it systematically map a set (or,
a function) to an individual? There is no question that when the combines with a
singleton set (i.e., a function that maps one and only one individual to 1) the
individual that it returns is the single member of the set. Consider the most
disobedient husky. Here [[most disobedient husky]]—in virtue of being a
superlative—already denotes a singleton set, and the happily combines with this
to yield the single member of that set. (In function terms, [[most disobedient
husky]] maps only one individual to 1, and that individual is the value of [[the most
disobedient husky]].)
But what is the (extensional) value of the planet that was made of green cheese
if indeed there is no such planet, or the beautiful dog if there is more than one?
The answer to this given by Russell (1905) is that the has a rather complex
meaning such that a sentence like (6) asserts the existence, the uniqueness, and
the baldness of the present King of France, and hence is simply false in our world
and time:
(6) The King of France is bald.

A sentence like (7) would also be false if uttered in 1379 in view of the fact that
there were two popes at that time—even if indeed both were bald:
(7) The pope of the Roman Catholic church is bald.

Strawson (1950) argued instead that both the existence and the uniqueness part
are presuppositions. Thus one might conclude (7) is true for a world/ time pair with
one and only one King if he is bald; false for a world/time pair with one and only
one King and he is not bald, and undefined for all other world/time pairs.
120 8. ADJECTIVES, NOUNS, DETERMINERS, AND MORE

8.3. THE

To decide just how to resolve this controversy, we begin with the observation
that the uniqueness requirement clearly has a rather different status from the
existence requirement. We are happy to utter sentences in which uniqueness is
violated and do so all the time—as long as it is clear to the listener what is being
referred to. There are two senators from Rhode Island, but if Sheldon Whitehouse
(one of them in 2013) raises his hand to be recognized by Joe Biden (the presiding
officer of the senate in 2013), there is nothing odd about Biden saying “The Chair
recognizes the senator from Rhode Island.” Arguably, then, “uniqueness” means
that [[the]] combines with some set and returns the single (or the most)
contextually salient member of that set.2 Thus it might be that the meaning of
[[the]] is a function from speech contexts to (ignoring worlds and times) a function
of type <<e,t>,e>. Chapter 2 noted other instances of words which require access
to the speech context (in particular, “indexicals” like I, you, etc.). The reason (7)
seems deviant is that it needs a richer context to select one of the popes; if such
a context were supplied there would be nothing deviant about (7). To formally
build the contribution of speech contexts into the semantics at this point is too
lengthy a project, and so will not be attempted here. A different approach to
narrowing down the set accessed by the is discussed later under the rubric of
domain restriction; see especially section 17.4.
But this still leaves the question of how to treat the existence requirement:
i.e., should the theory predict, with Russell, that (6) is false in our world and time
or, with Strawson, that it is undefined? Little in this book hinges on the decision,
but when needed we will take the position that [[the King of France]] at a world w
and a time i has no value (i.e., is undefined) if there is no King of France. We will
also assume that if the value of some expression is undefined at w and i then every
expression in which it is embedded also has no value at w and i. 11 Hence if [[the
King of France]] is undefined at our world and time, then [[the King of France is
bald]] is also undefined. 2 Section 17.4 contains further remarks of relevance, related to
the observations made in section 2.2. Recall that it was pointed out there that even

11 This is a huge oversimplification as will be obvious by a casual glance at the literature

on “presupposition projection.” Interested readers can consult Levinson (1983) for an
introduction.
 121
sentences like Every student got an A does not mean every student in the world, but only
those from some salient subset. The situation with the is similar: when Susan says “The
dog escaped” there is a contextually salient subset of dogs which is a singleton.

8.4. Prenominal adjectives and unary rules

8.4.1. An additional binary rule?

Section 8.1 dealt with predicative adjectives like tall. But what are traditionally
called “adjectives” also occur prenominally as noun modifiers (and they can
iterate):
(8) the disobedient husky; the intelligent seal; the intelligent obedient friendly elephant

In fact, there is a striking (near) generalization: any single word AP that can appear
predicatively can also appear prenominally.12 The situation is even more general:
any AP—even when complex—can appear as a noun modifier. But the complex
ones (eager to win the primary, happy that the economy did so badly) occur
postnominally.5 This is exemplified by NPs like the subjects in (9):
(9) a. The candidate eager to win the primary said something very stupid.
b. The candidate happy that the economy did so badly will pretend to beunhappy.

There are two broad classes of analyses for these cases. The first adds a new
binary combinatory rule. Rather than spell it out in the official notation, we give
first the syntax (and phonology) part as a phrase structure rule. We state this here
as technically two rules, but a more sophisticated treatment could collapse them:
(10) a. N ! AP N (where AP consists of a single word)

b. N ! N AP (where AP consists of more than one rule)

What is the associated semantics? It cannot be functional application: the types

are not right. Rather, a sensible semantics is intersection; red book is

12
There is a small class of exceptions. One is the adjective content (see Pope 1971).
Most of the others themselves form a class in that they all begin with the same bound
morpheme: awake, asleep, adrift, afloat, aghast, etc. 5 The discussion here and the rules
formulated later oversimplify the situation a bit. There are single-word adjectives that can
appear both pre- and postnominally where the different positions actually have different
semantic effects. See, e.g., Bolinger (1967).
 122 8. ADJECTIVES, NOUNS, DETERMINERS, AND MORE

the set characterizing everything that is both red and is a book. Thus, continuing to
use the space-saving notation, we write the full rule as follows: TR-9. N1 ! AP(single
word) N2 or N2 AP(complex); [[N1]]w = [[AP]]w \ [[N2]]w

Incidentally this has been phrased in Direct Compositional terms, but this general
type of solution is common in autonomous syntax views as well. Some
principle(s) of syntax will allow AP to be a sister to N (where single word adjectives
occur on the left), and some principle(s) will take the trees licensed by this rule
and interpret them by intersection. See, e.g., Heim and Kratzer (1998) for a version
of this.

8.4.2. Interlude: Unary rules (“type shift” rules;

morphological rules)

There is a second way to account for the noun modification cases of APs. This is
to adopt a rule which maps any AP into something of category N/N with no
phonological effect. To flesh this out, we digress to talk about the more general
notion of a unary rule, a concept which is often found in the literature under the
rubric of a type shift rule. Hence, recall that a linguistic expression is a triple of
<sound, category, meaning> and the rules (or rule schemata, or principles) in the
grammar take one or more triple as input and return a triple as output. The
fragment so far contains only binary rules: rules whose input is two expressions.
But there is no reason not to expect unary rules—i.e., rules whose input is a single
expression. Moreover, it is quite possible that sometimes the output expression
will be identical in one or more of the pieces of the triple to the input expression.
Such rules are actually common in discussions of morphology, although not
necessarily phrased in these terms. But note that (at least some) morphological
rules can be thought of as unary rules. They take as input a single triple (which
happens to be just one word or possibly what we are calling a lexeme), and yield
as output a word with a phonology which may or may not be different from that of
the input, a syntactic category which may or may not be different from that of the
input, and a meaning which usually is (though presumably need not always be)
different from that of the input.6
 8.4. PRENOMINAL ADJECTIVES AND UNARY RULES 123

6 The typical discussion of morphological processes in introductory linguistics textbooks

often conflates these three aspects of a morphological rule. For example,
While the output syntactic category can be the same as the input (as can be the
phonology) usually there is at least some fine-grained difference, as for example
the case of mapping singular nouns (or lexemes unspecified for number) to plural
nouns. Consider for example reduplication in the Philippine language Ilokano
(data from Hayes and Abad 1989). Noun plurals are formed by what Hayes and
Abad refer to as “heavy reduduplication,” meaning that the first consonant (if there
is one), the next vowel, and the next consonant of the noun stem are reduplicated,
as in [kaldíŋ] (goat)— [kalkaldíŋ] (goats); [púsa] (cat)—[puspúsa] (cats); and so
forth. (For those familiar with phonology, this means that the first heavy syllable is
reduplicated.) Pursuing the idea that lexemes are underspecified, this can be cast
as a rule whose input is any triple with a phonology, a syntactic category—
presumably N with no number feature—and a meaning of type <e,t>, whose
domain is the set of both singular and plural individuals. The output has the new
phonology described above, a new category which is N[PLURAL] and the
semantics is the function whose domain is only plural individuals. Of special
interest is also the fact that there are also morphological processes that do not
change the phonology (these are sometimes called zero derivation in traditional
works on morphology) but do change category and meaning. There is no reason
not to assume the existence of unary rules like this whose inputs are larger
expressions than just single words. The ones of relevance for the present
purposes are those which (like zero derivation in morphology) have no effect on
the phonology.
Before returning to the case of adjectives, it is worth making two terminological
points. First, those rules that map single words into single words are sometimes
referred to in the literature as lexical rules. We avoid that terminology here, and
reserve the term lexicon for that which is listed. Hence, whenever there is a
productive rule mapping one term to another,

one often finds a list of morphological processes which includes things like “reduplication,”
“suffixation,” “(noun) incorporation,” and “compounding.” The first two are descriptions of
what the phonology does. Noun incorporation generally refers to both the syntactic and
semantic effect of a process (the phonological process is simply the concatenation of two
morphemes), and “compounding” usually refers to a syntactic process (whose semantic
 124 8. ADJECTIVES, NOUNS, DETERMINERS, AND MORE

reflex differs from case to case). This kind of taxonomy obscures the fact that every
morphological process can be seen as a mapping from a triple to a triple.
the term lexical rule is an oxymoron—a form is either given by a rule or it is listed
in the lexicon. Second, it is also the case that in much of the semantics literature,
unary rules with no phonological consequences are often referred to as type shift
rules. This too is a term that will be avoided here. For one thing, this terminology
suggests rules whose output differs from the input only in meaning and semantic
type and not also in syntactic category. Given the premises of Categorial Grammar
as adopted here, any time an expression Æ is mapped to an expression with a
different type of meaning, will also have to differ in syntactic category from Æ.
Moreover, there is also nothing precluding a case where the output differs in
meaning without differing in type. The morphological rule prefixing un- to certain
adjectives is a case in point: unhappy obviously means something quite different
from happy but there is no reason to believe that the two differ in type or in
category. Type shift rules are thus just a special case of rules whose output differs
in meaning from the input, but have no privileged status here.7
Finally, note that any unary rule which does not change the phonology can
always be recast using what is often known as an empty operator which, in the
terms here, would be a lexical item with a syntactic category and a meaning but
no phonology. Given any unary rule which takes an expression Æ of the form <[Æ],
X, [[Æ]]> and returns an expression of the form <[Æ], Y, f([[Æ]])> where f is the
semantic operation performed by the rule, one can repackage this as a lexical item
of the form <[], Y/X, f>, where combines with Æ to give the result. (This is of course
phrased in CG terms but can be implemented in any other theory.) There is little,
if any, reason to choose between the unary rule and the silent lexical item
approach; they

7
We also note that we will eschew a slogan common in some literature to the effect
that “an item may type-shift only as a last resort.” This slogan makes sense only under the
autonomous syntax view whereby a syntactic representation is computed and fed to the
semantics. Under that conception, the “last resort” idea is that the rules of the grammar
work in such a way as to only combine the basic meaning of two sisters to give the value of
the mother. However, in a case where there are two sisters whose types are such that there
is no rule allowing these meanings to combine, there could be a rule allowing one to map
to a different type (and thus a different meaning). Under the architecture of the grammar
proposed here the “type-shift as a last resort” slogan makes no sense. But even under the
 8.4. PRENOMINAL ADJECTIVES AND UNARY RULES 125

autonomous syntax view this slogan is potentially quite misleading; it suggests that the
grammar has needs or goals, which of course it does not.
ultimately may just be different metaphors for the same thing. 8 In this text we will
adopt the phonologically inert unary rule solution, but should one prefer the silent
lexical item approach, readers are free to recast all of the unary rules in that way.
Here the term unary rule can be seen as a cover term for a class of solutions that
include unary rules or silent lexical items.

8.4.3. Back to adjectives

With this apparatus, it is easy to recast the new binary combinatory rule from
section 8.4.1. The rule is as follows, and will be referred to as mod:
R-2. For any expression Æ of the form <[Æ], AP, [[Æ]]> there is an expression of the form
<[Æ], N/N, [[]], where [[]] is a function (of type <<e,t>,<e,t>>) which takes a function
f and returns the function that characterizes the intersection of fS with [[Æ]]S>.
If [Æ] consists of a single word then the result is N/RN.

Note that R-2 covers the case of both complex APs and single-word APs, and
builds in the exceptionality of the single-word case. Those new items not covered
by the last clause in R-2 will automatically be marked as N/LN by the general word
order rules. The semantic part is stated awkwardly—a consequence of trying to
phrase complex model-theoretic objects purely in prose. The next chapter
remedies this. (One might notice that operation that maps the meaning of the
input expression to the meaning of the output
expression is essentially Curry’ed intersection.)

8
There is a tendency in some modern linguistic theories to favor theories which keep
the rules as simple as possible and put all “clutter” into the lexicon, and one sometimes
sees this used as an argument for the silent operator approach (it is, after all, a lexical item,
albeit a quiet one). But since the two solutions are arguably equivalent, the preference for
a silent lexical item over a unary rule seems irrational. Pursuing this further, one sometimes
sees claims to the effect that relegating all crosslinguistic variation and irregularity to the
lexicon helps explain how acquisition can take place. But there is no obvious reason to
believe that this hypothesis makes language acquisition any less mysterious (at least not if
there are silent items). Invoking a silent lexical item goes no further towards explaining how
a child could learn the relevant phenomenon than does invoking a unary rule, for the learner
has no more direct access to a silent lexical item than to a phonologically inert unary rule.
 126 8. ADJECTIVES, NOUNS, DETERMINERS, AND MORE

By way of example, take a universe with five individuals {a,b,c,d,e}, and a world
such that the lexical item disobedient (of category AP) has as its extension the
function mapping a,d, and e to 1 and b and c to 0. [[husky]] at the world in question
maps a and b to 1 and the others to 0; [[lab]] maps c, d, and e to 1 and the others
to 0. The derivation of disobedient husky, showing how the extensions are
computed, is shown in (11); we use the label “disobedient 1” to indicate the version
of that phonological form listed in the lexicon, and “disobedient2” to be the
homophonous form derived by R-2. Recall the notational conventions whereby for
some set S, Sf is the function characterizing that set, and for a function f, f S is the
set characterized by f. (Sf can be computed from S only if there are no partial
functions; in our example here there are none.)
(11) disobedient1, AP, function characterizing {a,d, e} listed in lexicon disobedient2,
N/LN, [[husky]] ! {a}F;, [[lab]] !{d,e}F by mod disobedient2 husky, N, {a}F
by l-app

The important step to understand here is the mapping from the meaning of
disobedient1 to the meaning of disobedient2. Put completely in set terms, the
latter takes any set, and returns the intersection of that set with the set
(characterized by) [[disobedient1]].

*8.4. Recast mod using the silent lexical item approach.

Is there any reason to prefer the unary rule (or silent lexical item) solution over
the binary rule solution given in 8.4.1? Each adds a rule, so neither is obviously
simpler. Moreover, while the rule in R-2 accounts for both the case of single-word
prenominal APs and complex post-nominal ones, we cannot really claim victory
for R-2 on that basis because we simply did not explore ways to make the
appropriate generalization for TR-9. There is, however, a difference between the
two solutions. The unary rule solution composes material like disobedient husky
in two stages and thereby predicts the existence of a new expression disobedient 2
(in this case of category N/LN). If there turns out to be some evidence that this is
an expression of the language, there would be evidence for the unary rule R-2 over
the new binary rule in (TR-9). And indeed there does seem to be a bit of evidence
in this direction, which is left for the reader to discover in the next exercise.
 8.4. PRENOMINAL ADJECTIVES AND UNARY RULES 127

8.5. Consider the NP the happy and honest candidate. Does the existence of
this NP provide evidence for the rule in R-2 over the solution in 8.4.1? You
should just concentrate here on the syntax of this expression. Since we have
not yet given a fully general definition for and, the tools are not yet available to
investigate how the semantics would work in each case, but this will be
revisited in Exercise 12.2. Note: The full set of relevant facts is actually a bit
complex (and the relevant judgments not entirely clear), so this question is
somewhat open-ended, but there is a clear difference in the predictions made
by each solution.

There are, moreover, other analogous cases in which there is evidence for the two-
step unary rule solution (see, e.g., Jacobson 1999). Hence we will generally prefer
this strategy in this book over adding new binary rules. But one should keep in
mind that in any given case there is often no strong evidence one way or another,
and in some cases the choice might truly might boil down to a matter of taste.

8.4.4. Are all noun modifier adjectives intersective?

Whether one ultimately adopts the binary or the unary rule solution, the claim is
that the semantics for something like red ball involves intersecting the set of red
things with the set of balls. But are all adjectives intersective in this way? After all,
big seems to have a very different value in a phrase like big flea than in a phrase
like big elephant. It is easy to conclude from this that
[[big]] does not pick out any kind of set but rather that [[big]] combines with a noun
to give the set of things that are big in comparison to other things in the noun set.
If this is correct, the semantics is not intersection; prenominal [[big]] would be a
function that maps a set to the subset of things that are big for that set.
The reasoning above turns out to be partly right and partly wrong. It is true that
[[big]] by itself does not pick out a set: it needs to be given a comparison class in
order to return the appropriate set. However, one can show that the existence of
the comparison class argument is independent of the question of how [[big]]
combines with the head noun. For the comparison class is needed for “gradable”
adjectives in any position (see Kennedy 2007 for much discussion). Thus even in
predicative position a comparison class is implicit; recall the discussion in
 128 8. ADJECTIVES, NOUNS, DETERMINERS, AND MORE

Chapter 2 centering on (4) Magic Johnson is tall. Here too one cannot really
determine the truth without fixing the comparison class for tall. When such an
adjective appears prenominally, there is a strong pressure for the comparison
class to be the one set by the noun: we tend to understand a tall basketball player,
for example, to mean someone who is tall for a basketball player. But this is not
necessary—it can refer to someone who is quite short for a basketball player but
tall compared to the general population. Similarly, let Dumbo be very small
compared to other elephants, but still pretty big in comparison to the animals we
usually see. We could point to Dumbo and say “Boy, look at that big elephant.” In
other words, the situation with prenominal adjectives is no different from the
situation with predicative adjectives: both need a comparison class. The
difference is only that in the prenominal case there is a strong tendency (perhaps
due to processing considerations?) to fix the class by the noun.
It is true, then, that the meaning of gradable adjectives is more complex than
given here. Such an adjective must always be supplied a comparison class in
order to pick out a set. Without giving a full formal account at this point, we adopt
a rough-and-ready account by hypothesizing that these all contain a free variable.
(The semantics for variables is elucidated in Chapters 9 and 15, and an alternative
way to model [[big]] without a free variable is discussed in section 15.5.2.) But for
now the informal intuition is clear enough: [[big]] means larger than average for an
X, where X picks up some contextually salient set. This is its lexical meaning; the
meaning of the homophonous noun modifier is exactly the elaborated version of
this that results from the application of R-2. The intersection semantics embodied
in R-2 remains, but the actual meaning of the lexical item big has this extra
parameter. For expository ease, we continue to ignore the extra parameter for the
comparison class until section 15.5.
Incidentally, it is not the case that all things that one might call prenominal
adjectives are intersective; things like former as in former senator crucially
combine with the intension of senator. It is, therefore, not intersective (there is no
set of things that are “former”), and it also (unsurprisingly) has no predicative
counterpart. (For that reason former is not traditionally called an “adjective,” but—
like prenominal tall—its syntax is indeed N/RN.) Former and related words are
briefly explored in Chapter 19.
It is also the case that superlative adjectives such as tallest are crucially
nonintersective (the set of tallest linguistics majors is not the intersection of two
 8.4. PRENOMINAL ADJECTIVES AND UNARY RULES 129

sets). Nor do superlatives occur predicatively (at least not without a fair amount
of contextual support), thus Nora is tallest is strange. 13

8.6. Return to the discussion in section 1.1.1 surrounding the sentences in (1)
in that chapter. To simplify the discussion, we took the tallest as a single lexical
item. You should now be able to go through the compositional semantics of
(1a) in somewhat better detail to show how it is put together. Of course, we
have given no semantics for superlative adjectives so your analysis will not be
complete.

8.5. Relational nouns

This chapter concludes with one more type of category: relational nouns. These
are nouns that take PP complements (where the preposition is always of ) as in:
(12) the mother of Romeo; the sister of Romeo; the wife of Bert; the author of Waverley; the
writer of that letter; the builder of the Acropolis; the mayor of Providence; the
student of Zellig Harris . . . )

Notice that expressions like mother of Romeo already denote singleton sets, but
this is not a necessary fact about relational nouns plus their PP complements
(witness the sister of Romeo: use of the here of course requires a unique
contextually salient sister but sister of Romeo itself need not be a singleton). And
these can of course occur with quantificational determiners (every student of
Zellig Harris, no friend of Sabrina, etc.). Note also that some of these are kinship
terms (mother, sister, etc.), but the class is certainly not restricted in this way. It
includes many other nouns (friend, colleague, student) and includes many forms
that are nominalized forms of transitive verbs (writer, builder, etc.).
Chapter 6 provided an account of those prepositions that function essentially
as case markers; of here is one of these. Assume that of Bert (in the wife

13 We do get sentences like Nora is the tallest, where there is an implicit set (“the
tallest of some set”).
 130 8. ADJECTIVES, NOUNS, DETERMINERS, AND MORE

8.5. RELATIONAL NOUNS

of Bert) is of category PP[OF] and of of category PP[OF]/NP. Assume further that—as

with the case of case-marking prepositions in general— [[of]] is just the identity
function on individuals. Thus mother of Romeo, wife of Bert, student of Zellig Harris,
and so forth themselves are complex nouns and have a syntactic distribution which
is the same as any simple noun (they occur with determiners like the to form an NP),
and like other nouns they cannot take subjects:
(13) *With Krugman writer of today’s op ed piece, you’ll have to get the New York Times.

Syntactically, then, the relational nouns are of category N/PP[OF]. Notice that—as
discussed in Chapter 6—the traditional morphological category of “noun” is defined
in terms of the result: a “noun” is anything whose final result is N. The morphology
needs to know about this since even though wife (of) and table are quite different
syntactically (wife taking a PP complement and table being an intransitive noun) both
of them take plural morphology.
Their semantic type must therefore be <e,<e,t>>. This comes as no surprise; the
very name “relational noun” suggests that these denote a two-place relation between
individuals. As with transitive verbs, English syntax treats this relation as Curry’ed;
unlike the case of transitive verbs the <e,t> function that results from applying a
function like [[daughter]] to its first argument never (directly) gets its second
argument.

8.7. Add to our fragment the following lexical items, whose extension at a world
w1 will be shown with the items. We are ignoring the fact that honest must be
supplied a comparison class, and treat it as a simple function of type <e,t>. The
“universe” will consist of four individuals
{a,b,c,d}.
<honest, AP, a 0, b 0, c 1, d 1.
<Lear; NP, d>
<daughter; N/PP[OF]; d a 1 b 1 c 1 d 0

all other individuals (a, b, and c) are

mapped to functions mapping all
individuals to 0.

Give the full syntactic and semantic composition of (i):

(cont.)
 131
(i) the honest daughter of Lear

In giving the semantics, show the types and also show the extension of each
expression at w1.

Of course many (all?) of these items can also occur without the PP complement
as ordinary intransitive nouns. Whether there is a predictable rule mapping items of
category N/PP[OF] into N with a regular semantic effect (and with no phonological
effect) or whether they are just listed as separate lexical items will be left open here.
In at least some cases the corresponding noun has additional aspects of meaning.
Thus, Barker (1995) notes that while child as an intransitive noun encodes youngness
(a child is necessarily young), child of does not. At the time of the writing of this book,
Chelsea Clinton is approximately 32 years old and hardly a “child,” but is still the child
of Hillary Clinton.
Finally, note that for many of these it is more natural to use constructions like
Romeo’s mother, Sabrina’s friend rather than the form with the of-PP.
These look much like constructions with ordinary genitive (or “possessive”)
NPs with normal intransitive nouns, as in Dorothy’s dog, Sarah’s key, etc. For the latter,
though, the semantic relationship between the noun (dog) and the genitive NP
(Dorothy’s) is completely free: Dorothy’s dog really means the dog that stands in
some contextually salient relation to Dorothy. (It could, for example, be the dog that
Dorothy is in charge of feeding at the local shelter; it could be the dog that was chasing
her in a scenario of several individuals each being chased by a different dog, etc.).14
For the case of relational nouns, however, the genitive can fill the argument slot
normally filled semantically by the OF-PP, thus Romeo’s mother means the same as
the mother of Romeo.

14This is similar to the case of noun compounds as in pumpkin bus, detailed in Downing
(1977), where the relation between the two nouns is completely free and specified by context.
A wonderful example illustrating the noun compound case was the term balloon boy coined
by newscasters in 2009 after a boy named Falcon pretended to be caught in a hot-air balloon
that was floating for miles. The entire episode turned out to be a publicity stunt engineered
by the parents. Balloon boy thus came to refer to a boy who was not in a balloon (the
contextually salient relation between boy and balloon being, perhaps, “pretending to be in”).
With a bit of imagination, one can similarly imagine the phrase Falcon’s balloon in the same
context.
 132 8. ADJECTIVES, NOUNS, DETERMINERS, AND MORE

8.6 SUMMARY: THE FRAGMENT SO FAR

(Since many of these relational nouns can also occur as ordinary intransitive nouns,
the genitive can also be similar to cases like Dorothy’s dog. Take, for example, an NP
like Bill Clinton’s child. This can refer to Chelsea Clinton. However, it of course can
also be analyzed as intransitive child. Imagine that we are in the context of a daycare
center at which Bill Clinton is employed. Each employee is assigned one child who
s/he must help with their coloring project. In that scenario, we can talk about Bill
Clinton’s child—note that, predictably, the referent of this NP must be young. (See
Partee and Borschev 2003 for much relevant discussion about the two kinds of
genitives.)
We will not deal with the cases like Dorothy’s dog (for interesting discussion see
Storto 2000), but it will be useful as we further construct our fragment to have an
analysis of the relational noun case (e.g., Romeo’s mother). Although not the most
elegant solution in the world, assume that all relational nouns can shift their syntactic
category and semantics by the following unary rule:
R-3. If Æ is an expression of the form <[Æ], N/PP[OF], [[Æ]]> then there is an expression of
the form <[Æ], NP/LNP[GEN], [[]] is a function of type <e,e> such that for all x, [[]](x) is
the unique (contextually salient) individual y such that [[Æ]](x)(y) is true>

(The semantics can be stated much more concisely after the next chapter.)

*8.8. As of yet, the system has no way to give an expression like Lear’s honest
daughter. See if you can see why not. You will be able to return to this after section
13.4.

8.6 Summary: The fragment so far

To the fragment in Chapter 6, add the following:

Lexicon:
“open-class” items
S[A]/NP; all of the intransitive adjectives like honest, red, tall, . . .
(S[A]/NP)/PP[OF]; fond N; table,
dog, cat, . . .
N/PP[OF]; mother, child, daughter, employer, . . .

“closed-class” items
<[of], PP[OF]/NP, identity function on individuals>
<[the], NP/N, the function mapping a set of individuals to the unique (or most) contextually
salient member of that set>
 133
<[is], S[Tense]/(S[A]/NP), identity functions on properties (i.e., on sets of type <e,t>)> Note: we
have given the tensed form here. The more general form of be might be S[V]/(S[A]/NP)
depending on how exactly the feature system is worked out.

Rules:
R-2. For any expression Æ of the form <[Æ], AP, [[Æ]]> there is an expression of the form <[Æ];
N/N; [[]] is a function (of type <<e,t>,<e,t>>) which takes a function f and returns the
function that characterizes the intersection of fS with [[Æ]]S>.
If [Æ] consists of a single word then the result is N/RN.
(Call this mod)
R-3. If Æ is an expression of the form <[Æ], N/PP[OF], [[Æ]]> then there is an expression of
the form <[Æ], NP/LNP[GEN], [[]] is a function of type <e,e> such that for all x, [[]](x) is
the unique (contextually salient) individual y such that [[Æ]](x)(y) is true>.
 9 Interlude: The semantics of
variables and the lambda calculus
9.1. First-order predicate logic 9.3. Naming model-theoretic
(modified) 134 objects 151

9.2. The lambda calculus 144

This chapter detours from the main project of building a semantics (and syntax)
for a fragment of English. Instead we pause to build a syntax and semantics for
two languages that look very different from English. The first is basically (with
modifications) first-order predicate logic which includes a treatment of
quantifiers. The second takes the first language and enriches it to what is called
the lambda calculus.
Since neither of these looks very much like English, one might wonder why
engage in such a project. This might seem especially puzzling because we will see
in the next chapter that English quantification is rather different from the treatment
of quantifiers in first-order logic, so why bother with the latter? There are two good
reasons to digress and build a semantics for these artificial languages. The first,
and most immediate, is that the ultimate language developed in this chapter (the
lambda calculus with quantifiers) provides a very useful way to name complex
model-theoretic objects. The reader has undoubtedly noticed that some of the
semantics in the preceding chapters has been stated quite awkwardly. This is fine
as long as it is clear: the language used to name model-theoretic objects is of no
consequence for the theory and is only a tool for the linguist (the grammar itself
simply pairs expressions with model-theoretic objects). But once the objects get
sufficiently complex (e.g., functions with functions as their domains as well as co-
domains) it is useful to have a simple way to write out these meanings without
excruciating and/or ambiguous prose, and the lambda calculus allows us to do
just this. We will, however, take the unusual step of trying wherever possible to
also write these out in prose (however excruciating) as a reminder that these name
 9.1. FIRST-ORDER PREDICATE LOGIC (MODIFIED) 135
actual model-theoretic objects and that the formulas have no theoretical
significance.
The second reason for a detour into the semantics of these artificial languages
is that the languages introduced here contain variables, and the semantics for the
interpretation of variables translates rather directly into a standard view of the
semantics of natural language pronouns. Variables (or their analogue) are also
central under one common view of the semantics of relative clauses. But a caveat
is in order: Part III develops two competing analyses of pronouns and relative
clauses. The view using variables is embedded in an autonomous syntax
approach, while the Direct Compositional alternative developed in Part III actually
makes no use of variables at all (except as a notational device). And since this
book is committed to direct compositionality, one might wonder why spend time
discussing variables in terms of the theory of natural language. The reason is that
the use of variables in semantics is so widespread that understanding the
semantics of variables is a matter of basic semantic literacy. But in any case, it is
time to develop more formally a metalanguage that we can use to name
modeltheoretic objects, since trying to continue with just English as the
metalanguage would become more and more difficult.

9.1. First-order predicate logic (modified)

9.1.1. Variables

To depart minimally from what has been done so far, we begin by building a
language a bit more English-like than is usual in treatments of first-order logic. We
adopt a Categorial Grammar syntax, but with only S and NP as primitives. The
correspondence between syntactic category and semantic type remains the
same, although intensions play no role here. Hence expressions of category S are
of type t, expressions of category NP of type e, expressions of category S/NP of
type <e,t>, etc. There is no use of directional features on the slash as the syntax
will put all functions before the arguments (making the language verb–object–
subject in terms of word order). The noun phrases are all simple with no internal
structure. To distinguish the expressions from actual English we notate them as
feed’, Fido’, etc.
Lexicon:
 136 9. VARIABLES AND THE LAMBDA CALCULUS

NP: Fido’, Porky’, Dumbo’, . . . (value of each is f, p, d, respectively)

S/NP: walk’, chew-gum’, squeak, . . . (each of which denotes some function in type <e,t>)
(S/NP)/NP: kiss’, kill’, feed’, . . . (each of which denotes some function in <e,<e,t>>)

Rule (to be revised directly): We number these NER-n (“Not-English Rule”) and since this
rule is temporary it earns a T.

TNER-1. Given an expression Æ of the form <[Æ], A/B, [[Æ]]> and an expression
of the form <[], B, [[Æ]]>, there is an expression ª of the form <[Æ()],
A, [[Æ]]([[]])>.

(Note that we have introduced parentheses into the phonological string in order to
make things easier to read. This may seem odd, but since there is no sense in
which this language actually has a real phonology, it is harmless.) Thus, TNER-1
combined with the lexicon above gives sentences like squeak’(Porky’), chew-
gum’(Dumbo’), feed’(Porky’)(Dumbo’), etc. It will also be useful to have rules
introducing the logical connectives especially since these will be convenient
when this language is used for writing meanings. We give the rules in phrase
structure rule:
NER-2: S1 à S2 & S3; [[ S1]] = 1 iff [[S2]] = 1 and [[S3]] = 1
NER-3: S1 à S2 v S3; [[ S1]] = 0 iff [[S2]] = 0 and [[S3]] = 0
NER-4: S1 à S2 ! S3; [[ S1]] = 0 iff [[S2]] = 1 and [[S3]] = 0
NER-5: [[S1]] = 1 iff [[S2]] = 0
S1 à ~ S2 ;
So far this is little more than a stripped-down version of the English fragment
built earlier, with a strange word order. But the first main difference is that we now
add in variables. So, in addition to Fido’, Porky’, Dumbo’, etc., the language also
contains an infinite number of NPs, each of which is an expression of the form x i
for some positive integer i. In order to give the semantics, we introduce the notion
of an assignment function— or, for short, an assignment. An assignment (function)
is a function which takes each variable name and maps it to some object. The set
of all assignments will be called G, and g will be used to refer to an arbitrary
assignment function. The semantics thus now has an extra dimension from what
we have seen so far. A variable name does not have a semantic value in and of
itself; it has a value relative to some way to map it to some object. In particular, for
any g, the value of xi on g is g(xi). Another way to say this is that the meaning of x i is
not an individual—it is a function from assignments to individuals (where the
assignment itself is a function from variables to individuals). Thus another way to
write what was said above is as follows: for any g, [[x i]](g) = g(xi). As is the case
with world and time parameters, the convention in much semantics literature is
 9.1. FIRST-ORDER PREDICATE LOGIC (MODIFIED) 137
to write the g as a superscript, thus [[x i]]g = g(xi) (where there is a hidden “for all g”
in front of this formula).1
The syntax now gives expressions like sings’ (x 3). So far, we have an
interpretation for the NP x3 but not for the whole sentence. Exactly the technique
used earlier for world and time arguments will be used here. Thus every
expression in the language actually has a value only relative to an assignment. Put
differently, the meaning of any expression is a function from the set of
assignments to whatever you thought its type of meaning was before reading this
chapter. NPs all have as their meanings functions from G to e. Ss denote functions
from G to t, and so forth. But the semantics is constructed in such a way that many
expressions denote constant functions from the set of assignments. This is true,
for example, for all of the lexical items above except the variables. Thus
[[Dumbo’]] is not d, but is a function

1
This particular definition of an assignment introduces an odd status to a variable: it is
both a syntactic expression and a member of a set which forms the domain of an actual
function used in the semantics. If we wish the semantics to be concerned with model-
theoretic objects only, then we have introduced variable names (a syntactic object) into our
toolkit of model-theoretic objects; a rather unnatural situation. This, however, is just an
artifact of the particular definition adopted above. One could revise this to take an
assignment to be a function from the set of positive integers (clearly unobjectionable
model-theoretic objects) to the set of individuals, such that for any g, [[x i]](g) = g(i). We will
not adopt this convention in the remainder of the discussion because section 9.2
introduces variables over other types besides just individuals. We would then have to
invoke fancier objects like ordered pairs of integers, ordered triples, etc. in the domain of
the assignment functions; this would be needed to distinguish all of these different kinds of
variables. All of that can be done but is unpleasant to read. We thus stick to the convention
of having an assignment be a function from the variables themselves to other objects.
from G such that for all g in G, [[Dumbo’]](g) = d. [[walks’]] is similar; it assigns
every g in G some function in <e,t>. Hence [[walks’]] is a function in <G,<e,t>>. (Of
course once we return to English, whose expressions denote functions from
world–time pairs to something else, we have one more layer in the functions:
[[walks]] is a function from worlds, times, and assignments to a function of type
<e,t>. It is constant on the assignment, but not on the world and time arguments.
But the language here is purely extensional.)
The semantics of TNER-1 is now to be revised in the obvious way, and we now
rename this as NER-1.
 138 9. VARIABLES AND THE LAMBDA CALCULUS

NER-1. Given an expression Æ of the form <[Æ], A/B, [[Æ]]> and an expression of the form
<[], B, [[]]>, there is an expression ª of the form <[Æ()], A, where for all g, [[ª]](g)
= [[Æ]](g)([[]](g))>. (Again, a more common notation for this last part is: [[ª]]g =
[[Æ]]g([[]]g).)

Cumbersome though this may seem, it is well worth getting the intuition for
what variables are doing here and how this aids in the compositional computation
of a formula like feed’s (x1) (x2) which contains two open variables. To illustrate, let
us shrink the language to have only two variables x 1 and x2. Moreover, let us shrink
the universe to only two individuals, p and f. Then there are four assignments as
follows:

(1) g1: x1 ! p g2: x1 ! p g3: xi ! f g4: x1 ! f x2 ! p x2 ! f x2 ! p x2 ! f

Let [[feed’]] on each g be the function in (2):

(2) p → p → 1 f →p → 0 f → 1f → 1

We—the humans looking at this—can see that [[feed’ (x1)(x2)]] should come out
false on g3 and true on all the others. But that is because we can see the entire
formula. We reason that for any g, [[feed’]](g)(f)(p) is false while [[feed’]](g)
applied to any other combination of individuals is true. So we only need to look to
see what assignment maps x1 to f and x2 to p, and that is g3. But the grammar is
“dumb” and does not get to see the entire formula: it needs a procedure to
compute the value in a fully compositional fashion. This is what is accomplished
by the method of treating everything as a function from assignments to something
else. To illustrate, the computation proceeds as follows:
(3) [[feed’]] = g1 ! the above; g2 ! the above, etc.

[[x1]] = g1 ! p, g2 ! p, g3 ! f, g4 ! f
To derive the value of the next larger expression feed (x 1) the compositional
procedure has access only to the value of feed’ on each assignment and to value
of x1 on each assignment. By the rule NER-1, it computes the value in (4):

(4) [[feed’(x1)]] = g1 → p → 1 g2 → p → 1 g3 → p → 0 g4 → p → 0
f→1 f→1 f→1 f→1

The interesting point is the next step. To get from here to the value of the full
formula, we could perfectly well erase the information in (2) and never look back
at it. That information gives the value of feed’. But the compositional computation
has no access to the meaning of smaller expressions like this, it “sees” only the
 9.1. FIRST-ORDER PREDICATE LOGIC (MODIFIED) 139
value of the two expressions that combine to give the full sentence, and these are
feed’ (x1) and x2. The value of feed’(x1) is given in (4), and the value of x2 is (5):

(5) [[x2]] = g1! p, g2 ! f, g3 ! p, g4 ! f

These two pieces of information are sufficient to compute the right result for the
full expression feed’(x1)(x2) and this is shown in (6):

(6) [[feed’(x1)(x2) ]] = g1 ! 1, g2 ! 1, g3 ! 0, g4 ! 1

The moral of this rather plodding exercise is to show that this technique allows the
meanings to combine in a fully compositional fashion.

9.1.2. Quantifiers

In addition to their potential use in modeling natural language pronouns, variables

are also used in first-order logic as a way to give a semantics for quantifiers like 8
(“for all”) and ∃ (“there exists”). Recall that the goal in this chapter has nothing to
do with a direct modeling of the English compositional syntax/semantics. The next
additions we make to our NonEnglish language is very un-English-like; and an
account of English quantification that respects the syntax will be the subject of
the very next chapter. But working through the semantics for the first-order logic
quantifiers again illuminates certain techniques (in particular, the notion of
“binding” of the variables) that will be useful later. Moreover, having an explicit
semantics for formulas with these quantifiers is useful in naming model-theoretic
objects.
Hence we add the following two rules:
NER-6. Given any expression Æ of the form <[Æ], S, [[Æ]]> there is an expression of the
form <[∃ xi [Æ]], S, for all g [[]](g) = 1 iff there is some assignment g’ just like g
except possibly for the value assigned to x i such that [[Æ]]g’ = 1>.

NER-7. Given any expression Æ of the form <[Æ], S, [[Æ]]> there is an expression of the
form <[8 xi [Æ]], S, for all g [[]](g) = 1 iff for all assignments g’ just like g except
possibly for the value assigned to x i [[Æ]]g’ = 1>.

Syntactically, these rules prefix a quantifier and a variable to a sentence. (The

brackets [ . . . ] are used in two senses in the rules above. The outermost brackets
in [∃ xi [Æ]] are meant, as usual, to refer to the “phonology” of the expression—
 140 9. VARIABLES AND THE LAMBDA CALCULUS

rather artificial here since this language is not pronounced— but nonetheless
retained from the discussion of English to keep the differences minimal. The inner
brackets around Æ are meant as actual parts of the string, just as parentheses
were introduced in NER-1. The addition of these brackets ensures that each string
is unambiguous.) NER-6 and NER-7 allow for any indexed variable to be
introduced. Notice also that the rules are not formulated to require that there be
an instance of xi within the sentence to which the quantifier and variable are
prefixed. There is no particular reason to complicate the rule to require such a
variable within the S; if there is none, the effect of these rules on the semantics
will just be vacuous—the meaning of is simply the same as the meaning of Æ.
(This is thus referred to as vacuous quantification.)
Let us walk through the semantics. NER-6 ensures that a sentence of the form
∃xi S is true on an assignment g if there is some g’ which differs only on the value
assigned to i such that the lower S is true on g’. What does it take to find such an
assignment? Informally it requires there to be some individual a who can be
slotted in to the position held by x i which would make the sentence true. In other
words, if the sentence were ∃xi [walks’(xi)] and we were evaluating this on g, we
would look to find some individual a such that walks’(x i) is true on the g’ where x i
is mapped to a. In turn, this means that there has to be some individual who did
indeed walk. Note that the phrasing “g’ just like g except possibly for the value
assigned to xi” means that g’ could of course be g itself. For the case of universal
quantification, the semantics ensures that walks’(x i) is true for every assignment
which differs from g only in the value that it assigns to x i. Suppose for example that
there are 23 individuals. Then there are 23 assignments that vary only on the value
assigned to xi (for any i). (One of these is guaranteed to be g itself.) Since this must
be true on all of these, it is ensured that every individual does indeed walk.

*9.1. Show that application of NER-6 and NER-7 in cases where the input
sentence does not contain an instance of the variable introduced in these rules
has no semantic effect. (In other words, show that vacuous quantification is
indeed vacuous.)

There is another way to phrase the semantics which is equivalent. First we

introduce a notational convention. For any g and any individual a, let g x-i ! a mean the
assignment just like g except that x i is assigned the value a (where of course this
 9.1. FIRST-ORDER PREDICATE LOGIC (MODIFIED) 141
again could be g itself ). Then the semantic parts of NER-6 and NER-7 could be
rewritten as follows:
NER-60. . . . for all g, [[]](g) = 1 iff there is some individual a such that [[Æ]](g x-i ! a) = 1.
NER-70. . . . for all g, [[]](g) = 1 iff for every individual a, [[Æ]](g x-i ! a) = 1.

These amount to the same thing as given above. If there are n individuals, then
there are n assignments just like g except possibly for the value assigned to x i. Take
the set of all assignments which differ from g only on (possibly) the value assigned
to xi and call that set g/xi. (A caution on the notation: gx-i ! a is a single assignment;
g/xi is a set of assignments.) There is a one-toone correspondence between the
set of individuals and the set g/x i. NER-60 ensures that [[]] will be true on an
assignment g if we can find an individual a such that [[Æ]] is true on g x-i ! a. But if
that is the case, then it of course follows that there is a g’ in g/x i such that [[Æ]] is
true on g’, and this is what NER-6 asks for. Conversely, if there is such a g’, then it
follows that there is some individual a such that [[Æ]] is true on g x-i ! a. Similar
reasoning holds for NER-7 and NER-70. The key is in noticing that there is a one-
to-one correspondence between the set of individuals and the set g/xi. Some
terminology will be useful below. Consider a set A, and imagine “slicing it up” into
a set of subsets such that each member of A occurs in one and only one of these
subsets. This is called a partition of A, and each subset is called a cell in the
partition (see the Appendix to Chapter 1). We will use the term cellmates to refer
to members of A which are in the same cell. The relevance of this here is that each
variable xi induces a partition on G (the set of all assignments); each cell in the
partition is the set of assignments that differ only on the value assigned to x i. As
noted above by one example, the number of members in each cell will be exactly
the number of individuals (and each cell has the same number of assignments in
it). Thus g/xi defined above is a cell in the partition induced by x i. We will see this
played out by example below, but we can note here that the semantics of NER-6
and NER-7 are such that all members of the cell of a partition defined by the
variable introduced by the relevant rule will receive the same value.
By way of concrete illustration, consider the composition of the two sentences
in (7), given the world set out earlier (two individuals: {p,f} and [[feed’]] as shown
in (2) and repeated here):

(7) a. ∃x1[8x2 [feed’(x1)(x2)]]

b. 8x1[8x2 [feed’(x1)(x2)]]
 142 9. VARIABLES AND THE LAMBDA CALCULUS

(2) p → p → 1 f →p → 0 f → 1f → 1

We—the all-seeing human beings—know that (7a) should come

out true regardless of the assignments (there is indeed someone that everyone
fed, namely p) while (7b) should come out false (it is not the case that everyone
was fed by everyone; p was fed only by p). But recall again that the compositional
computation is not smart and has no access to the meanings of deeply embedded
expressions (such as feed’). In each step, this computation has access only to the
semantics of the input expressions, and—having added in NER-6 and NER-7—it
also has access to the structure of the assignments. We will take the assignments
as given in (1) (using again only two variables):
(1) g1: x1 ! p g2: x1 ! p g3: x1 ! f g4: x1 ! f
x2 ! p x2 ! f x2 ! p x2 ! f

We already showed the compositional computation of feed’(x 1)(x2), the final step
in that was (6), repeated here:
(6) [[feed’(x1)(x2) ]] = g1 ! 1, g2 ! 1, g3 ! 0, g4 ! 1

Looking at (7a), the next step is to compute [[ 8x2 [feed’(x1)(x2)] ]] on each g. Here
we need access to the structure of the assignments. Computing this for g 1, we
look to at its value on all g’ just like g1 except for the value assigned to x 2. In other
words, we look to see if it is true for all members of g/x 2. That set is g1 and g2. (The
partition on G induced by x2 is { {g1, g2}, {g3, g4}}.) Looking only at (6), we see that
the value of [[ [feed’(x1)(x2)] ]] is 1 on both g1 and g2. Hence, the value of the larger
sentence 8x2 [feed’(x1)(x2)] maps both g1 and g2, to 1. Turning now to the value of
the larger sentence on g3 and g4, they too form a single cell, and each will receive
1 if and only if the expression in (6) assigns 1 to both assignments. It does not; it
assigns 0 to g3.
From this we derive the following:
(8) [[ 8x2 [feed’(x1)(x2)] ]] = g1 ! 1, g2 ! 1, g3 ! 0, g4 !0

The next step is to compute [[ ∃x1[8x2 [feed’(x1)(x2)]] ]] on each g. Consider g1. Here
the relevant set of assignments g/x 1 is the set {g1, g3} so those two assignments
are guaranteed to have the same value. g 1 will of course get the value 1. But so will
g3 even though the line just above assigned 0 to g 3. At this step, however, g3 no
longer cares only about itself—it merely needs to find some cellmate in the
partition induced by x1 that is assigned 1, and g1 fits the bill. The same will apply to
 9.1. FIRST-ORDER PREDICATE LOGIC (MODIFIED) 143
the computation on g2 and g4. They are cellmates and hence look to each other; g2
is assigned 1 by the value of the inner S shown in (8), and so both g 2 and g4 will be
assigned 1. The sentence thus comes out true on all assignments: it is technically
a function from assignments to truth values but is a constant function (and so we
tend not to detect the assignments, and don’t think of it as
assignmentdependent).

9.2. Show the last line of the computation of (7b).

Notice that at each step shrinks the space of possible ways the assignments
can diverge in their value. At the step of interpreting 8x2 [feed’(x1)(x2)], we have an
equivalence class that partitions the assignments into {{g1, g2}, {g3, g4}}. At the next
stage we have the partition {{g 1, g2}, {g3, g4}}. By the end we are guaranteed a
constant function: g1 will have to have the same value as g 2 by the first step, but
will have to have the same value as g 3 by the second step. And g3 and g4 must have
the same value by the first step. So at the end, all four have the same value.
Incidentally, we can now give a (rough) semantic definition of what it means for a
variable to be “bound”: xi is bound within an expression if all members of g/x i
receive the same value. Put differently, it doesn’t matter what value one chooses
for xi—the set of assignments in g/x i in fact differ on just that—but because they
are all in the same cell and look to each other, the fact that they disagree on the
value of xi is irrelevant.
Otherwise xi is free (and so the value of the expression containing it still depends
on what value is assigned to x i). We say “rough” here simply because there are
cases that correspond to the intuitive notion of a variable being free but where all
assignments are nonetheless assigned the same value (even if we were to fold in
worlds) as, for example, the case of a tautology. Thus a statement like x3 is x3 (given
a suitable definition for be here) is assigned the same value (1) on all assignments
and in all worlds, but—should we really care about a precise definition of a “bound
variable”—we would probably want to take x 3 to be “free” here. But this is a
problem only if we really ever need a precise definition. 2 As we will see later, this
is a notion for convenience only. If the grammar never needs to refer to a “bound
variable” (or, in the case of English, a “bound pronoun”) then our loose definition
is perfectly adequate since it is intended just for convenience in discussion.
Indeed, the fragment to be developed for the rest of this book has no need for the
 144 9. VARIABLES AND THE LAMBDA CALCULUS

grammar to refer to a notion of “binding.” 3 And this is equally true for the version
that makes use of assignment functions to model pronoun meanings as for the
variable-free version.

2
Heim and Kratzer (1998) develop a very interesting system using partial functions.
Their system is meant to model English (rather than the artificial toy language here) but the
same points would hold. Thus each expression has a value relative to some assignment,
but the assignments themselves can be partial functions from the set of variable names to
individuals. Moreover, the value of some sentence S is defined only for a subset of
assignments (only for those whose domain is the set of variables which occur in what we
intuitively think of as “unbound” within S). Space precludes discussion of this here, but it
does allow for a definition of “bound” vs “free” variables (or pronouns) that avoids the
problem noted above. However, as mentioned above, there will turn out to be no need for a
formal semantic definition of “bound” vs “unbound” pronouns (or variables).
3 There is much work within the linguistics literature that uses the term “binding” to state

various constraints relevant to the distribution and readings of pronouns and reflexives. The
goal here under Direct Compositionality is to have the empirical generalizations that are
packed into constraints on “binding” follow from the way the system is set up without the
grammar needing any such definition. This book will not have space to discuss so-called
binding constraints, but see Chapter 15 for a discussion of what is known as Weak
Crossover effects.
 9.1. FIRST-ORDER PREDICATE LOGIC (MODIFIED) 145

9.3. Take a universe with three individuals {p,f,k}. Take the language given above
but with just the two variables x1 and x2.. We are concerned here with just one
world, which is such that on any assignment g, [[kiss]]
(g) is:
p àp à 1 f à p à1 kàpà0
fà0 f à1 fà1

kà1 g f g kà0 f g kà0

(a) First, show the nine assignments.

(b) Then show the compositional computation for each of the
following, showingeach step as a function from the domain of assignments to
some other value:
(i) 8x1[∃x2[ kiss’(x2)(x1)]]

(ii) ∃x2[8x1[ kiss’(x2)(x1)]]

(iii) 8x2[kiss’(x2)(x2)]]

Note that once you have shown the computation of kiss’(x 2)(x1) for (i) there is
no reason to repeat this for (ii); just recycle what you have done.
(c) Consider (iv):
(iv) 8x1[kiss’(x1)(xI)]

The part within the brackets will not be the same as the corresponding part for
(iii). But (iv) (regardless of the structure of the world in question) is guaranteed
to have the same value as (iii). State why this is so.
9.2. The lambda calculus

9.2.1. Lambda abstraction

Continuing with the digression into languages which are not English, we now take
the language above as the point of departure and enrich it, developing what is
known as the lambda calculus. As noted earlier, this gives us a precise way to
name complex model-theoretic objects. Note, though that developing this
language has no theoretical status in that regard—one could continue to use
 146 9. VARIABLES AND THE LAMBDA CALCULUS

prose to name the object. However, our prose has been getting rather clumsy—
and would get even more so more complex objects are developed.
The first addition to the language is an infinite set of variables over any kind of
object (and with every syntactic category). There are variables whose category is
S/NP and semantic type <e,t>; we will name these P1,
P2, etc. There are variables over propositions (category S) which we will
 9.2. THE LAMBDA CALCULUS 147

label p1, p2, etc., and variables of type <e,<e,t>> (category (S/NP)/NP) which we
will label R1, R2, etc. Recall that, for convenience, we took the domain of the
assignments to be the variable names themselves, giving variable names a kind of
double status (syntactic objects, and objects in the domain of assignments) and
we continue with this convention here. We add the rule in NER-8; X and Y are
variables over any category, and y semantic type of expressions of category Y (for
any g and any expression of category Y, [[]](g) is a member of y).
NER-8. If Æ is an expression of the form <[Æ], X, [[Æ]]> and v is a variable of the form <[v],
Y, [[v]]> then there is an expression of the form <[ºv [Æ] ], X/Y, for any g, [[]](g)
is a function from the set of objects of type y, such that for each a in y, [[]](g)(a)
= [[Æ]](gv à a)>.

As was the case with NER-6 and NER-7, we are using square brackets in two ways
in the “phonology” part of NER-8. The outer brackets indicate “the phonology of”
the new expression; the inner brackets around Æ are actually introduced by the
rule (to keep the strings from being ambiguous).
Let us walk through the semantics part in prose. First, let x be the semantic type
of expressions of category X and y the type of expressions of category Y. Then our
new expression has some meaning which is a function from assignments to
something in <y,x>. On any assignment g, this new expression takes every
member a of the y set, and assigns to it the value that the inner formula (Æ)
assigned to the assignment just like g except where a is the value of the variable v.
Prefixing a º and a variable v to some formula is called lambda-abstracting over
that variable, but the important point is to see its semantic effects. One effect is it
takes a formula which is non-constant on g/v and turns it to a formula whose value
is the same for all members of g/v. The other effect is that it takes an expression
which, on any assignment, has as its value something in set a, and returns an
expression which (on any assignment) is a function from some set b to a. Hence
the expression Æ to which is being prefixed ºv has as its value a function in <G,a>.
Let the variable v be of type b, then the full expression ºv [Æ] has as its value a
function in <G,<b,a>>.
By way of example, take a universe with two individuals {p,f} and put just two
individual variables x1 and x2 in our language. The four assignments are as given
earlier in (1):
(1) g1 : x1 ! p g2 : x1 ! p g3 : x1 ! f g4 : x1 ! f
x2 ! p x2 ! f x2 ! p x2 ! f
 148 9. VARIABLES AND THE LAMBDA CALCULUS

Suppose that we are considering a world in which [[squeaks’]] for all assignments
is p à 1 and f à 0. Let us compute the value of lx 1[squeaks’(x1)]. First, we compute
[[squeaks’(x1)]] on each g and this gives:

(9) g1 à 1, g2 à 1, g3 à 0, g4 à 0

The next step is to compute the value of the whole formula. Once again, the
compositional procedure does not need to consult anything about the inner
formula; the computation computes the value of ºx 1[squeaks’(x1)] on each
assignment purely by looking at (9) and at the structure of the assignments. In
particular, the variable that is “bound” here by the º is x 1; the relevant partition is
thus {{g1, g3}, {g2, g4}}. It is guaranteed that for any given cell, each member in that
cell will receive the same value; this is because they “look to each other.” For
example, consider the computation of the value of lx 1[squeaks’(x1)] on g1. First,
note that the value of this on g 1 is going to be a function of type <e,t>, assigning 1
or 0 to p and 1 or 0 to f. The determination of what value is assigned to p is made
by looking at the value of the inner formula on the assignment just like g 1 except
that x1 is mapped to p. This is g1, so (8) tells us that p will map to 1. For f, the
relevant assignment is g3 so (9) tells us that f maps to 0. Exactly the same remarks
hold when we want to find the value of ºx1[squeaks’(x1)] for g3. Since g1 and g3 are
in the same cell, they are guaranteed to be assigned the same value by the full
formula. The same holds for g2 and g4. But since there is no x2 at issue, these will
come out the same as g1 and g3, and in the end we have the constant function from
assignments: on any g the value of lx 1[squeaks’(x1)] is p à 1 and f à 0.
The reader might have noticed that the value of lx 1[squeaks’(x1)] is exactly the
same as the value of squeaks’, and was destined to be so. By rethinking the
lambda expressions as ways to name functions (which is ultimately the way we
will be using them), this is not surprising: the full formula is (on any assignment) a
function which asks for an individual, and gives the value of squeaks’ applied to
that individual. So of course this is just a fancy way to write (semantically) exactly
the same object as squeaks’. Note that of course lx2[squeaks’(x2)] is also the same
object. As a second example, consider the formula in (10):

(10) ºx1[ºx2[feed’(x1)(x2)]]

The semantics of this is exactly the same as the semantics of feed’. Again, thinking
of these as ways to name functions, one can convince oneself of this by noting
that for any assignment g, this is a function that takes one individual and then
another and returns the value that feed’ returns when applied to the first individual
 9.2. THE LAMBDA CALCULUS 149

and then the second. And that, of course, is exactly the feed’ function (a function
of type <e,<e,t>>).

9.4. Given [[feed’]] as shown in (2) (and continuing with a language with only
two variables and a universe with only the two individuals {p,f}, demonstrate
the truth of the remarks above by showing the full compositional computation
of (10). Remember that the end product should be a function from assignments
to functions of type <e,<e,t>>.
9.5. Like the quantification rules, NER-8 does not require there to be an
instance of the variable introduced by this rule within Æ. If there is none, the
rule is not vacuous as in the case of the quantification rule, as the meaning of
is not even the same type as the meaning of Æ. State what the meaning of is
in general if there is no occurrence of the relevant variable within Æ.

In view of these equivalences one might wonder what the point is of this
notation. It is because other kinds of values can be written in this notation.
Take for example (11):

(11) ºx1[ºx2[kiss’(x2)(x1)]]

This does not have the same meaning as any other lexical item in the language
(keeping our lexicon to be just a subset of English), but corresponds to something
that we can roughly paraphrase as was-kissed-by’. We return to this below.

9.2.2. Lambda conversion

Since the ultimate interest is in the lambda calculus is as a way to name various
functions, it is also useful to understand ways in which one can simplify various
formulas in the lambda calculus. By “simplify” we mean naming the same object
in a shorter way. We can actually give proofs of the equivalences that we show
here, but to save space and possible tedium, we will try only to give the intuition
behind why the equivalences hold.
Take any formula of the form ºu[Æ](c). Call Æ/u à c the formula derived from
ºu[Æ](c) except where all occurrences of u in Æ are replaced by c. In general (but
 150 9. VARIABLES AND THE LAMBDA CALCULUS

not always—the exception to be discussed directly) these two formulas will

always have the same value. This is called lambda conversion.
We have already seen a few instances of this. Thus (12a) and (12b) are guaranteed
to have the same value; note that (12b) is the result of applying lambda conversion
to (12a):
(12) a. ºx1[squeak’(x1)](Porky’) = b. squeak’(Porky’)

The sign “=” here means has the same value (on, of course, all assignments).
Another example is (13):

(13) a. ºx1[ºx2[feed’(x1)(x2)]](Porky’) = b. ºx2[feed’(Porky’)(x2)]

We also know from the sort of equivalences discussed in 9.2.1 that (b) is
equivalent to feed’(Porky’). Similarly, the pair in (14) are equivalent:
(14) a. ºx1[ºx2[feed’(x2)(x1)]](Porky’) = b. ºx2[feed’(x2)(Porky’)]

Informally, both of these (on all assignments) denote the set of individuals fed by
Porky. We can show cases of two lambda conversions, as in (15):
(15) a. ºx1[ºx2[feed’(x2)(x1)]](Porky’)(Dumbo’) =
b. ºx2[feed’(x2)(Porky’)](Dumbo’) =
c. feed’(Dumbo’)(Porky’)

(In doing lambda conversions, one needs to make sure to pay attention to the
brackets. In (15a) Porky’ must convert in for x 1 because it is the argument of the
function shown in (11).)

9.6. Simplify the following by lambda conversion:

(i) ºx1[ºx2[feed’(x2)(x1)](Porky’)](Dumbo’)

However, it is not always the case that ºu[Æ](c) is equivalent to Æ/uà c. Rather
than giving a full explanation of when and why the equivalence does not hold, we
will simply give a recipe for determining when lambda conversion is not valid (i.e.,
when it is not guaranteed to result in an equivalent formula) along with an example
to help elucidate why. This will be followed by a brute-force recipe for allowing
conversion on a slightly different formula.
Hence we first give a purely syntactic definition of the notion of a
 9.2. THE LAMBDA CALCULUS 151

“bound” variable. (Recall that earlier we gave a rough semantic definition; a

variable xi is bound in a formula if the value of that formula is the same on all g/x i.
But for the purposes of the recipe, it easiest and most practical to see the point
with a syntactic definition of the notion bound. Leaving aside contradictions and
tautologies, the two definitions converge on the same result.) Hence we will
define bound as follows:

(16) Let O (for operator) stand for any of º, ∃, and 8

Given any formula of the form Ox[ . . . ], anything within the brackets is within the
scope of O (for x a variable).
A variable x with the name vi is bound by an operator O iff:
(i) the variable next to O in the formula is v i;
(ii) x is in the scope of O; and
(iii) there is no other O0 whose adjacent variable is v i such that x is in the scope
of O0 and O0 is in the scope of O.

Notice that this definition allows for variables which are free within some
subformula even though they may be bound within the main formula. For example,
x1 is bound within the formula ºx1[squeaks’(x1)] but is free within the subformula
squeaks’(x1). Then whenever there is a formula of the form ºu[Æ](c) one cannot
perform lambda conversion (and be guaranteed to preserve equivalence) just in
case c is or contains a variable which has the same name as one which is bound
within Æ. Intuitively, this is because the variable c or the variable within c will be
accidentally “caught” by the operator inside Æ.
As an example, take the pair in (17), where (17b) is the result of the illegitimate
lambda conversion:

(17) a. ºxi[∃x2[feeds’(x1)(x2)]](x2) b. ∃x2[feeds’(x2)(x2)]

The truth of (17a) is assignment-dependent—it depends on the value assigned to

x2. Paraphrasing it in prose, it says that x 2 is such that there is someone who feeds
it. This is because the lambda expression within (17a) is the (assignment-
dependent) function that maps an individual to true just in case there is someone
who feeds it. The whole formula then says that this function truly applies to x2, but
whether or not that is true depends on the value we assign to x 2. (17b), on the
other hand, is constant on all assignments—on any g, it simply asserts that there
is someone who feeds themselves. What has gone wrong is that the x 2 bound by
the existential quantifier in (17a) and the x 2 outside the lambda expression are not
really the same. The bound one inside the lambda expression could as well have
 152 9. VARIABLES AND THE LAMBDA CALCULUS

been x1, x3, or anything else: it is bound so its name no longer matters. The one
outside the expression is not bound, so its name does matter. Put differently,
ºxi[∃x2[feeds’(x1)(x2)]] has the same value on all g/x 2. But the outermost x2 does
not have the same value on all g/x2, and so neither does the full formula in (17a).
There is also a simple way to avoid mistaken lambda conversions such as those
shown in (17). Simply find some new variable that appears nowhere else, take all
of the bound occurrences of the colliding variable inside Æ (including the
occurrence that sits next to the operator that binds it), and rename it with the new
variable. One should be able to convince oneself that once a variable is bound its
name no longer matters, so renaming it in this way will have no effect on the
meaning of Æ. Then the lambda conversion can proceed. Given any expression Æ
with a variable a bound within it, call Æ’ the new expression in which every
occurrence of a is replaced by another variable b. Æ and Æ’ are said to be
alphabetic variants, which means that they are guaranteed to have the same value
on all assignments.

*9.7. Show that one cannot instead rename the free variable within the
expression (or which is the expression) which is being converted in. You can
show this by example and/or discuss more generally why the equivalence is not
guaranteed to hold in if one renames the free variable.
9.8. Take each of the following formulas, and simplify them as much as possible
(i.e., write them in a shorter form that preserves their meaning). Recall that P1
is a variable of type <e,t> and R 1 is a variable of type <e,<e,t>>.

(i) ºx1[kiss’ (x1)(x2)](x2)

(ii) ºx1[kiss’ (x1)(x2)](x1)
(iii) ºP1[8x1[P1(x1)]](feeds’(x1))

(iv) ∃x2[ºR1[R1(x2)(x2)](ºx3[ºx1[grooms’(x3)(x1)]])]
(v) ºP1[ºx1[P1(x1)]](chews-gum’)
 153
9.3. NAMING MODEL-THEORETIC OBJECTS

9.3. Naming model-theoretic objects

The advantage of having the explicit notation above—along with the principles for
rewriting lambda expressions as simpler (syntactically) but equivalent (semantically)
expressions—is that it provides a tool to write out complex functions which are
difficult to state in prose. It also allows a much less clumsy semantics for some of the
earlier rules. Notice that although the meanings of lambda formulas are stated with
reference to assignments and variables, we can remain for the moment
noncommittal as to the status of these for natural language semantics and still use
the relevant formulas. The reason is simple but quite important: for now, at least, all
of the model-theoretic objects that we will name using the lambda notation contain
no unbound variables within the formulas. Thus if assignments were an actual part of
the machinery, all of the objects of concern at this point are constant functions from
the set of assignments. But if everything is a constant function from the set of
assignments, we can just as well do without the assignments. Thus the use of
assignments, variables, lambda conversion, and so forth is important to understand
to make sure that we do correctly represent the meanings of the natural language
expressions. But these tools are (so far) not part of the linguistic semantic system.
Whether or not something like variables and lambda abstraction do play a role in the
semantic theory is addressed in Part III.
To give a few examples of the usefulness of this notation for writing meanings, we
will slightly modify the conventions above. First, let us use feed’ not just as a word in
the non-English language, but let us also use it to represent the meaning of English
feed. In other words, it is the same as [[feed]]. The reason for having both ways to
represent the meaning of feed has to do with the fact that each notation has certain
advantages in terms of readability, and so we will alternate back and forth between
the two notations as convenient. Second, the convention of numbered variables is
sometimes difficult to read. We will thus use x, y, z, etc. for variables of type e; P and
Q for variables of type <e,t>; p and q to be variables of type t, and others as needed.
When introducing variables whose type is not clear we will subscript them, as for
example in the following:

(18) ºR<e,<e,t>>[R(Fido’)(Dumbo’)]

This expression names a function in <<e,<e,t>>,t>: it characterizes a set of two-place

relations. In particular, it maps a relation R to true just in case R(Fido’)(Dumbo’) is true
or, informally, just in case “Dumbo Rs Fido.” Or, take the following (here we omit the
subscript, keeping R to be of the same type as in (19)):
 154 9. VARIABLES AND THE LAMBDA CALCULUS

(19) ºR[R(Porky’)]

This is a function of type <<e,<e,t>>,<e,t>>: it maps any two-place relation R to a

function that characterizes the set of individuals who “R Porky.” If this function were
applied feed’ (or, [[feed]]) it would yield the set of Porky-feeders.
We can write out the meanings that we gave above for to and for be. The former,
being the identity function on individuals, is (20a) and the latter is (20b):
(20) a. ºx[x] b. ºP[P]

Or, consider the account of passive in English according to which (21) is not derived
by any kind of movement rule, but where there is a productive (morphological) rule
mapping an active verb into a passive verb, changing its syntax (and phonology) and
its meaning:
(21) The squirrel was chased by Mitka.

We will not formalize the rule here (the reader can easily do so), but we can note that
its output would be the verb killed with a category roughly (S[Part]/NP)/PP[BY] (we
leave it open as to the best feature specification here on the result) and with the
meaning shown in (22a) or—just using a slightly different notation—(22b) (whose
visual awkwardness reveals the rationale for the prime notation):
(22) a. ºx[ºy[ chase’ (y)(x)]] b. ºx[ºy[ [[chase]] (y)(x)]]

Note that the rules for º-conversion are now quite useful as we build up the
compositional computation for (21).

9.9. To show the full computation of (21) requires a treatment of the auxiliary.
Pretend, then, that the passive verb is actually the single verb was-chased with
category (S/NP)/PP[BY] and meaning as given above. With that, show the full
syntactic and semantic computation of (21), doing º-conversions wherever (and
as soon as) you can to simplify the way in which the semantics is stated. (Note: It
is always helpful in terms of the exposition to do the º-conversions as soon as
possible.)
*9.10. Assume the Wrap analysis of three-place verbs discussed in Chapter 5 and
explicitly folded into the fragment in Chapter 6. That is, assume that all verbs of the
form ((S/NP)/NP)/X (for X any category) are
9.3. NAMING MODEL-THEORETIC OBJECTS
 155

S/NP)/WNP)/X. (This is actually not crucial for the main point of this
since the implementation of the question requires some decision
ucture of three-place verbs, we will make that assumption.)
est Transformational Grammar, it has been noted that there are a
r of verbs that occur both in the V-NP-PP[TO] configuration and in
configuration, and where sentences of the form in (ia) have the
onditions as those of the form (ib):

NP2 to NP3 b. NP1 V NP3 NP2

this are shown in (ii)–(iii) and this is generally referred to as the

ation:
na gave the toy to Fido.
ina gave Fido the toy.

ge told the Cinderella story to the children.

rge told the children the Cinderella story.

s of this (beginning with Oehrle 1976) treat these alternations as a

apping words to others. Take the verbs in the (a) form to be basic.
e rule in question, showing how the output expression differs from
yntactic category and in meaning. (The point, of course, is that the
kes it easier to state the meaning of the output verb in terms of the
he input verb. You do not need to use any ºs to state the meaning
verb; the meaning of the lexical item give in (iia), for example, is
]].)
sting sidelight: It is often claimed that the rule is not really
nd regular, for there are many verbs that do occur in the (a)
but not in the (b) configuration; examples include contribute,
ate, etc. These contrast with hand, toss, tell, give, and many others
cur in both configurations. Leaving aside the verb promise, state
ation of the verbs that do undergo the rule in phonological terms,
out how one can state the rule so that it is indeed a productive

One final example is the Curry’ed meaning for and (when it combines two expressions
whose meanings characterize sets), which was given earlier in terms of set
intersection. We can equivalently give its meaning as follows, where a is meant to be
any type:

(23) ºA<a,t>[ºB<a,g>[ºza[A(z) & B(z)]]]

 156 9. VARIABLES AND THE LAMBDA CALCULUS

Finally, this notation also allows us to state the semantics of many of the earlier
rules in a much more elegant fashion. R-3 (from Chapter 8) serves to illustrate this,
but first one more piece of notation. Given an individual x and a function P of type
<e,t>, we will use the notation Øx[P(x)] to mean the unique (or most) contextually
salient individual such that P(x) is true. (Often the iota operator shown here is used to
mean the unique such individual; here we use it in the same way that English the is
used.) Note then that any expression of the form Øx[Æ] denotes an individual. With
this notation, we restate R-3:
R-3. If Æ is an expression of the form <[Æ], N/PP[OF], [[Æ]]> then there is an expression of
the form <[], NP/LNP[GEN], ºxØy[ [[Æ]](x)(y)]]>.

9.11. Use the º-notation to restate R-2 from Chapter 8.

9.12. Give the meaning of didn’t in º-notation.
PART II Enriching the Domain
 10 Returning to English:
Generalized quantifiers
10.1. “NPs” with 10.4. Interlude: Scalar implicature 168
quantificational determiners 157
10.5. Generalized quantifiers
10.2. Generalized quantifiers 173
162 and conjunction
10.3. The determiners 178
167 *10.6. Domain restriction
10.1. “NPs” with quantificational determiners

The fragment so far contains no NPs with quantificational determiners, such as

the subjects in (1)–(4):
(1) Every husky howled.

(2) Some pig ate Old MacDonald’s hay.

(3) No student passed the phonology exam.

(4) Few donkeys appreciate their owners.

Recall that ordinary NPs like the disobedient husky, the ornery pig, and Porky all
have the same syntactic distribution and denote individuals. The types of
expressions shown in (1)–(4) also all have the same distribution as ordinary NPs
(they occur in subject position, they occur as objects of transitive verbs, etc.). But
what about their meaning?
10.1.1. First pass: Individuals?

Since other expressions with this distribution denote individuals, the

correspondence between syntactic categories and semantic types laid out in
Chapter 6 would lead to the conclusion that these do too. But this seems to defy
common sense: they surely are not individuals on any ordinary understanding of
that term. Of course, we have taken a fairly broad-minded view of what it means
to be an “individual,” allowing for abstract individuals such as modesty, individuals
that name kinds such as donkeys, and mass individuals such as water. So perhaps
these are just even fancier kinds of individuals. In this case, though, common
sense happens to be correct; calling these “individuals” will wreak havoc with the
semantics. Consider the intersection semantics for VP conjunction in light of (5):
(5) No student passed the phonology exam and got an A on the semantics final.
 160 10. GENERALIZED QUANTIFIERS
If [[no student]] is a fancy individual and that individual is in the intersection of
[[pass the phonology exam]]S and [[got an A on the semantics final]], then [[no
student]] is in each of these sets. But then that predicts that (5) is synonymous
with (6):
(6) No student passed the phonology exam and no student got an A on the semantics
final.

Clearly the (at least primary) meaning of (5) is not the reading in (6); it merely says
that there is no student who did both. Note that if we were to replace no student
in (5) and in (6) with an ordinary individual-denoting subject NP such as Sarah,
then synonymy between the case of VP conjunction and S conjunction does hold.
This was discussed earlier, and is the straightforward result of the fact that if x is
in A \ B then x is in A and x is in B.

10.1. a. Construct an argument to show that [[every student]] also cannot be an

individual.
b. Do the same for [[some student]].

In both cases your arguments can be similar to the one above, but you will need
to do more than mechanically switch no student in the above sentences to
every student or some student.
10.1. “NPs” WITH QUANTIFICATIONAL DETERMINERS

Hence, our semantics for conjunction would be doomed if one tried to force these
expressions to be individuals. We will thus for the moment leave aside the
ultimate goal of giving these a type of meaning which they share with ordinary NPs
like Babar, and will concentrate in this chapter on finding some meaning for these
expressions which captures the appropriate truth conditions. Chapter 11 returns
to the task of unifying these and ordinary NPs. In fact, for now we give quantified
expressions a different syntactic category from ordinary NPs—call their category
GQ (this stands for generalized quantifier, which concept is explained in 10.2).

10.1.2. Second pass: Sets of individuals?

Could the relevant expressions denote sets of individuals (or, equivalently,

functions of type <e,t>)? This is tempting at first glance. Let [[every husky]] be the
characteristic function of the set of all huskies. An immediate drawback with this
 161
hypothesis is that it necessitates a new semantic combinatory rule. Consider (1).
Since [[howled]] is a function of type <e,t> a new rule would be needed to
combine the two functions to give a result of type t. For the particular case at hand,
the semantics here could be as given in (5) (we use the phrase structure
rule/semantic rule format for this, and it will go in the category of a fictional rule):

FR-4. S à GQ VP [[S]] = 1 iff [[GQ]]S [[VP]S

This gives the right truth conditions when the GQ has every as its determiner.
But the attempt to treat GQs as denoting sets of individuals along with a rule like
FR-4 breaks down in other cases. Take [[no husky]]. What sort of set could this
possibly be? The first temptation that springs to mind is that [[no husky]] is the
empty set. But then what is the meaning of no Siamese cat? It too must be the
empty set. Yet this makes the disastrous prediction that the two sentences in (7)
are synonymous:
(7) a. No husky used the litter box.
b. No Siamese cat used the litter box.

Alternatively, one might try allowing [[no husky]] to be the set of all nonhuskies.
But then FR-4 gives the wrong meaning: no husky is howling would be true if all
the non-huskies are howling—obviously an incorrect result. Of course, one could
posit a different rule for this case (here, let the VP set be a subset of the GQ set).
But now there is no uniform treatment for how VPs combine with GQs, and the
rules would need to refer to the internal structure of the GQ (as each determiner
would have a separate rule). This is surely not a desirable state of affairs.
Moreover, it is not clear at all how to extend the set idea to GQs like few huskies
(what set of individuals would that be?), most huskies, etc.

10.1.3. Third pass: English as first-order logic? Straw Man makes a

cameo reappearance

At this point, Straw Man reappears. Straight from his logic class where he has
learned all about quantifiers, variables, and assignment functions, he is eager to
apply his new-found knowledge to the semantics of English. Why not, he says,
assume that the grammar maps the sentences in (1) to representations like the
language developed in section 9.1? Every husky is howling is mapped to (8a),
Some husky is howling is mapped to (8b), and No husky is howling to either (8c)
or (8c’) (these are equivalent):
 162 10. GENERALIZED QUANTIFIERS
(8) a. 8x1[husky’(x1) à howl’(x1)]

b. ∃x1[husky’(x1) & howl’(x1)]

c. ~∃x1[husky’(x1) & howl’(x1)] c’. 8x1[husky’(x1) à ~ howl’(x1)]

The idea, then, is that the grammar assigns meanings by the two-step process of
mapping English sentences into LFs like those in (8), and these in turn are
compositionally interpreted by the procedure given in Chapter 9.
What is the price? First of course, by abandoning Direct Compositionality,
Straw Man is advocating a more complex model of grammar. The syntax
computes representations and there is an additional series of rules mapping
those into representations of the sort in (8). Moreover, the consequence is that
although every husky, some husky, etc. behave syntactically like constituents,
they have no meaning. Notice that in (8a), for example, there is no subformula
which corresponds to every husky. The “meaning” of every husky is simply the
contribution that it makes in terms of a direction for how to build the relevant LF.
Of course, Straw Man is not committed to Direct Compositionality, and so is
not particularly moved by these arguments. But there are other reasons to
question this solution. First the rules are complex: the rules mapping the English
sentences to the requisite LFs need to be formulated, and this is not
10.1. “NPs” WITH QUANTIFICATIONAL DETERMINERS

that simple (we leave it to the reader to see how to state these). Note too that the
relevant rule cannot be stated completely generally; the rule mapping (2) its LF
(8b) introduces “&,” whereas the rule mapping 1 to (8a) introduces “à.” So there is
no single rule mapping sentences with expressions of the form [Quantifier N] into
the relevant LF.
And it gets worse. Consider (9):
(9) Most Siamese cats have lovely voices.

Can we map this to some formula of first-order logic to give the right truth
conditions? First, we note a tangential problem: most is arguably vague— one
might be unsure as to whether (9) requires 51% of the Siamese cats to have lovely
voices, or perhaps more (and if so, how much more). Fortunately, this question is
not central to the point below, so simply assume that a sentence like (9) is true
provided that more than 50% of the Siamese cats have lovely tails. We will
introduce a new quantifier M that we can read informally as for most. Following
the strategy for all and some, we might try to assign to (9) either the representation
 163
in (10a) or in (10b). The semantics of these will be analogous to the semantics for
quantifiers given in Chapter 9, except that the full formula is true on an assignment
g just in case the inner formula is true for more than 50% of the assignments in
g/x:
(10) a. Mx[Siamese-cat’(x) & have-lovely-voices(x)]
b. Mx[Siamese-cat’(x) à have-lovely-voices(x)]

It is immediately obvious that (10a) is incorrect; this is true if more than 50% of
the individuals are Siamese cats and have lovely voices. The problem with (10b)
is more subtle, but it too is incorrect. (9) is true if it’s the case that for most things,
if they are Siamese cats then they have lovely voices. But for (10b) the inner
formula is true on any assignment g for which g(x) is not a Siamese cat. Thus
consider a universe with 100 individuals, 90 of which are not Siamese cats, and
where only three of the actual Siamese cats have lovely voices. Then (9) is false,
but our semantics assigns true to (10b). Of course we have only tried formulas
with the two connectives “&” and “à”—perhaps we just need to be clever and try
a different one. It need not even correspond to any of the standard ones for which
we have names (i.e., &, v, and à) as long as it can be defined. But it is quite easy to
show that no possible connective will do. There are only sixteen logically possible
ones and none gives the right result.
10.2. Show the sixteen logically possible connectives (remember, these all
take as input two values in {1,0} and return one value in {1,0}, and show that
none of them gives the right truth conditions.
*10.3. Although somewhat off the main topic, once you have constructed the
logically possible connectives asked for in Exercise 10.2, it becomes obvious
that many of these are “stupid”—i.e., it is not surprising that only a few of these
remotely correspond to any actual natural language words. Discuss which
ones would be rather “stupid” for a natural language to encode as a lexical item.

Finally, once again facts centering on conjunction challenge this solution. For
quantified expressions happily conjoin with each other (the facts are similar for
or):
(11) a. Every coyote and most wolves howl at the moon.
b. No pig and no turkey remained on the farm when the gates were left open.

One would hope to subsume these under the cross-categorial definition

(syntactically and semantically) for and (and or). But under Straw Man’s view,
 164 10. GENERALIZED QUANTIFIERS
every coyote is not a meaningful expression and so there is no way to directly give
a semantics for expressions like every coyote and most wolves.
Once again, then, he would be reduced to positing the hidden S conjunction
analysis (see Chapter 5). Unlike the case of VP conjunction, the 2-S analysis here
is not disastrous for the semantics: (11a) for example does happen to have the
same meaning as (12):
(12) Every coyote howls at the moon and most wolves howl at the moon.

Still, this requires the adoption of additional rules, and it obviously would be
desirable to account for (11) with the more general treatment of conjunction. We
thus part ways with Straw Man.

10.2. Generalized quantifiers

10.2.1. The semantics

The remarks much earlier in Chapter 5 concerning the Curry’ing of transitive verbs
showed that it is not necessarily difficult to have the semantics
 165
10.2. GENERALIZED QUANTIFIERS

truly respect the syntax. Sometimes the first way that springs to mind concerning
the meaning of an expression won’t work, but this might only mean that the
semantics has actually packaged the meaning in slightly more sophisticated
ways. Unless these ways turn out to be torturously complex, a reasonable
hypothesis is that this is indeed how the semantics works. Hence, let us return to
taking seriously the notion that every husky, most Siamese cats, no wolf, etc. all
have meanings. Consider where the discussion left off before Straw Man
reappeared. We eliminated the possibility that these expressions denote
individuals, and the possibility that they denote sets of individuals. What if we try
just one level of complexity beyond sets of individuals, i.e., sets of sets of
individuals? Happily, this turns out to be exactly what is needed and accounts for
all of the facts considered so far. (Initially we will show the point in set terms; it
will be easy enough to then convert this into function terms.) Indeed, this solution,
proposed for English in Montague (1973), was highly influential, and has been
accepted in most work in formal semantics ever since. Any set of sets of
individuals (or, equivalently, a function of type <<e,t>,t>) is called a generalized
quantifier.
Take every pig. In pure set terms, let the every-pig set (given a world and time)
be the set of sets that have the [[pig]] set as a subset. Thus [[every pig]] S = {S|
[[pig]]S S}F. Returning to function terms, [[every-pig]] is of type

<<e,t>,t>; here the domain of [[every pig]] is the set of functions of type <e,t>, the
co-domain is just t. Thus it maps a function P (of type <e,t>) to true iff for all
individuals x such that [[pig]](x) is true, P(x) is true. (Assume that it is false
otherwise.) Using the notation developed in Chapter 9, we can write the meaning
(extensionally) in any of the ways shown in (13)—they all say the same thing (we
use the prime-notation rather than double brackets for readability):

(13) a. [[every pig]] = ºP<e,t>[ pig’S PS ]

b. [[every pig]] = ºP[8x [x pig’S à x PS]]

c. [[every pig]] = ºP[8x [pig’(x) à P(x)]]

Incidentally, (13c) contains within a first-order formula which might seem odd
since the first-order rendering of the meaning is exactly what was argued against
above. But in fact nothing here contradicts that discussion. (13c) is used only as a
way to write the model-theoretic object. Any of the other notations shown above
 166 10. GENERALIZED QUANTIFIERS

(or the prose description) will do just as well. And a sentence containing every pig
is not mapped into an LF using this first-order quantifier; it is just part of the
notational system for writing the meaning.
In order for this meaning to do any work, this needs to fold into the rest of the
compositional semantics. Consider (14):
(14) Every pig grunted.

Since every pig does not denote an individual, (14) cannot be put together in the
way already seen for a sentence like Porky grunted. In fact, recall that we are
temporarily ignoring the commonality of ordinary NPs and expressions like every
pig (this is the subject of Chapter 11) and we are calling the category of every pig
a GQ. We get exactly the right truth conditions if we reverse the function argument
structure from what happens in the case of a VP combining with an ordinary NP.
That is, since the subject in (14) is a function of type <<e,t>,t> and the VP of type
<e,t>, one would expect that [[every pig]] applies to [[grunted]]. Indeed, this gives
exactly the right result.
It is easiest to see this at first in set terms. The sentence is true if the
[[grunt]] set is a member of the set of sets denoted by every pig. What does it take
to make that true? [[every pig]] is the set of all sets with [[pig]] set as subset; and
so if the [[grunt]] set is in the [[every pig]] set, then the [[pig]] set is a subset of the
[[grunt]] set. And indeed this is what it takes to make (14) true. (Of course the
sentence need not be about every pig in the universe but only every pig within
some contextually salient domain. But this is a separate issue, postponed until
section 10.6.) One other issue that we will not deal with is whether or not (14)
presupposes that there are pigs (this is reminiscent of the situation with the). The
meaning written out above in (13) predicts not: if there are no pigs then [[every
pig]] maps every VP meaning to true. Whether or not this is a good prediction is an
open question. It might be that (14) should be undefined in a pigless world (or
context), or it might be that conditions of language use (i.e., the pragmatics) make
(14) deviant in such a world/context (and hence the existence requirement here
is not a semantic matter). Nothing in the subsequent discussion crucially hinges
on the outcome of this question (for the meaning above could easily be modified
if pig existence were a presupposition), and so we leave the semantics as stated.
10.2. GENERALIZED QUANTIFIERS
 167
10.2.2. The syntax

What about the syntactic category of every pig? For convenience we have called
this GQ. But given the Categorial Grammar syntax, one would hope that the fact
that it can combine with a VP is subsumed under the general rule schemata given
earlier rather than introduce a new rule combining GQ with S/NP. This is easily
done: assume that the category of every pig is S/(S/NP) (the directional slashes
will be discussed directly). This preserves the correspondence set up earlier
between syntactic categories and semantic types, and with suitable directional
slashes allows the combination of a VP with a subject like every pig to be
subsumed under R-1.1
The astute reader might notice two obvious problems with this solution.
Both of these problems will disappear in Chapter 11, but at least the first one
needs to be given a temporary fix before proceeding. And this concerns the
directional features on the slashes. Notice that the general word order rules
spelled out in (3) in Chapter 6 predict that the more fine-grained category here
should be S/L(S/LNP). Obviously this gives the wrong word order (allowing for, e.g.,
grunted every pig). The desired category for every pig is S/ R(S/LN). So the word
order rules must be overridden. For now we will help ourselves to the notion that
the lexical specification overrides the default rules and just build the
exceptionality into a lexical entry. Of course every pig is not listed in the lexicon,
so the exceptional part comes from the lexical entry for every which we
(temporarily) take to be (S/ R(S/NP))/N. (Although this works it is somewhat
embarrassing: for this lexically specified R-slash just happens to give the same
word order one would get if every pig were an ordinary NP. Again this
embarrassment disappears later.) The rest of the slash-directional features will
be correctly specified by the rules, ultimately giving every the ultimate category
(S/R(S/LNP))/RN, which means that every pig is S/ R(S/LNP). Second, we surely
don’t want to have every pig (or every) defined as a “verb.” But if the result category
is S, it is. A suitable use of features could circumvent this, but actually the
discussion in Chapter 11 eliminates this worry as well. We will, incidentally, often
continue to use the term GQ as a shorthand for
S/R(S/LNP).

1 A different solution is taken in Partee and Rooth (1983) where they treat these as NPs,
and allow a single syntactic category (here, NP) to correspond to more than one semantic
type. However, their solution is not rooted within a Categorial Grammar.
 168 10. GENERALIZED QUANTIFIERS

10.4. Consider a universe with six individuals {a,b,c,d,e,f} and a world w 1 in

which [[pig]]S = {a,b}, [[elephant]]S = {c,d,e}, and [[wolf]]S = {f}.
Moreover, [grunt]]S = {a,b,e}. Take the following sentence:

(i) Every pig grunted.

First, show the extension of [[every pig]] S in w1 in set terms. (That is, show the
set of sets in [[every pig]]S, keeping in mind that this is not just sets that we have
names for (such as grunt) but all of the relevant sets. Show the syntactic
computation of (i), and how the extensions are put together, although you need
not show the extension of [[every]]; just begin with [[every pig]]

10.2.3. Extending to the other quantified cases

The beauty of this solution is that all of the GQs can be given meanings of the same
type. Take some pig. The set version of this is the set of sets which have a non-
empty intersection with the [[pig]] set. The function version can be notated in—
among others—any of the following ways:
(15) a. {S| pig’S \ S 6¼ }F

b. ºP[pig’S \ PS 6¼ ]

c. ºP[∃x[pig’(x) & P(x)]]

Take a sentence such as Some pig ate. In set terms, this is true if the [[eat]] set is
in the above set of sets, and for that to be the case there must be some individual
which is in both the [[pig]] set and the [[eat]] set. And that of course gives exactly
the right truth conditions. [[no pig]] is similar; it is (in set terms) the set of sets with
an empty intersection with the [[pig]] set. [[three pigs]] is the set of sets whose
intersection with the [[pig]] set has a cardinality of at least 3; we will write this as
shown in (16):

(16) ºP[ j pig’S \ PS j 3 ]

(Note that we have used at least 3 rather than exactly 3; we return to this in 10.4.
But nothing here hinges on that decision; should Three pigs stayed home mean
that exactly three pigs stayed home, one can trivially revise
(16).) Expressions with complex quantificational determiners like at most
 169
10.3. THE DETERMINERS

three pigs are also amenable to this treatment; [[at most three pigs]] is shown in
(17):
(17) ºP[ j pig’S \ PS j 3 ]

Expressions like many pigs and few pigs are also generalized quantifiers, but the
details are postponed until section 16.3.2.
Recall that giving a semantics for sentences such as (9) Most pigs have lovely
voices (we let the Siamese cats mutate into pigs here) was one of the problems
for Straw Man. It is no problem here. The meaning for most pigs is shown in (18)
(this of course is also the meaning for at least half of the pigs):

(18) ÎP pig’s∩ Ps > pig

2

Notice incidentally that all of the above generalized quantifiers have been defined
in terms of intersection with the exception every. But this too could be defined this
way, since for any two sets A and B, A is a subset of B if and only if the intersection
of A and B is A.

10.5. Take the universe as given in Exercise 10.4, and w 1 as laid out there. Show
the extension of [[no elephant]], [[most elephants]], and [[two pigs]] at w 1.

10.3. The determiners

So far the discussion has centered on the meanings of the full expressions every
pig, no pig, etc. But of course these ultimately need to be compositionally
computed from the meanings of the lexical items, and so we now pull this apart.
The quantificational determiners themselves combine with nouns to give
generalized quantifiers, so their syntactic category is (S/ R(S/NP))/N and their
meaning is of type <<et,>,<<e,t>,t>>. This type looks complex but it really is not:
put in set terms, these determiners take a set and map it into a set of sets. This
means that each determiner is simply a Curry’ed relation between sets of
individuals (more technically, they are Curry’ed relations between functions
which characterize sets of individuals). Take for example [[every]]. In set terms,
 170 10. GENERALIZED QUANTIFIERS

this is nothing more than the Curry’ed version of the subset relation. [[some]] (in
set terms) is the Curry’ed relation between two sets A and B which can be
paraphrases as the “has a non-empty intersection” relation. Put slightly more
formally, it takes a set A and then a set B and returns true iff the intersection of A
and B is not empty.2 Using the lambda notation, one can write [[every]] and
[[some]] as in (19):

(19) [[every]] = ºP<e,t> [ºQ<e,t> [PS QS]]

[[some]] = ºP<e,t> [ºQ<e,t> [PS \ QS 6¼ ]]

10.4. Interlude: Scalar implicature

The above discussion concerning some used singular forms such as some
student. The points to be made below are more naturally exposited using plurals
like some students. Hence, while there may well be differences between some
student and some students, we will overlook these here and switch to the plural
form. Thus consider the meaning given above for some and consequently for
sentences like (20):
(20) Some students (absolutely) love formal semantics.

Given our semantics, this is true as long as the intersection of the student set and
the set of lovers of formal semantics is not empty. This includes the situation in
which all of the students love formal semantics. Yet at first glance that often
strikes speakers as incorrect; on hearing (20) in an actual conversational context
most listeners would conclude that there are also students who don’t love formal
semantics. This, incidentally, is true even if we are given a restricted domain. If I
am talking about the students I have taught over the years (and hence the domain
is restricted) and I utter (20), the listener will conclude that there are some
students within the relevant domain (i.e., those who have taken formal semantics
from me) who do not love formal semantics.
It might therefore seem that [[some]] should be revised as follows:

(21) [[some]] = ºP<e,t> [ºQ<e,t> [PS \ QS ¼6 & PS 6 QS]]

2 There is an extremely rich literature exploring possible constraints on determiner

meanings cross-linguistically: while there are many relations between sets that one could
 171
write out, there are hypotheses that only some of these can actually be realized as single
lexical items. See, e.g., Keenan and Stavi (1986).
10.4. INTERLUDE: SCALAR IMPLICATURE

But there turns out to be good reasons to reject this revision. (The discussion
below is just a small sample of the vast amount of work under the rubric of
Gricean pragmatics, scalar implicature, etc.). First, if this truly were the meaning
of some, then (22a) would be a contradiction; note that the meaning given above
is the meaning for the complex determiner some but not all, and indeed (22b) is a
contradiction:
(22) a. Some students love formal semantics, in fact they all might!
b. Some but not all students love formal semantics, in fact they all might!

In addition, ordinary sentences like (23) would be odd. Why add but not all if some
had the same meaning as some but not all?
(23) Some but not all students love formal semantics.

A temptation might be to claim that some is ambiguous. Suppose there are two
homophonous lexical items some1 and some2, where some1 has the (weaker)
meaning shown in (19) and some2 has the stronger meaning shown in (21).
Before addressing this possibility, note that the discussion here is parallel to the
discussion of the meaning of or in Chapter 3. Recall that at first glance, a sentence
of the form S1 or S2 seems to require at least one but not both sentences to be true;
informally or seems to have “not and” as a part of its meaning. But we showed that
this was not part of the truth conditions.
(24a) is not a contradiction while (24b) is:
(24) a. Sabrina will go skydiving on her vacation or she will go bungee jumping on her
vacation and (in fact) she might do both.
b. Sabrina will go skydiving on her vacation or she will go bungee jumping on her
vacation but she won’t do both; and (in fact) she might do both.

Nor is or ambiguous; this was demonstrated in Chapter 3 with respect to course

prerequisite scenario where the student who had taken both Phonetics I and
Phonology I could not be banned from registering in Phon/Phon II. (See the
discussion surrounding (3) in section 3.2.)
The same point can be repeated with some. Suppose that—instead of (3) in
Chapter 3—the linguistics department in question had the following description
of the prerequisites:
 172 10. GENERALIZED QUANTIFIERS

(25) Prerequisites are strictly enforced. Everyone who has taken some Phonetics course
is allowed to enroll in Phonology I, and no one else is.
Once again the instructor in charge can hardly ban a student who had taken all of
the Phonetics courses from enrolling in Phonology I. And this is not just a matter
of saying that the context biases in favor of the some1 meaning (the weaker
meaning), because we can add background assumptions that should bring out the
some2 (stronger) meaning if it were available. Suppose the linguistics department
in question is such that every Phonetics course covers a certain amount of
phonology, that having taken all of the Phonetics courses one might well have
covered just about everything that is covered in Phonology I, and that the
department does not like students to get credit for courses whose material they
already know. The instructor of Phonology I might well frown upon the enrollment
of the student who has taken all of the Phonetics courses, but she simply cannot
use the language of (25) as a reason to ban the student from the course. Yet if
some were ambiguous, she would be within her rights to do so.
Thus the situation with or and the situation with some are quite parallel: or does
not have the exclusive or meaning and some does not have the “some and not all”
meaning. Just as or often—but not always—suggests “not and,” some often—but
not always—suggests “not all.” As shown in the course catalogue scenario, the
suggestions at issue disappear in the same sorts of environments (we have seen
only one such place here, but others will be discussed in Chapter 17). Incidentally,
most speakers are quite convinced that even if or really has only the inclusive
reading, either . . . or surely is exclusive, as in (26):
(26) Either Mitka or Kolya has been digging in the yard.

There is a very strong suggestion here that both have not been digging. But even
with either . . . or the exclusive reading goes away in the kinds of contexts we have
been considering, as evidenced by the following:
(27) Everyone who has taken either Phonetics I or Phonology I is allowed to enroll in
Phon/Phon II (and no one else is).

Once again this has no exclusive or reading; our instructor cannot ban the student
who has taken both courses.
The fact that the situation with or and the situation with some are similar is not
entirely surprising; there are parallels between or and some just as there are
between and and all, as can be discovered in the exercise below.
 173
10.4. INTERLUDE: SCALAR IMPLICATURE

10.6. Show that or is to and and some is to all. One way to think about this is to
imagine a finite domain, and think of possible paraphrases of Every student left
and Some student left.

In fact these similarities give another reason to doubt the hypothesis that these
words are actually ambiguous: it would be a strange accident if or and some were
both ambiguous, and in just the same ways. And so we turn now to a pragmatic
explanation for why or often suggests “not and” while some often suggests “not
all”—an explanation rooted in how language is used rather than in the literal
meanings of the words. And ultimately the kind of pragmatic explanation to be
developed immediately below combined with a closer look at the semantics will
provide an explanation for just where the stronger understandings do emerge and
where they don’t, as will be shown in section 16.6.
The explanation here is quite standard, and is rooted in what is known as
Gricean pragmatics. This is based on the observation from Grice (1975) that there
is a distinction between the literal meanings of linguistics expressions and their
use. Humans are rational beings engaged for the most part in cooperative
conversation, and hence able to draw certain inferences in various conversational
contexts that go beyond literal entailments. Some of these inferences can be
drawn in virtue of what is not said, as well as what is. One of the relevant principles
of cooperative conversation is that speakers in general will be as informative as
possible (without being overinformative); Grice dubbed this the Maxim of
Quantity. Given two sentences S1 and
S2 where the first entails the second, the first is stronger; i.e., it gives more
information about the world. Suppose then a speaker uttered S 2 when it would
have been just as easy to utter S1 (by this we mean that, for example, S 1 is not
significantly longer, more awkward, etc.). A listener will reason that the speaker is
communicating in a rational and cooperative way, and so given that the speaker
did not utter the stronger sentence S1 it must be because s/he is not in a position
to truthfully utter S1. Perhaps S1 is false, or perhaps it is simply the case that the
speaker does not have enough information to be sure that S 1 is true. These
inferences are known as scalar implicatures, a term first introduced in Horn
(1972) and studied in detail in many subsequent works (see Horn 2004 for a
survey).
 174 10. GENERALIZED QUANTIFIERS

Applying this to the case at hand, consider the relationship between (28a) and
(28b):
(28) a. Some students (absolutely) love formal semantics.
b. Every student (absolutely) loves formal semantics.

Given the semantics in section 10.2.1, (28b) does not—strictly speaking— entail
(28a). For the (b) sentence can be true in a world (or context) with no students,
while the (a) sentence cannot. But assuming that we are looking at these in a
context where speaker and hearer agree that there are students, then (28b)
combined with background assumptions about the context does entail (28a).
From now on we will speak a bit sloppily and say that every entails some, keeping
in mind that this is a shorthand for: a sentence with every combined with the
background assumption that the domain is not empty entails the corresponding
sentence with some. Hence on hearing (28a) (in the nondefective context where
there are students), a hearer will conclude that a rational and cooperative speaker
would have uttered (28b) if they were in a position to. If we add to this an
assumption on the part of the hearer that the speaker is in a position to know
whether (28b) is true or false, the conclusion follows that it must be false. Hence
the illusion that some means “some but not all.” The same logic holds for the case
of and and or; hearing (29a), a hearer can conclude that the speaker was not in a
position to say (29b). Add to that the assumption that the speaker is in a position
to know whether (29b) is true or false, one can conclude that it must be false:
(29) a. Kolya was playing or fighting.
b. Kolya was playing and fighting.

Finally, we have also been assuming that the meaning of three pigs is as given in
(16), which means that Three pigs stayed home is true even if four did. The reason
that this suggests that exactly three pigs stayed home is the same. 3
Two points about scalar implicatures are worth stressing as they are often
oversimplified. First it is technically incorrect to speak of sentences having
implicatures; implicatures are drawn only in the context of actual utterances.
(Sentences have truth conditions, but not implicatures.) The remainder of this
book will at times also be guilty of using the terminology of a 3 Recently, however, the
issue of whether ordinary numbers like three really mean at least three or exactly three has
been revisited; the facts regarding numerals are much less clear than the facts discussed
above with respect to some and or.
 10.5. GENERALIZED QUANTIFIERS AND CONJUNCTION 175

sentence implicating something, but keep in mind that this is just shorthand.
When claiming that a particular sentence has an implicature, what one really is
saying is that the implicature arises in the plausible contexts that one supplies
when no overt context is given. Second, the direct implicature from, for example,
(29a) (given a suitable context) is only that the speaker does not know (29b) to be
true. The stronger implicature that (29b) is false comes from the additional step
of assuming speaker knowledgeability (and therefore is not present in contexts
where one can reasonably assume that the speaker is not in a position to know
whether (29b) is true). Finally, there are many open questions about scalar
implicatures. For example, one must be careful about just how the competing
utterances are computed. Consider (30):
(30) a. Chris used to have a dog.
b. Chris used to have a pug.

(30b) is stronger than (30a), but surely on hearing (30a) a listener will (generally)
not conclude that (31b) is false (or that the speaker does not know that (30b) is
true). This inference would arise only in contexts where the dog breed is relevant,
as for example, if I asked you whether Chris used to have a pug and you answered
with (30a). (In that case, I would conclude that either you don’t know if the dog
was a pug or else that it wasn’t.)4

10.5. Generalized quantifiers and conjunction

10.5.1. VP Conjunction

We now return to the two conjunction problems facing Straw Man. First, recall
from Chapter 5 that Curry’ing the meanings of transitive verbs allows

4
The pragmatic picture developed above has been challenged in recent work such as
Chierchia (2004), Fox (2007), and Chierchia, Fox, and Spector (2012). These authors take
the view that the computation of implicature is not just a matter of human reasoning, as is
taken in the Gricean program. Rather they posit that the compositional computation in the
grammar itself computes not just the literal meanings but also a strengthened meaning; i.e.,
the “implicatures” are part of the system of grammar. A similar although not identical
position is put forward in Levinson (2000). We will not address this here. The interested can
 176 10. GENERALIZED QUANTIFIERS

consult the above works, and see also Russell (2006) and Geurts (2009), among others, for
critiques of the grammatical computation of implicature view.
for a very simple semantics for (31); the conjunction of the two VPs in (31a) is just
set intersection, and in (31b) is set union:
(31) a. Sam fed the pig and meditated.
b. Sam fed the pig or meditated.

Recall further that the only option available to Straw Man was to conclude that
these actually involve hidden S conjunction. But, as discussed earlier, this
(without invoking yet additional apparatus) makes the wrong predictions about the
meanings of (32a) and (33a), which have meanings different from the meanings of
the two-sentence conjunctions in the (b)sentences:
(32) a. Some pig squealed and grunted.
b. Some pig squealed and some pig grunted.

(33) a. Every pig squealed or grunted.

b. Every pig squealed or every pig grunted.

The earlier discussion in Chapter 5 left it open as to how to account for the
meanings of the (a) sentences as we had no semantics at that point for subjects
with quantificational determiners.
But we can now show that the semantics here succeeds in giving the right
meanings to these sentences. Take (33a). [[squealed or grunted]] is the union of
the squealers and the grunters. The full sentence says that this set is in the set of
sets denoted by [[every pig]]. In order for that to be true, the [[pig]] set must be a
subset of the [[squeal or grunt]] set. But notice that all that requires is that each
pig is a squealer or a grunter—it is not necessary for all the pigs to have the same
property. Consider the pigs Strawbuilder, Stickbuilder, and Brickbuilder. If
Strawbuilder squeals and Stickbuilder and Brickbuilder both grunt, the sentence
is true, which is exactly as our semantics predicts. (The secondary reading, which
is brought out with the continuation of but I don’t remember which, is accounted
for in
Chapter 11.) (32a) comes out exactly right as well: a single pig must do both. The
semantics here requires the intersection of the squeal set and the grunt set to
have a non-empty intersection with the pig set. Thus there must be at least one
pig in that VP intersection set; but that itself contains only those individuals that
do both.
 10.5. GENERALIZED QUANTIFIERS AND CONJUNCTION 177

10.7. Take the universe and w1 as shown in Exercise 10.4. Add to that the
following: [[howl]]S = {c,e}. Show the syntactic and semantic computation
(showing the extensional value of each expression at w 1) for both (i) and (ii) (as
usual, simply ignore is and the progressive morphology). Do not show the
actual extension of [[no]], [[and]], and [[is]]; this would be unbelievably
laborious. You can write these out using the ºcalculus. These remarks hold for
subsequent exercises as well.

(i) No pig grunted and howled.

(ii) No pig grunted and no pig howled.

10.5.2. Generalized quantifier conjunction

In addition to VP conjunction, we have seen that generalized quantifiers can also

conjoin with each other, a fact which should be subsumed under the generalized
lexical entries for and and or given in the fragment summary in Chapter 6. Recall
that we have not yet finalized the meanings of these items, but have concluded
that whenever they conjoin two set denoting expressions, the result is intersection
and union. The semantics of sentences like
(34) thus works out exactly as expected:
(34) a. Every student and some professor meditated.
b. Every cat and no wolf stayed behind.
c. Three phonology students or every syntax student will come to thecolloquium.

Take (34a). [[every student]]S is the set of sets with [[student]]S as subset, and
[[some professor]]S is the set of sets with a non-empty intersection with
[[professor]]S. Since these are conjoined by and, the full GQ will denote the set of
sets that occur in each of these. Thus in order for (34a) to be true, [[meditated]] S
must be both in [[every student]]S and in [[some professor]]S. This in turn means
that [[student]]S is a subset of [[meditated]]S (and hence indeed every student did
meditate) and that [[mediated]]S also has as member at least one professor.
These are the right truth conditions. For (34b), we will display the full syntactic
and semantic composition, keeping in mind that and has the Curry’ed meaning
written out in º-notation in (23) in Chapter 9. We begin with lexical items already
fully specified for directional slashes, and until the last step we will use GQ as an
 178 10. GENERALIZED QUANTIFIERS

abbreviation for S/r(S/LNP). It will also be clearest to recast the semantics for and
into the function version (given in (23) in Chapter 9). In reading this, one should
make sure to understand each line, and how the lambda conversions shown here
help to simplify the formula.

(35) no, GQ, ºP<e,t> [ºQ<e,t> [PS \ QS = ]] wolf, N, wolf

no wolf, GQ, ºP<e,t> [ºQ<e,t> [PS \ QS = ]](wolf ’) =

(by the º-conversion equivalence)

ºQ<e,t> [wolf ’S \ QS = ]

and, (X/LX)/RX, ºA [ºB[ºz[A(z) & B(z)]]]

X is instantiated by GQ, so the next step is:
and no wolf, GQ/LGQ,
ºP<<e,t>,t> [ºQ<<e,t>,t> [ºP<e,t>[P(P) & Q(P)]] (ºQ<e,t> [wolf ’s \ QS = ]) = ºQ<<e,t>,t>

[ºP<e,t>[ºQ<e,t> [wolf ’S \ QS = ] (P) & Q(P)]] = ºQ<<e,t>,t> [ºP<e,t>[wolf ’S \ PS = &

Q(P)]]

every, GQ/RN, ºV<e,t> [ºT<e,t> [VS TS]] cat, N, cat’ every cat, GQ, ºV<e,t> [ºT<e,t>
[VS TS]](cat’) = ºT<e,t>[cat’S TS] every cat and no wolf, GQ, i.e., S/R(S/LNP)
ºQ<<e,t>,t> [ºP<e,t>[wolf ’S \ PS = & Q(P)]](ºT<e,t>[cat’S TS]) = ºP<e,t>[wolf ’S \ PS =
& ºT<e,t>[cat’s TS] (P)] = ºP<e,t>[wolf ’S \ PS = & cat’s PS]

stayed behind, S/LNP, stayed-behind’ every cat and no wolf

stayed behind, S, ºP<e,t>[wolf ’S \ PS = & cat’S PS](stayed-behind’) =
wolf ’S \ stayed-behind’S = & cat’S stayed-behind’S

10.8. Take (34b) again. As best you can, describe in prose how the semantic
computation proceeds. This is similar to what is done in the text for (34a)
except that in the informal prose description there we made our lives simpler
by de-Currying and. Your task then is to try to state in prose the meanings of
each of the complex expressions (no wolf, and no wolf, every cat, every cat and
no wolf, and finally the whole sentence).
 10.5. GENERALIZED QUANTIFIERS AND CONJUNCTION 179

10.9. Using the model in (35), take the sentence below in (i) and show its
syntactic and semantic composition (take attended the rally as a single
unanalyzed VP to avoid busywork irrelevant to the point here):
(i) Three students and no professors attended the rally.

Show that this turns out to be synonymous with (ii):

(ii) Three students attended the rally and no professor attended the rally.

*10.10. The sentence in (i) below involves both GQ and VP conjunction: (i) No

cat and no wolf howled and screeched.

Work through the full syntactic and semantic computation along the lines
shown in (35), doing lambda conversions wherever possible (and as soon as
possible). Note that this is synonymous with (ii) and not with either (iii) or (iv).

(ii) No cat howled and screeched and no wolf howled and screeched.
(iii) No cat and no wolf howled and no cat and no wolf screeched.
(iv) No cat howled and no cat screeched and no wolf howled and no
wolf screeched.

Try to get the intuition for why this is so. One might mechanically work through
the derivations of all four sentences to see that our semantics does indeed give
the right results, but the best thing to do would be to try to get an intuition of
how the patterns are working here, and why.

A fork in the road

At this point in the text there is a choice. Readers now have enough background to
proceed directly to Chapter 16 which takes up the distribution of so-called
Negative Polarity Items (as well as additional material on scalar implicatures).
This might be especially rewarding for readers who are anxious to see the formal
tools developed so far pressed into service for making order out of what at first
glance appears to be a rather chaotic domain. However, one can also just move
on to Chapter 11 and postpone the material in Chapter 16. (Another reasonable
path is to insert Chapter 16 after Part II.) In any case, the next section in this
 180 10. GENERALIZED QUANTIFIERS

chapter—on domain restriction—should be read before Chapter 16 and also

Chapter 17; it is also useful for section 13.2.2 which itself is a starred section). It
merits an asterisk before it at this point because readers who are not planning to
look at Chapter 16 at this point can skip it for now. But unlike the other starred
sections (which can be skipped altogether with little or no loss for later material),
this one will be crucial later. Thus it is included now both for readers planning to
move directly on to Chapter 16, and also of course for readers interested in the
topic.

*10.6. Domain restriction

It has been noted from time to time in this text the interpretation of expressions
with quantificational determiners is limited to some domain. If we are farmers and
I turn to you mid-morning and say Every pig was fed, you would certainly not
interpret this to mean every pig in the world, or even every pig in the region, but
every pig on our farm. Similarly if I say No pig ate this morning we again understand
this as no pig from the relevant set—in this case once again the pigs on our farm.
The particular way in which the domain is restricted of course depends entirely
on context. But an interesting question arises. Is the fact a GQ is interpreted
relative to some (contextually set) domain built into its meaning itself? Or is this
simply a consequence of the fact that sentences are always interpreted not with
respect to full worlds but rather smaller situations? The first point of view (that the
semantics of a GQ has a slot for a domain restriction as part of its meaning) has
been taken by von Fintel (1994), Stanley and Szabo (2000) and others; the second
has been argued for by
Kratzer (2004). Here we consider the first view, in fact one argument for this (put
forth originally in von Fintel 1994) will be explored in section 16.4; another
potential advantage emerges in Chapter 17.
Saying that the domain restriction is part of the semantics means that a QP like
every pig does not actually have quite the meaning given earlier, since it has an
extra piece in its meaning which is for the domain restriction. Of course the
particular domain cannot be part of the meaning since that varies wildly from
context to context, but the expectation that there will be a contextually set domain
can be part of the meaning. At this point, we will informally use the notion of a free
variable in order to flesh out the domain
 181
*10.6. DOMAIN RESTRICTION

restriction proposal, even though we have yet to really give a full semantics for
variables in natural language. (In fact, Chapter 17 recasts this without use of a
variable.) In other words, suppose that [[every pig]] is not as shown in (36a) but rather
(36b):
(36) a. ºP[pig’S PS ]

b. ºP[pig’S \ D PS ]

Although a discussion of variables is postponed, it is fairly easy to get an intuitive

understanding of (36b). D is a free variable over sets of individuals, and its value is
ultimately set by the context (note that this is reminiscent of the understanding of the
free variable over comparison classes in the meaning of adjectives as was discussed
in section 8.4.4).
Assuming that (36b) is the meaning for every pig, the next question is how to arrive
at this from the compositional semantics. We will follow von Fintel (1994) in treating
the domain restriction variable as part of the meaning of the determiner. Thus [[every]]
is actually:
(37) ºQ[ºP[QS \ D PS ]

Suppose further that all determiners come with a domain restriction as part of their
meaning. We have already seen that domain restriction is at play in the interpretation
of sentences with no and with every, and it is easy to see that it will also be in play for
sentences like At most three pigs ate this morning. Moreover, a sentence like Some
student passed the exam is not understood as saying some student anywhere, but
rather this asserts that some student out of some relevant set passed. So [[some]] is:
(38) [[some]] = ºQ[ºP[QS \ D \ PS 6¼ ]]

(We have not put brackets around the three sets above because intersection is
associative and so the bracketing would make no difference.) Note that it is also
reasonable to think that there is a domain restriction as part of the meaning of the.
Thus, as noted already in Chapter 1, a sentence like (39) might be taken to be about
the tallest linguistics major at Brown
University:
(39) We didn’t order a gown long enough for the tallest linguistics major.

In view of this, the meaning of the might be revised as (40):

 182 10. GENERALIZED QUANTIFIERS

(40) [[the]] = ºP[Øx[x PS \ D]]

For the moment, little in the discussion in this text hinges on whether or not the
lexical entries here for some, every, etc. do contain the domain restriction piece. It is
easier to work with these meanings if we ignore this part, and so we will generally do
just that. The official fragment includes both versions of the lexical entries for the and
for quantificational determiners (with and without the domain restriction).
 11 Ordinary NPs and type lifting
11.1. Montague’s solution 182 11.4. The lowest types hypothesis 189

11.2. Partee and Rooth’s solution: Lift 11.5. Resolving some syntactic
as a unary rule 185 loose ends 194
11.3. Generalizing the rule 186 11.6. Summary: The fragment so far
196
Chapter 10 began with the observation that expressions like every pig, some pig, etc.
have the same distribution as ordinary NPs like Porky and the mayor of Providence.
We then proceeded to give them meanings quite different from those of ordinary NPs,
and also renamed their syntactic category. The fact that GQs—like ordinary NPs—
occur in subject position has been taken care of in virtue of their full category
S/R(S/LNP), and their ability to occur in object position is a complex topic that forms
the subject of Chapter 14. But there is another way in which ordinary NPs and GQs
behave alike: they can conjoin with each other (hereafter we use the term conjunction
to mean combining with either and or or):
(1) a. Dumbo and every seal escaped from the circus.
b. Obama and some Republican senator will appear at the fundraiser.
c. Obama or some Republican senator will appear at the fundraiser.

Our syntax does not yet admit of these: and and or are of category (X/X)/X but GQ
and NP are different categories. And even if conjunction of unlike categories were
allowed, there remains the more pressing question: how to assign a meaning to these
conjoined expressions?1 We saw that the

1 Indeed there are cases of apparent coordination of different categories; see Chapter 6, n.
9. As noted there, Bayer (1996) shows that these can be reduced to coordination of like
categories in a Categorial Grammar.
184 11. ORDINARY NPs AND TYPE LIFTING

conjunction of two GQs can be handled as instances of set intersection or union,

but ordinary NPs do not denote sets.
In fact, the problem is somewhat independent of the cases in (1). For ordinary
NPs themselves can conjoin:
(2) Bachmann and Perry spoke at the Tea Party rally.

While our syntax predicts these to be well-formed, the semantics has nothing to
say about them: it makes no sense to talk about the intersection of individuals!
Now in the case of (2), one might suspect that and in (2) is a different lexical item
from the and that intersects sets. Perhaps it takes two atomic individuals and
returns a plural individual (such that Bachmann and Perry is quite similar to the
candidates). In fact, there may well be a second and that does just this. But
ordinary NPs also conjoin with or and no similar story could be told for that case.
Again or has been taken to involve set union, but the NPs in question do not (as
we have set things up so far) denote sets:
(3) Bachmann or Perry spoke at the rally.

So the tasks are as follows: (a) to provide a complete account of the identical
distribution of NPs and GQs; (b) to have the syntax be such that GQs can conjoin
with ordinary NPs, and (c) to provide a semantics for (1)–(3). As mentioned above,
(a) will be postponed until Chapter 14; the other tasks will be addressed here.

11.1. Montague’s solution

Montague’s analysis (Montague 1973) provides an ingenious solution. Once again

it relies on the idea that perhaps the meanings are not always what they appear:
perhaps the semantics has packaged things in clever ways. Thus, while one can
retain individuals as basic building blocks in the semantic ontology, Montague’s
hypothesis is that ordinary definite NPs (including proper names) don’t actually
denote (world/time-dependent) individuals, but rather they have a fancier
meaning. The solution rests on the observation that every individual can be
repackaged as a set of sets that contain that individual; the individual r whom we
are used to thinking of as the individual denoted by Romeo can be repackaged as
the set of sets containing r. In function terms, this is the function of type <<e,t>,t>
that maps any function P of type <e,t> to true if P(r) is true. This is shown in (4) in
both set and function notation:
 185
11.1. MONTAGUE’S SOLUTION

(4) [[Romeo]]S = {S| r S}

[[Romeo]] = ºP[P(r)]

One might worry that potentially allows two NPs to “collide”—are we guaranteed
that there could not be two individuals with all the same properties? The answer
is yes, we are: the property of being identical to r is true only for r. In set terms, the
singleton set {r} is the Romeo set of sets and not in the set of sets corresponding
to any other individual. Thus any individual is mapped to a unique generalized
quantifier, and that generalized quantifier corresponds to no other individual. The
function that maps any individual to the corresponding generalized quantifier is
called the lift function (and notate this as l). For any generalized quantifier
generated by an individual there is a function lower that takes the set of sets and
recovers the individual (it is the only individual in every set in the GQ). Of course,
lower is not defined for every GQ; if it were we would have had no need for GQs in
the first place. (Recall that there was no sensible way to take no husky to be an
individual.) Thus l is a total function from individuals to generalized quantifiers,
and lower is a partial function from generalized quantifiers to individuals; it maps
any GQ which is l(x) to x. Note incidentally that now the extension of an NP like
Romeo differs from world to world (as it should since the property of being in love
with Juliet might hold for r in one world but not another), but the idea that proper
names are rigid designators has not gone away. For at the center of the whole thing
remains the individual r who is constant across worlds; and [[Romeo]] in any
world w is the set of properties that that individual r has at w.
The above deals just with proper names, but of course definite NPs like the
disobedient husky will also have lifted meanings. Let m be the individual who is
indeed the (most or only) contextually salient disobedient husky, then the [[the
disobedient husky]] will be ºP[P(m)]. Under this hypothesis, the—like the other
determiners—would be listed in the lexicon with category (S/ R(S/NP)/N and with
meaning of type <<e,t>,<<e,t>,t>>.

11.1. Give the meaning for the under this account. To write it out, you can use
notation that includes the iota-operator in section 9.3 in the discussion
surrounding R-3.

This solves the problems elucidated above. Ordinary “NPs” of course have the
same distribution as GQs because they are GQs. The two always also have the
same semantic type, namely <<e,t>,t>. Since “NPs” like Obama and Boehner both
186 11. ORDINARY NPs AND TYPE LIFTING

denote sets of sets, it is not surprising that they can conjoin. This is just like any
other case of conjunction with and and or of set-denoting expressions; the
semantics is intersection or union. And of course the same holds true for cases of
conjunction of these with the kind of GQs discussed in the preceding chapter. We
will informally walk through just one example here shown in (5) (this is more
natural with but than with and; we assume that the two have the same semantics
for our purposes):
(5) Mitka and/but no cat escaped.

[[Mitka and no cat]]S is the set of sets which is the intersection of [[Mitka]] S and
[[no cat]]S. The former is the set of sets with m as member, and the latter is the set
of sets that have a non-empty intersection with the [[cat]] set. Thus in order for
[[escaped]]S to be in the set of sets denoted by the subject, it has to be both in
[[Mitka]]S and in [[no-cat]]S from which it follows that m must be a member of
[[escape]]S, and that [[escape]]S has an empty intersection with [[cat]]S. These are
exactly the right truth conditions.

11.2. Consider a universe with four individuals {a,b,c,d} and a world w 1. Let
[[Babar]] in w1 (and of course in all worlds) be ºP[P(b)] and [[Dumbo]] be
ºP[P(d)]. Let [[seal]] in w1 be {a,c}, [[elephant]] be {b,d}, and [[intelligent]] be
{a,b,c}.

(a) Show the full compositional computation extensionally for w 1 of the

intelligent elephant. (Thus show the extension of each expression at w 1.)
Remember to fold in the shift of a predicative adjective to the prenominal
adjective.
(b) Show the computation of this for any world. In other words, this question
asks you to write out [[the intelligent elephant]] using the lambda notation,
to show what the value would be regardless of what are the facts of the
world in question.
(c) Show the extension of Dumbo at w1 (i.e., show the actual set of sets, rather
than naming the object as a formula).
11.2. PARTEE AND ROOTH’S SOLUTION

11.3. Add the following to the facts above: [[Silkie]] = ºP[P(c)] (at w 1 and, of
course, for all worlds), and [[escaped]] in w1 = {b,c}. Show the compositional
computation for the following sentences. In each case, first show the value of
 187
each expression extensionally for w 1, and then also show in general how the
semantics is put together using º-notation.

(i) Silkie and the intelligent elephant escaped.

(ii) Dumbo and every seal are intelligent.
(iii) Dumbo or Silkie
(iv) Dumbo and no seal

11.2. Partee and Rooth’s solution: Lift as a unary rule

But Montague’s solution does have drawbacks. First, individuals are part of the
ontology and part of the basic building blocks for meanings, yet there are no
linguistic expressions whose extensions are just individuals. The least complex
type of meaning would be the meaning of a VP (which is a function of type <e,t>).2
(Under CG, the syntax has a similar peculiarity: NP is a basic category and yet
there are no linguistic expressions of category NP.) Moreover, following some
observations in Keenan and Faltz (1985), note that there are proper names for
individuals, but no proper names for other types of GQs. For example there is no
name Droloryman picking out the set of sets which is the union of {S| d S} and {S|
m S}. Were proper names listed in the lexicon as having type <<e,t>,t>, the fact
that they always correspond only to individuals does not follow.
Partee and Rooth (1983) thus proposed an interesting modification of
Montague’s analysis. They suggested that ordinary definite (both proper names
and those with the as determiner) do indeed denote individuals in their basic
meaning. In other words, a proper name like Barack Obama is listed as picking out
(in all worlds), some individual—call it o. And the has the meaning given in Chapter
8: it is of type <<e,t>,e> and maps a set to the unique (or most) contextually salient
member of the set. Thus [[the disobedient husky]] is, as discussed earlier, an
individual. (And, in the CG syntax, its

2
A different solution to this anomaly was proposed in Bennett (1974) where he indeed
takes VP-meanings as the lowest type in the ontology.
category would be just NP/RN.) However, they also proposed a unary rule (which
they dubbed a type shift rule) which allows these to map with no phonological
effect into the generalized quantifiers of Montague. Hence an expression like the
disobedient husky has two meanings; it can denote an individual, or it can denote
the generalized quantifier built from that individual.
188 11. ORDINARY NPs AND TYPE LIFTING

Partee and Rooth were not assuming a Categorial Grammar syntax, but
adapting their proposal into the framework here, we posit the following unary rule
(to be generalized below):
TR-10. Let Æ be an expression of the form <[Æ], NP, [[Æ]]>. Then there is an expression of
the form <[Æ], S/R(S/NP); ºP<e,t>[P([[Æ]])]>.

11.4. Return to the world and lexicon as laid out in Exercises 11.2 and 11.3.
Given the Partee and Rooth proposal, we of course must modify the meanings
for the lexical items given there; now [[Babar]] (in all worlds) = b, [[Dumbo]] =
d, and [[Silkie]] = c. Redo your answers to (11.3i) and (11.3iii) using the Partee
and Rooth proposal.

11.3. Generalizing the rule

It turns out that generalizing TR-10 will have nice consequences as we built the
system further. First, we generalize the definition of l (lift). For any member x of a
set X and any set of functions F from X to Y, there is a function l(x) from F to Y, such
that for all f in F, l(x)(f) = f(x). Put informally, l(x) is something which takes any
function g as argument (provided that x is in the domain of g), and gives back that
which one would have gotten had g applied to x. A somewhat subtle point is that
as defined here, the domain of l(x) is every single function which has x in its
domain—not just those functions whose domain is some particular set of
interest. Similar remarks apply to the co-domain. Since it is often convenient to
consider just those functions in l(x) which are functions from some specific set X
to some specific set Y, we will use the notation l <X,Y>(x) when needed (but
subscript on the l operation will be omitted when not necessary). Thus l <X,Y>(x) is
a function of type <<X,Y>,Y>. The original lift of individuals to functions of type
<<e,t>,t> is just one instance of the
11.3. GENERALIZING THE RULE

general l operation. Given an individual a, l <e,t>, (a) is the function that takes any
<e,t> function P and returns the value of P(a); it is ºP[P(a)]. We can now generalize
TR-10 as follows (the syntax will be further generalized in 11.5):
TR-11. Let Æ be an expression of the form <[Æ], A, [[Æ]]>. Then there is an expression of
the form <[Æ], B/R(B/LA) or B/L(B/RA), ºf<a,b>[f([[Æ]])] (i.e., l<a,b>([[Æ]]) )>, where a
is the type for expressions of category A and b for expressions of category B.
 189
Of course the syntax here looks a bit dismal. In the first place there are two
possible categories for the output expression (depending on the slash directions).
And we surely should be embarrassed about what looks to be an accidental fact
that the directionality on the slashes just happens to give the new expression a
category that preserves word order. In other words, it just happens to ensure that
the new lifted expression combines with an expression of category B/A to give just
what would have resulted in the syntax from using the unlifted version.
Temporarily, we will live with this embarrassment.
Consider a simple example like Mitka howls. Notice that we now have an
infinite number of derivations for this sentence. The first is the obvious one; the
lexical meaning of Mitka is the simple individual m is the argument of [[howls]] (of
type <e,t>). The second involves lifting Mitka to the GQ with meaning ºP[P(m)] and
that takes ordinary [[howls]] (still of type <e,t>) as argument. But now we have
introduced a third derivation, in which [[howls]] is lifted to be of type <<<e,t>,t>,t>;
that is, it takes a generalized quantifier as argument.

11.5. Show this derivation (showing both the syntactic and semantic
composition). Do as many l-conversions as you can to show that the meaning
is the same as the meaning that results from the two it is equivalent to the two
derivations given above.

But there is no harm in allowing in these extra derivations: they all converge on
the same meaning. Indeed this is guaranteed by the very definition of l; for any
function f l(a)(f) is destined to give the same result as f(a), for that is exactly how
lift is defined.
What, then, could possibly be the advantage of incorporating the generalized
rule in TR-11 which allows for all of these extra derivations? The advantages come
from cases where the output of the lift operation itself “does something” other
than immediately taking the would-be function as its argument—as, for example,
when it conjoins with another expression of the same category. As one such case,
recall the extra reading of a sentence like (6) discussed in Chapter 5 (the material
in parentheses is there to bring out the relevant reading):
(6) Every wolf howled or sang (but I can’t remember which).

This reading is now predicted; it is the reading that results if howl and sing are both
lifted over GQ meanings. That is, each of these denote sets of GQs; they combine
with or (to give the union of the relevant sets), and then they take the subject GQ
190 11. ORDINARY NPs AND TYPE LIFTING

as argument. The derivation is spelled out in (7), using function rather than set
notation throughout. We leave off those slash directions which are predictable
from the word order rules given in section 6.3, and put in only those which are
stipulated either as part of the lexical category (in this case, of every) or are given
by TR-11 (those stipulations themselves will be removed in section 11.5). The
derivation in (7) also simplifies the full analysis of or; since or in actuality is
Curry’ed, the expression howl or sing involves two steps, but we collapse these
here into one:
(7) howl, S/NP, howl’ (of type <e,t>) ! by TR-11 howl, (S/L(S/R(S/NP)), ºP<<e,t>,t>[ P (howl’)]

sing, S/NP, sing’ (of type <e,t>) ! by TR-11 sing, (S/L(S/R(S/NP)), ºQ<<e,t>,t>[ Q (sing’)]

howl or sing, (S/L(S/R(S/NP)), ºR<<e,t>,t>[ºQ [Q(sing’)](R)] ∨ºP[P (howl’)](R)]

= ºR [R (sing’) ∨R (howl’)] every wolf, S/R(S/NP), ºP[8x[wolf ’(x) ! P(x)]]

every wolf howl or sing, S, ºR [R (sing’) ∨R (howl’)](every-wolf ’) = every-
wolf ’(sing’) ∨ every-wolf ’(howl’) = ºP[8x[wolf ’(x) ! P(x)]](sing’) ∨
ºP[8x[wolf ’(x) ! P(x)]](howl’) =

8x[wolf ’(x) ! sing’(x)] ∨ 8x[wolf ’(x) ! howl’(x)]

11.6. The derivation shown in (7) derives the relevant reading by lifting each of
the VPs separately and then combining them with or. What would be the
reading for the derivation in which they are first conjoined with or (that is the
unshifted lexical items conjoin), and then the resulting VP howled or sang lifts
by TR-11?
 11.4. THE LOWEST TYPES HYPOTHESIS 191

11.4. The lowest types hypothesis

11.4.1. Some additional predictions

Partee and Rooth made the interesting speculation that in processing a sentence,
a listener will first interpret expressions in their lowest types (that is, the types
given by the lexicon without additional rules), and will posit the higher type
meanings only if there is evidence for this. (Notice that this is phrased as a
principle of the processor and not of the grammar. Some modern work recasts
this as a grammatical principle, we return to this in 11.4.2.) The idea is that the
processor will recompute and use higher types if there are either grammatical or
pragmatic reasons for this. For example, in the sentence in (6) the continuation
but I don’t remember which forces the more complex reading. Or, there could be
other pragmatic or contextual cues leading a listener to the more complex
reading. This also means that in processing an ordinary sentence like Porky
grunted a listener computes the meaning using the individual-denoting meaning
of Porky. However, in hearing a sentence like Porky or Strawbuilder grunted or
Porky or every wild boar grunted the existence of the conjunction causes a
reanalysis whereby the lifted meaning of Porky is accessed. A nice consequence
of this proposal is that it succeeds in predicting that the extra reading for every
wolf howled or sang is less available and is a secondary reading, which accords
with most people’s (admittedly unsystematically collected) intuitions. 3
A second example of the advantage of both generalizing the lift rule and
adopting the lowest types hypothesis concerns the ambiguity of sentences like
(8):
(8) Every student didn’t pass the exam. 3 This hypothesis is reminiscent of some other

proposals in the syntactic processing literature according to which listeners compute

the “minimal” derivation available such as the “Minimal Attachment” hypothesis of

Frazier (1978). In some of these other cases, however, it is unclear how this actually

could take place for no algorithm is provided (nor is any algorithm obvious) that would

guide the processor to the minimal derivation. The lowest types hypothesis, however,

is quite clear and avoids this problem: the processor initially accesses the lexical

meaning and then recomputes only when needed.

 192 11. ORDINARY NPs AND TYPE LIFTING

On one reading (arguably the most accessible one), the sentence claims that all
the students failed. But it has an additional reading: one which is often referred to
under the rubric of not scoping over every; in this reading (8) says only that at least
one student failed. This is not as immediately accessible but can certainly be
brought out with the right context (and/or intonational contour). To show this with
a slightly different example, it used to be standard protocol on Amtrak trains
running between Boston and Washington that—as the train approached smaller
stations such as New London, New Haven, etc.—the conductor would announce
(for example):
(9) Next station stop, New Haven. All doors will not be open.

Surely the passenger scheduled to disembark in New Haven would not say
“But conductor, my ticket is to New Haven—how can I possibly get off this train?”
(Fortunately, should there have been a pragmatically challenged passenger the
conductor did generally follow this announcement with something like “Please
exit at the rear of the train.”)
One kind of account for the “wide-scope negation” reading is the one
embedded in a theory that maintains that a sentence is mapped into an LF which
inputs to the model-theoretic interpretation. Under this view, not can raise and
take scope over the full sentence giving the secondary reading. (Such accounts
generally treat the primary reading in a somewhat different way than has been
done here, but we will develop that here.) But notice that the relevant reading can
also be accounted for under the Direct Compositional model without invoking a
level of LF.
First is a small housekeeping detail. In the fragment in Chapter 6, didn’t has as
its category (S/NP)/(S/NP). It was not given a cross-categorial definition (the way
and and or were) and so can combine with VPs but not necessarily with members
of other categories. Indeed, it should not have a completely general category since
it does not combine with NPs, APs, and so forth. However, to account for the
secondary reading we do need to assume that it can combine both with ordinary
VPs and with lifted VPs (and hence it also has the category
(S/(S/(S/NP)))/(S/(S/(S/NP))). Although the existence of this extra category does
not follow from anything, recall that we are oversimplifying in any case: didn’t is
not really a single lexical item and the actual negation is either not or n’t which
follows (and in the case of n’t is cliticized to) the first in a string of auxiliaries. We
have avoided a full treatment of negation in English because a treatment of the
syntax of auxiliaries is far beyond the current scope. Thus simply assume that the
 11.4. THE LOWEST TYPES HYPOTHESIS 193

lexical entry for didn’t from Chapter 6 is revised as follows: <[didn’t], X/X,
ºF<a,t>[ºAa[~F(A)]], where X is restricted to S/NP or S/(S/(S/NP))>. With this, the
secondary reading is accounted for; showing this is the task of the next exercise.

11.7. Show the simplest (lowest types) derivation (both the syntax and the
semantics) for (8) and then show the derivation which yields the weaker
meaning to the effect that at least one student failed the exam (but not
necessarily all). You can treat pass the exam as a single lexical VP (with
meaning of type <e,t>) for simplicity.
11.8. Show a third derivation in which didn’t first combines with the <e,t>
meaning of pass the exam and then that VP lifts over generalized quantifier
meanings. What are the truth conditions that result from this derivation?

There remain some unsolved problems with the adoption of the fully
generalized lift rule in TR-11, for indeed there are many cases in which the
additional readings do not exist. For example, the relevant reading is impossible in
(10):
(10) Some student(s) failed syntax and forgot to come to the phonology exam.

The extra reading—analogous to the secondary reading for (6)—would be one in

which some students failed syntax and some forgot to come to the phonology
exam, but they need not be the same students. (This, of course, is the reading one
would expect also under the Conjunction Reduction analysis discussed and
rejected in Chapter 5.) Notice that even when the context is loaded so as to give
this reading every chance to emerge, it is absent. Suppose for example we are
discussing where students go on their spring vacations, and I report:
(11) Some students go skiing for the week and stay home and study.

The only reading this sentence has is the contradictory one in which the same
students did both; there is no way to force this into the pragmatically sensible
reading. Similarly, there does not seem to be a second (“wide-scope negation”)
reading for (12):
(12) No student didn’t go skiing.
 194 11. ORDINARY NPs AND TYPE LIFTING

The only available reading is that every student skied; there is no weaker reading
which simply asserts that at least one student skied. Incidentally the problem
here is not a problem just for the type-lifting account; any other account of the
ambiguity of (6) and (8) will need to ensure that the mechanisms for that do not
generalize to give additional readings for (11) and (12).

*11.4.2. The lowest types hypothesis: Processing vs grammar

The lowest types hypothesis (or variants thereof ) is sometimes phrased in terms
of a constraint in the grammar within theories that do not maintain a Direct
Compositional architecture. In particular, suppose that the input to the semantics
is a syntactic representation (whether it be the surface or an LF representation).
Consider a compositional interpretation of an expression of category A containing
two daughter expressions of category B and C. (We are not using a Categorial
syntax here.) Suppose that [[B]] is a function of type <c,a> and [[C]] is of type c.
The idea is that the only possible semantic derivation is the one in which [[B]]
applies to [[C]]. Thus C cannot map to the lifted meaning which is the output of
(the semantic part of) TR-11. This means that, for example, the meaning of Mitka
howls can be put together in only one way. There is, however, a type-lifting rule
along the lines of Partee and Rooth, which can apply “as a last resort.” That is, if
the semantics encounters an expression like [Mitka and every wolf] the only way
to combine these meanings is for the expression Mitka to lift, and in that case this
is allowed. Thus here the claim that the meaning of Mitka howls is put together
using the lowest types is not a claim about processing, but indeed the grammar
does not allow the derivation where Mitka lifts to take howls as argument (nor any
of the higher derivations). In fact, more generally a number of researchers have
posited that any type shift rule (a unary rule shifting the meaning of an expression
into a new meaning with a different type) applies only “when needed” or “to repair
type mismatch.”
The “type lift (or, more generally, type shift) as a last resort” view as a principle
of grammar (rather than processing) makes no sense under Direct
Compositionality. Nor does the idea that such rules exist in order to “repair” type
mismatches (another slogan that one sometimes sees). This rubric presupposes
a view where there is an already computed object which the semantics is “trying”
to interpret. Note of course that this view means that the additional readings
 11.4. THE LOWEST TYPES HYPOTHESIS 195

discussed above for (6), (8), and (9) cannot be accounted for in the way suggested
here (using type lifting). This is because the grammar has a perfectly good way to
put together the meanings without this, and if type lifting is allowed in the grammar
only when there is no other way to compose the meanings, it will not be allowed
here. This illustrates a key difference between the last-resort strategy as a
principle of processing vs a principle of grammar. A human being processor (such
as our poor Amtrak traveler wishing to de-train in New Haven) can (hopefully!)
reason that the lowest types meaning is pragmatically inappropriate, and
recompute. The grammar, however, is not “smart” in this way; it is a purely
algorithmic device with no ability to reason about trains and stations and
conductors. Thus a theory that maintains “type lifting as a last resort in the
grammar” would need a different account of, for example, the Amtrak reading.
The Direct Compositional view of grammar taken in this book has no notion of
unary rules (such as type lifting) existing “in order to repair mismatches” and no
notion that these rules can apply only when needed. For there can be nothing to
repair; the syntax and semantics work together to prove expressions well-formed
and assign them a meaning. There can be no notion of a rule needing to apply, for
the grammar itself has no needs or goals. Only the human processor does: the
goal of the processor is to interpret the sentence, but the grammar is a purely
static system predicting the well-formedness of expressions and pairing them
with meanings. Thus, with Partee and Rooth, we will assume that the lowest types
hypothesis is a hypothesis processing, not grammar. 4

4
Sometimes one sees claims that the existence of unary (or, type shift rules) is
explained by the slogan that they “exist to repair type mismatches” (this goes along with
the notion that they apply only as a “last resort”—i.e., when need to repair type
mismatches). But it is unclear what—if anything—this is supposed to explain. First, why
would one want an explanation for the existence of these rules as opposed to any of the
other sorts of rules in the grammar? Perhaps the driving intuition is that because these rules
do not affect the phonology of an item, their existence is peculiar and needs “explanation”
in a way that other rules don’t. But even if we grant this, there are problems with the claim
that the last-resort principle explains the existence of these rules. For one thing, there are
many unary rules (especially ones mapping single words into others) that do not clearly
have this function. Consider for example the verb give which can appear with both two NP
complements and an NP and
 196 11. ORDINARY NPs AND TYPE LIFTING

11.5. Resolving some syntactic loose ends

We can now clean up some of the aspects of the syntax which seemed
problematic. First, recall that the generalized lift rule given in TR-11 had the
embarrassing property that the syntactic category of the derived expression was
such that it happened to preserve word order. That is, when the lifted version of
Porky (which is S/R(S/LNP) combines with a VP like grunted (of category S/LNP) the
lift rule ensures that it takes its VP argument to its right. In the first place this was
just stipulated as part of the rule. We might have expected the lift rule itself to give
an underspecified category. But then the general word order directional slash
rules (given in (3) in
Chapter 6) would make the category S/ L(S/LNP). So we need an extra stipulation
here to override the general case: this is reminiscent of the discussion earlier
about the lexical entry for every (see section 10.2.2). Worse, this stipulation
makes it a complete accident that the lifted material takes the function as its
argument to give exactly what we would have gotten had the function expression
taken the original expression as its argument. And of course we also have the
possibility of output expressions of the general form A/L(A/RB). This is the general
form of the category of lifted VPs (thus these are S/ L(S/R(S/NP)) and here we need
to ensure that they take the generalized quantifier to their left (just as the
generalized quantifier would have taken the unlifted version to its right), again
making it looks like an odd coincidence that the lifted VP combines with a GQ to
give exactly what we would have gotten had the GQ combined with the unlifted
VP.
But it actually is no accident that it works this way if lift is an operation on
syntactic categories as well as on the meanings. Recall that section 6.5 posited
that any category of the form A/DB (for D some directional feature) corresponds to
a (Curry’ed) function taking two strings as argument and returning a third. Given
that, there is a natural connection between the semantic operation l and the
syntactic operation performed by the lift rule. Speaking informally, what it means
to be a lift of a category A means to

PP[TO], where one can derive one meaning from the other (see Exercise 9.10). But since
both meanings are of type <e,<e,<e,t>>>, a type mismatch hardly “explains” the difference
in meaning. Second, one must still predict which shifts exist and which don’t. And third,
under this sort of teleological view of the grammar, there is no explanation as to which
 11.4. THE LOWEST TYPES HYPOTHESIS 197

mismatches are allowed by the syntax in the first place (nor, of course, any explanation as
to why the syntax would allow these mismatches).
 198 11. ORDINARY NPs AND TYPE LIFTING

11.5. RESOLVING SOME SYNTACTIC LOOSE ENDS

return a category B/(B/A) which is such that when an expression in this category
combines with an expression of category B/A, the result is the same as when the B/A
combines with the homophonous (unlifted) expression in B. To formalize this, we
define L as an operation on syntactic categories where L is defined in terms of the
functions that correspond to each category.
(13) Definition:
For any category A, LB(A) is any category such that for all expressions x of category A and
all expressions y and z such that [x] = [z], FLB(A)([y])([z]) = FCAT(z) ([z])([x]).

This looks slightly different from ordinary lift for the simple reason that the strings
themselves are not functions (unlike the meanings of those strings). Rather, the
category of a string corresponds to a function and it takes one argument more than
does the corresponding meaning of the string, and this fact is reflected in the above
definition. To see the connection between this we can state this in prose: the lift of
any category A is a category with the following property: for any x in A, the new
category is such that when applied to an expression y and then an expression z
(whose phonology is the same as that of x) the result is the phonological string that
one gets by taking the function associated with the category of z and applying it to x.
Since x is an expression whose category is A, it follows that z must have as its
category B/A (for some B), and that x must therefore be of category B/(B/A). Notice
also that for any category A, LB(A) is not unique (hence LB is not a function).
For NP, for example, LS(NP) is both S/R(S/LNP) and S/L(S/RNP).
This might seem like quite a bit of trouble to generalize the two outputs of
L(A) rather than simply listing them as B/ R(B/LA) or B/L(B/RA), but it has at least three
advantages. First, it makes the details of the lifted categories nonaccidental,
capturing the idea that the fact that word order is preserved is not stipulated but is
part of the definition of the operation. Second, it naturally extends if there are
additional operations such as Wrap. (See the extended fragment in Chapter 6.) Thus
if there are items of category A/ WB whose corresponding function is as defined in the
fragment in Chapter 6 there needs to be one additional slash which is A/IB (I for infix);
the definition of FA/ B can be supplied by the reader. Thus giving a general definition of
I

L allows an A to lift to B/I(B/WA) and so forth. The third advantage concerns the lexical
category of quantificational determiners, to which we return directly.
From here on we will use the convention of bold italics to notate semantic
operations (operations on model-theoretic objects); bold capital letters for
operations on syntactic categories, and bold lower-case non-italics for names for the
 199
full rules. Moreover, whenever there is need to specify the ultimate result category of
some category L(A), we notate that as LB(A), as was done above. As usual, subscripts
will be omitted when not needed.
Armed with this, we can recast the lift rule so that it does not stipulate the slash
directionality on the output—it simply says that the input category A maps to L(A).
Hence we revise TR-11 to rule R-4, which will also be referred to as l:
R-4. Let Æ be an expression of the form <[Æ], A, [[Æ]]>. Then there is an expression of the
form <[Æ], LB(A); ºf<a,b>[f([[Æ]])] (i.e., lb ([[Æ]]))>.

(Since LB(A) is not unique, R-4 actually allows an expression to map to more than one
output expression.)
We now also have a solution to the problem of the lexical category of
quantificational determiners. Recall that we have listed these as (S/ R(S/NP))/N where
the R feature was stipulated in order to override the default rules. But we can now
solve this by saying that the category itself in the lexicon is not spelled out as above,
but is simply LS(NP)/N. Note, incidentally, that this means the category can be spelled
out in two ways, either as (S/R(S/LNP))/RN or as (S/L(S/RNP))/RN, hence expressions
like every dog are both S/R(S/LNP) and S/L(S/RNP). Finally, there is no problem with the
grammar being “tricked” into thinking that an item like every is a verb and undergoing
verbal morphology. Although its ultimate result category is S, it is not specified in the
lexicon the way verbs are.

11.6. Summary: The fragment so far

The following lexical items and rule are added to the fragment from Chapter 8:
Lexicon: because these are all constant functions from worlds, we leave off the world
parameter:
The determiners without domain restriction:

<[every], LS(NP)/N, ºP[ºQ[PS \ QS = PS]]>

<[some], LS (NP)/N, ºP[ºQ[PS \ QS 6¼ ]]>

<[no], LS(NP)/N, ºP[ºQ[PS \ QS = ]]>

<[three], LS(NP)/N, ºP[ºQ[ |PS \ QS| 3]]>

<[most], LS(NP)/N, ºP[ºQ[ |PS \ QS| > |PS|]]>

2
 200 11. ORDINARY NPs AND TYPE LIFTING

11.6. SUMMARY: THE FRAGMENT SO FAR

The determiners with domain restriction (temporary; to be revised in Chapter 15

without use of a free variable for the domain):

<[every], LS(NP)/N, ºP[ºQ[PS \ D \ QS = PS]]> and so forth

for all of the others above.

In addition there are complex determiners like exactly three, at most three, less than
half, but since these ultimately need to be compositionally derived from their parts,
we will not include them here:
<[didn’t], X/X, ºF<a,t>[ºAa[~F(A)]], where X is restricted to S/NP or S/(S/(S/NP))>

Definition of l: For any x (of any type) in a set X and any set Y, l Y(x) is a function in <<X,Y>,Y> such
that for all f in <<X,Y>,Y>, = lY(x) = f(x).

Definition of L: For any category A, LB(A) is any category such that for all expressions x of
category A and all expressions y and z such that [x] =
[z], FLB(A)([y])([z]) = FCAT(z)([z])([x]).
Rule l:
R-4. Let Æ be an expression of the form <[Æ], A, [[Æ]]>. Then there is an expression of the
form <[Æ], LB(A), ºf<a,b>[f([[Æ]])] (i.e., l<a,b>([[Æ]]))>.
 201

12 Generalized conjunction
12.1. Generalizing and and or 198 12.3. Summary: The fragment
so far 203
12.2. Remarks on noun conjunction 201

12.1. Generalizing and and or

The last several chapters featured instances of and and or conjoining expressions
of a variety of categories, where the semantic contribution of and and or is often
intersection and union respectively (more accurately, as the Curry’ed versions of
these two binary operators). But to say that the semantics is always
intersection/union is not quite right. In the first place, as was hopefully discovered
in Exercise 5.6, one cannot collapse S conjunction and VP conjunction in this way.
In the case of S conjunction, intersection is of the set of worlds characterized by
each of the daughter Ss, which means that intersection crucially is of intensions.
The case of VP conjunction involves intersection only once the world argument
has been supplied; here it is intersection of the extensions. Moreover, there are
cases of more complex expressions which can conjoin. Take for example the case
of transitive verbs:
(1) a. Br’er Fox chased and caught Br’er Rabbit.
b. Tim watched and hated the movie.

Transitive verbs have meanings of type <s,<e,<e,t>>>. Even considering just their
extensions, they are of type <e,<e,t>>—they do not characterize a set, and so they
do not denote the kind of object that can be intersected. Still, one might have the
intuition that these all involve intersection “at the end of
12.1. GENERALIZING AND AND OR

the day.” In other words, [[watched and hated]] (ignoring the world argument) is
ºx[watched’(x)S \ hated’(x)S]. For that matter, one can see everything as ultimately
tied in to ordinary first-order propositional &, for to be in the intersection of two
sets A and B is to be in A and to be in B. All of the remarks above hold for the
parallel case of union with respect to or.
This is exactly the idea that Partee and Rooth (1983) pursue in giving a single
cross-categorial definition and meaning for and and or. They first recursively
define a binary operator meet, notated . is defined for any two members of the set
of truth values, and for any two functions f and g, f g is defined if for all x, f(x) g(x)
is defined. Moreover, for the base case—where takes two truth values and returns
a third—its semantics is just &. This is given in (2i). The recursive definition allows
it to take two complex functions (of the same type), each of whose ultimate result
is a truth value and return a function of the same type as the inputs, and the
definition is given in (2ii) (here we use the notation (a,b) for ordered pairs):
(2) i. Given two truth values a and b (i.e., members of {1,0}) a b = a & b (i.e., = (1,1)
à 1, (1,0) à 0, (0,1) à 0, and (0,0) à 0)
ii. Given two functions f and g such that f g is defined, f g = ºx[f(x) g(x)].

Then and can be listed in the lexicon as being of category (X/X)/X, with meaning
ºA[lB[A B]]. This meaning is just Curry’ed .
A similar treatment is available for or. First we recursively define join, notated

q. As above, it is defined for any two members of the set of truth values, and for

any two functions f and g, f q g is defined if for all x, f(x) q g(x) is defined. The base

step again is the same as first-order v, and the recursive step is parallel to the

definition of :

(3) i. Given two truth values a and b (i.e., members of {1,0}) a q b = a v b (i.e., q = (1,1)
à 1, (1,0) à 1, (0,1) à 1, and (0,0) à 0)

ii. Given two functions f and g such that f q g is defined, f q g = ºx[f(x) q g(x)].

The full lexical entry for or, then, is <[or], (X/X)/X, ºA[ºB[A q B]]>. Both of these,
then, are given a generalized meaning that applies to a variety of semantic types,
and are given a cross-categorial syntax.
Notice that, since English contains no expressions that are of type t (the lowest
type are expressions of type <s,t>) the lowest meaning of and is ºp <s,t>[ºq<s,t>[p q]]
which is the same as ºp<s,t>[ºq<s,t>[ºss[p(s) & q(s)]]], and this is found in ordinary S
conjunction cases such as Porky grunted and
 203
200 12. GENERALIZED CONJUNCTION

Trickster howled. However since we are systematically ignoring intensions, we

can for practical purposes in this text take the lowest type meaning of and to be
just ºp[ºq[p & q]] and this is what is used in conjoining S. For the case of VP
conjunction (keeping everything extensional again, as always), the recursive
definition gives and the meaning ºP<e,t>[ºQ<e,t>[ºxe[P(x) & Q(x)]]] which is the
meaning we’ve been using all along. Thus the intersection of two sets is everything
that is in both sets, which is equivalent to ºP<e,t>[ºQ<e,t>[PS \ QS]].
Now consider a more complex case, such as the case of transitive verbs as in
(1a). We show this using the full syntactic and semantic composition in (4). We
add the directional slashes throughout, but of course in all cases these are not
part of the lexical entries but are supplied by the word order rules given earlier in
(3) in Chapter 6.
(4) caught, (S/LNP)/RNP, caught’ and, (X/LX)/RX, ºA[ºB[A B]] and caught,
((S/LNP)/RNP)/L(S/LNP)/RNP)), ºR<e,<e,t>>[caught’ R] chased and caught,
(S/LNP)/RNP, caught’ chased’ = ºx[caught’(x) chased’(x)] =
ºx[ºy[caught’(x)(y) chased’(x)(y)]] =
(since we are ignoring the world argument) ºx[ºy[caught’(x)(y) & chased’(x)(y)]]
Br’er Rabbit, NP, r chased and caught Br’er Rabbit, S/LNP, ºx[ºy[caught’(x)(y) &
chased’(x)(y)]](r) = ºy[caught’(r)(y) & chased’(x)(y)]
Br’er Fox, NP, f
Br’er Fox chased and caught Br’er Rabbit, S, ºy[caught’(r)(y) & chased’(x)(y)](f) =
caught’(r)(f) & chased’(r)(f)

In prose, what and does when it conjoins two transitive verbs such as chased and
caught is create a new complex transitive verb which is the relation that holds
between two individuals a and b just in case a chased b and a caught b. Notice
that in showing the composition of (4), we rewrote caught’ P chased’ by two steps
to the equivalent lx[ly[caught’(x)(y) & chased’(x)(y)]].
This has the opposite effect, as does º-conversion; here a formula is being
rewritten as a longer one. But it is useful as it is more readable. Once again one
should bear in mind that converting one formula to another is just a way to write
meanings in a way which might be easier to read, but it is not necessary (except,
of course, when asked for in an exercise).
 12.2. REMARKS ON NOUN CONJUNCTION

12.1. Show the full syntactic and semantic computation for (i) (treat briar
patch) as a single unanalyzed noun. Use the model in (4).

(i) Every rabbit or some fox entered the briar patch.

12.2. a. Show the full syntactic and semantic computation for (i):
(i) The rabbit is happy and healthy.
b. Now do the same for (ii). Caution: Keep in mind the word order rules for
adjectives that modify nouns. (ii) the happy and healthy rabbit

12.2. Remarks on noun conjunction

Noun conjunction presents an interesting problem. Consider (5):

(5) Every man and woman voted in the election.

There are two points to notice here. The most striking is the fact that the lowest
type meaning is absent: this does not have a “hermaphrodite” meaning, where it
claims that everything which is both in the [[man]] set and in the [[woman]] set
voted. On the other hand, (5) is quite easily understood as synonymous with every
man and every woman voted in the election.

12.3. Show how the grammar developed so far predicts the existence of the
reading in which (5) means the same as every man and every woman voted in
the election. You will need to think about how to type-lift man and woman to
get the appropriate meanings.

As shown by Exercise 12.3, it is not surprising that every man and woman can
have a meaning which is the same as every man and every woman. But what is
surprising, perhaps, is the lack of the hermaphroditic meaning. Now in this
particular case it is tempting to place the entire burden on the pragmatics. Such a
reading is obviously absent (or at best difficult to access) because our world does
not contain hermaphrodites. One might push this reasoning further. Recall the
remarks in Chapter 8 that many of the
202 12. GENERALIZED CONJUNCTION
 205

concepts that get lexicalized as nouns (at least as those which are often referred
to as basic-level nouns) tend to be sets that are mutually exclusive— a tiger is not
also a lion and a table is not also a chair. Thus in many cases the pragmatics will
rule out the intersection of the lowest types reading (that is, the intersection of the
sets which correspond to the basic lexical meaning of the two nouns will
necessarily be empty, and hence a listener assuming a cooperative speaker will
recompute to the higher types). Unfortunately, this cannot be the whole story. For
even when we choose nouns whose sets are very easy to intersect, and even
when we stack the context to bring out this reading, the relevant reading still
seems to be absent or, at least, extremely difficult.1 For example, take the context
of a club for both students and faculty who enjoy any kind of outdoor activity,
including skiing, hiking, and bungee jumping. If we saw a notice that the club would
be hosting a meeting that was open only to every bungee jumper and professor,
we would be unlikely to conclude that the invitees to the meeting are just those
who are both professors and bungee jumpers. A student bungee jumper and a
professor hiker would seem to be invited as well. 2 (The judgments are, however,
delicate, and the interested reader is invited to construct other examples and
perhaps do some systematic informant work in order to elucidate the full pattern.)
Similar facts hold with relational nouns. Thus expressions of the form
RN1 and RN2 generally lack the reading: ºx[RN1’(x)S \ RN2’(x)S] (for RN an
abbreviation for N/PP[OF]). This is again true even when the pragmatics makes
such a reading quite available, as in (6).
(6) a. Every cousin and fellow student of John is coming to his party.
b. Every professor and employer of Maria will write
her a very good recommendation.

1 There are cases where this reading does emerge, as in (i) (discussed in Partee and
Rooth 1983 and attributed to unpublished work by R. Cooper):
(i) Every wife and mother is underpaid.
2
Further complexities arise when the determiner is a; here it often becomes much easier
to get the “missing” reading, as in:
(ii) A student and bungee jumper is running for election.
(ii) is most naturally understood with the lowest type meaning where a single person who
is in the intersection of students and bungee jumpers is running.
12.3. SUMMARY: THE FRAGMENT SO FAR

We thus leave this as an interesting and unsolved problem. (See also n. 2.)
Interestingly, or is quite well behaved in this regard. Note the following:
(7) a. Every professor or bungee jumper is invited to the meeting.
b. Every man or woman will be allowed to vote in the special election.
c. No student or bungee jumper came to the demonstration.

Certainly the primary reading of, for example, (7b) is the reading obtained by
having [[every]] take as argument the union of the men and the women sets.

12.4. i. Show the full derivation for this reading (showing only the derivation of
the GQ every man or woman) and show that this GQ is equivalent to
the GQ every man and every woman.
ii. Show the full derivation using the simplest types for no student or
bungee jumper.
Not surprisingly, (7b) also has a secondary reading which can be brought out with
the continuation “but I can’t remember which” (see Exercise 12.3 for the
derivation of this reading), but in this case this is definitely the less accessible
reading, as the lowest types hypothesis predicts.
Looking ahead, section 13.4.4.2 will entertain the possibility of adopting a slight
variant of the Partee and Rooth proposal, in which and and or are both listed in the
lexicon with their simplest meanings and categories (where they conjoin only
sentences) and where the extra meanings and categories are derived by a unary
rule.

12.3. Summary: The fragment so far

To the fragment in Chapter 11, add the following lexical items:

<[and], (X/X)/X, ºAx[ºBx[A B]]> (for x the semantic type of expressions of category X)
<[or], (X/X)/X, ºAx[ºBx[A C B]]>
PART III Relative Clauses, Scopes,
and Binding: Some Theoretical
Controversies
 13 Relative clauses: Sketching
two accounts
13.1. Preliminaries 207 13.4. A Direct Compositional
(and variable-free) analysis 221
13.2. What do relative clauses modify?
209 13.5. Appendix: Reassuring the
13.3. An account with movement and
Appalled Syntactician 240
traces: Assignment functions in
grammar 13.6. Summary: The fragment so
214 far 242
13.1. Preliminaries

This chapter and the following two approach three phenomena (relative clauses,
GQs in object position, and pronominal binding) from two different points of view.
For each case we present a relatively standard non-Direct Compositional account
of the domain, followed by a Direct Compositional account. There are several
reasons for developing two different accounts of each phenomenon—not the
least of which is to introduce some of the theoretical controversies in the
literature. Beyond that, seeing what the different points of view have in common
gives a deeper understanding of the domains and of the generalizations that hold
across theories. And of course an understanding of more than one account
enables one to read literature written from a variety of theoretical perspectives.
That said, a reader approaching all of this material for the first time might worry
about a possibly confusing overload. Such a reader might want to initially work
through the technical details of just one of the approaches across the three
chapters while skimming the other, and return to the details of the other account
only on a second pass.
The concern in this chapter is the syntactic and semantic composition of NPs
and GQs which contain relative clauses, such as those in (1):
(1) a. the house which Jack built
b. the house that Jack built
c. the house Jack built
d. every rat which the cat chased
e. no mouse who chased Calico
 13.2. WHAT DO RELATIVE CLAUSES MODIFY? 211
Some initial caveats are in order. Both theories developed here are intended as
sketches, for we ignore a number of interesting issues and details have been
amply documented in the syntactic literature. For example, no account will be
given of cases involving the phenomenon known as pied-piping. This term refers
to examples in which the front of the relative clause contains not just a single
relative marker, such as that or which, but a more complex phrase containing a
relative pronoun (which, who, etc.) within it (in all the examples below, the pied-
piped phrase is italicized); the last example is (slightly modified) from Ross (1967)
and shows that the pied-piped material can be quite complex:1

(2) a. the house into which Jack ran

b. the rat whose mother Calico chased
c. the king the cousin of whom the Duchess was engaged to
d. the reports the height of the lettering on the covers of which the government
proscribes

Moreover, for both theories developed below, the treatments of the actual relative
pronoun (which, who, whose, etc.) are almost certainly incorrect. In both cases,
the treatments are designed purely for expository convenience. In fact there is
little point in worrying too much about how to fold in the contribution of the relative
pronoun without also giving an analysis of the cases in (2) since an adequate
treatment must generalize to the pied-

1
The colorful terminology comes from Ross (1967). His observations were framed in a
theory in which the wh-word and the material surrounding it “moves” to the front, and the
metaphor involves the wh-word as the Pied Piper and the rest of the material that tags along
with it as either the rats or the unfortunate children of Hamlin.
piping cases.2 Relatedly, we will say little about relative clauses with that (as in
(1b)) and with no relativizer (as in (1c)). Finally, we ignore many of the syntactic
complexities that are well known in the literature under the rubric of island
constraints. The goal is only to develop two different approaches to the basic
semantic composition coupled with (in each case) a general sketch of how the
syntax works. Interestingly, the two approaches do converge on the meaning of
the relative clause itself, even though they get there by very different methods.
And those different methods themselves will figure in quite prominently in the
discussion in the next two chapters.
 212 13. RELATIVE CLAUSES: TWO ACCOUNTS

13.2. What do relative clauses modify?

13.2.1. The noun modification argument

But before considering the internal composition of relative clauses, we turn to a

question hinted at briefly in section 1.1: does a relative clause combine the noun
or with the entire NP or GQ? (This question is independent of the theoretical
debate elucidated later in this chapter.) In other words, is a phrase such as every
table which Martha refinished put together as illustrated in (3a) or as illustrated in
(3b)?
(3) a. every [table which Martha refinished]
b. [every table][which Martha refinished]

13.1. As shown above in (1), relative clauses occur both within expressions
where every is the determiner and within those where the is the determiner.
Thinking only about the syntax, does this fact bear on the choice of the two
possibilities illustrated in (3)? Can both theories handle this fact with no
additional complexities in terms of the syntactic combinatory rules (or, in CG
terms, the category of a relative clause)?

2
Accounts of this phenomenon compatible with the general framework assumed here
can be found in Sharvit (1998) and Jacobson (1998), both of which should be accessible
after Chapter 15. A somewhat different type of account goes under the rubric of
“reconstruction” (see Chomsky 1977) and a more recent version of this known as the “copy
theory of movement” (see Fox 2002, 2003).
The question was discussed in syntactic terms at least as early as Stockwell,
Schachter, and Partee (1972) (much earlier work simply assumed the structure in
(2b); see, e.g., Ross 1967).3 But in a seminal paper, Partee (1975a) pointed out that
the semantics can be used to help answer this question; the semantic
composition is much more straightforward under (3a) than under (3b). 4
To show this, we make the assumption that the relative clause which Martha
refinished denotes (relative, of course, to a world and a time) the characteristic
function of the set of things that Martha refinished. (Or perhaps it is a fancier
packaging of this, as in the treatment of prenominal adjectives in section 8.4.3.
But since that will not affect the discussion, we use the simpler type for
exposition.) The semantic composition of (3a) is thus straightforward: the relative-
 13.2. WHAT DO RELATIVE CLAUSES MODIFY? 213
clause meaning—a set of individuals— intersects with the noun meaning (another
set of individuals). The determiner then combines with a complex noun (which
denotes a set of individuals) in the perfectly routine fashion. So, whether the noun
happens to be complex (e.g., table which Martha refinished) or simple (e.g., table),
it denotes a set. Obviously fancier cases like every table which Martha refinished
which Jack sold are equally straightforward—the complex noun table which
Martha refinished combines with another relative clause to give another complex
noun and the semantics is again intersection.
Now consider how the semantic composition would have to work under (3b).
For the sake of discussion, we quite crucially ignore the possibility that [[every
table]] has a domain restriction slot (see section 10.6), and take [[every table]] to
be the set of sets of individuals that have the table set as a subset. 5

3
Prosodic evidence would also seem to favor the account in (3b) since there is
generally a prosodic break after table. However, as pointed out at least as early as Chomsky
(1965) the prosodic evidence can be quite misleading.
4
These remarks assume a Direct Compositional approach to at least this construction.
That is, the semantics gives evidence only under the assumption that (at least in this case)
the syntactic composition and the semantic composition are the same. In a theory in which
they are divorced and where a syntactic representation is mapped into a different LF, one
could break these apart, and one could assume that the syntax is as in (3b) and mapped
into an LF more like (3a). As far as this author is aware, this has not been explored in the
literature.
5
This is not a trivial simplication. As will be noted in 13.2.2, the argument elucidated in
13.2.1 actually needs to be rethought if the domain restriction “slot” (which we have
represented so far as a free variable) is part of the meaning of the GQ.
If the material is put together as in (3b), we need to define an operation that maps
the set of sets having [[table]] as subset to the set of sets having as a subset the
set of things that are both tables and are things that Martha refinished. In other
words, the model-theoretic object shown in (4a) (given here in function terms)
needs to be mapped to the object in (4b):

(4) a. ºP[8x[table’(x) à P(x)]]

b. ºP[8x[[table’(x) & refinished’(x)(martha)] à P(x)]]

The term model-theoretic object is stressed here to ward off the temptation to
think that the solution lies in mapping the formula in (4a) into the more complex
 214 13. RELATIVE CLAUSES: TWO ACCOUNTS

one in (4b). Obviously this can be done: simply define an operation that takes any
string “S1” on the left of the arrow and within the narrowest bracket [ and converts
that into the string “S1 & S2” where “S2” is the representation of the relative clause
in some symbolic representation such as the non-English language defined in
section 9.1. But this is not an operation on meanings but rather on syntactic
formulas, and thus not a possible way for the semantics to work. Since (4a) and
(4b) both name functions of type <<e,t>,t>, the task is to formulate the operation
that maps the (4a) function to the (4b) function. Or, returning to set terms, the
operation will—for any set A (the denotation of the relative clause)—map (5a) to
(5b):
(5) a. {B| table’S B}

b. {C| table’S \ A C}

There is such an operation but it is complex, and so is left as an advanced exercise

for the ambitious.

*13.2. Be ambitious. In other words, show the operation that will derive the set
of sets [[every table that Martha refinished]] from [[every table]] and more
generally will derive for any set S [[every-N S]] from [[every-N]]. A hint to get
started: think first about how—given the set of sets—one can recover the set
of tables from [[every table]], i.e., how one can map [[every table]] into [[table]].

Regardless of the complexity of the operation, the implausibility of this solution

becomes evident upon noticing that different quantificational determiners require
different operations. Thus things would have to either be very non-compositional
or the grammar would need to assign a different category to, e.g., some as it
assigns to every (and yet other categories to other determiners). To show this,
notice first that the operation for every (i.e., the operation that maps (5a) to (5b))
involves mapping some set P (a set of properties) into some superset of P. To
convince oneself of this, note that any property P in the set [[every-table]] (i.e.,
every property P that holds for every table) certainly holds for every table that
Martha refinished (since, after all, each such table is a table). But the reverse does
not hold. If every table that Martha refinished is made of wicker then [[is made of
wicker]] is in (5b) but not necessarily in (5a). Thus there are properties in the (5b)
set that are not in the (5a) set. (Readers who have already read Chapter 16 can
 13.2. WHAT DO RELATIVE CLAUSES MODIFY? 215
note that this is the same as the observation that [[every table that Martha
refinished]] is weaker than
[[every table]].)
It is now easy to see that the operation that derives the meaning in (5b) from
(5a) cannot possibly be the same operation that derives [[some table that Martha
refinished]] from [[some table]]. For clearly [[some table]] has more sets in it;
there might be a crooked table (hence [[is crooked]] is in [[some table]] but it is
not necessarily in the set [[some table that Martha refinished]] (she could be quite
careful to only refinish uncrooked tables). So here the operation that combines
the relative clause meaning with the GQ meaning would have to map a set into a
subset. (Again, for readers familiar with the notion of strength in Chapter 16, one
can note that here [[some table which Martha refinished]] is stronger than [[some
table]], reversing the pattern for every.)
Thus (given the assumptions made here about the meanings of the GQs and the
relative clauses), the only way to have two different operations at work without
abandoning compositionality would be to say that some table and every table are
actually of different syntactic categories. This would allow the grammar to use a
different operation when each of these combined with the relative clause. And yet
other operations would be needed for the case of other GQs like no table. At this
point it is quite clear that the compositional semantics is much simpler under the
analysis in (3a); the set denoted by the relative clause simply intersects with the
noun set, as in other cases of noun modification.
*13.3. State what the operation would need to be for combining [[some N]] with
a relative clause meaning and do the same for [[no N]]. Once again begin by
considering how one could map [[some N]] (for any N) into [[N]], and similarly
for the case with no as determiner.

*13.2.2. But—not quite so fast . . .

Like so many other things in semantics, the story does not end here. (Since the
further twist has no effect on the subsequent discussion in this book, this section
can safely be skipped.) Bach and Cooper (1978), drawing on evidence from
Hittite, argued that the grammar must be able to put things together in the way
shown in (3b) and that, moreover, there are ways that this can be done without
positing different categories for each GQ. 6 We will neither give their Hittite
 216 13. RELATIVE CLAUSES: TWO ACCOUNTS

evidence here nor spell out their proposal in detail, as it requires some complex
use of variables, but we will sketch the idea. Recall that the discussion directly
above crucially ignores the possibility that GQs such as every table come with a
domain restriction slot or “variable” as part of its meaning; see the discussion in
section 10.6. Bach and Cooper’s proposal makes use of the existence of this
piece of meaning: it combines the semantics of the relative clause with the
semantics of the GQ in such a way as to have the relative clause satisfy the
domain restriction. This is not enough, for there must be a new slot or variable for
a domain restriction even after the relative clause combines with the GQ. There
are two reasons for this. First, additional relative clauses are possible (so-called
stacked relatives) as in:
(6) a. every table which Martha refinished which was brought to the Antiques
Road Show
b. every table which Martha refinished which was brought to the Antiques Road Show
which was appraised for more than a thousand dollars etc.

6
But we really don’t need to consider only Hittite to suspect that language does allow
the semantics to put things together in the way shown in (3b). And this is because English
contains lexical GQs (everybody, nobody, someone, etc.) and these of course can combine
with relative clauses, as in nobody who voted for the proposal or everybody who loves the
snow, etc.
Each subsequent relative clause presumably would be filling the domain
restriction in the GQ with which it combines, and therefore a new domain
restriction piece must be added with each subsequent relative clause. Secondly,
even when a GQ contains a relative clause, it is still understood relative to some
contextually set domain. Thus if I tell you that every student who took formal
semantics got an A you once again will not interpret this as meaning every student
in the world who took formal semantics but more likely would restrict this to the
students in my formal semantics class this year (or some other contextually
relevant domain). Thus an analysis in which the relative clause fills the domain
restriction piece must be supplemented with something that introduces a new
slot or variable each time. See Bach and Cooper (1978) for one proposal and
Jacobson (1983) for an alternative cast without use of variables.
For the remainder of this chapter, however, we will adopt the analysis in (3a),
since supplying the details of an alternative analysis is beyond our scope. As
noted above, nothing in the subsequent discussion hinges on this, for the concern
 13.2. WHAT DO RELATIVE CLAUSES MODIFY? 217
for the rest of the chapter is with the internal semantic (and syntactic)
composition of the relative clause. And either way, the relative clause itself
denotes a set of individuals (or perhaps a fancier packaging of a set).

13.3. An account with movement and traces:

Assignment functions in grammar

We now return to the question of how the meaning of the relative clause is put
together. The first account to be considered here is relatively standard, and is
generally phrased in terms of a theory with a non-Direct Compositional
architecture where the syntax computes representations which are interpreted by
the semantics. For the case at hand, it suffices to view “surface structure” as what
is interpreted (but see section 14.2), and where that structure itself is derived by
a series of rules mapping one syntactic representation into another. In particular,
this theory assumes movement rules in the syntax. For a relative clause as in (7),
the relative pronoun which moves from its original position as object of built to the
front of the relative clause (as indicated by the lines). Thus the syntax computes
these representations in a series of steps and the representations serve as the
input to the semantics. This account is also usually coupled with the
 218 13. RELATIVE CLAUSES: TWO ACCOUNTS

(logically independent) assumption that each NP comes with an index (a positive

integer), and that when it moves it leaves a phonologically empty “trace” in the
original position, where that trace bears the index of the original item. We have
arbitrarily picked the index 8 here:
(7) every house [which8 [Jack built t8]]

The semantics thus interprets a structure roughly like (7). Since the goal in this text
is not to develop an explicit fragment within this framework, we will not actually
formulate the movement rule in any generality; a variety of versions can be found
in almost any introductory syntax text within the group of theories that go under
the rubric of Government and Binding, Principles and Parameters, or Minimalism
(as well as many other earlier theories maintaining transformations).
We will, however, elaborate the internal structure a bit for the purposes of
developing the semantic composition. Thus assume that the structure after
movement (and which inputs the semantics) is actually the more elaborated
version in (8). (Part of the motivation here for some of the particular details is to
be able to generalize this to another case discussed in Chapter 14.) Note that this
assumes that when which moves, its index is copied onto the node labeled here,
(8) also assumes that the index on which remains after movement, but since it
plays no role in the semantic computation it actually makes no difference whether
it remains or not. Very often we will omit the index on the moved wh-item. We use
label RC for the relative clause; the name is inconsequential:
(8) N

N RC

house which 8 Λ8

NP VP

Jack TV NP

built t8
There are, incidentally, a variety of slightly different notations that could have been
adopted. Following, for example, the treatment of a related case in Heim and
Kratzer (1998) one could treat “8” as a phonologically null element which is a
 13.3. AN ACCOUNT WITH MOVEMENT AND TRACES 219

sister to S rather than recording it as a feature on the node labeled . The choice is
purely a matter of notation.
Since we are not using the Categorial Grammar fragment here, we need to say
something about how structures of this sort in general are interpreted. Leaving
aside for the moment the contribution of the trace, and leaving aside the
interpretation of the node labeled , we can essentially take the fragment up
through Chapter 5 and recast the semantics as rules that work compositionally
(hence, bottom-up) to interpret local trees (as was done in Exercise 5.1). In other
words assume that there is a rule to the effect that an S with two daughters NP
and VP will have as its value (extensionalizing) [[VP]]([[NP]]). All of this can be
cast more naturally in a type-driven framework (see section 6.5): for any node A
with two daughters B and C, if [[B]] is a function in <c,a> and [[C]] is a member of
c, then the denotation of the mother (A) will be [[B]]([[C]]). Following Heim and
Kratzer (1998) assume that another type-driven principle is that for any two
sisters B and C which are such that they both have meanings of type <a,t> (for any
type a), the meaning of the mother is the characteristic function of the intersection
of the sets that each daughter characterizes. (See the relevant discussion
surrounding APs in section 8.4.1.)
But the full interpretation of (8) involves one major new piece of apparatus that
our treatment of English has so far not required: the use of assignment functions.
To add this to the semantic toolbox involves essentially the moves that were
introduced earlier in Chapter 9 for the two non-English languages. Thus assume
that in addition to worlds, times, individuals, and truth values we have a set of
assignment functions (or simply assignments) G. For now, take each assignment
g to have as its domain the set of positive integers, and as its co-domain the set of
individuals. Assume further that the interpretation of any actual linguistic
expression is relative to assignments. Put differently (and somewhat non-
standardly) the meaning of any expression is a function from the set of
assignments to something else. (For example, a sentence denotes a function from
G to functions of type <s,t>; a VP is a function of type <G,<s,<e,t>>> and so forth.)
As seen earlier in the artificial languages with variables, any expression that
contains nothing “unbound” within it is a constant function from G. Thus [[every
dog]] is a function G to a function of type <<e,t>,t> but it is a constant function
from G, and each g in G receives the by now familiar value.7
As is undoubtedly obvious, the role of the assignments is of course to give an
interpretation to the trace (and later to pronouns). Thus while “normal”
 220 13. RELATIVE CLAUSES: TWO ACCOUNTS

expressions like Jack, build, house, the, and so forth have the same value on all
assignments, traces do not. They function just like the variables introduced in
Chapter 9; their value varies with assignments, and depends on just what the
assignment assigns to their index. (Since all NPs come with an index, the reader
might wonder whether the same is true of ordinary NPs such as the house or, for
that matter, even Jack. For these too will have an index: and so one might wonder
whether [[the house]]8 is different from [[the house]]9? We will assume not and
will assume that the indices play a role only in the interpretation of traces and
pronouns. The function of the indices on the full NPs is relevant only when that NP
moves and leaves a trace.) Thus for any index n and any assignment g, [[tn]](g) is
g(n). We will use the more normal convention of putting the assignment as a
superscript: thus, for any g, [[tn]]g = g(n). This changes all the semantic rules, but in
a perfectly straightforward way. For example, the rule interpreting a local tree
whose mother is S and whose daughters are NP and VP now says that for all g,
[[S]]g = [[VP]]g([[NP]]g). The assignments are thus folded into the combinatorics
just like the world and time arguments are once the rules are fully intensionalized. 8
If everything is bound in some expression, its value is a constant function on the
set of assignments.
Consider, then, the tree in (8). The interpretation of the VP-node is a function

from assignments to <e,t> functions (as usual ignoring the worlds and time) such

that for all g, [[VP]]g = [[built]]g(g(8)). Note [[built]] is constant on all g, and that g(8)

is some individual; hence the VP is a function that maps any g to the set of

individuals who built whatever it is that g assigns to 8. Moving up the tree, for any

assignment g, the value of 7

Heim and Kratzer (1998) develop an interesting alternative where the assignments are
partial functions, and assign values only to those integers which correspond to indices on
“unbound” material within the relevant expression. This does have some advantages; see
Chapter 9, n. 2.
8
Hence the value of an expression is a function from worlds, times, and assignments
to something else. Incidentally, the order in which these different arguments are taken
makes no difference: just as the worlds and times are independent of each other so too are
the assignments independent of worlds and times.
 13.3. AN ACCOUNT WITH MOVEMENT AND TRACES 221

S is the proposition that Jack built whatever it is that g assigns to 8. This is a non-
constant function from assignments to propositions. The actual proposition that
it denotes depends on what assignment one uses.
The interesting action comes in the next step: the interpretation of the node
labeled 8. This is similar to the non-English rule NER-8 of Chapter 9, and is given
here in (9). Because this is not going to be a part of the ultimate Direct
Compositional fragment developed in the book, no rule number will be assigned
to it. Moreover, since this is embedded in a framework in which the input to the
semantics are trees, we will formulate the rule that way:
Λn (9)g

= λae [ [[S]]g: n→ a]
S

In prose, this says that for any assignment g and any local tree labeled n whose
daughter is the single node S, the value of n on g is the function from individuals to
propositions which maps any individual a to the value of the daughter S on the
assignment function just like g except that (if it differs at all) the integer n is
mapped to a.
To illustrate, suppose that Jack built the Vanderbilt mansion and the
White House, but did not build the Breakers. Suppose that g 1 assigns 8 to the
Vanderbilt mansion; g2 is exactly like gi except that it assigns 8 to the Breakers, and
g3 is exactly like the others except that here 8 is assigned to the White House. Then
the value of the S-node in 8 is 1 on g1, 0 on g2, and 1 on g3. Now we compute the
value of 8 in (8) on each assignment, beginning with g 1. This value is the function
that takes each of Vanderbilt, Breakers, and White House and assigns either 1 or
0. First we see that
Vanderbilt is mapped to 1 because [[S]] on g1 is 1. The reason that the value of S
on g1 is the crucial one to consider is that g 1 is the assignment just like g1 in all
respects except (if there is any difference) 8 is mapped to Vanderbilt. In this case,
it happens that that there is no difference; here the assignment that g 1 “looks to”
is itself. The Breakers is mapped to 0 on g1, because here g1 “looks to” g2 (which is
the assignment just like g1 except that in this case 8 is mapped to Breakers). The
value of the S on g2 is 0; for [[S]] on g2 is the proposition that Jack built g2(8), which
is the proposition that Jack built the Breakers (which is false). Similarly, White
House will be assigned 1. It is easy to see that g1, g2, and g3 are all assigned the
same value (a function of type <e,t>) at the level of the node labeled 8 since they
 222 13. RELATIVE CLAUSES: TWO ACCOUNTS

all “look to” each other. In this way, just like NER-6 in Chapter 9, what happens
here is that the value of the S-node is a (generally) non-constant function from
assignments to propositions, while the value of the -node is now a function from
assignments to <e,t> functions, and for all assignments that differ only on the
value assigned to the relevant integer, it is a constant function.
Moving up the tree, the next question concerns the contribution of which. We
have set things up here in such a way that which actually makes no contribution:
its ultimate contribution was just to mark where it had moved from in such a way
that that position (the object of built) is a trace. To use a technique which has
already been introduced, we will take the tack of having [[which 8]] (on any g) be
the identity function on functions of type <e,t>.15 (Note that there are actually an
infinite number of lexical items which since it has an index: even if the index is
removed after movement it has to be there before movement and so there must
be these different items. But they all have the same meaning as things are being
set up here.) Thus [[which8 Jack built t8]] is (on any g) the characteristic function of
the set of things that Jack built. Note too that so will be [[which7 Jack built t7]]; once
the trace is “bound” at the step its index does not matter.
Finally, the meaning of the expression house which Jack built t 8 is put together
by the intersection of [[house]]S and [[which8 Jack built t8]]S. Of course, this is
oversimplified in that neither of these two expressions denote just sets of
individuals, but rather have as their value functions from assignments to sets. In
both cases they have the same value on all assignments, so one can assume that
the semantics will specify that the value of the mother node N on an assignment
g is the value of [[house]]Sg intersected with [[which Jack built t8]]Sg. And of course
this need not be specified for this particular construction, but would follow from
a type-driven principle as discussed above.
Many of the details here could have been worked out slightly differently. But the
key point of this analysis is that it makes use of (a) movement of the relative
pronoun to leave an indexed trace; (b) the use of assignments in order to interpret
the indexed trace; (c) the rule that “binds” the trace—i.e., the rule which is the
interpretation of the node labeled , and (d) a view of grammar in which the input to
the semantics is representations (and in this case representations that are derived

15 An alternative is to simply have which have no meaning, and have a convention that
for any local tree with two sisters A and B where one is meaningless, the meaning of the
mother is the meaning of the meaningful daughter.
 13.3. AN ACCOUNT WITH MOVEMENT AND TRACES 223

in the syntax not just by something like phrase structure rules but also by
movement rules).

stacked relatives

N RC

N RC which 7 Λ7

house which 8 Λ8 S

S NP VP

NP VP Sarah sold t7

Jack built t8
 224 13. RELATIVE CLAUSES: TWO ACCOUNTS

13.4. A Direct Compositional (and variable-free) analysis

We now consider this domain in a Direct Compositional setting (and also a setting
without movement or indices). The basic idea for an account of constructions like
relative clauses that makes no use of movement goes back to at least Gazdar
(1982) and later work within Generalized Phrase Structure Grammar (see Gazdar,
Klein, Pullum, and Sag 1985). Several similar proposals were subsequently put
forth within the Categorial Grammar literature. Here we initially adopt one CG
version which is roughly that in Steedman (1987), although it will contain various
modifications.

13.4.1. Preliminaries: Function composition

First we introduce the operation known as function composition. Take a function f

from A to B and a function g from B to C (any two of those sets could happen to be
the same). Then the two functions can be composed into a new function notated
as gof which is a function from A to C, where for any x in a, g of(x) = g(f(x)). (Thus gof
= ºx[g(f(x))].) The basic intuition becomes clear with a single example. Let f and g
both be functions illustrated in (10), where the members of set A are in the left-
hand column, the members of B in the middle column, and the members of C in
the right column.

(10) f g
A B C

a1 b1 c1

a2 b2 c2

a3 b3 c3

a4 b4

b5

Then g of maps a1 to c2, a2 to c1, a3 to c3, and a4 to c1. Since f and g are functions
(i.e., each member of A is assigned a unique value in B by f, and each member of
B is assigned a unique value in C by g), it is easy to see that the g of is also destined
to be a function.
 13.4. A DIRECT COMPOSITIONAL ANALYSIS 225

We now (temporarily) add two rule schemata to the fragment. In section 13.4.4
it will be shown that each of these can be unpacked into a different rule followed
by the ordinary application schemas (R-1a and R-1b). But even after making this
move, the rules here are convenient to refer to as a shorthand for two rules, and
so we will call them CR-5a and CR-5b (Collapsed Rule) and will refer to them as
the function composition schemata (we give these just extensionally):
CR-5a. Let Æ be an expression of the form <[Æ], A/RB, [[Æ]]> and is an expression of the
form <[], B/RC, [[]]> then there is an expression ª of the form <[Æ-], `/RC,
[[Æ]]o[[]]>.
CR-5b. Let Æ be an expression of the form <[Æ], A/LB, [[Æ]]> and is an expression of the
form <[], B/LC, [[]]> then there is an expression ª of the form <[-Æ], `/LC,
[[Æ]]o[[]]>.

Suppose there are three expressions: a of category A/ RB, b of category B/RC, and
c of category C. Notice that the grammar now allows two different ways to
compose the string [a-b-c] and prove this a well-formed expression of category A.
One is the normal “application-only” derivation shown in (11a) and the second is
a derivation involving function composition followed by application as shown in
(11b) (we will represent these with trees, showing the rule used for the
composition of each of the larger expressions):

(11)
a. A (by R-1a) b. A (by R1a)

A/RB B (by R-1a) A/R C (by CR-5a) C

a B/R C R B/RC c
C A/ B
b c a b

Both modes of combining these expressions give rise to the same ultimate string
[a-b-c]. They also give the same meaning. Thus note that the first application-only
derivation puts the meanings together so that the ultimate string has the meaning
a’(b’(c’)). (The prime notation is used here rather than brackets for readability;
recall that these are interchangeable.) The derivation represented in (b) puts the
meanings together as follows: a’ob’(c’). But by definition of function composition,
for all x a’ob’ is the same as a’(b’(x)), and so the meaning of the final result in (11b)
is a’(b’(c’)). Thus the two different ways to put the string together converge on the
same word order and the same meaning.
 226 13. RELATIVE CLAUSES: TWO ACCOUNTS

One might notice that so far, it is just a stipulation that CR-5a and CR-5b happen
to preserve word order. That is, CR-5a for example stipulates that the expression
of category B/RC will go to the right of the expression of category A/ RB, but as
written here this is an accident. It would have been just as easy to write a rule
which put the B/RC expression to the left of the other. Similarly, it appears to be an
accident of the way we have written the rules that the output is A/ RC and not, say,
A/LC. Either possibility would have wreaked havoc with the fragment, and gives all
sorts of strange word orders that do not exist. But in fact, we need not think of this
word-orderpreserving property of the function composition schemas as an
embarrassing accident. (This is reminiscent of the remarks made earlier with
respect to type lifting.) Recall that for any expression x of category X, there is an
associated function FX such that FX([x]) is a function from strings to strings. Hence
the actual syntactic categories in CR-5a and CR-5b can and ultimately should be
restated in terms of actual function composition. After all, the very definition of
function composition ensures that the string put together by the combinatorics
aob(c) has the same word order as the “application-only” derivation a(b(c)). More
specifically, for any string x of category A/ RB, y of category B/RC, and z of category
C, FA/ B([x]) o
R

FB/ C([y])([z]) = FA/ B([x])(FB/ C([y])([z])) so it follows that the composite of the two
R R R

relevant functions is the function corresponding to the category A/ RC (similar

remarks, of course, hold for the left-slash cases). For this reason, this view also
ultimately allows CR-5a and CR-5b to be collapsed as one function composition
rule (just as R-1a and R-1b can be collapsed). However, for present purposes we
simply continue to list the two composition schemata separately (and
stipulatively), but number them in such a way as to indicate that they can be
collapsed.
Consider now some actual English examples such as those in (12):
(12) a. Every fox chased Br’er Rabbit.
b. Calista loves model-theoretic semantics.

Every fox in (12a) is a GQ; its full category is S/ R(S/LNP) (recall that we arrive at this
category by listing every in the lexicon as being of category L S(NP)/N; see the
discussion in section 11.5.) Thus (12a) can be derived not only in the usual
application-only way but also as shown in (13) (for simplicity, no domain
restriction piece is included as part of the meaning of the GQ):

(13) every fox; S/R(S/LNP); ºP[fox’S PS] chased’ (S/LNP)/RNP; chased’

every foxed chased; S/RNP; ºP[fox’S PS] o chased’ (by CR-5a) =

 13.4. A DIRECT COMPOSITIONAL ANALYSIS 227

ºx[ºP[fox’S PS](chased’(x))] = ºx[fox’S

chased’(x)S]

in prose: this is the set of individuals x such that the fox set is a
subset of the chasers of x
(hence: the set of individuals that every fox chased)
Br’er Rabbit; NP; b every fox chased Br’er Rabbit; S, ºx[fox’S
chased’(x)S](b) = fox’S chase’(b)S

Should the last line appear confusing, one can convert that into the following
equivalent but perhaps less confusing way to write it: 8x[fox’(x) à chase’(b)(x)].
Since an ordinary “NP” like Calista can lift to be a GQ, a parallel derivation exists
for (12b). We leave this as Exercise 13.5; note that [[Calista loves]] is the set of
individuals that Calista loves.

13.5. Show the full syntactic and semantic composition for (12b) under the
derivation where Calista first combines with loves. Treat modeltheoretic
semantics as an unanalyzed NP as if it were listed in the lexicon (i.e., you need
not give its internal composition).

13.4.2. Interlude: Coordination

Enter a thoroughly appalled syntactician: “But wait! How can one possibly be as
naive as to propose that sentences like (12a) or (12b) have a structure in which
the subject and the transitive verb form a constituent rather than the transitive
verb and object forming a VP constituent? Don’t we learn from our most
elementary syntax books that there is a VP constituent and that therefore the
transitive verb must always combine first with its object?” Actually, we
(presumably) do know that the verb chased can combine with its object first to
form a VP, but it does not follow from this that this is the only possible derivation
(i.e., “structure”) for these sentences.
Recall the discussion in section 5.1. What is the evidence for a VP constituent?
The most robust piece of evidence concerns coordination; we find conjoined
cases like (14):
(14) Calista loves model-theoretic semantics and hates syntax.
 228 13. RELATIVE CLAUSES: TWO ACCOUNTS

But this shows only that there is one derivation in which loves combines first with
model-theoretic semantics (to conjoin with hates syntax); it does not show that
there can’t also be a second derivation of (12b). (Readers familiar with other
standard arguments for constituent structure might also be worried about those;
we address some in the Appendix to this chapter.)
Conjunction cases like (14) show that the derivation for (12b) that fits the
compositional picture illustrated in (11a) does in fact exist, and this of course is
the evidence that there is a “VP” constituent in English. But is there any evidence
that the derivation illustrated in the abstract in (11b) also might exist? That is, is
there any evidence that Calista loves can also be a constituent in (12b)? Quite
strikingly, the answer is yes: we need only turn the conjunction argument around.
For there are also sentences like (15), which have traditionally gone under the
rubric of Right-Node Raising:
(15) Calista loves and Gollum hates model-theoretic semantics.

It is well worth pausing for a moment to consider the common wisdom on

sentences like (15). Many textbooks argue that only constituents may conjoin—
as has been assumed throughout this book—and that conjunction is therefore a
test of constituency. But under “standard” accounts there is no way for Calista
loves and Gollum hates to be treated as constituents. Faced with such apparent
counterexamples to the generalization that “only constituents conjoin,” various
additional mechanisms are proposed for cases like (15). For example, one
common view is that (15) underlyingly consists of two sentences (Calista loves
model-theoretic semantics and Gollum hates model-theoretic semantics), and
the two instances of the object NP modeltheoretic semantics move to the right
and merge into a single NP. And there are other proposals as well (see, e.g.,
Williams 1978; McCawley 1987), but all involve additional complexities in the
grammar. In fact, it is somewhat ironic that many introductory textbooks contain
the adage (accepted here) that conjunction is a test for constituency, and then go
on to call cases like (15) and other related cases instances of “non-constituent
conjunction.” Assuming that conjunction is indeed a test for constituency, we
must conclude that Calista loves and Gollum hates are both well-formed
constituents (i.e., well-formed expressions). This is part of the beauty of the CG
account of these: the addition of the function composition schema CR-5a is
sufficient to predict that Calista loves is an expression of category S/RNP as is
Gollum hates. That the two can conjoin is completely expected by the cross-
categorial treatment of and. The semantics works out exactly as it should; this is
left for the reader to verify in the next exercise.
 13.4. A DIRECT COMPOSITIONAL ANALYSIS 229

13.6. Give the syntactic and semantic composition of (15), adding nothing to
the fragment except for the function composition schemata in CR5a and CR-
5b.

Before returning to relative clauses, it is worth noting that there are many other
cases of so-called “non-constituent coordination” that are accounted for by the
addition of FC. Consider for example each of the cases in (16):
(16) a. Calista loves and believes that Gollum hates model-theoretic semantics.
b. Mitka dashed into and destroyed the chicken coop.

We have not yet folded prepositions like dash into our fragment, nor have we yet
incorporated embedded sentences (as in simple cases like Barbara believes that
Paul hates model-theoretic semantics). But an interested reader can easily figure
out the syntactic categories involved for all of the items above within a CG syntax,
and can show just how it is that the syntax will allow these. The semantics will
also be straightforward once we have the semantics for basic items like believe
(which involves incorporating intensionality into the fragment and hence is
postponed) or into (which involves a more thorough treatment of prepositions),
but again an interested reader can see how the semantics would work for these
additional cases of “non-constituent” coordination. In each case, the term “non-
constituent coordination” is incorrect; these simply involve the coordination of
constituents.

*13.7. An especially interesting case of “non-constituent coordination” which

falls out under the analysis here is (i), which is discussed in Dowty (1988):
i. Captain Jack served lobster yesterday and bananafish today.

Assume first that yesterday and today are VP modifiers (in CG terms, they are
of category VP/LVP, where the L feature is given by the word order rules in (3) in
Chapter 6. To give their meanings clearly requires an intensional semantics
since they each take time-dependent sets as their input and crucially return
new time-dependent sets. For the purposes of this question, then, we need to
oversimplify (having not yet fully folded in intensionality); simply assume
(counterfactually) that [[yesterday]] and [[today]] are each of type
<<e,t>,<e,t>> and don’t worry about the details of their meaning.
 230 13. RELATIVE CLAUSES: TWO ACCOUNTS

With that oversimplification, show that (i) can be derived from the fragment
developed here (with the addition of the function composition schemata), and
show that its meaning will be the same as the meaning of (ii):

ii. Captain Jack served lobster yesterday and Captain Jack served bananafish
today.

Hint: Feel free to help yourself to the use of the fully generalized typelifting rule
R-4 in Chapter 11, especially as concerns the object NP.

13.4.3. Returning to relative clauses

13.4.3.1. The basic account

Other than the coordination facts, is there any other evidence that expressions like
Barbara loves are well-formed expressions in the language? Do they exist in any
other environments? The answer again is yes: these are exactly the sort of
expressions that combine with relative pronouns in relative clauses, and with
questions words in wh-questions (as in, for example, embedded questions like I
wonder what Barbara likes). In short, these occur in traditional extraction or wh-
movement constructions. We can thus speculate that relative pronouns and
question words themselves are listed with some category like X/(S/ RNP), where X
here is a placeholder (it will have one value in the case of questions and another
in the case of relative clauses).
Before fleshing out the full details of the syntactic and semantic composition
for the relative-clause case, we should note that expressions of category S/ RNP
(such as Calista loves) do of course have a far more limited distribution than
ordinary “VPs”—i.e., expressions of category S/ LNP (as in loves model-theoretic
semantics). The latter (with varying morphologies) occur as complements of
auxiliary verbs like will, have, be; and as complements of verbs like like, etc.
(17) a. Calista will [love model-theoretic semantics]
b. Calista has [loved model-theoretic semantics] (for at least 30 years)
c. Calista is [loving model-theoretic semantics]
d. Calista likes [teaching semantics]

Auxiliaries thus presumably are of category (S/ LNP)/R(S/LNP) (in each case the
complement is specified for a certain morphological feature, and the reason that
auxiliaries and verbs like want are not treated by the grammar as modifiers is that
 13.4. A DIRECT COMPOSITIONAL ANALYSIS 231

the fine-grained category for which they select is not identical to the result
category). There are, however, no auxiliary-like items which ask for an S/ RNP
complement; we do not find hypothetical auxiliaries like fung that occur as
follows:
(18) *Sally fung [Calista love]

But this is not surprising. After all, note that the auxiliaries in the lexicon merely
say that they take complements of the form S/NP. The fact that these
complements are S/LNP and not S/RNP is a fact about the general word order rules
spelled out in (3) in Chapter 6. If the auxiliary has, for example, is listed simply
(S/NP)/(S[PERF]/NP) then the general word order rules will spell this out more
fully as (S/LNP)/R(S/LNP). Thus it is predicted that expressions like Calista love(s)
occur only with items that specifically override the default rules or in conjunction
which is entirely cross-categorial.
Presumably, then, relative pronouns and wh-question words have the special
status that they are listed in the lexicon as combining with S/ RNP. (They also
combine with expressions of category S/ LNP as in every mouse who/which
chased Calico. A more sophisticated treatment of these words would remedy this;
here we just list two separate syntactic categories for these in the fragment.) We
can now turn to the full composition of every table which Martha refinished. Recall
that the movement account developed in 13.3 posits the structure in (8), where
the trace has an index and where the S Martha refinished t8 has as its interpretation
a proposition relative to assignments. But it shifts its meaning (by the rule in (9)) to
denote on each assignment the set of things that Martha refinished. This is
constant on all assignments and so, for all intents and purposes, the expression
Martha refinished t8 (as a constituent of category 8) denotes a set. We formulated
things in such a way that [[which Martha refinished t8]] also is (on all assignments)
a set.
The Direct Compositional version sketched here—which makes no use of
traces and no use of indices or assignments—ends up with the same result but
gets there in a different way. Notice that function composing the lifted (GQ)
[[Martha]] with [[refinish]] directly results in an expression Martha refinished
which denotes the set of things that Martha refinished. The derivation of this is
exactly analogous to the derivation of Calista loves which was worked out in
Exercise 13.6. There is no need to assume that refinish combines in the syntax
with an actual NP (such as a trace) and no reason to assume that [[refinish]]—of
type <e,<e,t>>—needs to actually find an argument of type e to combine with.
Indeed it does not find such an argument, but is composed with the GQ subject.
 232 13. RELATIVE CLAUSES: TWO ACCOUNTS

As in the view sketched in 13.3, we need to say something about the syntactic
and semantic contribution of which. Keeping in mind that we are simplifying for
the sake of exposition, assume a new primitive syntactic category, which we will
label RC whose semantic type is <e,t>.10 The syntactic function of the relative
word who or which is to map an S/ RNP into an RC, so we give this the category in
the lexicon RC/(S/RNP). (The fact that which takes the S/ RNP to its right follows
from the general slash-directional rules in (3) in Chapter 6.) As done in 13.3,

10
In standard movement views this is not the usual category label for relative clauses—
they are usually considered to be some kind of “CP,” where CP is the label for embedded
clauses in general. But they do not have the same distribution as ordinary CPs such as the
complement of think in Chris thinks that the world is flat.
Nor do they have the same internal structure, and—as will be shown below—they do not
have the same kind of meaning either. There are various ways in the standard view to
account for those differences; one is to have a feature on the CP node which distinguishes
relative clauses from ordinary CPs. In the end, then, some difference must be found and
there seems to be no real reason to think of these as CPs. But the label is arbitrary, and
should one wish, one could take “RC” to be really something like CP[REL]. This is merely a
matter of nomenclature.
we take its meaning by the identity function over functions of type <e,t>.
Thus [[which]] is ºP[P].
The remaining step is to allow an expression of category RC to combine with the
noun to give a noun. As was the case with noun-modifying adjectives discussed in
Chapter 8, there are two possibilities. One is to simply add a new rule schema to
the effect that an RC can occur to the right of a noun, and the meanings of the two
combine by set intersection. This, incidentally, is essentially parallel to the tack
taken in “standard” approaches.11 The other is to posit a unary rule shifting an RC
to an N/N with the semantics exactly like that of the shift from AP to N/N. We will
choose the latter here for this can be achieved with very little that is new. Since
the output category is the same as that in R-2 and the semantics is also the same,
we need merely slightly amend R-2 (mod) as follows:
(R-2) Let Æ be an expression as follows: <[Æ], AP or RC, [[Æ]]>. Then there is an expression
: <[Æ], ˝/˝, ºP<e<e,t>>[ºx[P(x) & [[Æ]](x)]>.
 13.4. A DIRECT COMPOSITIONAL ANALYSIS 233

13.8. Show the full syntactic and semantic computation of (i) and (ii) under the
Direct Compositional approach:

(i) every table which Martha refinished

(ii) the woman who Calista saw the cousin of
(iii) every course which Calista loves and Gollum hates

11
Thus some rule or principle is needed to ensure that expressions of the form which
Mabel refinished etc. combine with nouns. It is often suggested in elementary discussions
of X-bar theory that the fact that certain kinds of “adjuncts” (or modifiers) occur with certain
items need not be stated on a case-by-case basis, but follows from a general X-bar theory
rule of the form X-Bar à X-Bar YP, where the relative-clause category simply instantiates YP
here. But of course this is not enough; the theory needs to be supplemented with
rules/principles specifying what instances of YP go with what categories instantiating X. In
other words, let CP [REL] be the category for relative clauses; it can instantiate YP when X
is instantiated by N but not if X is instantiated by, e.g., V. In the end, the X-bar rule schemata
are not schemata in the sense here; they do not fully specify exactly what particular rules
do and do not exist.
Finally, it is well known that relative clauses can be indefinitely complex in terms
of embedded sentences; we find, for instance, examples like (19):
(19) the planet which Chris believes that Sarah said that Calista saw

Consider first a simple case like (20):

(20) Chris believes that Calista saw Jupiter.

We have actually not introduced embedded sentences yet because (as already
noted) a full semantic treatment requires the full intensional apparatus, for the
semantic object which combines with [[believe]] is not a truth value but a
proposition. Hence [[believe]] is of type <<s,t>,<e,t>>. But we can simplify at this
point and take [[believe]] and [[say]] to be of type <t,<e,t>>. We also introduce a
new primitive syntactic category CP which is the category for that Calista saw
Jupiter and we take expressions of this category to have meanings of type t
(ultimately, of course, of type <e,t>). Let that be listed in the lexicon as <[that];
CP/S; ºpt[p]> and let believe be listed <[believe]/; (S/NP)/CP; [[believe]]>. (The
listing for say is analogous.) With this apparatus, (19) can be derived.

13.9. Show the syntactic and semantic computation of (19).

 234 13. RELATIVE CLAUSES: TWO ACCOUNTS

13.4.3.2. Syntactic loose ends

There are some awkward details in the account here. In particular, the category RC
is (borrowing terminology from Gazdar and Sag 1981) a phantom category. It is the
category for which Mabel refinished, but there is no item (at least not in the
fragment developed here) that asks for such a category as argument. In fact it
shifts to N/N and only then combines with the noun, and so it seems a bit silly to
posit this intermediate category. Why not simply have which and who take the
S/RNP and directly give the noun modifier? In other words, why not list which as in
(21) (and similarly for who):
(21) <[which], (N/N)/(S/NP), ºP<e,t>[ºQ[ºx[P(x) & Q(x)]]]>

In fact, this is the solution in Steedman (1996). We have resisted the temptation
to do this (at the expense of needing the strange category RC) for several reasons.
First, it does not capture the commonality between the semantics of modification
by APs and by RCs (here R-2 takes both AP and RC as input to yield N/N). Second,
in the long run, the category RC can be revised to be simply the label for S with a
certain feature on it, a feature that indicates that it has a pronoun within it, making
it a less odd category. In the variable-free treatment of pronouns that will be
discussed in Chapter 15 this is quite natural and the fact that the type of this
expression is of category <e,t> is also automatic. And third, under the variable-free
view the semantics of the relative word can be subsumed under the semantics of
pronouns in general, which ultimately allows a treatment of the piedpiping cases
(see Jacobson 1998). The category and meaning for which in (21) extends in no
obvious way to cover the pied-piping cases. This book, however, will not have
space to discuss pied-piping (nor the semantics of relative pronouns) any further,
and so here we will just live with the awkwardness of the category RC and the
(perhaps somewhat unintuitive) idea that the relative word has as its meaning
ºP<e,t>[P].
There are other loose ends. We have provided no account of relative clauses
where the “gap” is internal, such as those in (22) (where ___ indicates the gap site):
(22) a. the book which Bill believes [__ pleased George]
b. the story which George read __ to the children

Indeed the function composition schemata put things together in such a way that
the “missing argument” can only be on the right edge of an expression or on its left
edge. There are many actual and potential variants of the proposal discussed here
designed to remedy this. One would be to extend the whole system to incorporate
Wrap slashes. Another is to allow some additional schemata (see, for example,
 13.4. A DIRECT COMPOSITIONAL ANALYSIS 235

Steedman 1996). Still another strategy is proposed in Oehrle (1990) which

combines some of the insights of the Steedman (1987) proposal discussed here
with ideas that are much more similar to the (related) proposal in Gazdar, Klein,
Pullum, and Sag (1985). We leave it to the interested reader to pursue some of this
literature, which should be accessible at this point. Finally we have said nothing
about various ill-formed cases—in particular the cases known in the literature as
island constraint violations (Ross 1967). The above literature contains some
discussion of relevance, and there is a vast literature questioning whether such
violations really should be accounted for in the grammar or are instead derivable
from facts about information structure (see, e.g., Kuno 1975) or other processing
considerations (see, e.g., Kluender 1998).

*13.10. One of the constraints discussed in Ross (1967) is the Coordinate

Structure Constraint; in movement terms this was phrased as “no movement
out of a coordinate structure.” This was designed to predict the illformedness
of expressions like:
(i) *the table which Sally built chairs and Mabel refinished __

There has, however, been considerable discussion in the literature as to

whether this is really a syntactic problem or simply violates various constraints
on information structure. Advocates of the non-syntactic explanation point to
perfectly well-formed cases like those in (ii):
(ii) a. the beer which Sam went to the store and bought __
b. the piece of property which either Sally is going to win the lottery or she will
have to sell __

Does the account of coordination here predict the ill-formedness of (i) or not?
Are there assumptions about and that one could make that would change the
answer? Note: Your answer will depend on exactly what it means to say that
and is cross-categorial and is listed as (X/X)/X. For the purposes of this
question, take that statement to mean that X is an underspecified category in
the lexicon which is more fully specified before and combines with anything.
Thus, for example, and can receive the full specification (S/S)/S or (VP/VP)/VP
and these items then combine by the normal rules. You might want to revisit
this question after reading
13.4.4.2.
13.4.4. A refinement: The Geach rule

13.4.4.1. The rule and its application to relative clauses (and coordination)
 236 13. RELATIVE CLAUSES: TWO ACCOUNTS

The function composition analysis presented above can be modified slightly to

recast function composition as a unary rule; this has some advantages which will
be shown below. The revision involves what is known in the CG literature as the
Geach rule (see Geach 1970) or alternatively Division, and is a very simple
mapping from functions to functions. In particular take any function f of type <a,b>
and any set c. Then f can be mapped to a function which we will call g(f) of type
<<c,a>,<c,b>>, where g(f) = ºC<c,a>[ºxc[f(C(x))]]. In prose: the new function g(f )
takes as argument any function C from c to a, and then any member x in c, and
returns the value that f assigns to C(x). Of course, as with other operations that we
have discussed, g is really a family of operations, because for any function f of type
<a,b>, g is not uniquely defined; it can yield a function of type <<c,a>, <c,b>> or
<<d,a>,<d,b>>, etc. We can use g just to mean the set of all of these functions,
and when we need to distinguish one from the other we can use the notation, for
example, gc(f ). The interesting thing to notice is that this is nothing more than a
unary version of the function composition operator. (Recall that any binary
operator such as function composition which simultaneously takes two
arguments—in this case two functions— to return some value can be Curry’ed to
take its argument one at a time. Thus g is the Curry’ed version of the binary function
composition operation.)
One can easily get the intuition of g combined with an appreciation of its value
for the Direct Compositional program by an example. Ignoring worlds and times,
consider the relationship between the first-order propositional negation ~ and the
meaning of English didn’t which is VP negation.
The latter can be defined in terms of the former as follows:

(23) [[didn’t]] = ºP<e,t>[ºx[~P(x)]]

Notice that [[didn’t]] is g(~) (more technically, it is g e(~)); in this case g maps a
function of type <t,t> into one of type <<e,t>,<e,t>>. We are not suggesting that
didn’t is derived from ~ by a productive rule in the grammar here; rather, its lexical
meaning boils down to g(~). This example also illustrates the usefulness of this
notion for the Direct Compositional program. At first glance, the contribution of
the negation in Porky didn’t grunt would seem to correspond to ordinary first-order
logic ~ which is a propositional operator. In the syntax, on the other hand, didn’t
combines with a VP to give a VP. (Of course, as pointed out earlier, even this is an
oversimplification, since the real negation is n’t; we return to this point directly.)
One temptation, then, for reconciling this apparent mismatch between the syntax
and the semantics is to posit that the negation raises at LF so that the relevant
representation for this which inputs the semantics is [not [Porky grunted]]. But the
 13.4. A DIRECT COMPOSITIONAL ANALYSIS 237

Direct Compositional strategy is simply to posit that the obvious meaning is just
packaged in a slightly fancier way, and in this case taking [[didn’t]] to be g(~) does
the job. The moral can be applied to many other cases as well: whenever there is
something that at first glance seems to scope over sentences in the semantics
but takes smaller scope in the syntax, one can posit that it is actually the Geach’ed
version of the obvious meaning.
In fact, as noted repeatedly, positing a primitive lexical item didn’t is still not
quite true to English grammar; negation is expressed by the word not or the clitic
n’t which actually combines with auxiliaries. But one can simply apply another
round of the g operation to give the actual meaning for n’t. (For simplicity, we will
leave our fragment as is with didn’t as the negation, but the revision would be
simple and an interested reader can work this through; part of this is suggested in
the next
exercise.)

13.11. Ignoring the fact that each auxiliary selects for a VP with a certain kind
of morphology, auxiliaries are of the category VP/VP (i.e., (S/NP)/ (S/NP)) with
meaning of type <<e,t>,<e,t>>). The syntax of English must also be set up in
such a way that only the first auxiliary in a chain of these combines with n’t (an
interested reader might try to find ways to ensure this via a judicious use of
features and giving auxiliaries more finegrained categories in terms of the
featural composition of the VPs that they select).
The task here, though, is not to give a full account of the syntax, but to show
that the meaning of n’t can be given in terms of the meaning of ~ using two
applications of g. Show how this can be done. (Keep in mind that—ignoring
intensions—auxiliaries are of type <<e,t>,<e,t>>.) Note too that the syntactic
category of n’t is such that it takes an auxiliary to its left to produce an auxiliary.
Of course, more needs to be said to account for the fact that n’t is not an
independent word. Finally, the English auxiliaries crucially refer to the world
and time parameters of the VPs with which they combine; will for example
obviously really combines with expressions of type <i,<e,t>> since it says
something about times, and can, may, etc. crucially combine with expressions
of type <w,<e,t>> as statements containing these are statements about
possible worlds.
Hence the semantics you give above will be just a sketch without folding
 238 13. RELATIVE CLAUSES: TWO ACCOUNTS

(cont.)

in intensions, but the basic sketch should be clear using the simplified
extensional types. Simply take as your task to have the sentence in (ia), for
example, have the same truth conditions as the paraphrase in (ib), and similarly
for other auxiliaries:12

(i) a. The crocuses won’t bloom in February this year.

b. It is not the case that the crocuses will bloom in February this year.

Similar remarks hold for other items that seem to have semantic scope over
sentences but in the syntax have a much smaller scope such as past-tense
morphology.
Returning to the analysis of relative clauses, notice that since g is nothing more
than the unary (Curry’ed) version of the function composition operator, the two
rules in CR-5a and CR-5b can be recast instead as unary rules; we do so in R-5:
R-5a. Let Æ be an expression of the form <[Æ], A/RB, [[Æ]]>. Then there is an expression
of the form <[Æ], (A/RC)/R(B/RC), gc([[Æ]])> (for c the semantic type of expressions
of category C).
R-5b. Let Æ be an expression of the form <[Æ], A/LB, [[Æ]]>. Then there is an expression of
the form <[Æ], (A/LC)/L(B/LC), gc([[Æ]])> (for c the semantic type of expressions of
category C).

12 There is actually an interesting complication that we ignore here. In general, an auxiliary

with n’t attached at the end gives rise to a meaning where the negation scopes over the
auxiliary, as is shown in the paraphrases in (i) and (ii). Hence can’t, for example, means
(roughly) “not can”; You can’t go home that way can be paraphrased as It is not the case
that you can go home that way. But must has the opposite behavior: You mustn’t drive
without a license doesn’t mean “It is not the case that you must drive without a license” but
rather “You are obligated to not drive without a license.” This has the odd result that can’t
and mustn’t mean more or less the same thing. To make matters even stranger, this fact is
not general: the auxiliary-like item need—which on the face of it has a meaning similar to
must—is well behaved. This can be observed by thinking about the meaning of “You needn’t
drive without a license.” The observation about must and can is due to Horn (1972).
We will refer to these rules as the g-sl schemata (“sl” here is for “slash” and the
reason for this name will be clarified in Chapter 15 when a second g rule is
 13.4. A DIRECT COMPOSITIONAL ANALYSIS 239

introduced). As with the case of CR-5a and CR-5b and the earlier case of type
lifting, the actual directional features on the slashes in the syntax need not be
stipulated under the view that the syntactic category of a string x corresponds to
a function F, such that F([x]) is a function from strings to strings. We will not work
through the details here, but see the discussion above with respect to CR-5. Since
the g operation is just a Curry’ed version of the function composition operator,
those remarks extend straightforwardly to this case. Should there be slash
features using wrap, this too can be incorporated. Moreover, these two rules will
ultimately be collapsed into one under that view; again we list them separately
here for expository convenience.

13.12. Redo the derivations for (15) in the text and for (ii) from Exercise 13.8
using R-5a instead of CR-5a. If ambitious, do the same for (i) in Exercise 13.7.

Is there any particular reason to prefer the two-step version (g-sl plus app) over
direct function composition? Indeed there is, and this is reminiscent of other
cases where it turns out that a unary rule rather than a new binary combinatory
rule was useful. That is, the two-step solution gives the intermediate step (i.e., a
step in which the relevant expression has just undergone g-sl), which might be the
input to other rules. And one of those rules that could apply to the intermediate
step is g-sl itself; in other words with the intermediate step the rule can apply
recursively. The advantage of this is the subject of Exercise 13.13. It is also
interesting to note that the revised solution here is actually somewhat closer to
the spirit of the similar analyses in the Generalized Phrase Structure Grammar
literature and the HeadDriven Phrase Structure Grammar literature (see the slash-
passing conventions in these theories).

13.13. Show that recursive application of this rule predicts the possibility of
coordination in cases like (i) (you can give just the syntax; you need not give the
semantics, and refer to the category CP and the category of

(cont.)

believe given above in the discussion surrounding (20). (Notice that this
sentence is more natural with but instead of and, but for simplicity we give the
 240 13. RELATIVE CLAUSES: TWO ACCOUNTS

version with and.) You don’t need to worry about the internal structure of was
corrupt.

(i) Raymond admitted that some and denied that every candidate was corrupt.

*13.4.4.2. Generalized and as generalized g-sl?

There is an obvious way to generalize the g operation such that it applies to n-place
functions, and “opens up” each of the argument slots to themselves be functions.
Without giving this in fullest generality, we illustrate with the case of the
generalization to a function f of type <a,<b,c>>. This can be mapped to something
that we will call gen-gd(f ) which is a function of type <<d,a>,<<d,b>,<d,c>>> such
that:

(24)
gen-gd(f) = ºX<d,a>[ºY<d,b>[ºDd[f(X(D))(Y(D))]]]
It should be obvious that this can be generalized for functions of any number of
places.
What is especially interesting here is that the recursive definition of — and q
given in (2) and (3) in Chapter 12 involves precisely successive applications of
gen-g to the Curry’ed first-order & and v (i.e., the operators of type <t,<t,t>>. To
show this, we begin with the base step of Curry’ed & which we will call AND, which
is ºp[ºq[q & p]]. Then the case of conjoining two expressions of type <e,t> to give
a new one of type <e,t> involves the generalized operator shown in (25a), which is
exactly (25b):

(25)
a. ºP<e,t>[ºQ<e,t>[ºx[AND(Q(x))(P(x))]]
b. gen-ge(AND)

The meaning of and used to conjoin two generalized quantifiers is gen-

g<<e,t>,t>(AND). This item takes AND of type <t,<t,t>> and maps it to a new meaning
of type <<GQ,t>,<<GQ,t>,<GQ,t>>> (where GQ is shorthand here for <<e,t>,t>).
Since conjunction always works on expressions whose final result is a truth value
(notice that the generalized definition of — ensures this), every instance of [[and]]
can be derived from AND by a succession of application of gen-g.
Does this have any interest beyond the sheer fun of noticing the relationship
between the various tools discussed so far? Actually, it might (the remarks here
are speculation and are designed to invite the reader to think more about these
issues). Thus, one might entertain the possibility that and and or are not listed
 13.4. A DIRECT COMPOSITIONAL ANALYSIS 241

cross-categorically in the lexicon. Rather, suppose there is a single lexical item

and as follows: <[and], (S/S)/S, AND>. (This of course cannot be exactly right since
the lowest type for and is actually of type <<s,t>,<<s,t>,<s,t>>> while AND is of
type <<t,t>,t>, but—as usual—the appropriate refinement is straightforward.)
Similar remarks hold for or. Then we can derive other versions of and and or by
application of gen-g to these items. Notice that we have left off the slash
directionality in the above; if this is the right way to think about the cross-categorial
status of and and or the rules need to be stated more fully to give the appropriate
directional features on the slashes. We leave further discussion of this for the
intentionally quite open-ended exercises below.

*13.14. Assume that the above suggestion is correct as an account of the

polymorphicity of and and or (i.e., there is actually just one lexical entry and all
other uses are derived by a generalization of the g-sl rule). Then first, state the
appropriate rule or rules, and state what needs to be assumed about the slash-
directional features. Second, notice that it was pointed out that the simple g-sl
rule (like function composition) need not stipulate the slash-directional
features under a view of grammar where the syntactic categories themselves
are functions. The definition of g-sl to such a function will predict the output.
The generalized g-sl rule, however, does not have quite the same property and
this might well militate against thinking of this as a real generalization of g-sl.
Discuss why this is so. (Note: this is truly an advanced exercise.)
*13.15. One potential problem with the suggestion above concerns cases
involving noun conjunction. In other words, in addition to and and or of category
(S/S)/S we might need additional basic lexical items of category

(cont.)

(N/N)/N. Are these additional items necessary? Make sure to revisit the
discussion of noun conjunction and the answers to Exercises 12.3 and 12.4.
The answer here is actually rather complex, so don’t expect a simple “yes” or
“no” answer.

13.5. Appendix: Reassuring the Appalled Syntactician

Many people familiar with the standard arguments in syntax regarding constituent
structure are initially shocked that anyone would suggest that Calista loves can be
 242 13. RELATIVE CLAUSES: TWO ACCOUNTS

a constituent in Calista loves model-theoretic semantics. In fact, the addition of

g-sl (or the function composition schemata) allow for all sorts of substrings of
sentences to be analyzed as constituents (i.e., as well-formed expressions with a
category and a meaning) which are quite standardly thought to resist all the
standard tests for constituency. We have already addressed some of the
objections that one might raise against this non-standard notion of constituents,
but since elementary linguistics and syntax textbooks contain a battery of other
arguments about constituent structure, it is worth a brief discussion of this here.
Obviously, space precludes going through each of the textbook arguments
against the sort of odd constituents allowed here, but we briefly comment on the
logic of these arguments to reassure the appalled syntactician that the fragment
developed here is not insane. Consider one classic argument concerning
“movement” (parallel arguments concerning “deletion” and “substitution by
anaphoric elements” can be answered in similar ways). Here, then, is an argument
found in literally almost any introductory linguistics or syntax textbook.
The textbook: Certain expressions may move to the front of a sentence, as
shown in (27), where (27a) is the version of the sentence with no movement. (The
construction here is known as the Topicalization construction and needs special
discourse conditions and/or special prosody to make it natural sounding; we
surround the relevant sentences with extra material to make them more natural).
The moved material is put in italics and the sentences that we are primarily
interested in are underlined:
 243
13.5. APPENDIX: REASSURING THE SYNTACTICIAN

(26) a. Kelly predicted that Chris would adopt the Siberian husky puppy.
b. The Siberian husky, Kelly predicted that Chris would adopt;

the malamute, she predicted that he would leave.

c. Kelly thought that Chris would adopt the Siberian husky, and indeed,
adopt the Siberian husky he will.

In (26b) an NP is found at the front of the clause rather than in normal object position;
in (26c) an entire VP (which occurs with an auxiliary verb) is found at the front. The
logic then goes on to assume that the (b) and (c) sentences are derived by a
transformation moving material to the front of a sentence, and that only constituents
can move in this way. Thus the ungrammaticality of things like (27b) is used to argue
that the material at the front cannot ever be analyzed as a constituent. We will not go
to the trouble to try to improve this sentence by embedding it in relevant discourse
contexts, since it is fairly clear from the outset that it is hopeless:
(27) a. Kelly believed that every dog howled.
b. *Dog howled, Kelly believed that every.

Of course the argument itself is phrased in terms of a theory with movement rules,
which are not being adopted here. But that fact does not undermine the potential
force of the argument since any theory does need to account for the existence of
cases like (26b) and (26c) while ruling out (27b). The cases in (26) are not difficult to
accommodate in the framework here: simply assume (for example), that there is a
unary rule which takes any expression of category S/ RX and shifts it to one of category
S/LX (for X a variable here over categories). Then even though the textbook argument
for the non-constituency of dog howled would need to be rephrased, the argument
would still seem to survive. One might assume that there must be no category X such
that dog howled is of that category, and/or no way to analyze Kelly believed that every
to be an expression of category S/ RX (for some X).
So the very generous notion of constituent structure assumed here appears to be
in trouble: for it predicts that dog howled can indeed be analyzed as a single
constituent, as can Kelly believed that every. The (ambitious) reader should be able
to show that dog howled can be analyzed as an expression of category S/ LDet, where
Det is an abbreviation for the category of quantifiers, which is (S/ RVP)/RN. (To show
this, you will need to lift howl to take GQs as argument.) Moreover, Kelly believes that
 244 13. RELATIVE CLAUSES: TWO ACCOUNTS

every can be analyzed as a well-formed expression of category S/ R(S/LDet) (with the

help of some fancy footwork involving g-sl and l). So why can’t the category S/ LDet
instantiate X in our Topicalization schema? Or, to use the more traditional movement
terms, why can’t this category “move”?
The answer is quite simple. The way the textbook argument is phrased assumes
that constituency is both a necessary and sufficient criterion for “movement” (or for
an ability to appear in this fronted position). But the latter assumption is incorrect:
under no theory is it the case that constituents of any category can occur at the front
of the Topicalized construction. In movement terms, this means that constituenthood
is a necessary condition for movement; it is not a sufficient one. In the terms of the
non-movement Direct Compositional theory developed here, the Topicalization unary
rule that shifts S/RX to S/LX is stated too generally: there have to be restrictions on
what can instantiate X. After all, there are many cases of perfectly “standard”
constituents which cannot occur at the front in a topicalization construction, as in:
(28) a. *Man who built The Breakers, Mary hired the.
b. *Who built the Breakers, Mary hired the man.
c. *Is flat, Mary thought that Chris believed that the world.

Tensed VPs, complex nouns (man who built the Breakers), relative clauses, and many
other expressions are disallowed in Topicalized position. To be sure, a fully adequate
theory would have something more to say about just what categories can and can’t
“Topicalize” (in the terms here, what categories can and can’t instantiate X in the
topicalization unary rule), but since it can be shown independently that there are
many such restrictions, there is no real reason to conclude that the fact that many of
these “odd constituents” do not occur in this construction threatens the conclusion
that they are indeed constituents. Similar logic holds for the traditional arguments
which are usually framed in terms of “substitution by an anaphoric item” and
“deletion.”

13.6. Summary: The fragment so far

To the fragment in Chapter 12 add the following:

Lexicon:
<[which], RC/(S/RNP) and RC/(S/LNP), ºP[P]>
<[who], RC/(S/RNP) and RC/(S/LNP), ºP[P]>
 245
13.6. SUMMARY: THE FRAGMENT SO FAR

Rules:
R-2 (mod) (revised). Let Æ be an expression as follows: <[Æ], RC, [[Æ]]>. Then there is an
expression : <[Æ],˝/˝, ºP<e,t>[ºxe[P(x) & [[Æ]](x)]>; if [Æ] consists of a single word, the
result is N/RN.
R-5a. Let Æ be an expression of the form <[Æ], A/RB, [[Æ]]>. Then there is an expression of the
form <[Æ], (A/RC)/R(B/RC)>, gc([[Æ]]) (for c the semantic type of expressions of category
C).
R-5b. Let Æ be an expression of the form <[Æ], A/LB, [[Æ]]>. Then there is an expression of the
form <[Æ], (A/LC)/L(B/LC)>, gc([[Æ]]) (for c the semantic type of expressions of category
C).

Comments:
• R-5a and R-5b should ultimately be collapsed into one, and the directional features on the
slashes would not be stipulated. That rule might ultimately be generalized further to
Generalized g.
• The treatment of which and who should ultimately be revised, and revised in such a way as
to account for the pied-piping cases. See Jacobson (1998) for some discussion.
14 Generalized quantifiers in object
position: Two approaches
14.1. A tempting but inadequate 14.4. Some comparisons 263
solution 245
14.5. Generalized quantifiers in
14.2. Quantifier (Lowering)/Raising 248 embedded sentences 272

14.3. A Direct Compositional 14.6. Summary: The fragment so

far 273
solution 255
Chapter 10 focused on the syntax and semantics of GQs like every wolf, no goose,
and some professor. In particular, the hypothesis that these have meanings of type
<<e,t>,t> and are of syntactic category S/(S/NP) (or simply L S(NP)) provides an
elegant account of how these combine in the syntax and semantics with VPs.
Moreover, it was noted that GQs have the same distribution as ordinary NPs (such as
the big bad wolf and Porky). This is partially accounted for by the assumption that
ordinary NPs can input the lift rule R-4, which means that, for example, an expression
with the phonology [porky] can also be a GQ. And so, it follows that wherever a GQ is
allowed, so is an ordinary NP. But this gives only half the story: nothing so far
precludes the possibility of an environment which allows ordinary NPs but not GQs.
Yet no such environment exists.
In more concrete terms, the fragment so far has no account of GQs in, for example,
object position, as in (1):
14.1. A TEMPTING BUT INADEQUATE SOLUTION

(1) a. Romulus loves every wolf.

b. Carolyn saw no stars.
c. Wolfie chased every pig.

Assuming that [[love]], [[saw]], [[chase]] etc. are of type <e,<e,t>>, some extra device
is needed to allow for these GQs in object position. To this end, this chapter will
explore two different types of solutions which have been proposed. The first (in
section 14.2) is not Direct Compositional and makes use of a level of Logical Form as
the input to the semantic interpretation, and where LF is distinct from the pronounced
sentence. Hence, this approach posits a rule which moves the object GQ to a different
position at LF. In other words, as far as the semantics is concerned, GQs do not in fact
occur in object position. The second solution (in section 14.3) is Direct Compositional
 247
and requires no revisions to the architecture assumed here. Rather than adding a rule
mapping one representation into another, it adds instead a unary rule, mapping a
function that has an e-type argument position to one asking for an <<e,t>,t> argument.
After developing the two views, section 14.4 provides some comparison.

14.1. A tempting but inadequate solution

But first we should dispense with a simple hypothesis which—while elegant in that it
requires adding nothing more to the fragment developed so far— does not give the full
range of facts. This rests on the observation that one possible derivation for (1a), for
example, involves combining (lifted) Romulus with loves in the way discussed in
Chapter 13, to give the expression Romulus loves of type <e,t> and of category S/ RNP.
Moreover, recall the speculation that every is listed in the lexicon as having category
LS(NP)/N. This means that this category can be more fully fleshed out as either
(S/R(S/LNP))/RN or (S/L(S/RNP))/RN. And that in turn means that every wolf—in
addition to the category S/R(S/LNP) that it has when it occurs in subject position—also
has the category S/L(S/RNP) and can take the S/RNP Romulus loves to its left. The
reader can verify that the derivation sketched here not only gives the result that (1a)
is well-formed, but also gives (1a) the right meaning as well. Indeed, this derivation
does exist according to the fragment here.
14.1. Show the full syntactic and semantic details of the derivation sketched
above.

But this cannot be the whole story about quantified NPs in object position, First,
the grammar still needs something more to account for cases of coordinated VPs with
GQs in object position (in one of the conjuncts or in both) as in (2):
(2) a. Carolyn extinguished every lamp and (still) saw no stars.
b. MacDonald milked every cow and fed every pig.
c. MacDonald milked Bessie and fed every pig.

A derivation analogous to the one noted above allows Carolyn extinguished to be an

expression that can combine with a GQ to its right, but that provides no way to
coordinate extinguish every lamp and saw no stars, which is needed for (2).
Second, consider the following sentences:
 248 14. GQs IN OBJECT POSITION: TWO APPROACHES
(3) a. A doctor has examined every patient.
b. Some wolf chased every pig.

These exhibit what is known as a scope ambiguity. Take (3a). On one reading, a doctor
has wide scope over every patient; in this reading there is a single doctor who did all
of the examining. The second reading is weaker; it requires only that every patient has
been examined by some doctor, but not necessarily the same doctor for each one.
Thus here every patient has wide scope over a doctor. 16 The second reading is
definitely harder to get but is brought out with the proper context. Imagine a medical
clinic wishing to
14.1. A TEMPTING BUT INADEQUATE SOLUTION

shut down for the night, but whose rules do not permit closing until every single
patient has been seen by at least one doctor. The receptionist announces that at last
we can close, for indeed “A doctor has (by now) examined every patient.” (We are
using a doctor here interchangeably with some doctor.) (3b) is similar; some find it
difficult to get the wide-scope reading on every pig, but it emerges with appropriate
context (and with the help of the right prosody). But consider what happens according
to the account under consideration here. While (3a) can be derived, it has only one
meaning: the meaning in which a doctor has narrow scope.

14.2. Show the full syntactic and semantic details of the derivation of (3a) in which
a doctor combines first with examine (ignore the auxiliary here) by function
composition (or, by g-sl plus application). Show that it results in the meaning where
every patient has wide scope.

16 In the case of a sentence containing two GQs one of whose determiner is some and the
other every, it is difficult to definitively show that there is a genuine ambiguity rather than a
vagueness. This is because one of the possible readings happens to entail the other. Thus in
(3b), if the wide-scope subject reading is true, it is also the case that every wolf is chased by
some pig—it just happens to be the same pig. However, any solution based on giving a
semantics for this in which the sentence has only one reading (the weaker one, in which every
wolf was chased by some pig) will then give the wrong semantics for the corresponding
sentence in which the quantifiers are reversed, as in:
(i) Every pig chased some wolf.
 249
14.3. Construct a context to bring out the wide-scope reading of every pig in (3b).

Thus not only does this solution give only one reading for these scopally ambiguous
sentences, but it also gives the least preferred reading, not a happy consequence. So
something more is needed.2 Thus we now turn to

In other words if one were to set up the semantics for (3b) such that the only reading is wide
scope on the object, then (i) should also have only the wide-scope object reading. But that
makes the incorrect prediction that (i) is false in a world in which each pig did some wolf-
chasing though the pigs did not chase the same wolf.
2
It is also not clear that the above type of solution can fully generalize to GQs in all positions,
although this depends in part on the status of Wrap operations. The data at issue concerns the
fact that GQs combine with three-place verbs as in (i) (and we get both scope possibilities):
(i) George read every child a story about goats.
This type of observation was made originally in Hendriks (1993) with respect to Dutch and
German where, in embedded sentences, both the subject and the object precede the verb (and
where either can be GQs, with the same scope ambiguities as we find in English). But it should
be noted that cases like (i) (and the Dutch/German cases involving subjects and objects) are
not necessarily fatal for the type of solution under consideration here. The ability of that solution
to extend to these cases depends on exactly how and if Wrap operations are folded into an
expanded
two different solutions. Like the two accounts of relative clauses in Chapter 13, one
(in section 14.2) does not maintain Direct Compositionality; the second (in 14.3)
does.

14.2. Quantifier (Lowering)/Raising

The first solution considered here is embedded in a theory with a level of

representation (LF) which is input to the compositional semantics, and at which only
a trace or pronoun (of type e) is actually in object position at that level. This view also
crucially makes use of variables and assignments as part of the semantic machinery
for natural language; traces (and pronouns) have indices in the syntax, and thus they
are analogous to the variables in the artificial languages developed in Chapter 9. 3 (In
fact, this view, as usually implemented, claims that all NPs have indices in the syntax;
as will be shown below, the index on full NPs plays no role in their interpretation.)
Thus to say that the trace/pronoun “is of type e” is to say that its semantic value is an
 250 14. GQs IN OBJECT POSITION: TWO APPROACHES
assignment-dependent individual. Crucially, then, as far as the semantics is
concerned, GQs actually do not exist in direct-object position—all direct objects at
the level of semantic interpretation are expressions with meanings of type e.
One of the earliest versions of this view was put forward (without full semantic
detail) within the framework of Generative Semantics (Bach 1968; McCawley 1971).
The fundamental premise of Generative Semantics was that the phrase structure
rules define a level of LF which is the level relevant for semantic interpretation, and
subsequent transformations map that into a different, surface level. Thus all
quantified NPs (i.e., all GQs) were actually generated by the phrase structure rules not
in subject, object, etc. position, but rather were higher in the tree. The normal NP
positions at LF are filled with silent “variables” (or, if you will, silent pronouns, or what
is now known as a

fragment. But in any case, we have seen other reasons that additional apparatus is needed to
fully account for the existence of GQs in object position. 3
Keep in mind that the Direct Compositional fragment developed so far uses variables and
assignments merely for the metalanguage being used as a convenient way to name model-
theoretic objects (as discussed in Chapter 9). It does not use them as an actual part of the
semantic machinery for natural language, and they could be dispensed with entirely with any
other system for naming these objects (such as the use of prose descriptions, done here
whenever possible).
 14.2. QUANTIFIER (LOWERING)/RAISING 251

trace), and the full material such as every pig, no wolf, etc. is lowered into those
positions by a transformation known as Quantifier Lowering.
A variant of this was subsequently taken up in Chomsky (1976) and May (1977).
Because this variant is much more common in recent literature it is the one to be
considered here, and—at the risk of overlooking various revisions which have been
explored since these original proposals—will be referred to as the modern LF
view.4 This variant shares with Generative Semantics the assumption that the level
relevant for interpretation is LF, that quantified NPs are not in object position at
that level, and that the surface position and the LF position of GQs are related by
a movement rule. The difference lies in the general architecture of the grammar. In
the modern LF view, the syntax proves well-formed surface representations by a
series of rules (including perhaps movement rules of the sort moving relative
pronouns that were discussed in Chapter 13). But in addition to the movement
rules that happen “in the overt syntax,” there are movement rules that map a
surface representation to an LF, and this (as in Generative Semantics) is the input
to the semantics. (By a rule applying “in the overt syntax” we mean one that
applies in such a way as to have an effect on the actual surface, pronounced
sentence.) In other words, the main difference between this view and the
Generative Semantics view of relevance to GQs is in the direction of the rules. In
Generative Semantics there is a rule of Quantifier Lowering (since the rules map
LFs to surface representations); in the modern view the rule is known as
Quantifier Raising (since surface representations are mapped to LFs).17

4
Revisions of the general picture of the organization of the grammar have been explored
in a number of works that have gone under the rubric of “Minimalism” (see, e.g., Fox and
Nissenbaum 1999). But the differences primarily concern the question of just how to state
the rules that give LFs as output; the LFs in the end and the interpretation of these are
generally assumed to be something like the view that will be exposited here. And different
researchers also advocate slight differences in the actual form of the LFs from that are

17 There is actually an earlier Direct Compositional variant of the Quantifier

Lowering/Quantifier Raising account due to Montague (1973) and known as the
Quantifying-In solution. Montague’s particular proposal contained additional assumptions
to be discussed below. But leaving those aside, a version of his proposal would be much
like the Generative Semantics view except it makes no use of an actual level of Logical
Form. The syntax and semantics work in tandem; surface strings are directly built, and
each expression is interpreted as it is “built” in the
 252 14. GQs IN OBJECT POSITION: TWO APPROACHES

looked at here, but these differences have no consequence for the discussion in this
chapter.
Incidentally, the LF view is perfectly compatible with the movement theory of
relative-clause interpretation developed in section 13.3 although at first glance it
might appear that it is not. For in section 13.3 it was assumed the input to the
compositional semantics was the surface representation, derived through a series
of rules applying in the syntax (including one which moved the relative pronoun to
the front and left an index trace). In reality, though, the full model in which that view
is embedded takes LF (not the surface representation) to be the input to the
compositional semantics. But there is no need to revise the presentation in
Chapter 13: simply assume that in the cases discussed there nothing of interest
happens to map the surface representation to the LF. Thus, what was taken there
to be the surface representations are identical to the LFs and so again it is an LF
which is interpreted.
Returning to the case at hand, the idea is to posit a rule (most easily seen as an
optional rule) which allows both quantified and simple NPs to move from their
surface positions to higher positions. This is much like the rule moving a relative
pronoun in the “overt” syntax. (Since this theory is generally not coupled with a
Categorial Grammar syntax, it posits a single category both for ordinary NPs like
Porky and for generalized quantifiers like every pig. We will continue to call this NP,
although in much literature it is called DP.) The rule in question is known as
Quantifier Raising (QR for short). The difference between QR and the rule which
moves relative pronouns is that QR applies in the component of the grammar that
maps surface representations to LFs. Put differently, it has no phonological
consequences, and so is often called covert rather than overt movement.
Incidentally, while the rule is known as Quantifier Raising, it is formulated here to
allow any NP to raise, including ordinary definite NPs like the book. The rationale
for this generalization will be pointed out in 14.4.2.1. Assume

syntax. Where the Quantifying-In solution differs from the Direct Compositional approach
to be discussed in 14.3 is in using a more complex set of syntactic operations than are being
used here. In particular, this approach builds up an expression like Romulus loves he5 (or
one can use Romulus loves t5) whose meaning is the same as the meaning of the node
labeled S in (4) below. This then is converted to the sentence Romulus loves every wolf by
a rule substituting the GQ onto he5; the semantics performs the two operations that
correspond to the interpretation of the higher portions of the tree in (4). The point, then, is
that there is no actual level of representation in which the GQ is higher in the tree (as in
 14.2. QUANTIFIER (LOWERING)/RAISING 253

Generative Semantics and in the “modern” LF view) and thus it is Direct Compositional, but
it does require a more complex view of the syntax.
further that each NP comes with an index and that QR leaves an indexed trace in
the position from which the NP raised. While we adopt a convention according to
which the index remains on the raised NP, it should be noted that the index on that
NP plays no role in the interpretation. We will not formalize the rule here because
its precise formulation depends on a variety of theoretical assumptions. It suffices
to assume that (1) can have an LF as in (4) (other LFs are also possible if, for
example, the subject NP also raises):
(4)
S

NP6 Λ6

every wolf S

NP VP

Romulus V NP

loves t6

To take stock: this solution requires the addition of QR and, of course, does not
maintain Direct Compositionality. The grammar keeps track of representations
which are interpreted, there are various levels of representation, and there is
movement without phonological consequence as well as movement with
phonological consequences. That said, the apparatus set up in section 13.3 (see
in particular the rule in (9) in Chapter 13) is now sufficient to interpret this LF and
nothing more needs to be said. The step-wise semantic interpretation of the
expression labeled 6 is akin to the interpretation of 8 shown in (8) in Chapter 13, and
so we leave the details to be spelled out in the exercise directly below. Notice that
this, of course, also requires the use of variables and assignments, but the
(informal) non-Direct Compositional fragment in section 13.3 already added that,
so nothing new is needed here if this is embedded in the same general theory as
that explored in section 13.3. Note that [[ 6]] will have a constant value on all
assignment functions (as it contains no other unbound traces or pronouns within
it), and on all assignments it is the characteristic function of the set of things that
 254 14. GQs IN OBJECT POSITION: TWO APPROACHES

Romulus loves. This then is taken to be the argument (on any assignment g) of
[[every wolf]]g.
14.4. Work through the full semantic details for the interpretation of (4) showing
the value for each node in the tree.

The fact that (3b) is ambiguous between both scope possibilities also follows.
For the wide-scope object reading, the most straightforward way to derive this is
via an LF where only the object has raised. That LF is exactly like the one in (4)
except that there is an NP of type <<e,t>,t> in subject position; it is shown in (5a).
It is also possible to raise both the subject and the object but have the object have
wider scope than the subject; this results in exactly the same reading as if only the
object raised. (See Exercise 14.6.) However, another possibility is for both the
subject and the object to raise, but the subject to have wider scope than the
object; that LF is shown in (5b). Below these two LFs we work through the
semantics for (5b). We have numbered all of the nodes in (5b) purely for
convenience to refer to them; the numbering is not part of the system itself.
Moreover, here we use chase’ to mean not just the semantic value of the word
chase (since that is a function from assignments to functions of type <e,<e,t>> but
rather to refer to the value that this receives on all assignments (since it is a
constant function). (Similarly for pig’ and wolf’.)
(5)
a. S b. S (1)

NP Λ6 NP (2) Λ 7 (3)

every pig S some wolf S (4)

NP VP NP (5) Λ 6 (6)

some wolf V NP every pig S (7)

NP (8) VP (9)

t7 V (10) NP (11)

chased t6
chased t6

for any g:

[[11]]g = g(6)
 14.2. QUANTIFIER (LOWERING)/RAISING 255

[[10]]g = chase’
[[9]]g = chase’(g(6))
[[8]]g = g(7)
[[7]]g = chase’(g(6))(g(7))
[[6]]g = ºxe[chase’(x)(g(7))] (this is by Rule 9 given in Chapter 13)
[[5]]g = ºP<e,t>[8y[pig’(y) ! P(y)]]

[[4]]g = ºP<e,t>[8y[pig’(y) ! P(y)]](ºxe[chase’(x)(g(7))]) =

8y[pig’(y) ! ºxe[chase’(x)(g(7))] (y)] = 8y[pig’(y) ! chase’(y)(g(7))]

[[3]]g = ºxe[8y[pig’(y) ! chase’(y)(x)]]

[[2]]g = ºP<e,t>[∃z[wolf ’(z) & P(z)]]

[[1]]g = ºP<e,t>[∃z[wolf ’(z) & P(z)]](ºxe[8y[pig’(y) ! chase’(y)(x)])=

∃z[wolf ’(z) & ºxe[8y[pig’(y) ! chase’(y)(x)]](z)]] =

∃z[wolf ’(z) &8y[pig’(y) ! chase’(y)(z)]]

14.5. Using the above as the model, show the computation for the meaning
associated with the LF in (5a).
14.6. Construct a second LF for the wide-scope object reading in which both
subject and object are raised, but the object is above the subject and show the
computation of its meaning.
14.7. To the extent possible, take each of the lines shown in the semantic
computation of (5b) above, and state the model-theoretic objects in prose. For
example, [[6]] is, on any g, that function that characterizes the set of objects
that g(7) chased.
14.8. It was assumed above that ordinary NPs like Romulus can also undergo
QR. We leave it open in this exercise as to whether in the theory under
consideration in this section such an NP can only have a meaning of type e (its
lexical meaning) or whether it can also type-lift to have a meaning of type
<<e,t>,t>. The conclusion to be drawn from this exercise will turn out to be the
same either way.
Note that—unlike the sentence in (3b)—the corresponding sentence with an
ordinary NP such as Romulus in (surface) subject position is not ambiguous:
 256 14. GQs IN OBJECT POSITION: TWO APPROACHES

(i) Romulus chased every pig.

Since QR is the mechanism used to account for scope ambiguities, a (naive)

conclusion that someone might be tempted to draw form the lack of ambiguity
in (i) is that QR should not indeed be able to raise the NP Romulus out of the
sentence at LF.

(cont.)

Show why this conclusion is incorrect and that there is no harm in allowing it
to raise. One can show this using either the e-type meaning for Romulus or the
<<e,t>,t> meaning, so show it both ways. (If it has the etype meaning, general
type-driven principles will presumably work so that the meaning of its sister
node labeled is the function which takes the individual r as argument.)

One might have observed that—as presented here—the minimal derivation

(i.e., the one involving the least steps in the computation of the LF) is the derivation
resulting in (5a) which gives the wide-scope object reading. Since this reading is
greatly dispreferred to the narrow-scope object reading, one might take this as a
criticism of this approach. This, however, would be hasty, for two reasons. First,
there might be an independent explanation (perhaps having to do with processing)
for the strong preference for subject wide scope. Second, the QR view is often
embedded in a theory or theories with additional assumptions that we are ignoring
here but which make the situation more complex. Suffice it to say here that
readers familiar with the VP-internal subject hypothesis can explore how this
would interact with more fully fleshed out LFs and how this would interact with the
scope predictions. There also remains the question of how (or whether) this basic
approach accounts for the coordination cases in (2) (without positing that VP
coordination always involves hidden coordination of two Ss). This is left for the
reader to explore in the following exercise. (Further interactions between scope
and coordination are discussed in section 14.3, and Exercises 14.16 and 14.18
will return to how to think about those in the QR approach.)
 14.2. QUANTIFIER (LOWERING)/RAISING 257

*14.9. As noted in Exercise 13.10, it is often assumed (since Ross 1967) that
there is a constraint according to which material may not move (in the syntax)
out of a coordinate structure (this is known as the Coordinate Structure
Constraint). It is further often assumed in the general framework under
discussion here that QR is subject to the same constraints on movement as are
the other movement rules (such as the rule fronting
 258 14. GQs IN OBJECT POSITION: TWO APPROACHES

which) that are posited within that framework. However, the assumption that
QR should also be subject to the Coordinate Structure Constraint is
independent of the basic approach towards GQs in object position being
exposited here. For the purposes of this exercise, then, ignore this assumption.
Given this, show how the QR approach can give a representation for the
sentence in (2) (there are in fact two possible representations) without having
to posit that the two conjoined VPs are really “hiddenly” Ss.6 (That is, use the VP
conjunction analysis that has been assumed throughout this book.) Then,
carefully work through the truth conditions of your representation(s) and see if
they are correct. Note: This is subtle and challenging, so work through it
carefully. The best method is to work through all the semantics making all
possible º-conversions until you arrive at a first-order logic formula for each
representation, and then compute the truth conditions for that formula.

14.3. A Direct Compositional solution

The fact that generalized quantifiers occur in object position (usually thought of as

an e-type argument position) combined with the existence of the scope

ambiguities is often used as the poster child evidence for the claim that the

semantics does not work in tandem with the syntax, but is instead read off a level

of LF at which items are rearranged. But such a conclusion is premature: we have

seen over and over that fancier ways to package meanings allow a theory in which

the semantic composition 6

Readers familiar with the VP Internal Subject Hypothesis (VPISH) might be tempted
to solve this problem using that hypothesis. Note, though, that this still requires something
extra, and actually is very similar to simply saying that the conjoined VPs are actually two
“sentences.” That is, under VPISH there are two subject positions (even though they are
called “VP subjects” rather than “S subjects”) and the two moved or raised subjects merge
into one. This is thus not substantially different, for the present purposes, from some
version of a Conjunction Reduction analysis.
 14.3. A DIRECT COMPOSITIONAL SOLUTION 259

respects the syntax and works in tandem with it. Actually, there are quite a number
of available ways to accommodate the distribution (and readings) of GQs in a
Direct Compositional fashion; just one such strategy will be explored here. This
involves a unary rule operating on the verbs (and on other items that combine with
NPs).7
By way of background, we turn first to an account of GQs in object position
proposed in Montague (1973). The general strategy throughout much of
Montague’s work is to “fancy up” the types of the basic meanings of lexical items
(i.e., the meanings that they are listed as having in the lexicon). What would this
strategy lead to? It would lead to the assumption that we have been wrong all
along about the meaning of ordinary transitive verbs. They are not of type <e,<e,t>>
but rather of type <<<e,t>,t>,<e,t>>—they expect generalized quantifier meanings
in object position. This hypothesis is, of course, in no way incompatible with the
fact that ordinary NPs like Mitka and the big bad wolf can appear in object position,
since these too can have the lifted meanings of type <<e,t>,t>. (And recall that—
under Montague’s strategy of packing all of the complexity into the lexical
meanings—definite NPs have only meanings of this type. The fragment here
revises this so that the lexical meanings of definite NPs are of type e, but they can
lift to GQ meanings by the rule l.) The corresponding syntactic category—under
the Montague treatment— would of course be parallel; the lexical category of
chase would not be (S/NP)/NP but rather (S/NP)/(S/(S/NP)).
Just what is the meaning of chase in this view? To exposit, let chase *’ be the
ordinary Curry’ed two-place relation “chase” between individuals—that is, the
meaning that we have assumed for chase until this point. Then let actual [[chase]]
be the following, which we from now on call chase’MONT:

(6) ºP<<e,t>,t>[ºx[P (ºy[chase*’(y)(x)])]

To get the intuition of the relation between the ordinary (Curry’ed) two-place
relation chase*’ of type <e,<e,t>> and chase’MONT (of type <<<e,t>,t>,<e,t>>) it is
helpful to say the above in prose. (6) is a function that first takes a GQ-type
meaning and then some individual x as argument,

7
Some other Direct Compositional accounts can be found in Montague (1973) (see n.
5), Moortgat (1997), and Barker (2001) which uses quite a different set of techniques.
 260 14. GQs IN OBJECT POSITION: TWO APPROACHES

and returns true if the property of being ordinary-chased by x is in the set of sets
mapped to true by the GQ. For example, if chase’MONT combined with [[every pig]]
it would return the function that maps an individual to true just in case the property
of being chased by that individual is in the [[every pig]] set. Put differently, it returns
the function that maps an individual x to true just in case the set chased by x has
the [[pig]] set as a subset. And this of course gives the right meaning for the
expression chased every pig. The syntax is straightforward; given the revised
category for chase, it can of course combine with any GQ in object position.
But this revision of the lexical meaning of chase is still not enough. For this gives
only the wide-scope subject meaning for (3b), as shown in (7):
(7) [[chased every pig]] = ºx[every-pig’ (ºy[chase*’(y)(x)])] =

ºx[ºP<e,t>[8z[pig’(z) ! P(z)]]( ºy[chase*’(y)(x)])] =

ºx[8z[pig’(z) ! ºy[chase*’(y)(x)](z)]] = ºx[8z[pig’(z) !
chase*’(z)(x)]]

[[some wolf chased every pig]] = some-wolf ’(ºx[8z[pig’(z) ! chase*’(z)(x)]]) =

ºP[∃y[P(y) & wolf ’(y)]](ºx[8z[pig’(z) ! chase *’(z)(x)]]) =

∃y[ºx[8z[pig’(z) ! chase*’(z)(x)]] (y) & wolf ’(y)] =

∃y[8z[pig’(z) ! chase*’(z)(y)] & wolf ’(y)]

We have laboriously worked through all the steps here, but it may be easier to get
the intuition using prose. [[chase every pig]] characterizes the set of individuals x
which are such that the set x chases has all the pigs in it. Combining this with the
subject says that this set has a non-empty intersection with the wolf set. And
hence there must be at least one wolf who chased all the pigs. This solution, then,
takes us only halfway there. It does not give the wide-scope object meaning
(which in this case is the weaker meaning) whereby each pig is chased by some
wolf but not necessarily the same wolf. Incidentally, it will not help to invoke the
derivation in which some wolf combines first with chased (by g-sl on some wolf
plus application) and then combines with every pig. Recall that g-sl + application
is equivalent to function composition. But notice now that by the very definition of
function composition, [some-wolf ’ o chaseMONT’](every-pig’) is equivalent to
somewolf(chaseMONT’(every-pig’)). So this too results in the narrow-scope reading
for the object.
 14.3. A DIRECT COMPOSITIONAL SOLUTION 261

But there is an easy fix. The next step is to take the strategy proposed by Partee
and Rooth (which is consistent with their general strategy discussed in Chapter
11). Assume that ordinary transitive verbs like chase are indeed listed in the
lexicon with the ordinary meaning of type <e,<e,t>> but map by a unary rule to a
homophonous item that allows instead for a generalized quantifier in object
position. In other words, assume that our original understanding of [[chase]] is
correct: it is simply chase*’ (we henceforth omit the subscript asterisk since we
are using chase’ and [[chase]] interchangeably to be the meaning of chase). But,
according to the Partee and Rooth proposal, it maps to chase’ MONT by a unary rule
given below in TR-12. In the Categorial syntax assumed here this would be
accompanied by a syntactic part as well; for the moment we state just the
semantic part as the rule will be generalized below):
TR-12. Let Æ be an expression whose meaning is of type <e,<e,t>>. Then there is a
homophonous expression whose meaning is of type <<<e,t>,t>.<e,t>> where [[]]
= º P<<e,t>,t>[ºx[P (ºy[Æ’(y)(x)])]]

For the data considered so far we could stop here. The reading of (3b) where
the object has wide scope is accounted for in the way discussed in section 14.1,
and the reading where it has narrow scope is accounted for in the way shown
above. But Hendriks (1993) noted that the Partee and Rooth rule can easily be
generalized to allow any e-type argument slot to “lift” to allow for a GQ instead,
and this also allows for both scope possibilities. (Hendriks argued for this solution
on the grounds that the function composition approach will not help in the case of
a language like German where the two NPs are on the same side of the verb; but
as noted in n. 2, this depends on whether and how a Wrap operation is folded into
the system. Still we pursue the Hendriks’ generalization here as it is quite
interesting and has advantages with respect to coordination.) Although the
formalism for stating the generalized rule is a bit clumsy (merely because we need
to introduce some new notation), the rule itself is just a fairly natural generalization
of the Partee and Rooth rule. Modifying Hendriks’ terminology slightly, we will refer
to the generalization as Argument Slot Lift (argslot-l).
It is initially easiest to state the rule in prose. Let f be any function that takes n
arguments (including 0) of any type, then an e-type argument, and then m other
arguments (including 0) of any type to return 1 or 0. Then argslot-l(f ) is a function
which also takes the same n arguments, then a generalized quantifier argument P,
and then the same m arguments. It returns 1 just in case P maps to 1 the set of
individuals x which are such that had f applied to the same n arguments and then
 262 14. GQs IN OBJECT POSITION: TWO APPROACHES

to x and then to the same m arguments, f would have returned 1. To put this in
notation we need to introduce a few notational conventions. Let K ! be any
sequence of semantic types (of any number, including none), similarly for L !. Let
XK! be any sequence of variables each of the same type as the sequence notated
by K! (again of any number, including none), and similarly for X L!. Finally let ºXK! be
a sequence (of any length, including 0) of ºs and variables of the relevant type
(similarly for ºXL!). Then we define argslot-l as follows:
(8) Definition of Argument Slot Lift (argslot-l):
Let f be any function of type <K!,<e,<L!,t>>>. Then argslot-l(f) is a function of type
<K!,<<<e,t>,t>,<L!,t>>>, where argslot-l(f) =
ºXK! [ºP[ºXL! [P(ºxe[f(XK!)(x)(XL!)])]]].

As formulated here, argslot-l is not uniquely defined; if there is a function with

more than one e-type argument it can apply to either one. When needed, we will
use the notation argsloti-l to refer to the operation that lifts the ith argument slot
(where the object slot of a transitive verb, for example, is the first slot; lifting the
subject position would be argslot2-l). Thus note that the Partee and Rooth rule is
simply argslot1-l applied to a function of type <e,<e,t>>. Armed with argslot-l, we
can add to the fragment the full unary rule in R-6; A! here means any sequence of
syntactic categories as does B! ; once again in both cases the sequence can be
empty.
R-6. Let Æ be an expression of the form <[Æ], ((S/A !)/NP)/B!, [[Æ]]> where A is of length n.
Then there is an expression of the form <[Æ], (( S/A!)/GQ)/B!, argslotn+1-l([[Æ]])>.

Because the notation is indeed clumsy it might be obscuring the more important
model-theoretic intuition of what argslot-l does; the reader is thus encouraged to
revisit the prose description given above to internalize the operation.
Now consider the case of a transitive verb like chase. As has already been
shown, when this inputs argslot1-l, this gives chase’MONT. This allows a GQ in
object position, and if this verb is taken as argument of the subject GQ, the result
is the narrow-scope reading on the object for a case like (3b).
*14.10. Another way to derive (3b) is by application of argslot 1-l followed by
argslot2-l. That is, the object argument slot is lifted first to a GQ slot, and then
the subject slot is lifted. This gives exactly the same reading as one would get
by lifting the object slot and then applying the result to the every pig to give the
VP chase every pig and then having that be the argument of the subject. In other
words, this gives the same reading as the derivation shown above in (7) and
 14.3. A DIRECT COMPOSITIONAL SOLUTION 263

results in a narrow-scope object. Work through the full syntactic and semantic
details to show this.
*14.11. Related to the discovery above, consider a function of type <e,t>.
argslot1-l maps this to a function of type <<<e,t>,t>,t>. Work through the
semantics of argslot1-l(f). What other operation gives exactly the same result?

But there is an additional derivation: one in which both argument slots lift but
where the subject slot lifts first and then the object slot does. And this gives wide
scope to the object:
(9) chase, (S/NP)/NP, chase’ ! (via argslot2-l) chase,
(S/GQ)/NP, ºye[ºP [P (ºxe[chase’(y)(x)])]] call this
expression chase2
! via argslot1-l applied to chase2:

chase, (S/GQ)/GQ, ºQ [ºR [Q (ºze[chase2’(z)(R)])]]

chase every pig, S/GQ, ºQ [ºR[Q (ºze[chase2’(z)(R)])]](every-pig’) =
ºR[every-pig’(ºze[chase2’(z)(R)])] = (by unpacking chase2’)
ºR[every-pig’(ºze[ºye[ºP [P (ºxe[chase’(y)(x)])]](z)(R)])] =
ºR[every-pig’(ºze[ºP[P(ºxe[chase’(z)(x)])] (R)])] = ºR[every-
pig’(ºze[R (ºxe[chase’(z)(x)])])]
some wolf chased every pig, S,
ºR[every-pig’(ºze[R (ºxe[chase’(z)(x)])])](some-wolf ’) = every-pig’(ºze[some-wolf
’(ºxe[chase’(z)(x)])])

In prose, this says that the pig set is a subset of the set of individuals z who are
such that the set chasing z has a non-empty intersection with the wolf set. Or, put
more simply: the pig set is a subset of the set of individuals each of whose chasers
include a wolf. And this is the reading we are after here.
14.12. (This exercise is parallel to the one in 14-8 framed within the LF theory.)
The Direct Compositional fragment here allows the NP Romulus to lift to a GQ
with a meaning of type <<e,t>,t>, and thus allows one derivation of (i) exactly
analogous to the derivation shown in (9):

(i) Romulus chased every pig.

 264 14. GQs IN OBJECT POSITION: TWO APPROACHES

Yet (i) is not ambiguous in the way that (3b) is. Does this pose any problem for
the assumption that Romulus can type-lift? Show why or why not.

We will not give further examples, but it is clear that this strategy will generalize
to cases of three-place verbs such as those in (4), and indeed will allow any scope
possibilities between different NP positions of the same verb. But we should note
that generalizing this strategy to account for the existence (and readings) of GQs
in all NP positions is not automatic; a bit more needs to be said. First, and most
important, notice that AS-L is defined in the semantics only for those functions
whose final result is a truth value (see Exercise 14.13 to discover why this is so).
But prepositions (at least some of them) have been taken to be the identity
function on individuals and so would need to also be given a meaning in which
they are the identity function on GQs, since we of course get expressions like read
the book to every child. Perhaps they should be listed as the identity function on
NP and lift(NP), where those two categories can be collapsed in some way.
Moreover, how exactly should we generalize the syntax in R-6 so that PP as well
as NP argument positions can lift? A plausible answer is that the category PP is
really simply the same as the category NP but with the addition of a special feature
(in other words, this is just like a case-marked NP). We will leave the full details
open, acknowledging that some (hopefully minor) modification is needed to
account for the full range of GQs under this solution while the LF solution covers
all of the cases straightforwardly. (Of course, this is true only given some actual
formulation of QR, but assume it can be formulated to raise any NP from any
position whatsoever.) Finally, the rule R-6 applies only to expressions whose
syntactic category has S as the final result. The syntax can (and probably should)
be generalized to include any category whose final result category is of semantic
type t (as, for example, the category N).
*14.13. The syntactic categories in R-6 can be generalized in various ways, as
suggested in the discussion above. However, the semantic definition of argslot-
l cannot be generalized to apply to functions whose final result is e rather than
t. The basic idea crucially requires the input function to be one whose final
result is t. Think through why this is so. (Hint: it has to do with the type of GQs
themselves.)
 14.3. A DIRECT COMPOSITIONAL SOLUTION 265

There are some interesting predictions of this general approach with respect to
coordination. Consider a sentence like (10):
(10) A/Some doctor examined every inpatient and released every outpatient.

The salient reading for this is one in which the subject has wide scope. There is no
problem getting this sentence with this reading. Here both examine and release
simply undergo AS1-L to give what we have been calling examine MONT and
releaseMONT; each take the relevant GQs as argument and the two VPs are
conjoined in the normal way.

14.14. Show the details of this derivation. To save steps, feel free to deCurry
and and combine the two VPs in a single step.

But it is predicted that there should also be a reading in which every inpatient
and every outpatient both scope over the subject. This would involve applying
argslot1-l to both verbs. The two are of the same type (and category); both VPs are
of type S/GQ and should conjoin. It is simply not clear (to this author at any rate)
whether this is a good or bad prediction; informants seem to vary on whether they
accept this reading. To help bring out this reading (if it exists), we can set up the
following context. The chief resident at the local hospital is allowed to go home
for the night only after every outpatient has been released by some doctor and
every inpatient has been checked over by a doctor. The resident announces that
she is now able to go home, for, after all, Some doctor (or other) has examined
every (single) inpatient and (has) released every last outpatient. Again we leave it
open as to whether this reading does emerge. Of course that reading is expected
in any case to be quite difficult to access. First, it is generally the
 266 14. GQs IN OBJECT POSITION: TWO APPROACHES

case that wide-scope objects are dispreferred. Combining this with coordination
might just push things too far to make the reading accessible. The reader is invited
to try to construct other such cases and also to explore the parallel facts with or
(in which the wide-scope reading in some cases actually seems easier to get).8

14.15. Show the details of the derivation of (10) under the reading in question.
Again, feel free to de-Curry and to save steps.
*14.16. Assuming that there is no constraint precluding QR from out of a
coordinated VP, show that the LF account also allows both readings discussed
here (using only VP conjunction).

14.4. Some comparisons

14.4.1. The generality of the solutions

The two different approaches considered here make quite different commitments
to the organization and architecture of the grammar. Under the first approach,
representations (i.e., trees) are referred to in the grammar, and there are rules
mapping one representation to another. The input to the semantics is a
representation computed by a series of steps “in the overt syntax” followed by a
series of steps (including so-called covert movement)

8
Notice that it is predicted that this should also have a “mixed” reading, one which can
be paraphrased as “For every patient, there is some doctor who examined him/her, and
there is some particular doctor who released every patient.” This is because there are two
forms of examine and of release. In one, argslot 1-l is followed by arglost2-l and the other
reverses these. But the two end up being of the same syntactic category (S/GQ)/GQ and
the same semantic type and so should be able to conjoin. It is quite clear, however, that
this reading is completely hopeless and will not emerge no matter how much one tries.
Hopefully this can be explained by some sort of processing principle. Note that a version
of this problem potentially arises in the LF view as well. There too there will be a mixed
reading if the quantifiers can raise out of conjoined VPs (as is necessary if there is no
additional hidden subject for each VP). The actual mixed reading there is slightly different,
but it is equally impossible.
 14.4. SOME COMPARISONS 267

yielding an LF representation. We did not explicitly formulate the rule of QR since

some of the details depend on other theoretical assumptions, but some
formulation is needed. The resulting LF is compositionally interpreted. The
second approach adds a rather symbol-heavy type shift rule, but maintains a
Direct Compositional architecture and adds no new devices. One can hope that
the symbol-heaviness of the rule is in part an artifact of the notation; the model-
theoretic basis of the rule, while not trivial, is not that difficult to see.
There are, however, arguments that are rather traditionally taken to support the
QR (or Q Lowering) view which should be addressed here. The first is that the
device needed to allow quantifiers in object position in the first place generalizes
to give both scopes. But this true on the second approach as well: Hendriks’
generalized version of the Partee and Rooth rule which allows GQs in object
position allows both scope possibilities. A second argument sometimes given for
the QR approach is that the apparatus developed for the quantifier scopes
automatically generalizes to account for pronominal binding in general. We return
to this in the next chapter. To anticipate: it is indeed the case that given the
apparatus of the QR approach, pronominal “binding” can also be accounted for
quite easily with suitable lexical entries for pronouns. But this ends up being a
mixed blessing. For it will be shown in section 15.2.2 that using the QR approach
for pronominal binding overgenerates, and needs to be reined in by an additional
constraint.9 One area where the LF approach arguably does

9
Another quite standard argument for the QR approach, dating back to Generative
Semantics (see, e.g., Lakoff 1971; Postal 1974), is that rules which presumably move
things like relative pronouns to the front are subject to a set of constraints, and the rules
moving GQs into (in Generative Semantics) or out of (in the QR view) a sentence are subject
to the same constraints. Of course, this is phrased in theory-internal terms: the view of
relative clauses developed in section 13.3 does not assume a movement rule. But it could
be rephrased more theory-neutrally. The claim is that the gap (or “missing argument”) in a
relative clause cannot be in certain syntactic environments (known as islands; see Ross
1967) and that the same is true of the position from which a GQ “raises.” In reality, however,
the generalizations are quite murky and there is actually only partial overlap between the
environments disallowing wide-scope quantification and those allowing “extraction.” (The
former is much more limited than the latter.) Moreover, our understanding of the principles
that give rise to island effects is quite hazy, and there is a good deal of controversy as to
whether these have any syntactic basis at all.
 268 14. GQs IN OBJECT POSITION: TWO APPROACHES

have a small advantage is that it is easy to generalize it to account for both the
syntax and the semantics of GQs in all positions that can be occupied by an
ordinary NP. (Of course the rule remains to be formulated, but it is easy enough to
formulate it to raise any NP leaving an e-type trace in that position.) As noted
above, the particular Direct Compositional solution explored here does need
some refinements to extend to the case of PPs; how serious a problem this is
remains open.
On the other hand, the LF approach faces some difficulties concerning the
existence of GQs in object position in coordinated VPs (see Exercise 14.9; readers
who did not do that exercise can guess that incorrect truth conditions result from
the most straightforward solution). And along the same lines, the LF account
needs something additional (beyond the treatment of didn’t developed here) in
order to account for scope interactions with GQs and negation, as can be
discovered in the exercises below. Nothing extra is needed for these scope
interactions under the Direct Compositional fragment developed here. Again, this
is left for the reader in the subsequent exercise.

14.17. Notice that (i) is ambiguous between a reading in which the negation has
scope over the GQ and one where the scopes are reversed (the reading where
every pig has wider scope than the negation is more difficult to get, but one can
construct a context to bring it out):

(i) MacDonald did not feed every pig.

How are the two readings accounted for in the Direct Compositional fragment
developed so far?
*14.18. What will the LF account need to say in order to account for the two
readings?

14.4.2. Antecedent-Contained Ellipsis

14.4.2.1. The “textbook” argument

 14.4. SOME COMPARISONS 269

There is another very interesting argument for the QR approach worth

considering—not only to answer it, but also because it provides an illuminating
illustration of the interconnectedness of various theoretical assumptions.
It also allows a brief introduction to another rich empirical domain (ellipsis). The
argument was originally put forth in Sag (1976) albeit in slightly different terms
(based on observations from Bouton 1970), and has become (literally) a textbook
argument for QR; see, e.g., Heim and Kratzer (1998). By way of background,
consider first the phenomenon known as VP Ellipsis or VP Anaphora. This involves
sentences in which an auxiliary occurs without its complement but where the
“missing” complement is (usually) understood as the meaning of some other
overt VP in the discourse context:
(11) a. Bode will ski the course in four minutes. Lindsay will too.
b. Sally said she would complete the marathon, and indeed she did.

The second sentence in (11a), for example, is most naturally understood as

conveying the proposition that Lindsay will also ski the course in four minutes. We
henceforth use the term VP Ellipsis (VPE) in an analysisneutral way to refer to the
phenomenon in question.
Whether the missing meaning needs to be supplied in virtue of being the
meaning of some other VP which is overt in the discourse context or can instead
be supplied by some sort of inference in a given context is the subject of a long-
standing debate. For the present purposes, it is sufficient to observe that at least
generally the missing meaning is the meaning of some other overt VP in the
discourse context; call that the antecedent VP. 10 Given that, the phenomena of
VPE can be used as a tool for determining the exact meaning of the antecedent
VP.
Before proceeding further, more needs to be said about how the grammar
accounts for these cases where an auxiliary is missing its complement and the
missing meaning is supplied via some antecedent VP. Does the sentence Lindsay
will too in (11a) actually have as its representation at some level Lindsay will ski
the course in 4 minutes too where the VP ski the course in 4 minutes is silent or
deleted? If so, is this material present and deleted “in

10
This term is a bit misleading in that the VP which supplies the meaning can actually
follow the ellipsis site, as in the following discourse (a full discourse is given here to
highlight the point that the missing meaning does not come from any other sentence):
 270 14. GQs IN OBJECT POSITION: TWO APPROACHES

(i) Speaker A. Did Sarah finally decide which candidate to endorse?

Speaker B. Yes. Once Todd did, she finally endorsed Newt.
(the first sentence can be understood as “Once Todd endorsed
Newt”)
the overt syntax” or is it copied in during the computation of LF? Or might we
instead think of these as containing no VP like this at any level, but where instead
the function [[ski the course in 4 minutes]] is somehow supplied as part of the
understanding of the sentence? For example, one might think of the semantics of
(11a) as containing a “free variable” of type <e,t>, whose value is supplied by the
discourse context.
Unfortunately, space precludes doing justice to these questions here. And so
we will just pick one way to think about this construction where the choice is
motivated by two considerations. First, the view chosen here is compatible with
the general assumptions made throughout this book. And second—although this
view is not shared by most researchers who do advocate LF—it does not in any
way impact on the argument to be developed below and is therefore a safe and
theory-neutral way to exposit the textbook argument for QR. Assume, then, that
there is no linguistic VP (at any level); for now we can think of the semantics as
containing a “free variable” of type <e,t> as argument of the auxiliary, where the
value of this variable is ultimately supplied by context. More concretely, assume
that an auxiliary such as will is listed in the lexicon as having the category
(S/NP)/(S/NP) and has a meaning of type <<e,t>,<e,t>>.18 Call this will1. Assume
further that there is a unary rule mapping all auxiliaries to homophonous forms
that occur with no VP complement; for the case of will call this second item will 2.
There are, incidentally, a number of ways in which those items traditionally called
“auxiliary verbs” share syntactic properties, so it is reasonable to posit a feature
[AUX]; items with this feature input the rule in question. (Lest it appear to the

18 Here is a case where times are actually crucial: obviously will’ maps a timedependent
set into another time-dependent set and so its meaning is really of type
<<e,<i,t>>,<e,<i,t>>>. Its full meaning is not difficult to spell out: it takes a function P of
type <e,<i,t>> (a time-dependent set), then an individual x and then a time i and returns
true if there is a time i’ later than i such that P holds of x at i’. Since adding in the extra
time layer makes the exposition more cumbersome and is irrelevant to the basic point,
we will ignore it here.
12
For example, under some accounts an auxiliary may select for an “empty operator”
E which occurs as sister to the VP and suppresses its phonological content;
 14.4. SOME COMPARISONS 271

reader that we are adding unary rules willy-nilly it should be noted that any theory
needs some device to predict that auxiliaries can occur without an explicit (or
overt, or pronounced) complement. The devices vary, but some such device is
needed no matter what.12) Thus will2 has as its category just S/NP, and its meaning
is [[will1]](P), where P is a free variable over functions of type <e,t> whose value is
ultimately fixed by picking up some salient <e,t> function in the discourse context.
(Chapter 15 will give a more explicit account of “free variables” in general,
including ultimately eliminating them in a variablefree context.) In (11a), then, the
relevant function that fixes the value of P is [[ski the course in four minutes]].
Recall that (for reasons left open here) in general the function that sets the value
for P is made salient by being the meaning of an actual VP which is uttered in the
discourse context.
With this background, consider the phenomenon known as
AntecedentContained Ellipsis first noted in Bouton (1970) and exemplified in (12):
(12) Sarah will read every newspaper which Katie will.

It is quite reasonable to suspect that the mechanisms for VPE in general extend
to this case. And so, at first glance, this leads to the conclusion that will in the
second clause is will2 and is of category (S/NP) but with meaning [[will 1]](P). In
more theory-neutral terms: whatever allows for a silent or missing VP meaning in
(11a), those same mechanisms should allow for the missing VP meaning within
the relative clause in (12). Hence under the set of assumptions made so far, the
value for P is the meaning of some other overt VP in the discourse context. Since
(12) is fine without elaborating the discourse context any further, this must be the
meaning of the matrix VP. Notice that a key premise in the development of the
argument is that the missing or silent meaning in (12) must be a VP type meaning.
This follows (in the textbook view) not only because this is crucially seen as VP
Ellipsis, but also because the textbook argument is always embedded in a theory
maintaining the view of relative clauses developed in section 13.3 (the
movement/trace account). Crucially, under that view, which combines with some
expression whose meaning is of type <e,t>.
In other words, take the non-elliptical version in (13):
(13) Sarah will read every newspaper which Katie will read.

Recall that which Katie will read is (under the movement/trace account) analyzed
syntactically as which [L-8 Katie will read t8], where [[Katie will read t8]] is—on any
 272 14. GQs IN OBJECT POSITION: TWO APPROACHES

assignment g—the proposition that Katie will read g(8). The meaning of [ L-8 Katie
will read t8] is—on any g—the set of things that

see Merchant (2001) for this type of account. And so something is needed to predict what
can and can’t select for E.
Katie will read. Moving up the tree, the meaning of which Katie will read t 8 is the
same, and this set (on any g) intersects with [[newspaper]] S (on any g). Thus, in
order to provide a semantic composition for (12) in such a way that all the types
can mesh, the ultimate understanding of the free variable P (which stands in for
the “missing meaning” that combines with the auxiliary) must be [[read t 8]]. That
is, it has to be a full VP meaning which contains in object position an “open
individual” (i.e., a non-constant function from assignments to individuals). If this
meaning is supplied, then the meaning of the relative clause in (12) can be put
together exactly like the meaning of the relative clause in (13) is.
But what VP can supply the missing meaning? As noted above, without any
further context this is perfectly interpretable (where (12) is understood as the
same as (13)). And so, the antecedent VP can only be the matrix VP. But as it
stands, it can’t be the matrix VP! After all, the matrix VP itself is read every book
which Katie will, whose meaning is [[read every book which Katie will P]]. But if
the <e,t> function to be assigned as value of P itself contains P, there is an infinite
regress. Or, as it is sometimes phrased, we arrive at the antecedent containment
paradox.
How can this paradox be resolved? Enter QR to the rescue. There is no problem
if the object NP every book which Katie will has undergone QR leaving an indexed
trace in the object position of the matrix. Under this view, the matrix VP can have
as its LF representation read t8 which is (relative to an assignment) a completely
sensible function of type <e,t>. This function thus supplies the value for P, and the
antecedent containment paradox disappears. The full details are left for the
subsequent exercise.

14.19. Show the LF after raising. For the purposes of how this has been
exposited here, assume that both which and every book which Katie will are
indexed 8, so both leave t8 in the relevant position. Then show how the matrix
VP is interpreted to have a meaning of the right sort to fix the value of the free
variable P.
 14.4. SOME COMPARISONS 273

Incidentally, we can now return to why QR was formulated to allow any

NP to raise (including definite NPs). This is because we also find Antecedent
Contained Ellipsis when definite NPs occur in object position as in (14), and the
paradox would arise here too unless the candidate that Todd will is raised at LF so
that the matrix VP is [endorse ti] (for some i). (14) Sarah will not endorse the candidate
that Todd will.

14.4.2.2. Antecedent-Contained Ellipsis as Transitive Verb

Phrase Ellipsis
We have dwelt on this case in some detail not only because it is relevant to the
QR debate, but also because it dramatically illustrates the interconnectedness of
various assumptions, and shows how changing some fundamental assumptions
can significantly alter the conclusions. The astute reader will already have noticed
that the argument is framed within the trace/movement account of relative
clauses and relies on assumptions that are not shared by, for example, the Direct
Compositional account in section 13.4. In particular, the argument rests on the
assumption that will in the relative clause in (12) needs to find an <e,t> type
meaning—i.e., the meaning of some VP. 19 But consider again the non-elliptical
version (13) under the Direct Compositional account in section 13.4. The
compositional semantics puts this together by combining will only with an
expression of category (S/NP)/NP (which we will refer to henceforth as a
Transitive Verb Phrase or TVP), and combining its meaning with a two-place
relation of type <e,<e,t>>. In particular, ordinary will undergoes the g-sl rule so that
it maps to something looking for an <e,<e,t>> expression to give back a meaning
of that same type. If (12) can be put together in a fashion analogous to (13) there
is no need to assume that some actual full VP is required in the discourse context
to supply the missing meaning; all that is needed is TVP to supply the meaning.
And of course the lexical meaning of read is available. And so, the argument that
the object must have moved out by covert QR leaving a trace vanishes. This point
was first noticed in Cormack (1984) and developed further in Evans (1988) (see
also Jacobson (1992, 2003, and 2008 for elaboration and extension to a wide

19 The remarks here are oversimplified in that in many LF accounts it is assumed that
will raises to take a full sentence as its scope; its meaning is thus actually of type <t,t> (or,
more accurately, <<i,t>,<i,t>>). But the same point will go through: it ultimately needs to
find an expression of type t with which to combine, which will still lead to the conclusion
that there has to be a trace in object position.
 274 14. GQs IN OBJECT POSITION: TWO APPROACHES

range of cases). In a nutshell is that all that is missing here (and picked up by
context) is read’—this is exactly as expected under the account of relative
clauses developed in section 13.4.
It is worthy pursuing this in a bit more detail. Listed in the lexicon is will 1 whose
category is (S/NP)/(S/NP) and whose meaning is of type <<e,t>,<e,t>>. In the
derivation of (13) (the non-elliptical version), it maps by g-sl to what we can call
will3 of category ((S/NP)/NP)/((S/NP)/NP) or, for short, of category TVP/TVP. In
(13), it finds its complement read of category TVP (and with meaning <e<e,t>>);
the result is a TVP. The rest of the composition is left to the reader to work out; it
is exactly analogous to cases already discussed. What about (12)? The only
additional assumption that needs to be made is that the unary rule shifting
auxiliaries in such a way that they occur without their complements can easily be
stated to allow g-sl (will) (i.e., will3) to do the same. Thus corresponding to will 2
(which is a VP with meaning [[will]](P)) the generalization of the rule allowing an
auxiliary to occur without its complement will allow will 4 of category TVP and with
meaning [[will3]](R). (That meaning can be further unpacked as ºxe[will1’(R (x))]. To
be sure the actual rule allowing for “missing complements” has not been
formulated here; one set of explicit details is worked out in Jacobson (2003) (in a
variable-free framework), but the reader can on their own formulate a rule in such
a way as to accomplish the necessary generalization.)
The rest is straightforward. R must have its value supplied by being the meaning
of some overt expression in the discourse context, but in the case that expression
is just a TVP rather than a VP. There is one point worth noting. In the matrix
sentence, read must undergo argslot1-l in order to occur with the quantified NP in
object position, and so this item does not have a meaning of type <e,<e,t>> (rather,
it is of type <<<e,t>,t>,<e,t>>. But there is no problem with the analysis; this item
is listed in the lexicon as read of category (S/NP)/NP with meaning of type
<e,<e,t>> and so presumably the ordinary two-place relation read (or what we
called earlier read*) is indeed the meaning of this item (it just has the second
meaning as well).
An ambitious reader is invited to sketch out the full derivation of (11a) in both
theories. One will see that the QR/movement/trace account involves a series of
representations and thus rules mapping representations to representations. The
Direct Compositional account, on the other hand, involves application of a series
of unary rules mapping local expressions to others but makes no crucial use of
representations. (Note too that it is misleading to think of the unary rules as
changing the meaning of expressions; these rules are simply proofs that the
 14.4. SOME COMPARISONS 275

relevant output expressions are well-formed on the basis of the well-formedness

of input expressions.)

14.5. Generalized quantifiers in embedded sentences

We have sketched two accounts of generalized quantifiers in object position and

their scopal interaction with GQs in subject position, but have not looked at all at
the distribution of GQs (and their scope possibilities) in embedded sentences.
Consider for example a sentence like (15):
(15) Some candidate thinks that every proposal will be rejected.

The argslot-l solution to scope ambiguities developed in section 14.3 does not
readily allow for a reading where every proposal scopes over some candidate
here. This is because it involves the manipulation of argument slots and so
predicts (without additional machinery) that the sort of scopal interactions shown
in (3b) are restricted to what we might call co-arguments of the same verb.
Arguably, this is a good prediction. Although the facts are a bit murky, the
general wisdom is that (15) does indeed lack a reading in which every proposal
has wide scope over some candidate (see, e.g., May 1985). 14 Indeed much of the
literature within the QR tradition formulates QR to be “clausebounded,” meaning
that an NP within an embedded sentence can raise only to have scope over that
sentence and not higher. However, in the Direct Compositional fragment
developed here there is an interesting additional prediction: the reading with wide
scope on every gun control bill should emerge in cases like (16):
(16) Some senator plans to vote against every gun control bill.

Notice that we have switched here to a case where the embedded material is a
subjectless infinitive VP. This is because, for reasons which remain unclear, it is
extremely difficult to get the relevant reading with a full embedded tensed S. But
the existence of a wide-scope reading for every gun control bill in (16) is sufficient
to make the point. This reading should

14
Hendriks’ actual proposal did allow an embedded GQ to scope over a GQ in the
matrix for he posited an additional rule (called Value Lift) which accomplished this. We
will not explore this here.
 14.6. SUMMARY: THE FRAGMENT SO FAR

(and presumably does) emerge because—however one analyzes embedded infinitive

VPs like the one in (16)—the use of g-sl predicts that plans to vote against can
combine to give a complex transitive verb. That, then, can undergo the two argslot-l
steps exactly as sketched in (9).

14.6. Summary: The fragment so far

Add to the fragment in Chapter 13 the following:

Definition:
Let f be any function of type <K!,<e,<L!,t>>>. Then argslotn+1-l(f) is a function of type
<K!,<<e,t>,t>,<L!,t>>> (where n = the number of slots in K) where AS n+1-L(f) = ºXK! [ºP[ºXL!
[P(ºxe[f(XK!)(x)(XL!)])]]].

Additional rule: argslot-l (argument slot lift):

R-6. Let Æ be an expression of the form <[Æ], ((S/A!)/NP)/B!, [[Æ]]> where A is of length n. Then
there is an expression of the form <[Æ], (( S/A!)/GQ)/B!, argslot n+1-l([[Æ]])>.
 277

15 The interpretation of pronouns :

Two accounts
15.1. Introductory remarks 274 15.4. More on Weak Crossover 305

15.2. The LF (variable-ful) view 278 15.5. More on free “variables” 317

15.3. A Direct Compositional and 15.6. The i-within-i effect

variable-free approach 286 revisited: A solution 320

15.1. Introductory remarks

15.1.1. Variables or not?

The previous two chapters considered relative clauses and GQs in object position
from two points of view. The most striking difference between the two views is the
one that has been stressed all along in this book. One uses a Direct Compositional
architecture in which there is no notion of a level of representation that inputs the
semantics; indeed the grammar does not actually refer to representations (in the
sense of, e.g., trees). The other view—the LF view—makes crucial use of
representations, and the representation inputting the semantics is not the one
representing the pronounced sentence, but is an abstract LF in which a constituent
can move to a position other than where it is pronounced.
There is a second major difference between the two approaches sketched in
Chapters 13 and 14. The LF theory crucially uses assignment functions as part of the
semantic apparatus (and indices in the syntax). Thus, like the
15.1. INTRODUCTORY REMARKS

artificial language developed for convenience in Chapter 9, expressions have as their

semantic value a function from the set of assignments (G) to something else. The
Direct Compositional fragment makes no use of variables, indices, and assignments.
The semantic composition of a relative clause does not use an indexed trace whose
index is lambda-abstracted over. Rather, the grammar puts relative clauses together
using the g-sl rule (or function composition). What is seen as movement in the
standard theory—which is what leaves the indexed trace—is instead a case of a
function that happens not to combine with its expected argument. Rather, the
function in question itself is the argument of another function. So, for example, in
which Martha refinished, [[refinish]] does not combine with a variable of type e nor
with any object at all. Rather, (lifted) Martha undergoes g-sl and in turn takes refinish
as argument. The LF view and the Direct Compositional view also part ways on the
interpretation of GQs in object position. Under the LF view, such a GQ raises and
again leaves an indexed trace, allowing a transitive verb not only to find as its object
an expression of type e. Crucially, the value of that e-type expression is dependent on
assignments (as it is an indexed trace). And this trace corresponds to a variable which
is lambda-abstracted over in order to create an expression of type <e,t> to serve as
argument of the raised GQ. The Direct Compositional alternative developed in
Chapter 14 just maps the transitive verb to one expecting an <<e,t>,t> argument in
object position.
Thus (in addition to an arguably simpler architecture), the Direct Compositional
theory developed so far has the advantage of avoiding assignments as part of
semantic apparatus for the interpretation of natural language. It might be worth
stressing again that although we are using the º-language and variables as a way to
write out complex functions, this is just for notation. One can—as we have often
done—restate these model-theoretic objects in prose, and there are other logical
systems which could have been used to write these meanings which do not use
variables.1 But in the fragment so far, no expression of English has as its semantic
value an assignment-dependent object, and so assignments play no role.
As in the previous two chapters, this chapter also considers a phenomenon—
pronominal binding—from two points of view. The first makes use of both LF

1
One such system is the Combinatory Logic developed in Curry and Feys (1958) in which
all of the meanings that can be written using the lambda calculus can be written without
variables.
 279
and variables; the second is a variable-free and Direct Compositional approach. But
it is important to point out that the issue of Direct Compositionality is not crucially
tied in with the question of whether or not the grammar makes use of assignment
functions and variables (although the two positions do go rather nicely together).
There are Direct Compositional accounts which do use variables. But space—and
potential overload for the reader—preclude developing such an account here. For
interested readers, a fully Direct Compositional account using variables is available
on the textbook website at <http://sites.clps.brown.edu/compositionalsem>. After
developing the two different accounts (in sections 15.2 and 15.3) the rest of this
chapter returns to some observations of relevance to both theories (sections 15.4
and 15.5).
Before proceeding, one further remark is in order about the overall simplicity of the
two views developed here. Readers might be struck by the fact that the LF view with
variables (section 15.2) takes far fewer pages to exposit than does the variable-free
view in 15.3. This might well give the impression that the former is considerably
simpler. But this is only because this incorporates quite a bit of machinery developed
in earlier chapters which has no analogue in the Direct Compositional view, including
the use of LF, movement (QR) to get that LF (which rule was never explicitly
formulated), assignment functions as developed in Chapter 9, the use of an
interpretive rule akin to lambda abstraction in order to interpret the relevant LFs, and
so forth. Also, no explicit syntax was given in conjunction with the LF view; rather a
kind of informal syntax was just assumed. Hence an overall comparison of the two
theories requires full and explicit fragments for both theories—starting from scratch
and including both a syntax and a semantics. That cannot be accomplished in this
book; here we develop only the Direct Compositional fragment (and even then some
details are left out or are just hinted at). And of course the tools here are used only in
the service of an account of English, so an ultimate reckoning of the types of tools
needed in language in general is quite far off from what is done here. Hopefully,
though, the Direct Compositional fragment developed here can serve as a starting
point for theory comparison.

15.1.2. Pronominal binding: The empirical phenomenon

The empirical domain of immediate interest is the interpretation of pronouns. Imagine

asking someone who knows nothing about semantics: what
 280 15. PRONOUNS: TWO ACCOUNTS

15.1. INTRODUCTORY REMARKS

is the meaning of a pronoun such as he, she, they, etc.? The answer that a
semantically naive person usually gives is that a pronoun picks out some contextually
salient individual in a given context. A scenario like that in (1) makes this a perfectly
plausible conjecture; here he picks up the contextually salient individual Elmo:
(1) We are at a party and our friend Elmo makes his goodbyes and walks out the door,
forgetting to put on his hat. I turn to you and say “uh oh, he forgot his hat.”

But it is easy to show that pronouns don’t simply pick out individuals.
Sentences like those in (2)—whose intended reading is quite clear— immediately
defeat the naive hypothesis.
(2) a. Every womani invited the man who shei had met at the rally.
b. No fourth grade boyi called hisi mother (on Mother’s Day).

What is the individual referred to by she in (1a)? Even more strikingly, what could
possibly be the individual his in (1b)? In the face of (1), the immediate temptation is
to think of pronouns as having as their meanings variables, and this is the line pursued
in the next section. Of course, this alone would accomplish nothing unless coupled
with some notion of a semantics for variables. Otherwise, to say that a pronoun like
he corresponds to a variable merely restates the question (what does a variable
mean?). Indeed, though, we already have a semantics for variables—both for the
artificial languages in Chapter 9 and in the account of traces in Chapters 13 and 14—
so we can take that as the point of departure in section 15.2.
A notational point: recall from the very first chapter of this book that indices are
being used to indicate the intended reading. Whether or not indices are part of the
actual grammatical apparatus is of course part of what is at issue. To keep the
different uses of indices straight, when a pronoun or GQ is indexed with a subscript
such as i and j, this is intended only as a way to indicate a particular reading. When
indices are being used as part of the grammatical apparatus (in the theory that
maintains this), we continue to pick arbitrary integers for the indices. Also, a
terminological point: we continue to use the terminology of a “bound pronoun” and
the “binder of a pronoun” in a purely informal but hopefully clear way. Take for example
(2b). As will be discussed below, neither theory actually posits any direct relationship
between no fourth grade boy and his. We will nonetheless continue to refer to the
former as the “binder” of the pronoun for expository purposes. Moreover, the variable-
free view contains no real notion of “binding” in any sense, and as has been seen at
 281
various earlier points in the text, a definition of “binding” of a pronoun in a theory with
assignments is actually a bit complex (and is not really necessary). Again, then, the
term bound pronoun is just an expository convenience.

15.2. The LF (variable-ful) view

15.2.1. The basic account

Given the apparatus for the interpretation of traces in Chapter 13 and the
mechanisms for binding traces left by QR, the machinery needed to provide a
meaning for sentences like (1) is largely already in place. All that is needed is lexical
entries for pronouns. Thus assume that a pronoun has an index as part of its lexical
entry which plays a key role in its interpretation, and the bound reading of a pronoun
is a consequence of the fact that the NP which binds it can undergo QR. This means
that there are an infinite number of pronouns in the lexicon: he1, he2, etc., and there
must be some kind of schema over lexical meanings to the effect that for any lexical
item hen, and for any g, [[hen]]g = g(n). Incidentally, the fact that a pronoun has an index
is not just an automatic consequence of the more general fact discussed earlier that
(in the LF account) all NPs have an index. For the index on an NP like every third grade
boy is on the entire NP node, not on a single lexical item. Moreover, as shown in
Chapter 14, that index plays no role in the interpretation of the expression every third
grade boy (to highlight this fact, QR was formulated so as to actually remove the index
from that expression). Its sole purpose is to leave that index on the trace, where it is
interpreted. For a pronoun, the index is a key part of its semantics.
Consider (2b). Here his is a genitive pronoun which combines with a relational
noun; one can use R-3 introduced in section 8.5. (R-3 was formulated within a theory
using a Categorial Grammar syntax, which is not usually assumed under the LF view,
but this is orthogonal to the points here.) Thus, his has some index—say, 5. Then the
value of his5 is an assignment-dependent individual; on any g it is g(5). Similarly, the
value of his5 mother is also assignment-dependent; for each g, it is the unique
individual who is the mother of g(5). Now consider the subject NP no fourth
 282 15. PRONOUNS: TWO ACCOUNTS

grade boy. It can also be indexed 5 (if it has a different index, then his 5 remains free;
we return to free pronouns in 15.2.3). Since no fourth grade boy can undergo QR,
one LF for the sentence no fourth grade boy called his mother is the following:
(3)
S

NP Λ
5

no fourth grade boy S

NP VP

t5 V NP

called his 5 mother

This gives the bound reading on the pronoun, and the compositional details are
straightforward. As discussed above, the NP his 5 mother on any g denotes the
mother of g(5). The value of called his 5 mother on any g is the characteristic
function of the set of individuals who called the mother of g(5). The value of the S
[t5 called his5 mother] on any g is the proposition that g(5) called the mother of g(5).
The next node up—5—has a value which is constant on all assignments. It is (on
all assignments) the characteristic function of the set of individuals who call their
own mother (see the rule in (9) in Chapter 13). And that function is taken as
argument (on any g) of [[no fourth grade boy]]; the latter (on any g) is the set of
individuals with an empty intersection with the fourth grade boy set, and we arrive
at the correct truth conditions.
Thus given the apparatus of QR, the notion of indexed traces, and the use of
assignments for the material already covered, pronominal binding becomes
automatic, provided that pronouns come with indices in the lexicon and have the
same semantics as traces. Another consequence of this system is that it extends
immediately to the case of multiple pronouns and multiple binders, as will be
brought out through the next exercise.

15.1. In order to show the point above, we need to deal with an embedded
sentence. As has been noted multiple times, the full semantics of sentences
embedded under verbs like think requires giving an intensional semantics (the
object of someone’s thoughts is not a truth value but

(cont.)
 15.2. THE LF (VARIABLE-FUL) VIEW 283

something about the way the world might be; technically then it is a set of
worlds). But for simplicity, simply extensionalize here, and take think and hope
to both have meanings of type <t,<e,t>>. Then show the LF and the semantic
computation for each of the following sentences:

(i) Every fourth grade girli hopes that no third grade boyj thinks that heri dog
chased hisj cat.
(ii) Every fourth grade girli hopes that the boyj who shei invited will bring hisi dog
(to the party).

In (ii), omit will and to the party from the computation; these are put in to make
the sentence more natural but need not be dealt with here. Further, treat third
grade boy and fourth grade girl as single unanalyzed nouns. Finally, for her dog
and his cat we have no account of these in the fragment (recall that R-3 in
Chapter 8 dealt with genitives for relational nouns, but not for this kind of
genitive). So simply take something like Dorothy’s dog to be an individual which
we will just write out as [[thedog-of(d)]], keeping in mind that this is a
shorthand, and take it as being of type e. Using this as the model for writing out
her dog etc. should be straightforward.
15.2. Now consider a sentence like (i), again on the reading indicated by the
indexing:

(i) Every third grade boyi said that hei (already) fed hisi dog.

Assume that both pronouns here (and every third grade boy) are indexed
5 in the syntax. Then show that there are two different LFs which give the
meaning above; keep in mind that all NPs—including pronouns—can optionally
undergo QR. There are additional LFs where his dog raises, but this is not at
issue here; neither of the LFs of relevance for this exercise involves QR on his
dog.

As noted in section 15.1, one will often hear it said of an example like (2b) that
no fourth grade boy binds his. And here too that terminology will be used where
convenient as a way to indicate a certain reading. But this terminology is
potentially misleading, for under the account above (as well as in the variable-free
account in 15.3) there is no direct semantic relationship between the pronoun and
the subject GQ. Rather, the two are coindexed in the syntax and it is this fact which
 284 15. PRONOUNS: TWO ACCOUNTS

allows the trace left by QR to be coindexed with the pronoun. This is the only
sense in which there is a relationship between no fourth grade boy and his. In fact,
it was pointed out in Chapter 9 that the notion of the binding of a pronoun or a trace
with index n is most naturally defined as the point in the computation where an
expression shifts from one whose value varies on the assignments g in G/n to one
whose value is constant on all g in G/n. Hence here the binding of a pronoun he n
happens at the stage where the node labeled n is interpreted. Again, then, binding
has nothing to do with the GQ itself. Of course none of this really makes any
difference as it is mere terminology. The way the grammar has been set up here—
even in the view above making use of LF and indices—the grammar itself never
refers to a notion of binding.2

2
A common slogan in much of the literature that endorses something like the account
of pronouns developed just above is that “binding requires c-command.” (For a definition
of c-command, see n. 4, this chapter.) This is used in two different ways, depending on
whether one is talking about c-command at LF or c-command at some syntactic level
(such as the surface representation). The most common use of this slogan refers to LF, and
in this case it is really nothing more than an observation rather than any kind of principle
that must be stated in the grammar. That is, it means that the relevant reading in sentences
like those in (1) emerges as a result of QR, and consequently the “binder” (the NP which
undergoes QR) c-commands the pronoun at LF. But this is just a fact about the way the
semantics has been set up and not an additional principle in the grammar. The second use
of this slogan refers to a version of the Weak Crossover Constraint which is discussed
directly below in section 15.2.2. Here the idea is that there is an actual constraint within
the grammar. Again, though, it need not refer to any notion of “binding”; a version of this
constraint is given in (8) which does refer to c-command but not to a binder, and is a
constraint on the relationship between the trace left by QR and the pronoun.
Note also that the syntactic literature—especially that beginning with Chomsky
(1981)—is full of discussion of constraints on “binding” under the rubric of so-called
“Binding Theory.” Under this view, one might think that the grammar does access such a
notion. But the notion of “binding” at issue in these discussions has little to do with the use
of that term here. Rather the so-called “Binding Theory” refers to constraints on the
interpretation and distribution of reflexive and reciprocal pronouns (“Principle A”), and on
some principles constraining the interpretation of ordinary pronouns including principles
that would need to block certain understandings of free pronouns (“Principle B” and
“Principle C”). This book has not and will not deal with these phenomena, save to say that
there are Direct Compositional and variable-free accounts of the relevant phenomena.
See, e.g., Szabolcsi (1987) and
 15.2. THE LF (VARIABLE-FUL) VIEW 285

15.2.2. A wrinkle: Weak Crossover effects

Thus the QR approach generalizes quite readily to handle the semantics of bound
pronouns. But the ease with which it generalizes is a mixed blessing. For the ability
of a GQ to take scope over other material does not in fact correlate completely
with the distribution of binding. The consequence of this fact is that the approach
sketched so far needs to be supplemented with an additional principle relevant to
pronouns so as to rein it in. Otherwise, it is predicted that certain sentences have
meanings that they do not in fact have.
The phenomenon in question centers on the fact that although a GQ in object
position can generally scope over a subject, it cannot bind a pronoun in subject
position. Consider the following contrasts:
(4) a. No fourth grade boyi called hisi mother (on Mother’s Day).
b. *Hisi mother called no fourth grade boyi (on the first day of school).

(5) a. Every patienti was discharged by some doctor who had seen him i.
b. *Some doctor who had seen himi has discharged every patienti.

Notice that in (5b), for example, the problem is not one of scopes. We have already
seen that in a corresponding sentence like (6) it is possible to get a wide-scope
object reading as, for example, in the scenario where a clinic may close for the
night only after every patient has been discharged (and no single doctor needs to
sign off on all the discharges):

(6) We can finally close, since some doctor has discharged every patient.

The same scenario does not help for (5), verifying that the problem is located in
the binding of the pronoun. Nor is there anything odd about the intended readings
in (4b) and (5b); we can bring out the intended reading for (4b) by using a passive:

(7) No fourth grade boyi was called by hisi mother (on the first day of school).

This lack of the bound readings here is generally known as the Weak Crossover
effect (henceforth WCO). The term Crossover was coined originally

Steedman (1996) for two different accounts of reflexives both compatible with the general
architecture here, and see Jacobson (2007) for an account of so-called Principle B effects
within the framework here. None of these accounts require the grammar to make use of a
notion of “binding” (nor of indices).
 286 15. PRONOUNS: TWO ACCOUNTS

by Postal (1971) who discovered a similar asymmetry with respect to relative

clauses (and questions). Thus Postal noted the following contrasts (again we use
indices on the relative pronoun who, just to indicate the intended reading):

(8) a. No woman whoi had invited heri father (to her graduation) received an extra ticket.
b. *No woman whoi heri father had invited (to his retirement party) received a ticket.

(The use of no in the determiner here is not really central to the point, but is helpful
since substituting in the improves these, for reasons that will not be pursued here.)
Postal assumed a rule fronting the relative pronoun (as in the view in section 13.3);
he adopted the term Crossover to refer to a constraint to the effect that a wh-word
could not cross over a coindexed pronoun in its movement path. Since his original
work, it is generally agreed that whatever is responsible for the effect in (8) is also
responsible for the effect in (4) and (5).3 The primary focus here will be on the case
with GQs in object position, but we will show that the two accounts of WCO
effects discussed in this book do generalize to the relative clause case.
There are literally dozens of accounts of the WCO effect on the market, and

obviously space precludes comparing and contrasting them. The important point

is that something extra is needed to block the intended readings in (4) and (5); the

simple account using QR predicts that these readings should exist. One fairly

standard solution assumes that in addition to the apparatus developed so far in

the non-Direct Compositional theory (indices, traces, movement in the syntax,

and covert movement mapping surface representations into LFs) there are also

constraints on well-formed LFs. In other words, the rules that apply to yield LFs

(such as QR) can 3

The addition of the term “Weak” here is from Wasow (1972). The rationale for this is
that the effect is not entirely robust and can be ameliorated in a variety of ways. There is a
special case of the phenomenon known as “Strong Crossover” which is much more robust
and cannot be ameliorated, as in (ia) (the quantificational case) and (ib) (the relative-clause
case), where here the pronoun is not embedded in other material:
(i) a. *Hei praised no mani. (cannot mean: No man praised himself)
b. *No mani whoi he praised will win the modesty prize. (cannot mean: No man who
praised himself will win the modesty prize)
apply freely but there is an additional set of rules specifying that certain LFs are
ill-formed; call these LF filters. WCO effects can thus be accounted for if the
grammar contains the following LF filter:
 15.2. THE LF (VARIABLE-FUL) VIEW 287

(9) An LF which contains a trace which is coindexed with a pronoun is ill-formed unless the
trace c-commands the pronoun. 20

Incidentally, for the cases at hand it would have worked just as well to require that
the trace precede the pronoun (rather than c-command it).5 We use the c-
command version both because it is more common, and because it also lends
itself more easily to comparison with the account to be developed in section 15.3.
As the reader can easily verify, the LF needed to give the relevant (non-existent)
reading for (4b) and (5b) violates (9) since the GQ raises from object position and
hence its trace does not c-command the pronoun embedded in the subject. The
LFs for (5a) and (5b), on the other hand, are admissible. Notice that this
generalizes immediately to the relative-clause contrasts in (8); this is the subject
of Exercise 15.3.

15.3. Show the LFs for (4a) and (4b) to demonstrate that the remarks above are
correct.
15.4. Show the LFs for just the relative-clause portions of (8a) and (8b) to
demonstrate that the remarks above are correct. You can ignore the material in
parentheses and the auxiliary had.

15.2.3. Free pronouns

Recall the discourse in (1) above, repeated here:

(1) We are at a party and our friend Elmo makes his goodbyes and walks out the door,
forgetting to put on his hat. I turn to you and say “uh oh, he forgot his hat.”

Here the pronoun he (and also his) has nothing binding it, nor need it even have
any mention in the linguistic context of the name of the referent (here, Elmo). This

20 C-command is a notion defined in terms of trees. A node A c-commands a node B if

the first node above A (the “mother” of A) dominates B. Intuitively, then, if A c-commands
B this means that A is at least as high in the tree as B. Put in terms of things combining to
give well-formed expressions, the c-commanding material is introduced later (or at the
same time as) the c-commandee.
5 Most early accounts of this were actually based on linear order (see, e.g., Postal 1971;
Jacobson 1977) while many later accounts were based on c-command (see especially
Reinhart 1983). But there continue to be accounts based on linear order. See Barss and
Lasnik (1986), and an especially interesting account is found in Shan and Barker (2006).
 288 15. PRONOUNS: TWO ACCOUNTS

point can be made even more dramatically with the following kind of contexts
(based on observations of Lasnik 1976).

(10) We are at a party and a particularly obnoxious man who no one likes walks in the
door. Because everyone so dislikes this person, no one speaks to him all evening,
nor even utters his name. Eventually he takes the hint and leaves, and after he finally
walks out the door I turn to you and say “Thank goodness he left.”

(11) Exactly the same context as above, but when I turn to you I say “Boy,
everyone (sure) dislikes his style.”

How are we to understand the semantics of he/his in these cases? Clearly they
are not instances of bound pronouns. In early work in transformational grammar
(Postal 1966; Ross 1969; Langacker 1969) it was assumed that all pronouns are
represented as fuller NPs at the level that inputs the semantics and are reduced
to pronouns under linguistic identity with something else in the discourse context 6
(compare this to the remarks about VP Ellipsis in section 14.4.2). But since the
above contexts were deliberately constructed to ensure that there could be no
identical linguistic material in the discourse context, that cannot be the way to
think about the semantics of these pronouns either.
Notice, though, that under the theory developed in this section, nothing forces
the binding of a pronoun. Take (10). Suppose this is just he 5. Under the LF view,
there is no QR here, nothing happens, and the LF for this is just essentially [he 5
left]. Similarly for (11). In this case the subject everyone can indeed undergo QR,
but in the reading of interest it is not coindexed

6
Variants of this view have been resurrected in modern work as well; see especially
Elbourne (2005) who argues that all pronouns do contain hidden fuller linguistic material.
with his and so his remains free. Thus the final sentence has as its value an
assignment-dependent proposition rather than an ordinary proposition. But
presumably in ordinary discourse contexts listeners are looking for propositional
information. It is propositions that are used to convey facts about this world (they
denote sets of worlds, and in main clauses are taken to assert that this world is in
the relevant set). It is also propositional information that is used for inferences,
and so forth. Ultimately, then, the interpretation of a sentence with an unbound
pronoun is such that the listener applies the assignment-dependent proposition
to some actual assignment in order to extract propositional information. So in the
case of (10) (where the pronoun is indexed as he5), the context makes salient or
available an assignment in which 5 is assigned to Elmo. We will not pursue here
 15.2. THE LF (VARIABLE-FUL) VIEW 289

just how that happens. We will, moreover, often continue to speak informally of
the “pronoun picking up a contextually salient individual” even though technically
speaking what is picked up is an assignment. Notice, by the way, that all
sentences have as their values functions from G to propositions. However, in the
case of a sentence with an unbound pronoun there is no mystery as to how the
listener arrives at propositional information: the value of a sentence is a constant
function from assignments, so it does not matter which assignment is chosen.
Technically, though, an assignment must be supplied. We will have more to say
about so-called “free variables” in section 15.5.

15.3. A Direct Compositional and variable-free approach

As noted in the introductory remarks to this chapter, both quantification in object

position (and scope ambiguities) and relative clauses can be handled smoothly
under Direct Compositionality, and can be handled without use of indices,
assignments, and/or variables as part of the syntactic or semantic machinery. We
thus turn our attention now to an approach to the semantics of pronouns which
similarly requires no LF (and hence of course no QR) and also makes no use of
indices, assignments, or variables.
This approach has generally gone under the rubric of variable-free semantics,
and it has its roots in the variable-free logic of Curry and Feys (1958) (known as
Combinatory Logic), although that was not intended as a theory of natural
language semantics. The idea that natural language makes no use of variables was
perhaps first put forth (albeit in programmatic form) in
 290 15. PRONOUNS: TWO ACCOUNTS

Quine (1960). It has subsequently been explored by a number of researchers

particularly in Categorial Grammar and related traditions (Szabolcsi 1987, 1992,
Hepple 1990, Dowty 1992, Jäger 1996, Shan and Barker 2006, and a number of
papers of the present author, including Jacobson 1991, 1999, 2000). The
discussion here will focus on the particular implementation in especially
Jacobson (1999) (with some modifications). The proposals by the other authors
cited above include interesting alternative ways to implement the basic idea, but
obviously a comparison is beyond the present scope.

15.3.1. Pronouns and “binding”

Consider the underlined portions of the sentences in (12) under the reading
indicated by the indices there:
(12) a. Every third grade boy loves Jim’s mother.
b. Every third grade boyi loves hisi mother.

(13) a. Every male candidate thinks (that) Rick lost.

b. Every male candidatei thinks (that) hei lost.

In the LF view exposited above, the object NPs in both (12a) and (12b) are of the
same type: both have as values functions from the set of assignments (G) to
individuals. The object NP in (12a) is a constant function while the object in (12b)
is not. Moreover, in (12a) his has an actual index in the grammar. (13) is parallel; in
both cases the underlined portion is a function from G to propositions. By
contrast, the variable-free view takes Jim’s mother to denote an individual but his
mother does not—its value is a function from individuals to individuals. In general,
an expression with an unbound pronoun within it is a function from individuals to
whatever is the type of a parallel expression with no unbound pronoun within it.
Once again the term unbound (or free) is being used here in an informal way for
exposition, but its use is hopefully clear. If there are two unbound pronouns within
an expression then it is a function of type <e,<e,X>> (for some X).
To work out the details, three questions need answering. First, what is the
meaning (and syntactic category) of a pronoun itself? Second, consider a
sentence like (14):

(14) Every mani thinks that hisi mother lost.

 15.3. A VARIABLE-FREE APPROACH 291

The lexical meaning of lost is of type <e,e>. Since it was said above that his mother
denotes a function of type <e,e>, how can the two combine? (For that matter, how
it is that his mother has the meaning that it does?) And third, how do we ultimately
get the effect of what is informally being called binding? That is, how do we get the
intended reading in, e.g., (12b)?
To begin with the pronoun itself, it was hypothesized that any expression
containing a pronoun unbound within has as its value a function from individuals
to something else. Extending that to pronouns themselves, let a pronoun such as
he also have as value a function from individuals to individuals—hence of type
<e,e>. Since there are no indices, there is a single pronoun he in the lexicon, and
we assume that it denotes the identity function on individuals. (We ignore the
contribution of gender, and treat she and he alike, although eventually gender
information can be folded in. Moreover, strictly speaking, he stands here for a
lexeme without case marking; case-marking rules will more fully specify the
syntactic category and map the abstract lexeme into the various case-marked
versions.)
This also needs to be married to a syntax. Given the tight correspondence in
Categorial Grammar between the syntactic category and semantic type, he
cannot be an NP since NPs are of type e. Assume, then, that pronouns are NP-like,
but have a special feature (which will be notated as a superscript).
Hence we give a second recursive definition of categories as follows: If A is a

category and B is a category, then AB is a category. The semantic type of AB is a

function from B-type meanings to A-type meanings. Semantically, then, A/B and
AB are of the same type, but the latter does not actively want to combine in the

syntax with a B. Rather the superscript feature (in general) records that there is an

unbound pronoun within it, and records its semantic type. Hence a pronoun is
listed in the lexicon with category NPNP with meaning ºxe[x]. And, more generally,

any expression that has an unbound pronoun within it has a superscript in its

category.7 The category of his mother is also NPNP while the category of the

underlined material in (13b) is SNP. 7

However the reverse does not hold: not everything with a superscript has an actual
unbound pronoun within it. See 15.5 for additional uses of the superscript feature, and see
also Chapter 17. The analogue to this in variable-ful theories is that some expressions
contain unbound variables within them as part of their meaning even though there is no
 292 15. PRONOUNS: TWO ACCOUNTS

overt pro-form within these expressions. See, e.g., Mitchell (1986), von Fintel (1994),
Stanley and Szabo (2000), and others.
Now consider the semantic composition of the underlined expression in
(13b), where he combines with lost. Since lost is S/NP and [[lost]] is of type <e,t>,
it cannot directly combine with he which is of category NPNP and of type <e,e>. But
a simple unary rule—already familiar from R-5 (g-sl) in Chapter 13—will
accomplish the task. We need only a slightly different version of R-5; one whose
semantics is the same, but whose syntax has the effect of passing the superscript
feature from an argument to the result category. The fragment in the Appendix
actually suggests collapsing the two g rules, but for now we keep them distinct.
The new one is formulated in R-7 and will be referred to as g-sup:
R-7. (g-sup): Let Æ be an expression of the form <[Æ]; A/B; [[Æ]]>. Then there is an
expression of the form <[Æ]; AC/BC; gc[[Æ]]> (for c the type of expressions of
category C).

With this, he lost is put together as follows:

(15) lost, S/NP, lost’ !g-sup

NP NP
lost, S /NP , ºf<e,e>[ºx[lost’(f(x)]] he, NPNP,
ºy[y]
he lost, SNP, ºf<e,e>[ºx[lost’(f(x))]](ºy[y]) = [ºx[lost’(ºy[y](x))]] = ºx[lost’(x)] = lost’

Note then that he lost means the same as lost (modulo the contribution of
gender), but has a different category. As expected, its category is S NP.
It is worth pausing to observe an interesting similarity between the LF account
with variables developed in section 15.2 and the account here, which is that both
posit a commonality between “gaps” in constructions like relative clauses and
pronouns. The LF account treats one as a trace while the other is of course a
lexical item, but both have indices and the two have the same semantics. The
account here has nothing analogous to a trace, but treats the semantics of an
expression containing a gap and an expression containing an unbound pronoun
within it as a function of type <e,X> for some X. The argument slot is passed up in
both cases by a g rule which has two different versions in terms of the syntax—
one passing up a slash and one a superscript—but both have the same semantics.

15.5. Take his as listed in the lexicon as NP[GEN] NP[GEN] with meaning ºxe[x].
Show the syntactic and semantic composition of his mother lost.
 15.3. A VARIABLE-FREE APPROACH 293

The next step is to introduce machinery to get the effect of what we have
informally been calling binding. In other words, how is it that (12b) and (13b) have
the readings indicated by the indices there? This is accomplished by an additional
unary rule. Before giving its formulation, a comment is in order lest it appear that
unary rules are being added willy-nilly. For recall that the LF account also has a
unary rule (or something analogous) to effect binding: this is the interpretation rule
(9) in Chapter 13. This rule maps an open proposition (one which has different
values on G/i for some i) to one which is the same proposition on all members of
G/i. Its semantics is analogous to the semantics in Chapter 9 of lambda
abstraction. The rule to be introduced in the variable-free framework is different
in that it does not, of course, use assignments and variables. Another interesting
difference is that it is more local—a fact which will be quite advantageous in some
of the material to be considered later. But both accounts make use of a unary rule
(or something analogous) for binding.
Hence we first define an operation on functions which called z as follows (to be
generalized later):

(16) Let f be a function of type <b,<e,a>>. Then z(f) is of type <<e,b>, <e,a>>, and
is ºY<e,a>[ºxe[f(Y(x))(x)]].

The informal intuition is that z maps f to a new function whose first argument will
not be something in set a but rather will be a function from individuals to members
of a. Its next argument will be an individual, and it “merges together” the two
individual slots. Incidentally, there is no real reason to restrict z to functions of
type <a,<e,b>> rather than functions of type <a,<c,b>> (for any c); this is done
here for expository ease. We can then define the full unary rule at work as follows;
we refer to this rule as z:
R-8. Let Æ be an expression of the form <[Æ], (A/NP)/B, [[Æ]]>. Then there is an expression
of the form <[Æ], (A/NP)/BNP, z([[Æ ]])>.

By way of illustration, consider (12b). As already noted, [[his mother]] is the

function mapping each individual to his mother; call this the-mother-of function.
Now take love. Its lexical type is <e,<e,t>>, and so cannot combine directly with
his mother which is of type <e,e>. But one of two things can happen. One is that
love maps to g(love), whose meaning is of type <<e,e>,<e,<e,t>>>; we return to
that derivation below. Here we are interested in the bound reading; this results
from the mapping to z(love), whose meaning is of type <<e,e>,<e,t>>. It is easiest
 294 15. PRONOUNS: TWO ACCOUNTS

to get the intuition of the semantics of z informally. To z-love some function f (of
type <e,e>) is to be an x who ordinary-loves the value that f assigns to x. So in this
case, z-love combines with the-mother-of function, and to be an x who z-loves
themother-of function is to be an x who loves the-mother-of x. In function terms,
the meaning of the VP is ºx[love’(the-mother-of ’(x))(x)]; in set terms it is the set
of self ’s mother-lovers. And the full meaning for (11b) is derived by taking this as
argument of the GQ subject. The whole S, then, is true just in case the set of self
’s mother callers has the third grade boy set as a subset.
It is again worth comparing this to the LF theory with variables. In that theory
every third grade boy undergoes QR, so that the LF is (17):

(17) every third grade boy [8 [S t8 loves his8 mother]]

The value of the embedded S is—on any g—the proposition that g(8) loves the
mother of g(8). By the rule in (9), the value of the node labeled 8 is, on any g, the
set of self ’s mother lovers—exactly the same (modulo the use of the
assignments) as the meaning of the VP loves his mother in the variablefree
account. Thus the step that accomplishes binding happens at the S-level, where
there is a trace in subject position coindexed with the pronoun. In the variable-free
theory it happens through the mapping of love’ to z-love’. But the rest of the
composition is essentially the same. In the LF theory, [[ 8]]g is taken as argument of
[[every third grade boy]]g to give the final meaning. In the variable-free account, [[z-
loves his mother]] is taken as argument of [[every third grade boy]].
This discussion centered on the semantics only, but z has a syntactic side too.
Since his mother is not an ordinary NP but is of category NP NP it would also be
syntactically of the wrong category to be argument of ordinary loves whose
category is (S/LNP)/RNP. But z maps this to an expression whose category is
(S/LNP)/RNPNP—i.e., to something wanting to combine with a pronoun-containing
expression.
We further illustrate the z rule with (13b). Recall from (15) that he lost is of
category SNP with meaning [[lost]] (of type <e,t>). Now consider thinks. As usual,
we ignore the fact that thinks’ takes an intensional argument (i.e., it is of type
<<s,t>,<e,t>>) and simplify its lexical type as <t<e,t>>.
Then z-think is of type <<e,t>,<e,t>>, and to z-think some P of type <e,t> is to be
an x who thinks that P holds of x. So (13b) is put together as follows:
 15.3. A VARIABLE-FREE APPROACH 295

(18) thinks, (S/LNP)/RS, thinks’ !z thinks, (S/LNP)/RSNP, ºP<e,t>[ºx[thinks’(P(x))(x)]] he lost,

SNP, lost’ (from (3) above) thinks he lost, S/LNP, ºP<e,t>[ºx[thinks’(P(x))(x)]](lost’) =
ºx[thinks’(lost’(x))(x)]
this VP is then taken as argument of the subject GQ every male candidate

15.6. Show the full syntactic and semantic composition of (14). The first part is
already done in your answer to (15.5).
*15.7. By way of comparison, work through the derivation of (14) using the LF
account, seeing step by step where this differs from the variablefree account.
Try to see just where this is similar to the variable-free account and where it is
different.

Now consider a sentence like (19):

(19) Every first grade boyi thinks that Santa loves hisi mother.

Here the pronoun is not bound within the lower clause but only within the thinks
clause. But recall that the lexical verb loves can also map to g-loves (rather than
z-loves). Informally, this keeps the pronoun “free” and available for binding from
higher up. In this case, the lower VP is put together as follows:
(20) loves, (S/LNP)/RNP, loves’ ! g

loves, (S/LNP)NP/RNPNP, ºf<e,e>[ºx[loves’(f(x)]] his mother, NPNP, the-mother-

of loves his mother, (S/LNP)NP, ºf<e,e>[ºx[loves’(f(x)]](the-mother-of ) =
ºx[loves’(the-mother-of ’(x))]

The meaning of g-loves his mother, then, is of type <e<e,t>>.

A worry might arise: when this combines with the subject Santa (which denotes
the individual s), why is that not taken as argument of this function? If it were, the
sentence would be saying something about someone loving the mother of Santa,
obviously an unfortunate result. But in fact a derivation in which the above function
applies to the individual denoted by the subject is not countenanced by the
syntax. The lexical entry for Santa is an NP, and nothing allows this to directly
combine with an (S/LNP)NP. The derivation of (19) involves g on the GQ (type-lifted)
 296 15. PRONOUNS: TWO ACCOUNTS

version of Santa, as shown below; the remainder of the derivation of (19) is left for
the reader in Exercise 15.8.

(21) Santa, NP, s ! l

Santa, l(NP) (hence, S/R(S/LNP)), ºP<e,t>[P(s)] !g-sup

Santa, SNP/R(S/LNP)NP, ºR<e,<e,t>> [ºy[ºP<e,t>[P(s)](R(y))]] = ºR<e,<e,t>> [ºy[R(y)(s)]]

Incidentally, Exercise 15.12 will suggest a way to generalize the app rule (R-1) so
that the ordinary NP Santa can directly combine with loves his mother in
(19) (without use of l or g-sup). This gives s a much shorter derivation resulting in
the same string and meaning. Should there be some advantage to adopting a
theory allowing for shorter derivations, this option is quite possible, and an
interested reader is invited to explore this further on their own.

15.8. Show the rest of the syntactic and semantic computation for (19). (For
the treatment of that see the fragment at the end of Chapter 6.)

15.3.2. Free pronouns

Nothing in the system forces the application of z at some point in the derivation of
a sentence with a pronoun: there could be just applications of g “all the way up.”
Suppose, for example, a sentence like Santa loves his mother were derived using
the steps shown above (thus loves undergoes g rather than z), and where this is
not embedded further and is just a main clause. For that matter, a sentence like
Every boy thinks that Santa loves his mother also has a reading in which his is not
bound by either every boy or by Santa and is just a free pronoun. In that case, then,
we end up with the final “sentence” being of category SNP and with meaning of type
<e,t>.

15.9. Show the derivation of every boy thinks (that) Santa loves his mother
where his remains free.

Recall that a similar situation holds in the LF theory; there nothing forces
binding either. In the LF theory, a sentence like Santa loves his mother where his
is not bound denotes an assignment-dependent proposition, and its ultimate
 15.3. A VARIABLE-FREE APPROACH 297

propositional value is supplied not in the semantics proper. Rather, the listener
presumably applies this to some contextually salient assignment. The situation
here is a bit different, but analogous. In the variable-free account, the “sentence”
is ultimately of category SNP and not S. Its value is not a proposition but a function
from individuals to propositions. Here, then, the propositional information is
supplied by the listener applying this to some contextually salient individual
(rather than a salient assignment). Lest it seem odd to conclude that Santa loves
his mother (on the free reading) or a simple sentence like He left do not denote
propositions but rather functions into propositions, we need merely note that the
full value of such expressions is not a proposition in the variable-ful view either.
These are assignment-dependent propositions, which is equivalent to saying that
they are functions from assignments to propositions. Thus there is no proposition
unless some contextually salient assignment is picked. (In fact, arguably the
notion of an individual being contextually salient is somewhat less mysterious
than the notion of an assignment being salient.)
One might also wonder in the variable-free view why ordinary VPs like *Left are
not well-formed sentences in their own right in English. After all, this has the same
meaning (modulo the gender contribution of the pronoun) as He left. The reason
is purely one of syntax: S/NP and SNP are not the same syntactic category. The
latter but not the former can stand alone (and any theory needs some account of
what can and cannot be free-standing expressions). An obvious speculation is
that this might be the right analysis for socalled pro-drop languages; there
perhaps tensed VPs are allowed to stand alone, and have <e,t> meanings as their
value, just as he left does in English.8

*15.3.3. Generalizing the rules

This section generalizes some of the rules in the fragment. One reason for these
generalizations centers on cases with multiple pronouns to be discussed below.
Another is that in general the combinatorics need to be able to ignore
superscripts. The generalizations required are straightforward, and are easily
accomplished by defining (at least some of) the rules recursively, a move which
may well have other benefits throughout the system.

8
Note that wondering why it is that some languages allow tensed VPs as standalone
expressions and some do not is no different than wondering why some languages are pro-
drop and some are not.
 298 15. PRONOUNS: TWO ACCOUNTS

Consider (22):

(22) Every third grade boyi saw the man who hisi mother kissed (under the mistletoe).

The initial focus is on how it is that the relative clause who his mother kissed can
occur as a postnominal modifier. Recall that the fragment as developed so far
allows expressions like who Ms Claus kissed to occur as nominal modifiers via R-
2; this rule maps an RC to an N/N. Although that treatment might ultimately be
revised, that rule nonetheless serves well to illustrate the general point. The
details of the internal composition of who his mother kissed are postponed until
Exercise 15.11, but assume that it is (as we would expect) an expression of
category RCNP. The point is that R-2 is formulated to take as input RC and not RCNP
and so needs to be modified. Moreover, the output should not be N/N but rather
(N/N)NP. And, while we have so far dealt with cases containing only one pronoun,
any expression can of course contain any number of pronouns, such as the
relative clause in (23):

(23) the table which her mother built for his father

Here her mother built for her father is presumably of category (RC NP)NP and maps
to ((N/N)NP)NP.
This presents no serious problem. We simply generalize R-2 by defining the
syntactic and semantic mappings recursively. First, we define the base step,
which is the operation that shifts a function of type <x,t> to one of type
<<x,t>,<x,t>> with the associated semantics of intersection. Call this int0 and
define int recursively as follows (note the connection to the recursive definition of
[[and]] given earlier):
(24) a. Let f be a function of type <x,t>. Then int 0(f ) is of type <<x,t>,<x,t>> and
is ºg<x,t>[fS\gS].

b. Let f be a function of type <a,b> such that int i is defined for b (for some i). Then
intn(f ) = ºAa[intn-1(f(A))].

The same can be done with the syntactic side. Recall that some of the syntactic
operations on categories (such as l and g-sl) are related to the corresponding
semantic operations; hence the use of the same names. For R-2, however, this is
not the case; the syntactic mapping of RC (or AP) to N/N has nothing to do with
the intersection of strings, so we adopt a new name for the syntactic mapping—
call it NMOD. Thus we define this as follows:
 15.3. A VARIABLE-FREE APPROACH 299

(25) a. Let C be the category RC or AP (and perhaps some others). Then NMOD 0(C) is
N/N. If C consists of just one word, then NMOD0(C) is
N/RN.
b. Let C be any category XY such that NMODi (for some i) is defined for X. Then
NMODn(C) = (NMODn-1(X))Y.

R-2 can now be generalized as follows:

R-20: Let Æ be an expression of the form <[Æ], C, [[Æ]], for C any category for which NMOD
is defined>. Then there is an expression of the form
<[Æ], NMODn(C), intn([[Æ]])>.

Other unary rules can be generalized in exactly this way. (Since z will have a
different generalization, and since it is not clear that that there is any reason to
generalize z in this way, we will not extend this to z, although it would do no harm
to do so.) To save space, the generalizations will not be given here; the reader can
work through some examples via the exercises below. Incidentally, in the
particular case at hand, the numerical superscript is not necessary, but it is
included for exposition since the analogue is needed for some of the other rules.
This is because in some cases the semantics could work in more than one way if
the generalized rule did not keep track of the recursion. Take for example a
function f of type <a,<b,c>> and consider how a generalized g rule could apply to
this. There are two ways: it could map f to a function of type <<d,a>,<d,<b,c>> or
to a function of type <a,<<d,b>,<d,c>>>. But (for both the slash version of g and
the superscript version), the syntax of the two applications is different, and the
use of the numerical superscript will make sure that the syntax traces the
semantics. Of the two mappings of f noted above, the first is g o(f) and the second
is g1(f).

*15.10. Generalize g-sl (R-5), g-sup (R-7), and l (R-4) in the way shown above.

Armed with the answer to Exercise 15.10, the apparatus is in place to give the
full syntactic and semantic composition of who his mother kissed in (22).
We leave the demonstration of this for Exercise 15.11, but a few hints are in order.
It is helpful to first work again through the composition of who Ms Claus kissed,
and keep in mind that the composition of who his mother kissed should be (more
or less) parallel, just using the generalized versions of the relevant rules. There is,
however, one difference to be aware of. Once you have gotten his mother to be of
 300 15. PRONOUNS: TWO ACCOUNTS

the appropriate type, you should note that it cannot immediately take kissed as
argument, and so a few more steps are needed. Feel free to help yourself to l and
g. But, as will be explored in section 15.12, another possibility (which gives the
same result but in a shorter derivation) is simply to generalize R-1 in a way parallel
to the generalization of the unary rules. Since the exercises below are somewhat
technical, a less ambitious reader can skip them and simply take it on faith that
who his mother kissed can be composed into an expression of category RC NP.

*15.11. Show the syntactic and semantic composition of who his mother
kissed. Then show the composition of the man who his mother kissed. (The
details of the composition of his mother can be skipped as this was done in
Exercise 15.5.)
*15.12. Generalize R-1 in the way suggested above, and show how this
simplifies the derivation of his mother kissed. More generally, comment on just
how it is that superscript features are “passed” (in terms of functions vs
arguments) in the two different ways to do this. Go back to the derivation of
Santa loves his mother (in which his is free) shown in
(20) and (21) and show a shorter derivation for this with this generalization
(making sure that the meaning is the one intended in that discussion).

A slightly different generalization is needed for z, for R-8 gives it only for the
narrow case of binding a pronoun within some expression C by the argument slot
that is the very next one after C is introduced. But in fact, a pronoun can be bound
by any later argument slot. This is most obvious in the case of three-place verbs.
If the analysis of these in Chapter 6 is correct, the merging of the subject position
with the pronoun within the PP skips over the object position:

(26) No fourth grade boyi sent flowers to hisi mother.

Send is of category ((S/NP)/NP)/PP and it should map by an extended version of z

to ((S/NP)/NP)/PPNP. (See the discussion in 15.3.4 for one reason for adopting this
category.) Similarly, [[send]] is a function of type <e,<e,<e,t>>> and should map to
one of type <<e,e>,<e,<e,t>>> with meaning ºf[ºy[ºx[send’(x)(y)(f(x))]]]. It is
obvious that we need a generalization that allows binding across argument
positions. Incidentally, the need for a generalization of z which skips over
argument positions is independent of this particular analysis of send. The other
 15.3. A VARIABLE-FREE APPROACH 301

possibility is that it combines first with its adjacent NP and then with the PP and
so is of category
((S/NP)/PP)/NP. But then there would still be need for the generalized version of z
for sentences like No fourth grade boyi showed hisi picture to
Nora. Moreover, argument positions given as superscripts also need to be skipped
over; examples illustrating this point will be given later.
The rule should thus be generalized to bind across any argument slot. In the
syntax, the generalization is that it can skip over both ordinary slashes and
superscripts. To notate this, let A|B mean either A/B or AB. Then, borrowing the
notational conventions used in Chapter 14, we state the generalized z rule. (In this
case it is again not clear that there is any need for the grammar to keep track the
various versions of the rule, but to be safe we will distinguish them by using the
name of the result category.) Hence, we first generalize the semantics, and then
state the generalized rule:
Definition: Let f be a function of type <b,<c !n,<e,a>>> (where c!n means n number of

argument positions of any time, for n 0). Then za(f) =

ºY<e,b>[ºZ!n[ºxe[f(Y(x))(Z!n)(x)]]].

R-8 Generalized: Let Æ be an expression of the form <[Æ], ((A/NP)|C !n)/B, [[Æ]]>.
Then there is an expression of the form <[Æ], ((A/NP)|C !n)/BNP, za([[Æ]])>.

We will generally omit the superscript except where needed for clarity. While the
notation looks clumsy, this is arguably an artifact of the notational system;
conceptually the rule is straightforward. The semantics merely takes a function
that wants some argument in b, then any number of other arguments, and then an
individual argument, and maps it to one that takes an argument in <e,b>, then the
same other arguments and then an individual, and “merges” the last individual slot
with the individualargument slot in <e,b> argument. The syntax is similar; it
basically can skip over both slash and superscripts.

15.3.4. Weak Crossover

Recall the discussion in section 15.2.2 of Weak Crossover effects as exemplified

in (4):

(4) a. No fourth grade boyi called hisi mother (on Mother’s Day).
b. *Hisi mother called no fourth grade boyi (on the first day of school).
 302 15. PRONOUNS: TWO ACCOUNTS

The LF view there adopted the LF filter (9), which rules out any LF in which a
pronoun and trace are coindexed unless the latter c-commands the former. (We
should caution, though, that while this is a common strategy for accounting for
these effects, there are also many other accounts. Still the use of some sort of
filter is quite common.)
In the variable-free view developed here, the asymmetry between (4a) and (4b)
is built directly into the formulation of z: it maps a function of type <b<e,a>> to one
of type <<e,b>,<e,a>> (or—for the generalized version given in 15.4.3—it maps a
function of type <b,<c!n,<e,a>>> to one of type <<e,b>,<c!n,<e,a>>>) and it merges
the “later in” e-argument slot with the slot occupied by the pronoun. In other
words, in (4a) it maps call— which expects an ordinary e-type object—to
something wanting an <e,e> object, and it binds the newly introduced e-position
to its subject slot. It is automatic, then, that the effect of binding works in the
direction that it does. And so one nice consequence of the variable-free approach
is that nothing extra is needed to give Weak Crossover effects. The rules do not in
any sense require an additional filter to rule out certain representations or
derivations; the effect follows directly from the way the binding rule (z) works.
Of course the system developed here also allows subjects to first combine with
transitive verbs (see, for example, the discussion in section 13.4.1). Does this
interfere with the generalization above about the binding pattern? The answer is
no: the asymmetry is unaffected by the fact that the combinatory order is flexible,
because the binding step (the rule z) effects the argument structure of a very local
expression. To demonstrate, note first that indeed his mother can combine with
called to give his mother called of category (S/RNP)NP. This was demonstrated in
Exercises 15.10 and 15.11, but readers who skipped that section can simply take
this on faith. But the fact that his mother called can combine to be a single
expression does not disturb anything; the effect of binding does not take as input
a full expression of this sort. Rather the binding effect happens more locally—in
this case on call—and that gives only an output in which the subject slot binds a
pronoun in the object. In fact, call can first undergo z (before combining with the
subject), but combining this with his mother results in something of category
(S/RNPNP)NP and with meaning ºxe[ºf<e,e>[z-call’(f)(x)]]. Recasting that meaning in
slightly more readable form, this is ºx[ºf<e,e>[call’(f(x))(x)]]. Thus it takes an
individual x and a function f of type <e,e> and returns true if x called the value of f
applied to self. (And it is waiting for an NPNP in object position, but will “hold off”
on the superscript introduced by the pronoun in subject position.) But this does
not change the binding pattern. For example, this could combine with her dog in
 15.3. A VARIABLE-FREE APPROACH 303

such a way that his mother binds her dog, but this does not lead to the WCO-
violating reading. The details of these derivations are space-consuming and hence
left as an exercise for the ambitious reader; the important point here is that
recasting the effect of binding in this very local way means that the existence of
expressions like his mother loves (which are waiting for NPs in object position)
cause no problem for the account of WCO.9

*15.13. Show the details of the syntactic and semantic composition of (i): (i)

His motheri called herk dog.

Note that his is a free pronoun; once its value is set then his mother picks up
some individual. Therefore, the reading notated above is accessible via a
derivation in which her happens to pick up the same individual. We are
interested here, however, in showing how we get (i) on the bound reading.
15.14. Taking as point of departure the fact that his mother called has the
meaning and category shown above, show the composition of a relativeclause
case like:

(i) the woman who his mother loves

Convince yourself that this does not have the WCO-violating reading, but only
has a reading in which his is free (and could be bound from higher up).

It should, however, be noted that although the WCO effect requires no extra
constraint or stipulation, it does not follow from any deep principle of the general
theory here. For there could have happened to be the reverse binding rule s, 10
formulated below (using the FR designation, for fictional rule):

9
This is not to say that z always applies only to a single lexical item. Take, for example,
Every mani thinks that Mary loves hisi dog. One possible way to get the relevant reading is
for z to apply to thinks. But another possibility is that thinks that Mary loves combines (by
various applications of g-sl and r-app), and that then undergoes z. The result is the same
either way.
10
This is essentially the operation that Curry and Feys (1958) dubbed S. Hence the
reason for naming the binding rule z; it is a “backwards” version of S.
Definition of s:
 304 15. PRONOUNS: TWO ACCOUNTS

Given a function f of type <e,<x,y>>, s(f) is a function of type <e,<<e,x>,y>> such that s(f)
= ºx[ºZ<e,x>[f(x)(Z(x))]].
FR-5. Let Æ be an expression of the form <[Æ], (X/Y)/NP, [[Æ]]>. Then there is an
expression of the form <[Æ], (X/YNP)NP, s([[Æ]])>.

This does just the reverse of z: it maps an expression that expects both an NP and
a later argument of category Y into one which now expects the NP and a later
pronoun-containing Y, and semantically it merges the earlier slot with the pronoun
in the later argument. This is exactly what would give WCO violations.

15.15. Show how one can get the WCO-violating reading of (i) with FR-5: (i) *no

girl whoi heri father invited

15.16. Show how adoption of FR-5 also allows for a derivation of the WCO-
violating (ii):

(ii) *Hisi mother called no fourth grade boyi.

It might even be that s is available in a limited range of cases, for there are well-
known exceptions to WCO which are poorly understood (such cases would also
require exceptions to the LF filter in (9) under the variable-ful theory). Here we
continue to assume that only z exists and not s, leaving open the question of
whether there might be limited cases of s. It is, however, worth stressing again that
no extra piece of machinery such as a filter is needed.

15.3.5. What happens when there is more than one pronoun?

15.3.5.1. Multiple pronouns which are “the same”

One of the nice properties of the system with variables and assignments is that
cases with more than one pronoun—including cases where these are bound by
different things—present no difficulty. It is built into the system that these follow
just the way cases with just one pronoun do; Exercises 15.1 and 15.2 were
designed to elucidate that point. Hence, no competitor to a theory with variables
could possibly be taken seriously if it could not handle such cases equally well.
Having generalized the rules in the way done in section 15.4.3, it turns out that
such cases are also unproblematic in the variable-free view. The derivations
 15.3. A VARIABLE-FREE APPROACH 305

themselves can be space-consuming when written in the º-notation (as opposed

to being written using just the operations g and z), but no new apparatus is needed.
We will keep the discussion here somewhat informal, and let the interested
reader work through the details on their own. (Details can also be found in
Jacobson 1999, although the generalizations of the rules proposed in this chapter
differ slightly from the ones given there.) Moreover, although this section does not
quite merit an asterisk, the more casual reader wanting just the basic outline of
how the variable-free system works can safely skip this section. Little else—
including most of the further applications in Chapter 17—depends on a full
understanding of this material.
First, consider sentences with two pronouns where the two pronouns are—
speaking informally—“the same,” as for example (27) on the reading indicated by
the indices:

(27) Every third grade boyi thinks that hei should call hisi mother.

There are actually at least two different derivations that result in this reading. On
one, call undergoes z so its subject position binds his, while thinks also undergoes
z so that its subject position binds he. The LF variable-ful view also has a possible
analysis of this which is analogous; this is explored in Exercise 15.18.

15.17. Give the full details of the above derivation for (27). You can write this
out using just the z operator rather than in the º-notation if it makes it easier to
see what is going on. Also ignore should, and ignore internal details not relevant
(such as the full syntactic and semantic composition of every third grade boy).
15.18. If you did Exercise 15.2, discuss which of the derivations shown there
is analogous to the derivation in Exercise 15.17. Of course the theories are
different, but it is hoped that the notion of one being analogous to the other will
be clear.

But there is another way to get the appropriate meaning, which is best
illustrated by noting that there are cases where the position occupied by one
pronoun cannot bind the other; often these are referred to as cases where the two
pronouns are co-bound:
(28) Every third grade boyi thinks that the man who hisi mother loves (should) feed hisi
dog.
 306 15. PRONOUNS: TWO ACCOUNTS

This does not lend itself to the strategy above; neither pronoun is in a position
whereby z allows its slot to bind the other pronoun. As it turns out, (28) is
unproblematic. One meaning of the embedded S is such that the two pronouns
do not “know” they are the same; its meaning is (using an informal representation)
ºx[ºy[the man y’s mother loves (should) feed x’s dog]], and the category of the
embedded clause is (SNP)NP. But two applications of z on thinks will result in the
appropriate meaning. Readers working through the next two exercises can see
this in more detail. Notice that (27)—on the reading represented there—can also
be derived in a fashion exactly analogous to the derivation here of (28).

*15.19. Show the syntactic and semantic composition of the man who his
mother loves (should) feed his dog. Ignore should, and you can begin with the
man who his mother loves as already composed. Its meaning is ºy[the man
who y’s mother loves], of category NPNP. Similarly begin with his dog as
meaning ºx[the-dog-of ’(x)] of category NPNP.
*15.20. Show that two applications of z on think give the desired reading for
(28).

*15.3.5.2. And yet another possibility

Interestingly, there is another way to arrive at the reading in (28), which is to
generalize g-sub yet further in a way already explored for g-sl in section 13.4.4.2.
This is to allow g on an expression of category (A/B)/C to map to one of category
(AD/BD)/CD. The semantic operation is just as one would expect. Let f be a function
of type <c<b,a>>. Then this generalization maps f to a function of type
<<d,c>,<<d,b>,<d,a>>> with meaning ºX<d,c>[ºY<d,b>[ºDd[f(X(D))(Y(D))]]]. (And of
course presumably this should be generalized to apply to any number of argument
slots.) Were this to apply to feed, the lower clause would have as its meaning
ºx[the man who x’s mother loves should feed x’s dog] and be of category S NP and
a single application of z on thinks will give the reading in (28). (Note that the other
derivation is available too; that there can be two applications of z is an automatic
consequence of the system.) The difference between the derivation here and the
one sketched in the above exercises is that here the meaning of the embedded S
is such that the two pronouns are already “merged,” and the embedded S is thus
of type <e,t> rather than being of type <e,<e,t>>. As noted directly above, both
 15.3. A VARIABLE-FREE APPROACH 307

derivations are automatic in the system here. Notice that this derivation is
analogous to one in which the two embedded pronouns have the same index (if
you did Exercise 15.2, this is analogous to one of the derivations there).

*15.21. Beginning with the meanings of the subject NP and the object NP as
shown in 15.19, show the syntactic and semantic composition of (28) using the
generalization of g above.
*15.22. If you did Exercise 15.2, show which of the derivations is analogous to
the derivation in 15.21. The remarks in 15.18 about the notion of the finding
analogies between the two different theories hold here as well.

15.3.5.3. More than one pronoun—different binders

One can also construct cases of multiple pronouns which in the end are bound by
different positions. This presents no problem: these different readings arise
simply through combinations of z and g applied in different orders. We sketch just
one example to illustrate the point. Consider (29a) and (29b) (no man in (29a) is
recast as every man in (29b) to help make sure that the relevant binding pattern is
pragmatically salient in each case):
(29) a. Every boyi hopes that no manj will realize that hisi dog chased hisj cat.
b. Every boyi hopes that every manj will realize that hisj dog chased hisi cat.

As the interested reader can verify, his dog chased his cat has the meaning
ºx[ºy[y’s dog chased x’s cat]]. The key to the two different binding patterns lies in
what happens to realize (as usual, we ignore will here). Both involve an application
of z and of g to realize, but in different orders. The application of z binds a pronoun
to the subject position of realize while application of g “passes this up” for higher
binding. Therefore, if z applies first to realize followed by g, the subject position of
realize binds the instance of his corresponding to the dog-owner; the cat-owner
his is passed
 308 15. PRONOUNS: TWO ACCOUNTS

up as an open slot and is bound by z on hope. If g applies first then the dogowner
his is passed up. Binding the other his (the cat-owner) by the subject position of
realize involves the application of z that skips over the superscript that was
introduced by g; this is thus zS. The derivations are sketched more fully in (30)
which schematize both the syntax and semantics:
(30) a. every-boy(z(hopes)(g-sup(no-man)(g-sup(z(realize))(his-dog-chasedhis-cat))))
(= 29b)
b. every-boy(z(hopes)(g-sup(every-man)(zS(g-sup(realize))(his-dog-chasedhis-cat))))
(= 29a)

*15.23. Take the derivations shown above and recast them showing the details
both of the syntactic computation and also show the semantics using the º-
notation.

15.4. More on Weak Crossover

15.4.1. Definite NPs and WCO

15.4.1.1. An apparent problem (for both theories)

The contrasts shown earlier in (4) and (5) illustrate the WCO effect: a GQ in object
position cannot bind a pronoun within the subject. Under the QR/LF view, the LF
filter in (9) was posited to account for this; under the variablefree view this
followed from the formulation of z. While there are many other formulations of
Weak Crossover, the remarks below are of relevance to just about all of them.
Note too that the discussion in this section is relevant to both the LF variable-ful
and the Direct Compositional variablefree theory, and the discussion below
should therefore be of interest regardless of one’s theoretical persuasion.
The point of departure is that the empirical generalization regarding the WCO
effect would seem to be threatened by perfectly good sentences in which
quantified NPs are replaced by definite NPs or proper names. Thus, although
some speakers find sentences like the following slightly awkward, most agree that
these are clearly better than those in (4) and (5):
(31) a. The woman who had endorsed himi introduced McCaini (at the convention).
b. The woman who had endorsed himi introduced the winner of the Iowa caucusi (at
the convention).
 15.4. MORE ON WEAK CROSSOVER 309

And, while speakers vary in whether they accept cases like (32) where the
pronoun is in genitive position (and not within a relative clause as in the above
cases), here too it is generally agreed that these are far better than the
corresponding case in (4).
(32) a. Hisi mother called Jack Sprati (the minute he got to school).
b. Hisi mother called the winner of the spelling beei (almost immediately after he
won).

Since these have the readings indicated by the indices above, one might wonder
whether they pose a threat to the twin claims that proper names and definite NPs
can be GQs but that this kind of binding is blocked. If the reading here is a result of
binding—whatever mechanism one uses for that— why should these be good?

15.4.1.2. Not a problem after all

As it turns out, these are entirely unproblematic. Indeed, just about any theory
automatically predicts that the sentences in (31) and (32) can be understood in
the way indicated by the indices used there. Recall that— regardless of which
theory one chooses—there are pronouns which are free. This was shown by the
Elmo discourse in (1) at the outset of this chapter. With that in mind, then,
consider (33) under the understanding indicated by the indexing there:

(33) Jack Sprati called hisi mother.

There are two ways that (33) could have this understanding. The first is that his is
a bound pronoun, exactly as in the case of Every fourth grade boy i called hisi
mother. The details of course vary by theory, but both allow for a bound reading.
In the LF variable-ful theory, the subject NP can be coindexed with his and can
undergo QR. (Section 14.4.2.1 contained discussion of why the QR approach is
formulated in such a way as to allow ordinary definite NPs to undergo QR as well
allowing GQs to undergo QR.) In the variable-free approach, call undergoes z, and
the meaning of the VP is the function characterizing the set of self ’s mother
callers. Whether Jack Sprat in (33) is a GQ or an ordinary NP actually makes no
difference; the binding is a consequence of z on call.
But his in (33) can also be free. Under a theory with variables it has some index,
let us say 5. Assume that it happens not to be coindexed with Jack Sprat. Then the
value of the sentence is a non-constant function from assignments to
propositions. And presumably then there is an available g such that g(5) is Jack
Sprat. (This is just like the case of the context making available an assignment
 310 15. PRONOUNS: TWO ACCOUNTS

where the relevant variable is assigned to Elmo in the obnoxious guest scenario.)
This is often referred to in the literature as a case of accidental coreference
between [[Jack Sprat]] and the value assigned to free his. (It is also sometimes
called covaluation.) Analogous remarks hold in the variable-free version. Here call
undergoes g and the subject lifts and also undergoes g (although once again see
Exercise 15.12 for a way to get this via a shorter derivation). The result is a
sentence of category SNP with meaning of type <e,t>, and the function is applied
by the listener to the contextually salient individual Jack Sprat.
With this in mind, consider again (32a): His i mother called Jack Sprati. It is not
at all surprising that this is good. The binding derivation is blocked, just as there
can be no binding in something like *His i mother called every third grade boyi. But
the free reading is not blocked: here his remains free and can pick up the
contextually salient individual Jack Sprat. More technically: on the view with
variables it has an index (say, 5) and the sentence is understood on the
assignment g such that g(5) = Jack Sprat. On the variable-free view the sentences
denotes a function from individuals to propositions, and is applied to the
contextually salient individual Jack Sprat.

15.4.1.3. Independent evidence: Ellipsis and the notion

“sloppy identity”
Striking confirmation for the remarks above comes from ellipsis phenomena. This
will be exposited primarily using the variable-free approach, but similar remarks
hold under the variable-ful LF approach and some of the exercises are designed
to bring this out. Thus note that two key points emerged from the above
discussion. Point 1: In a sentence like (33), the understanding that Jack Sprat
called Jack Sprat’s mother comes about in two different ways (the pronoun can
be free or bound). In variable-free terms, call can undergo z (giving the bound
reading) or it can undergo g, where the entire sentence is then a function of type
<e,t> applied to a contextually salient individual, who can certainly be Jack Sprat.
Point 2: Whatever accounts for WCO effects in the case of quantified NPs
predicts that for a sentence like (32), the only way that the relevant understanding
can emerge is via a free pronoun. Ellipsis provides a nice probe to verify both of
these.
Now consider (34) on the understanding of the first sentence where what is at
issue is Jack calling Jack’s mother. Even keeping that constant, there is a well-
known ambiguity in the second sentence:
 15.4. MORE ON WEAK CROSSOVER 311

(34) Jacki won’t call hisi mother. But Bill will.

On one reading, it is understood as asserting that Bill—unlike Jack—will be a

dutiful son and will call Bill’s mother. This is especially easy to bring out in a
context where Jack and Bill are two boys at a boarding school, and it is Mother’s
Day. (Call that the Mother’s Day scenario.) The other way to understand the
second sentence is that once again Jack is not a dutiful son, but Bill will call Jack’s
mother. This easily emerges in the following scenario: Jack’s mother is in the
hospital, Jack and his mother have always been estranged, but Bill also knows
Jack’s mother and is quite fond of her. (Call that the Hospitalized Mother scenario.)
The first of these readings was dubbed the sloppy identity reading in Ross (1967)
and this terminology has been used ever since. The second reading is what Ross
dubbed the strict identity reading.
That the “missing” VP in (34) can be understood either way is exactly what one
would expect. Once again, concrete discussion of this point requires choosing
some theory of ellipsis (and of binding), and so we use the account developed in
section 13.4.2, though similar remarks hold under other accounts as well. There
it was hypothesized that what is generally known as VP ellipsis involves the
mapping of the lexical item will (of category VP/VP) to a second word that does
not actually expect to get a complement in the syntax. Continuing to use the
notion of a “free variable” as an expository device, the meaning of this second item
is will’(P). Thus P here is a free variable over <e,t> type meanings, and its value is
fixed by the context. Obviously in the full variable-free setting we need to modify
this since there could not, of course, be free variables; see section 15.6. But using
the notion of a free variable for exposition is convenient, so we continue with that
informal notion for now. Then a sentence like Bill will is ultimately [[will’(P)(b)]]
and something in the context supplies the value for P. In (34) it is obviously the
meaning of the VP in the Jack sentence.
Since the VP in the Jack sentence contains a pronoun, one possible meaning
for it is lx[x called x’s mother]. That is, it denotes the set of self ’s mother-callers.
And if this is picked up as the value of P, the result is the so-called sloppy reading.
The Bill sentence ends up saying that Bill has the same property, from which it
follows that Bill calls Bill’s mother. Notice, incidentally, that the term “sloppy
identity” is a misnomer: the meaning supplied in the second sentence is exactly
the same as the meaning of the first VP. And so the existence of this reading is
completely unsurprising.11 The parallel story under the LF approach is not exactly
the same, for here the VP itself in the first clause does not have as its value lx[x
 312 15. PRONOUNS: TWO ACCOUNTS

call x’s mother] but is, rather, just the assignment-dependent set such that for any
g its value is the set of individuals who called the mother of g(5). Nonetheless, a
similar point holds in the LF approach as well, and we leave the detailed
exposition of this to the reader in the exercise below.

15.24. Proponents of the LF theory usually make somewhat different

assumptions regarding VP Ellipsis, according to which there is some actual
level of linguistic representation of the second sentence containing material
which is identical to material in the first sentence. But because it is unduly
cumbersome to compare this to the view taken here, use the view of VP Ellipsis
discussed above (and in section 14.4.2) when doing this exercise. Thus, show
how the LF theory also accounts for the sloppy reading under the assumption
that the missing VP is understood as a free variable P, whose value is set by
being the meaning of some other expression in the discourse context.
Comment on the crucial role that LF is playing in supplying the “missing”
meaning.

The strict reading—brought out in the Hospitalized Mother scenario—is simply

the reading that results when his in the first sentence is free and picks up
contextually salient John. And so the understood value for P for the second
sentence is the set of individuals who call John’s mother. Incidentally, his can also
refer to any other salient person in the discourse context. Take, for example, the
following which we will call the Sam’s Mother scenario. Sam, Jack, and Bill are all
good friends, have been since childhood, and so they all know Sam’s mother.
Sam’s mother has been in an accident and is in hospital. We are discussing the
fact that Sam is out of the country (with no access to Skype) and wonder if anyone
will be able to call Sam’s mother. Here the natural understanding of the first
sentence is that

11
That the so-called sloppy identity reading can be analyzed this way was proposed at
least as early as Keenan (1971).
Jack will not call Sam’s mother. In that case (and without any further context to set
the value for P) the second sentence is understood as claiming that Bill will call
Sam’s mother. Notice that this reading—where his is Sam—and the strict reading
discussed above are basically the same. Whoever his happens to pick up in the
 15.4. MORE ON WEAK CROSSOVER 313

first sentence results in an <e,t> function which is picked up as the value of P. The
strict reading, then, is just a special case of the more general case where his picks
up someone in the discourse context. (This point is worth stressing, as it is
sometimes lost in discussions about these.)
The existence of the strict/sloppy ambiguity thus confirms Point 1 above: the
prediction that there are two ways to get the understanding of his in (33) shown by
the indexing there. Point 2—that in cases like (32a) there can be no binding—
cannot be shown by using VP Ellipsis for obvious reasons. But we can show this
by broadening the landscape of ellipsis phenomena. Before giving the relevant
sentences, consider the following scenario. Jack, Bill, and Sam are all childhood
friends and go off to boarding school on the same day. Their three mothers are all
worried that the boys will be lonely, and so they designate Jack’s mother to call
each boy some time during the first week. A and B are discussing whether or not
Jack’s mother has fulfilled her obligations yet, and they have the following
dialogue:
(35) A. Well, his mother did call Jack today. B. Yes,
but not Bill.

It should again be noted that some speakers find A’s sentence itself (on the
intended reading) to be somewhat marginal; it is most natural with stress on
Jack.12 (It also requires stress on did in this context.) But for most speakers it is at
least tolerable. The point to notice is that the B sentence is completely natural in
this scenario. Assume—as in the case of VP ellipsis—that a sentence fragment
as in B is understood by some process according to which the meaning of [[his
mother (did) call]] is supplied here. (Notice that in the CG fragment it makes
perfect sense to say that this has a meaning since this material can be composed
up to be a well-formed expression with a meaning. The LF view one would need
to say something a bit different about the understanding of the B-sentence, but
that debate would take us

12
This is known as focal stress or contrastive stress; a discussion of focus is contained
in Chapter 18. Many of the remarks in this section on ellipsis can be made also with respect
to focus and there is a well-known connection between ellipsis and focus (see, e.g., Rooth
1992).
too far afield. The ultimate point about the pronoun will go through regardless of
the exact theory of ellipsis.) So the B-sentence is fine, but it can only be
 314 15. PRONOUNS: TWO ACCOUNTS

understood on the strict reading. This follows if his in the A-sentence can be
understood as a free pronoun.
Strikingly, no sloppy reading is available. This can be confirmed by a simple
variant of the above scenario which is designed to bring out a sloppy reading. The
same boys go off to boarding school, the same mothers worry about their
loneliness. But instead of one mother being the designated caller, each mother
decides that at some point during the week she will call her own child. Once again
A and B are discussing whether or not all of the boys have yet gotten their phone
calls. (35B) cannot (at least not without great difficulty) be understood as
conveying that Bill’s mother did not call Bill, even though this scenario was loaded
in such a way as to try to bring out the sloppy reading.
This follows given the claim that his cannot be bound by the object position.
Putting this more technically in variable-free terms, his mother called can be
assembled into an expression with a meaning, but the way the binding rule z
works, this cannot have the meaning ºx[x’s mother call x]. That can only be
assembled if there were a backwards version of z that applied to call. The LF
theory—supplemented by the WCO filter in (9) in Chapter 13—also has no way to
compose up an LF for (35A) such that there is a part of that LF with that meaning.
In (35A) Jack can undergo QR, but if it was coindexed with the pronoun then its
trace would not c-command the pronoun, and the WCO filter would be violated.
Notice too that the attempted meaning of the dialogue in (35) in the relevant
scenario is possible if we simply change the structure of the sentences a bit, and
so (36) easily allows for the sloppy reading:
(36) A. Jacki was called by hisi mother today. B. Yes,
but Bill wasn’t.

This of course is not surprising, since his in A can be a bound pronoun.

Thus, summarizing the situation in theory-neutral terms, the fact that definite
NPs in object position can “co-refer” to a pronoun within a subject does not
threaten the WCO generalization, since this pronoun can be free. And the pattern
of strict and sloppy identity with ellipsis confirms the suspicion that such cases
only involve free and not bound pronouns.
15.25. Consider the following:

(i) Every third grade boyi thinks that hei should call hisi mother and that every
fourth grade boy should too.
 15.4. MORE ON WEAK CROSSOVER 315

This has a sloppy reading, where the third grade boys think that the fourth grade
boys should call their own (fourth grade boys’) mothers. Show how this is
predicted to be a possible reading using the variable-free approach and discuss
this (at least informally) with respect to the LF approach. Consult Exercises
15.17 and 15.18.
It is worth noting that this also has a strict reading even where the pronoun
is bound—i.e., a reading in which the second clause is understood in such a
way that each third grade boy thinks that each fourth grade boy should call the
mother of that third grade boy. We will not explore here how this is accounted
for in either theory. It is actually fairly easy to see what would be going on in the
LF theory, and the interested reader is invited to work that out. For the variable-
free theory, a more detailed account of ellipsis is needed; the account spelled
out in Jacobson (2003) allows for this.

*15.4.1.4. A more formal account of the strict reading

There is more to say about the strict reading, as an interesting issue arises (under
either theory). The key generalization is that however his is understood in the Jack
sentence, it has the same understanding in the Bill sentence. Under what we’ve
been calling the strict reading (as in the Hospitalized Mother scenario) his is
understood as Jack in both cases. In the Sam’s Mother scenario, his is understood
as Sam in both cases. As noted above, the strict reading is just a special case of
the more general free reading. But what is interesting is that there is no way for the
first his to be understood one way and the second a different way. (Of course that
is exactly what happens in the sloppy reading, so in order to show the point we
need to construct a case where we rule out the possibility of his being bound by
the subject. Once that is controlled for, we can show that there is no way to
understand the two occurrences of his differently.) To convincingly show the point
it is necessary to construct a scenario in which the non-existent reading is
pragmatically quite felicitous, but is nonetheless unavailable. Here, then, is an
appropriate scenario. Sam, Jack, Bill, and Tony have all been good friends since
childhood. Sam’s mother is Thelma, Bill’s mother is Louise, and Thelma and
Louise are also good friends. Thelma and Louise are in a car which plunges off a
cliff; both are injured and rushed to the hospital. Now imagine two people—A and
B—who are talking about this. There is a third person—C—who knows all of the
above facts. Moreover, C misses the first part of the conversation, but does
overhear the interchange in (37):
 316 15. PRONOUNS: TWO ACCOUNTS

(37) A. Well, Jack won’t call his mother. B. True.

But Tony will.

It is perfectly reasonable for C to conclude that A meant that Jack won’t call
Sam’s mother (Thelma) or to conclude that A meant that Jack won’t call
Bill’s mom (Louise). But whichever proposition speaker A’s utterance is meant to
convey, C assumes that B’s utterance is about the same mother. C cannot come
away from this conversation thinking that (on the basis of what he has heard) Jack
won’t call Thelma but Tony will call Louise.
Consider, now, the value of [[call his mother]] where his is free under either of
the theories discussed here. In the variable-ful theory, it is not strictly speaking a
function of type <e,t>, but an assignment-dependent function of this type: i.e., a
function of type <G,<e,t>>. In the variable-free theory, it is a function of type
<e,<e,t>>. Now consider what that entails in terms of the fact that some kind of
missing meaning is supplied in the B sentence. We have not spelled out in full the
details of just how it is that a missing VP-type meaning is supplied, and will
continue to be informal on that score. But regardless of the details, the following
question arises: is what is supplied as understanding in the second sentence a
function into <e,t> meanings (in variable-ful terms, the domain of that function is
G; in variable-free terms the domain of that function is e)? Or is it instead the case
that what is supplied is not literally the value of the first VP, but an <e,t> meaning
arrived at by applying the value of the first VP to some assignment (in the variable-
ful theory) or some individual (in the variable-free) theory?
The simplest way to ensure that Thelma and Louise do not get mixed up in (37)
is to assume the latter. That is, what is supplied as the “missing” meaning in the B
sentence is not a function from some domain into a function of type <e,t>, but
rather a full-blown <e,t> meaning itself. Interestingly, there is independent
evidence for the hypothesis that in this type of construction what is picked up is a
full-blown <e,t> rather than a function into such functions. This evidence comes
from the domain of indexical pronouns such as I, you, etc. Obviously their value
depends on the speech context: assume that an indexical pronoun denotes a
function from speech contexts to individuals.21 One way to work this into the full

21 This is oversimplified; beginning with Partee (1989b) it has been noticed that
indexical pronouns like I also function like “variables,” since “sloppy” readings are
available in cases like (i) (where the relevant reading is such that no one else called their
own mother):
 15.4. MORE ON WEAK CROSSOVER 317

theory of the compositional semantics is to take every expression as having a

value which is a function from contexts to something else. The value of a
sentence, for example, is a function from speech contexts to propositions—i.e.,
from contexts to functions of type <s,t>. Hence, for example, the proposition
expressed by a sentence such as I believe in variables depends on the context of
utterance. Obviously folding this into the full compositional apparatus is beyond
the present scope and so we will continue to ignore this context parameter for the
rest of this book, save to note its relevance to the analysis of VP Ellipsis.
So take a VP like call my mother tonight. Leaving aside the worlds and times, its
value is not a function of type <e,t> but rather a function from speech contexts to
such a function. Now consider the dialogue in (38), the implications of which are
discussed in Hankamer and Sag (1984):
(38) A: Jack is going to call my mother tonight.
B: Really? I’ll bet Bill won’t.

Here the understanding of the “missing” material is not the set of individuals who
call B’s mother, but rather those who call A’s mother. Yet since the value of call my
mother in A’s utterance is actually a function from speech contexts to sets, it
denotes the set of callers of A’s mother only once the speech context parameter
is supplied. If B’s sentence picked up the speechcontext-dependent set, then the
understanding would be that Jack would not call B’s mother. Hence indexical
pronouns provide evidence that in cases of ellipsis, what is supplied as the
missing meaning is not a function into <e,t> meanings, but an <e,t> meaning itself.
The case with free pronouns is analogous: what is picked up is the meaning of the
antecedent VP after it is applied to some assignment (in the variable-ful view) or
some individual (in the variable-free view). Thus Thelma and Louise are not mixed
up in (37) any more than A’s mother and B’s mother are mixed up in (38).

(1) Only I called my mother.

For relevant discussion of how to account for this in the variable-free view, see Jacobson
(2012); an account using variables is spelled out in, among others, Kratzer (1998).
 318 15. PRONOUNS: TWO ACCOUNTS

*15.4.2. A note on three-place verbs and WCO interactions

The discussion throughout this book has for the most part been restricted to
sentences with intransitive or transitive verbs, ignoring cases of three-place verbs,
as in give the milkbone to the puppy, read the story to the children, etc.
However—using only the CG terminology for the moment—it was noted in
Chapter 5 and 6 that there is some evidence that in cases like the above give first
combines with to the puppy and then the direct object NP (the milkbone)
combines with that string by Wrap. It was also pointed out in section 5.5 that a
similar strategy has been adopted in non-CG frameworks with movement (see,
e.g., Larson 1988). Under that view there is a level of representation at which give
to the puppy is a single constituent with the milkbone as its sister, and a
movement rule applies to give the final surface order. (See the discussion in
section 5.5 about various implementations of this.) But in the earlier discussions,
no actual evidence was given for these conclusions. We are now in a position to
give one piece of evidence for this: it has to do with WCO interactions. Since we
are focusing here on VPs of the form V-NP-PP, it will simplify the discussion (under
either theory) to think of PPs as just fancy NPs. That PPs with case-marking
prepositions like to have the same semantics as NPs has already been
established. Assume that their syntax is more or less the same too (with the
addition of the casemarker). In particular, in the Direct Compositional view,
assume that they can lift to generalized quantifiers (see the discussion in section
14.3), and assume that in the LF view they can undergo QR.
With this in mind, note that while (39) is good, (40) is degraded (the judgments
do not seem to be as sharp here as in the cases in (4) and (5); we leave it open as
to why not):
(39) a. Gail sent every packagei to itsi intended recipient.
b. Gordon gave every paintingi (in his collection) to someone who could really
appreciate iti.

(40) a. ?*Gail sent a package that contained his i favorite cookie to every soldieri (in that
battalion).
b. ?*Gordon gave a portrait of heri grandfather to every girli (in the class).
In order to conclude that this is an instance of a WCO violation, one needs to
ensure that when no binding is involved, the second NP can scope over the first.
Indeed the wide-scope reading for every emerges with no difficulty at all in (41):
(41) a. Gail sent a package to every soldier (in that battalion).
 15.4. MORE ON WEAK CROSSOVER 319

b. Gordon gave one of his paintings to every girl (in the class).

Let us first consider what this shows under the variable-free view, making use of
z. Under the formulation of z in R-8, the object slot (the slot occupied by every
package) can be merged with the pronoun slot only if the category of send is
((S/LNP)/WNP)/RPP. The z rule allows this to map to
((S/LNP)/WNP)/RPPNP with meaning ºf<e,e>[ºx[send’(f(x))(x)]] (which can be
equivalently written as ºf<e,e>[ºx[ºy[send’(f(x))(x)(y)]]]. The reverse binding pattern
is not possible as shown in (40). In other words, this is parallel to the subject–
object asymmetry with respect to binding. Under an account of WCO effects that
builds it in to z via the order of the arguments, the conclusion follows that verbs
like send and give have the Wrap category.
Similar remarks hold under the LF view. Given the WCO filter in (9) it must be
that—at LF—the trace of every package in (39) c-commands the pronoun within
the PP. On the other hand, should the PP raise in (40a) it must be that its trace
does not c-command the coindexed pronoun within the package-NP. The details
of just what the LFs look like depend on the precise analysis of the movement
rules at work in these cases, but the main conclusion is analogous to the
conclusion in the variable-free view.
Incidentally, one might take the binding patterns in (39) and (40) to show that
we have been wrong all along in our discussion of WCO effects—that the effect is
one based on linear order and not the order of argument introduction (or, c-
command). In that case, no wrap or movement would be necessary for the three-
place verbs. That WCO is about linear order was actually assumed in early
accounts of this construction, and has been revived from time to time (see, e.g.,
Shan and Barker 2006). If this turns out to be the case, the LF filter in (9) can easily
be modified to refer to precedence rather than c-command. In the variable-free
view, the requisite modification would involve a more general binding rule (where
any argument position could be “merged” with a pronoun within any other) and
where the rule was sensitive to the slash directions. We will, however, not pursue
this here, because WCO effects are not the only reason to adopt a
 320 15. PRONOUNS: TWO ACCOUNTS

15.5. MORE ON FREE “VARIABLES”

Wrap analysis (or its analogue in theories with movement). The distribution of
reflexives and reciprocals has also been used in support of this.

15.5. More on free “variables”

The preceding sections made precise the notion of free pronouns under both
views. In the variable-ful view, a sentence such as He5 left has as its value an
assignment-dependent proposition; for each g, it is the proposition that g(5) left.
Some principle predicts that this is applied (by the listener in a discourse context)
to some assignment g. In the variable-free view, the value of He left is a function
of type <e,t>, which the listener applies to a contextually salient individual. But
recall that throughout this book we have made use of other kinds of “free
variables” in addition to those that pick up individuals. For example, an adjective
such as big was hypothesized to contain as part of its meaning a “variable” of type
<e,t> which supplies the comparison class. In section 10.6 it was hypothesized
that quantificational determiners have as part of their meaning a “variable” that
restricts the domain. And in section 14.4.2 it was hypothesized that VP Ellipsis
involves a “variable” over properties (VP meanings) that occurs as argument of the
auxiliary. Incidentally, all three of these cases are variables over meanings of type
<e,t>, but many other kinds of free variables have been used in semantics. One
example—a variable over numbers—will be discussed in section 16.3.2 in
connection with the meaning of many. It is thus worth considering how these
additional kinds of things which have gone under the rubric of variables would be
handled in each of the accounts of pronouns developed in this chapter.

15.5.1. Free “variables” in the variable-ful account

Consider the meaning of big discussed in Chapter 8. There it was concluded that
this means “big for an X” where X is something of type <e,t>, it remains free, and
its value is set by the context. Putting this more formally, the value of Dumbo is big
is a non-constant function from G to a proposition, and G must have in its domain
not only the individual variables but variables of other types as well. Moreover, in
the informal discussion we used just X, but this too must have an index and so
 321
there must be an infinite number of variables of this type available. Thus consider
a sentence like (42) where the comparison class for big and the comparison class
for little need not be the same (most likely they are not):
(42) The big flea sat on the little elephant.

Were there just one variable of type <e,t> then the same comparison class would
have to be supplied for both. And of course we could make a more complex S with
more and more scalar adjectives, each getting a different comparison class. This
means that big must come with an index in the lexicon, and [[big 1]] would be “big
for an X1,” and so forth.
Similarly, suppose that quantificational determiners (every,some, etc.) come
with a domain restriction variable, such that [[every]] is ºP[ºQ[P S \ XS QS]] where
again X is of type <e,t>. (In section 10.6 the meaning for every was spelled out as
above except that XS was replaced with D—a variable over sets—the two are of
course interchangeable.) Here too X actually must be indexed (so that there are
items every1,every2, etc.) because one can have more than one quantifier in a
sentence where the domains are interpreted differently; the reader is invited to
construct relevant examples. Moreover, the <e,t> comparison class for big and the
domain restriction for every in (43) have to be able to be distinct, indicating that
the index of every and the index on big need not be the same:
(43) Every elephant ate a big peanut.

The restriction on every might be the set of animals traveling with the circus
troupe; the comparison class on big is most likely the set of peanuts.
What this means, then, is that the use of free variables as a way to model these
meanings requires that the assignment functions are more complex than has
been assumed so far. They are not simply functions from integers to individuals;
ordinary integers will serve for the case of variables over individuals but the
semantics (under this view) makes use of variables of other types. As one way to
work out the technical details, the variables of type <e,t> can be ordered pairs of
integers whose first member is any integer and second member is 1. (Thus the pair
(3,1) is a formal way to think of what we have been calling X 3.) Other variables can
be ordered tuples of other types. At the end of the day, then, a sentence like He 5
has taken every3 phonology course has as its value a function from assignments
to propositions, and the context will supply an assignment in which 5 is assigned
to
 322 15. PRONOUNS: TWO ACCOUNTS

15.5. MORE ON FREE “VARIABLES”

some individual, and (3,1) is assigned to some (characteristic function of a) set of

individuals (e.g., the set of phonology courses offered at Brown). 15.5.2. Free

“variables” in the variable-free theory

In the variable-free theory, these cases will be handled by assigning more complex
types to the lexical items involved: they all are functions from something else to
the type that was assigned in the original discussion. For example, [[big]] is not of
type <e,t> but of type <<e,t>,<e,t>> with meaning (informally) ºX[big for an X S].
(Since we are using X as a variable of type <e,t> rather than a variable over sets,
read the inner brackets as “big for something in the set characterized by X”.)
Moreover, its category contains a superscript. The category of the superscript
could, for example, be N, so the category of big would actually be AP N. Similarly,
[[every]]—if it does contain a domain restriction—is not just a relation between
sets but has as its meaning a function from sets to such a relation. In particular,
its meaning is ºX[ºP[ºQ[PS\XS QS]]]. And the syntactic category of every would be
revised to be (L(NP))N, which in turn means that every mouse would be L(NP) N.
In other words, what we have been informally calling free variables correspond
to extra argument slots. These slots are passed up in the way a pronoun slot is,
and so the meaning of a sentence like Every elephant escaped ends up being a
function from <e,t> meanings to propositions. The listener applies this to some
contextually salient set to get the domain restriction. Similarly, a sentence like He
has taken every phonology course is a function of type
<e,<<e,t>,t>>—it needs to have both the e-slot and the <e,t> slot filled and so will
be applied to some contextually salient individual and a contextually salient set
(again, perhaps, the set of phonology courses offered at Brown).
Both theories, then, posit more complex meanings for the lexical items. Other
than that, though, nothing extra is needed in the variable-free theory for these.
There is no change in the kinds of objects in the semantics; there just are fancier
functions than would be needed if there were no cases of “free variables.” The
apparatus for passing up these slots is already in place. Incidentally, passing up
the slots by use of g-sup and l makes for rather complex derivations, even though
no new complexities are introduced in the rules or machinery. However, that too
 323
can be simplified by defining R-1 recursively; see the discussion in 15.3.3 and
Exercise 15.12.
15.26. Given the meaning and category for big above, show the derivation (both
the syntax and the semantics) in the variable-free framework for (i), using a
combination of g-sup and l where needed.

(i) Dumbo is big.

*15.27. Having generalized R-1 to apply recursively as suggested in Exercise

15.12, show a shorter derivation of (i) than the one in Exercise 15.26 (unless
you already used that generalized rule in your answer to 15.26).

15.6. The i-within-i effect revisited: A solution

And now we come full circle to consider a solution to the mystery with which this
book opened: the i-within-i effect detailed in Chapter 1.4. Readers who skipped
that section at the outset should read it now. The solution here is framed within
variable-free terms, and also makes crucial use of the assumptions about the
syntax/semantics interface taken in CG and which have been maintained
throughout this book. It thus has three key ingredients: (i) binding is a matter of the
merging of argument slots; (ii) the syntax and semantics closely parallel each
other; (iii) nouns—while being of type <e,t>—do not have a syntactic subject slot,
and, concomitantly, relational nouns are of type <e,<e,t>> but of category
N/PP[OF] and hence have only one syntactic argument slot. Perhaps the
punchline is obvious from just these observations. But if not, we provide
elaboration.
Recall first the basic puzzle, sketched only briefly here as it is amply detailed in
Chapter 1. In the A-B-C party scenario—in which Betty and Bert were childhood
sweethearts—I can describe my encounter with Betty using (44a) or (44b), but not
(44c).
(44) a. I enjoyed talking to the woman who married her childhood sweetheart.
b. I enjoyed talking to the woman married to her childhood sweetheart.
c. *I enjoyed talking to the wife of her childhood sweetheart.

Beginning with (44a), there is nothing surprising (on any theory) about this. For
example, in the variable-free view, marry undergoes z and so its subject position
 324 15. PRONOUNS: TWO ACCOUNTS

binds the pronoun. (The lexical entry for who, given in Chapter 13, allows this to
combine with the VP to give a relative clause.)
15.6. THE I-WITHIN-I EFFECT REVISITED

(44b) is also unproblematic, although it does have interesting consequences

for the LF theory; see Exercise 15.28. We will assume that married to her
childhood sweetheart is an “AP” (should it turn out to be a slightly different
category nothing will change); but recall that “AP” is short for S[A]/NP. In other
words, “APs” combine in the syntax with subjects to give fuller expressions (what
are generally called small clauses). This claim is motivated by examples like (45);
see also the discussion in section 8.2 on this point:

(45) With Bill married to Calista, he’ll travel the world.

This in turn means that married is (S[A]/NP)/PP[TO] and, of course, its meaning is
of type <e,<e,t>>. It can thus undergo z; to her childhood sweetheart is a PP[TO] NP
and can thus occur as argument of z(married), and the subject slot of married will
bind the pronoun. This then maps to the nominal modifier by R-2. Notice that there
is no need to posit any silent subject of married in these cases; the fact that it has
a syntactic and semantic subject slot is sufficient to allow for the binding of the
pronoun.

15.28. Consider (44b) in the LF view. What does this view need to assume in
order to account for the binding of the pronoun here?

But now take (44c). Here wife is also of type <e,<e,t>>. Why then could this not
also input z? There would be nothing wrong with this semantically; z(wife’) is
ºf<e,e>[ºx[wife-of ’(f(x))(x)]] and her childhood sweetheart has a meaning of type
<e,e>. So z(wife-of ’)(her-childhood-sweetheart’) is the set of x’s who are women
and are married to x’s childhood sweetheart; exactly the reading that is asked for
in (44c). The i-within-i mystery is that (44c) lacks that reading, even though it
seems like a perfectly sensible meaning. And indeed, it is a sensible meaning. But
the key lies in the syntax under the assumption that the syntactic and semantic
operations are tightly coupled. The z rule (R-8) is formulated to apply only to items
whose syntactic category is (X/NP)/Y. The argument slot that binds the pronoun in
the semantics must also be present as a syntactic slot. And this formulation is not
a stipulation that is made up just for this case: it is just what would be expected
 325
given the tight correspondence in CG between the syntactic and semantic
operations. But recall that—even though ordinary nouns are of type <e,t>—they
have no syntactic subject slot and relational nouns similarly are just of category
N/PP[OF], where “N” is not an abbreviation for a more complex category. This odd
property of nouns was not stipulated for this case; it is independently motivated
in view of the fact that nouns do not occur in small clauses:

(46) *With Nora linguistics major, she’ll learn lots of languages.

*With Calista wife of Bill, he’ll travel the world.

Hence relational nouns, while of the right semantic type to undergo z, are just not
of the right syntactic category to do this. And so from these ingredients, the i-
within-i effect follows.
 Appendices to Parts I–III
The full fragment

Appendix 1: The basic fragment 323 Appendix 3: Informally discussed

addition for VP Ellipsis 330
*Appendix 2: Possible extensions of
the fragment for three-place
verbs using Wrap 329

Below is a summary of the full fragment developed to this point, along with some
of the open questions and tentative suggestions. This is included before the final
four chapters in Part IV (Further Topics) as we will not attempt to officially fold in
the treatment of focus in Chapter 18. Chapter 19 proposes one revision to Rule-1
(app) to account for intensionality. Some other modifications are necessary in the
rules to fully fold in intensionality, but this is straightforward, as will also be noted
in Chapter 19. (Thus the interested reader can tackle this as an exercise.)
Chapters 16 and 17 propose no changes to the basic rules, only some additional
lexical items.

Appendix 1: The basic fragment

Foundational concepts

Semantic types: A semantic type means something which can be the value of a
linguistic expression.
Basic types: The semantic types consist of the set e (individuals), t (truth values), w
(worlds), and i (times); s is used to abbreviate world/time pairs.
Recursive definition of other types: If a is a type and b is a type, then any function from
domain a to co-domain b is a type. The set of such functions is notated <a,b>.

Syntactic categories:
Basic categories: S, NP, PP, N, CP, along with a set of features on these categories. [A] and
[V] are examples of features that can occur on S, so S[A] is an adjectival small clause,
 327
and we use S[V, TENSE] for an ordinary S. Other features are relevant for verbal
morphology to give verbs in the progressive form (ending in -ing) or in the past participle,
etc.; these are not spelled out here.
“VP” is an abbreviation for S[V]/NP and “AP” is an abbreviation for S[A]/NP.

Recursive definitions of other categories:

(i) If A is a category and B is a category, A/ RB is a category.
(ii) If A is a category and B is a category, A/ LB is a category. (iii) If A is a
category and B is a category, AB is a category.

Correspondence of categories and types: (ignoring intensions)

Any expression of category A/B or AB has as its value a function in <b,a>.

All linguistic expressions are triples: <[sound], Category, [[meaning]]>.

The lexicon

For convenience, we distinguish between open- and closed-class items. The

former (in general) have values which vary according to the world and time, and
the latter in do not. The correlation is not perfect; names are obviously open-class
items but we continue to follow Kripke (1980) in treating these as rigid designators
(i.e., they have the same value for all worlds and times). However the fact that the
correlation is not perfect is of no consequence, because the distinction between
closed- and open-class items plays no actual role in the theory here and is largely
just a convenient way to organize the presentation of the lexicon. (Chapter 19
does explore a possible view in which the distinction between those items taken
here to have constant values across worlds and those which don’t does play a
role in the grammar; see section 19.2.) Note too that while numerical determiners
(e.g., three, three hundred and forty-eight, etc.) might at first glance appear to be
open class, they actually are not in that they are not basic lexical items. The basic
items include
APPENDIX 1: THE BASIC FRAGMENT

only individually listed words or morphemes for numbers between 1 and 20 and
the others are formed in predictable compositional fashion from these.
Because the open-class items generally have values that vary by world and
time, we will not spell out the meanings; and so list these with syntactic category
first, followed by their “phonology.” The closed-class items are listed in the usual
triple format, and are given extensionally.
 328 APPENDICES TO PARTS I–III
Open-class items:

NP includes: Mitka, Balto, Barack Obama, Porky, Farmer MacDonald, . . .

S/NP includes: walk, talk, grunt, . . .
(S/NP)/NP includes: kiss, likes, feeds, . . .
(S/NP)/PP[TO] includes: talk, speak, . . .
(S/NP)/CP includes: believe, think, . . .
S[A]/NP includes: honest, red, tall, . . .
Possible revision (in variable-free terms) to include comparison class argument,
exemplified with tall:
<[tall], (S[A]/NP)N, ºP<e,t>[“tall” for a PS]>
Mostotheradjectivesalsohaveacomparisonclassargument,thoughsomemaynot.
(S[A]/NP)/PP[OF] includes: fond
Might also need revision to include comparison class argument.
N includes: table, dog, cat, pig, . . .
N/PP[OF] includes: mother, child, daughter, employer, . . .

Closed-class items:
Definition of :
i. Given two truth values a and b (i.e., members of {1,0})a b = a & b (i.e., = (1,1) Ò 1,
(1,0) Ò 0, (0,1) Ò 0, and (0,0) Ò 0)
ii. Given two functions f and g such that f g is defined, f g = ºx[f(x) g(x)].

Definition of q:
i. Given two truth values a and b (i.e., members of {1,0})a q b = a v b (i.e., q = (1,1) Ò
1, (1,0) Ò 1, (0,1) Ò 1, and (0,0) Ò 0)

ii. Given two functions f and g such that f q g is defined, f q g = ºx[f(x) q g(x)].

<[and], (X/X)/X, ºAx[ºBx[A B]]> (for x the semantic type of expressions of category
X)
<[or], (X/X)/X, ºAx[ºBx[A q B]]>

Comment: an alternative possibility is to list and in the lexicon as being of category (S/S)/S
with meaning ºq[ºp[p&q]], where all other occurrences of it from the generalization of
g-sl given below. Similarly for or. See section 13.4.4.2.

<[didn’t], X/X, ºF<a,t>[ºAa[~F(A)]], where X is restricted to S/NP or S/(S/(S/NP))> Comment:

This is a placeholder. did is an auxiliary; English negation involves not or the clitic n’t
attached to an auxiliary. See Exercise 13.11.

<[to], PP[TO]/NP, ºxe[x]>

<[of], PP[OF]/NP, ºxe[x]>
<[that], CP/S, ºpt[p]> (p is actually something of type <s,t> in a fully intensional fragment)
<[the], NP/N, ºP<e,t>[ix[P(x)]]> (where here ix[P(x)] means “the unique or most contextually
salient x such that P(x) holds”)
 329
Comment: This might be revised as follows to include the domain restriction:
<[the], (NP/N)N, ºQ<e,t>[ºP<e,t>[ix[P(x) & Q(x)]]]>

<[is], S[Tense]/(S[A]/NP, ºP<e,t>[P]>

<[every], LS(NP)/N, ºP[ºQ[PS \ QS = PS]]>

<[some], LS(NP)/N, ºP[ºQ[PS \ QS 6¼ ]]>

<[no], LS(NP)/N, ºP[ºQ[PS \ QS = ]]>

<[three], LS(NP)/N, ºP[ºQ[ |PS \ QS| 3]]>

<[most], LS(NP)/N, ºP[ºQ[ |PS \ QS| > |PS| ]]>

——
2
Comments:
• All of the quantificational determiners above would have revised entries if it is correct
that the domain restriction slot is part of their semantics.
• In the variable-free framework, the lexical entries for every and some—if they include a
domain restriction—would be:
<[every], LS(NP)/N, ºT[ºQ[ºP[QS \ TS PS ]]]>

<[some], LS(NP)/N, ºT[ºQ[ºP[QS \ TS \ PS 6¼ ]]]>

<[which], RC/(S/RNP) and RC/(S/LNP), ºP[P]>

<[who], RC/(S/RNP) and RC/(S/LNP), ºP[P]>

Comment: The lexical entries for which and who are almost certainly incorrect as they do
not in any obvious way extend to pied-piping cases. See Jacobson (2008) for an
alternative.

<[he], NPNP, ºx[x]>

<[she], NPNP, ºx[x]>

Comment: The contribution of gender should ultimately be folded in. One way is to treat
these as partial functions. For example, the value of he is the identity function whose
domain includes only male individuals.

Word order rules

i. ...(S/X)... ii. ...(A/A)... iii. ...(A/B)...

L L R where (i) and (ii) must apply before (iii)

Comment: Rules of this general sort can be used for case features; see (4) in section 6.4.
 330 APPENDICES TO PARTS I–III

APPENDIX 1: THE BASIC FRAGMENT Unary and

combinatory rules

R-1 (app):
Definition of functions corresponding to syntactic categories:
a. FA/R B ([Æ])([]) = [Æ-]
b. FA/L B ([Æ])([]) = [-Æ]

R-1. If Æ is an expression of the form <[Æ], A/B, [[Æ]]> and is an expression of the form
<[], B, [[]]> then there is an expression g of the form <FCAT(Æ)([Æ])([]), A, [[Æ]]([[]])>
(where CAT(Æ) means the category of Æ).

Comment: This might be given recursively. Formalizing the recursion is left as Exercise
15.12, but note that the result is that, for example, an expression of category (A/ RB)C can
combine with an expression of category B (to its right) to give A C. (The same result is
possible by lifting B to A/L(A/RB); then apply gc-sup, and applying that result to (A/RB)C.
The suggestion above allows for a shorter derivation.)

R-2 (n-mod):
R-2. Let Æ be an expression of the form <[Æ], AP or RC, [[Æ]]>. Then there is an
expression of the form <[Æ], N/N, ºP<e,t>[ºxe[ [[Æ]](x) & P(x)]]>. If [Æ] consists of
a single word then the result is N/RN.

Comment: This should be defined recursively to apply also to, e.g., RC NP and other input
categories with superscripts. Hence we revise to R-2Rev.

Recursive definition of int:

a. Let f be a function of type <x,t>. Then int0(f) is of type <<x,t>,<x,t>> and is ºg <x,t>[fS\gS].

b. Let f be a function of type <a,b> such that inti is defined for b (for some i). Then intn(f) =
ºAa[intn-1(b(A))].

Recursive definition of NMOD:

a. Let CbethecategoryRCorAP(andperhapssomeothers).ThenNMOD0(C)isN/N.
b. Let C be any category XY such that NMODi (for some i) is defined for X. Then NMOD n(C)
= (NMODn-1(X))Y.

R-20: Let Æ be an expression of the form <[Æ], C, [[Æ]], for C any category for which
NMOD is defined>. Then there is an expression of the form <[Æ], NMOD n(C),
intn([[Æ]])>.

R-3:
R-3. Let Æ be an expression of the form <[Æ], N/PP[OF], [[Æ]]>. Then there is an
expression of the form <[], NP/LNP[GEN], ºxe[iy[Æ’(x)(y)]]>.
R-4 (l):
 331
Definition of l: For any x (of any type) in a set A and any set B, L <A,B> (x) is a function in
<<A,B>,B> such that for all f in <<A,B>,B>, l(x)(f) = f(x).
Definition ofL: For any category A, LB(A) is any category such that for all expressions x of
category A and all expressions y and z such that [x] = [z], FLB(A)([y])([z]) =
FCAT(z)([z])([x]).

R-4. Let Æ be an expression of the form <[Æ], A, [[Æ]]>. Then there is an expression of the
form <[Æ], LB(A), ºf<a,b>[f([[Æ]])] (i.e., l<a,b>(Æ))>.

Comment: This also should be recast as a recursive rule along the lines of the model above
for R-2. Spelling this out is left as an exercise.

R-5 (g-sl):
Definition of g: For any function f of type <a,b>, there is a function g c(f) of type
<<c,a>,<c,b>> such that gc(f) = ºX<c,a>[ºCc[f(X(C))]].

R-5a. Let Æ be an expression of the form <[Æ], A/ RB, [[Æ]]>. Then there is an expression
of the form <[Æ], (A/RC)/R(B/RC), gc([[Æ]])> (for c the semantic type of expressions
of category C).
R-5b. Let Æ be an expression of the form <[Æ], A/LB, [[Æ]]>. Then there is an expression
of the form <[Æ], (A/LC)/L(B/LC), gc([[Æ]])> (for c the semantic type of expressions
of category C).
Comments:
• R-5a and R-5b should ultimately be collapsed into one, and the directional features on
the slashes would not be stipulated.
• The rule (or rules) might ultimately be generalized further to Generalized g.

Example: Let f be a function of type <a,<b,c>>. Then gen-gd(f) is of type

<<d,a>,<<d,b>,<d.c>>> such that:
gen-gd (f) = ºX<d,a>[ºY<d,b>[ºDd[f(X(D))(Y(D))]]]
The syntactic mapping would map, e.g., (C/RB)/RA to ((C/RD)/R(B/RD))/R(A/RD).
This rule would allow and to be listed in the lexicon with the single category
(S/RS)/RS (and meaning of type <<t,t>,t>) where the other items are derived by
recursive application(s) of gen-g. Note, however, that gen-g is not a “word-
orderpreserving” operation in the same way that g is; see Exercise 13.14.
• g should be given a recursive definition along the lines given for R-2. This is left as an
exercise.

R-6 (argslot-l): Definition:

Let f be any function of type <K!,<e,<L!,t>>>. Then argslotn+1-l(f) is a function of type
<K!,<<e,t>,t>,<L!,t>> (where n = the number of slots in K) where AS n+1-L(f) = ºXK! [ºP[ºXL!
[P(ºxe[f(XK!)(x)(XL!)])]]].

R-6. Let Æ be an expression of the form <[Æ], ((S/A!)/NP)/B!, [[Æ]]> where A is of length n.
Then there is an expression of the form <[Æ], (( S/A!)/GQ)/B!, argslot-ln+1([[Æ]])>.
 332 APPENDICES TO PARTS I–III
*APPENDIX 2: POSSIBLE EXTENSIONS

R-7 (g-sup):
R-7. Let Æ be an expression of the form <[Æ], A/B, [[Æ]]>. Then there is an expression of
the form <[Æ], AC/BC, gc[[Æ]]> (for c the type of expressions of category C).

Comment: g-sup should be given a recursive definition along the lines given for R-2.
This is left. This is as an exercise for the reader.

Looking ahead (see section 17.3.2): This rule should be generalized to apply both to
ordinary slash and to superscripts. The requisite generalization is as follows:

Notation: Let A|B be either A/B or AB.

R-7 Generalized: Let Æ be an expression of the form <[Æ]; A|B; [[Æ]]>. Then there is an
expression of the form <[Æ]; AC|BC; gc[[Æ]]> (for c the type of expressions of
category C). (The value of | is of course held constant from the input to the output.)

R-8 (z) (simplified): Definition:

Let f be a function of type <a,<e,b>>. Then z(f) is of type <<e,a>,<e,b>>, and is
ºY<e,a>[ºxe[f(Y(x))(x)]].

R-8 (simplified): Let Æ be an expression of the form <[Æ], (A/NP)/B, [[Æ]]>. Then there
is an expression of the form <[Æ], (A/NP)/BNP, z([[Æ]])>.

Generalization of R-8:
Definition: Let f be a function of type <b,<c Òn,<e,a>>> (where cÒn means n number of

argument positions of any time, for n 0). Then za(f) = ºY<e,b>[ºZÒn[ºxe[f(Y(x))(ZÒn)(x)]]].

R-8 Generalized: Let Æ be an expression of the form <[Æ], ((A/NP)|CÒn)/B, [[Æ]]>.

Then there is an expression of the form <[Æ], ((A/NP)|C Òn)/BNP, za([[Æ]])>.

*Appendix 2: Possible extensions of the fragment

for three-place verbs using Wrap

Assume each expression of the language is a string with an infixation point. We write this
as [x|y] (but will suppress the infixation point when not relevant).
Add to the recursive definition of categories: If A is a category and B is a category, A/ WB is
a category and A/IB is a category. The lexicon
Assume each lexical item in English is a string of the form [x|].
Add to lexicon:
((S/NP)/NP)/PP[TO] — give, donate, . . .
 333
Word order rule Add:

((S/NP)/NP)/X

Comment: X here ranges over any category. One could also allow it to range over any
category including nothing; then ordinary transitive verbs would also be marked with
a wrap slash. In that case the effect of Wrap is vacuous.

Revision of the definition of functions corresponding

to categories of the form A/B
Let Æ be some expression whose string is [x|y]. Then

a. FA/RB ([Æ])([]) = [x|y-]

b. FA/ LB ([Æ])([]) = [-x|y]

Define a function appropriate for A/WB as follows. Let Æ be some expression whose string
is [x|y]. Then FA/W B ([Æ])([]) = [x|y]

Comment: These conventions ensure that the infixation point is inherited from the
infixation point of the function.

Appendix 3: Informally discussed addition for VP Ellipsis

Phrased here in variable-free terms:

Lexicon includes a set of auxiliary verbs (did, will, might, be, etc.) which are basically of
category (S/NP)/(S/NP), but in each case specify a particular morphological feature on
the complement. The class of auxiliaries is identifiable by the grammar, either in virtue
of the selection property noted above, or perhaps because they themselves all also have
a feature AUX.
Rule maps all auxiliaries of category (S/NP)/(S/NP) into one of category (S/NP) (S/NP), with
no change in meaning. These remain functions from VP meanings to VP meanings; the
difference is in their syntax. The new items encode that they are proVPs and pass up the
superscript feature, rather than expecting a complement in the syntax.
 PART IV Further
Topics
The final part of this book consists of four chapters designed to show how the
formal semantic tools developed here extend to give insightful treatments of an
even larger and richer fragment than has been developed so far. This includes the
distribution of so-called Negative Polarity Items, a number of further topics
related to binding, the treatment of focus, and intensionality.
The four chapters are logically independent and can be read in any order.
 16
Negative Polarity Items, semantic
strength , and scalar

implicature revisited
16.1. The notion of Negative Polarity 16.5. Open questions and further 358
Items 334 discussion
16.2. The distribution of NPIs 337 16.6. Semantic strength and scalar 360
implicatures
16.3. The first breakthrough 340
16.4. But why? And how? 349
The central mission of this chapter is to understand the distribution of a group of
items known as Negative Polarity Items. The beauty of this domain is that on the
one hand, it shows an extremely interesting application of the tools developed so
far (this requires only the tools through Chapter 10). On the other hand, there
remain both open questions about the full distribution of these items and
controversies as to the best characterization of this distribution, and so there is
work yet to be done. Nonetheless, the history of research on this domain is a
history of successes. Section 16.1 considers the relevant items, and 16.2 shows
that the distribution is quite subtle and appears almost random. One can imagine
ways to simply stipulate where these items occur with syntactic features, but a
further look (in 16.3) reveals that the relevant environments all share a semantic
property.1 Clearly the

1 The analysis of what these items have in common that is discussed in this book is
accepted to be correct by the majority of researchers in this domain (it stems from
distribution should follow in some way from the meaning of these items combined
with the property of these environments (and perhaps other principles), and
indeed 16.4 discusses one influential account along these lines. Meanwhile, the
tools elucidated here have another application: they allow for fine-grained
 337
predictions about just what scalar implicatures are found where, as is shown in
16.5.

16.1. The notion of Negative Polarity Items

16.1.1. any and ever

There are a number of items that have a very limited distribution. Consider, for
example, the contrast in (1):
(1) a. Porky hasn’t ever left the farm.
b. *Porky has ever left the farm.

That the culprit in (1b) is the word ever is easily demonstrated by the fact that
removing that word makes the sentence impeccable:

(2) Porky has left the farm.

And the only difference then between (1a) and (1b) is the presence of the negation
in (1a). So, at first glance it looks like ever can occur only in negative sentences,
and hence is commonly referred to as a Negative Polarity Item (NPI for short). We
will continue to use this terminology even though it will be shown shortly that
negation is just the tip of the iceberg.
Another NPI is any. Thus compare (3a) to (3b) and (3c); the contrast between
(a) and (b) gives a minimal pair which differ only the presence of the negation, and
the contrast between (b) and (c) gives a minimal pair differing only in the use of
any vs a:
(3) a. Tabby didn’t catch any mouse.
b. *Tabby caught any mouse.
c. Tabby caught a mouse.

work by Fauconnnier 1975 and Ladusaw 1979). Many of the controversies and open
questions concern exactly why the generalization holds, and exactly from what aspect of
grammar and/or pragmatics it follows. (See section 16.5.) But it should be noted that there
are dissenting voices as to whether the basic analysis exposited here is correct; see
especially Giannakidou (1997) and other works of Giannakidou’s.
16.1. THE NOTION OF NEGATIVE POLARITY ITEMS
 338 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES
But a brief digression on the full distribution of any is in order before continuing.
While often taken as the prime exemplar of the class of NPIs, any does have a
wider distribution than ever. In addition to its use in (3a), there is a use of any
exemplified in (4), which is generally referred to as “free-choice any.” This free-
choice reading emerges with modals such as can, could, would, and with also
imperatives:2

(4) a. Any cat can catch Minnie (because she is such a slow mouse).
b. Calico can catch any mouse (because she is such a skillful hunter).
c. Calico would catch any mouse that dared to cross her path.
d. Pick any card.

Of course, the question arises as to whether there really are two distinct lexical
items here: NPI any (as in (3)) and free-choice any (as in (4)). The wording above
assumes that these are distinct items, but a priori we do not know that. Indeed,
there is disagreement on just this point (for relevant discussion, see Horn 1972,
Carlson 1981, Kadmon and Landman 1993, Lahiri 1998, and others). Still, it will
simplify the discussion here to assume that free-choice any is a different lexical
item, and we will rely on the readers’ intuitions to be able to determine which any
we are dealing with.3 As one bit of evidence that the two are different (and as one
way to help bring out the difference), note that the free-choice any can be
replaced by any old, where old is not being used literally, whereas in the NPI usage
in (6), old is good only on its literal meaning:4

(5) a. Any old cat can catch Minnie.

b. Calico can catch any old mouse.
c. Calico can catch any old mouse that dares to cross her path.

Dayal (1995) points out that free-choice any can sometimes occur even in ordinary
2

non-modal sentences if the common noun is modified by a relative clause:

(i) Last year, Calico caught any mouse that dared to cross her path.
3 While the two might be separate items, they clearly are related. Evidence for this comes

from the fact that many languages use the same or closely related item to mean something
quite similar to both free-choice any in English and also NPI any in English. See, e.g., Lahiri
(1998) on Hindi.
4
Some speakers might be able to force this into the free-choice reading, the idiomatic
reading of any-old would require this. The NPI reading claims that she caught no old mouse
(perhaps her mouse-catching abilities were restricted to young ones). The free-choice
reading—to the extent one can force it at all—claims that it is not the case that any arbitrary
mouse would become her supper.
 339
(6) (Last year), Calico didn’t catch any old mouse.

*16.1. Another way to bring out the difference between NPI any and free-choice
any is with certain verbs. Thus notice that (i) below is impossible even though
imperatives generally allow for free-choice any:

(i) *Build any solar heated house.

(Compare this to Don’t build any solar heated house which contains NPI any.)
What are the verbs that block the free-choice reading, and why? What might
you say about (ii) which is good, but only if any pie is interpreted as any kind of
pie?

(ii) Bake any pie (you want). (Compare to Eat any pie (you want).)

This exercise is intended to be open-ended; a completely precise or formal

answer to the questions above is not expected.

Thus here we will treat free-choice any and NPI any as distinct lexical items, and
will construct examples that make the free-choice reading difficult or impossible.
Notice, incidentally, that there is no free-choice analogue to ever (compare (7a)
to (7b)):
(7) a. *Porky could ever leave the farm.
b. Porky could leave the farm at any time.

16.1.2. Minimizers

In addition to any and ever, there is also a large class of NPIs known as
minimizers—these are conventionalized or idiomatic expressions indicating a
minimal amount, and are exemplified in (8)–(11):
(8) a. Farmer MacDonald didn’t lift a finger to clean out the sty.
b. *Farmer MacDonald lifted a finger to clean out the sty.

(9) a. Robin Hood didn’t keep a red cent of what he stole.

b. *Robin Hood kept a red cent of what he stole.
 340 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES
16.2. THE DISTRIBUTION OF NPIs

(10) a. The disobedient pug didn’t budge an inch when called.

b. *The disobedient pug budged an inch when called.

Some of the (b) sentences here—such as (8b)—are good, but only on a very
strange literal reading. The class of minimizers is quite large, and includes also
very informal and/or idiomatic things like give a hoot (and other more colorful
variants of hoot), know (and other verbs) squat, drink a drop, spend a dime, etc.
For at least some speakers these seem to have a more limited distribution than
any and ever do, and we will not provide detailed discussion of these. (See Krifka
1995 and Zwarts 1998 for discussion.)

16.2. The distribution of NPIs

NPIs are a rich domain for illustrating how formal semantic tools shed light on very
commonplace and idiomatic facts about language. The relevant judgments are
quite robust; the existence of NPIs (especially of the minimizer variety) is
widespread cross-linguistically; and the feeling that these are quite “slangy” make
it clear that facts about their distribution could never have been acquired as the
end result of conscious instruction. But these turn out to have a rather subtle
distribution, and it is to this that we now turn.

16.2.1. First pass

The very name Negative Polarity Item suggests that the occurrence of these items
is limited to negative sentences (see, e.g., the good sentences above). But what
exactly is a negative sentence? Note first that these occur within a VP whose
subject is a GQ of the form no N whereas other GQs don’t allow this:
(11) a. No semantics student has ever climbed Mt Lafayette.
b. *Some semantics student has ever climbed Mt Lafayette.
c. *Every semantics student has ever climbed Mt Lafayette.

(12) a. No student turned in any correct answer on the exam.

b. *Some student turned in any correct answer on the exam.
c. *Every student turned in any correct answer on the exam.
 341
Perhaps this is not particularly surprising. While we have not given any actual
definition of a negative sentence, it is not too difficult to imagine that one could be
given that includes (11a) and (12a) and not the (b) and (c) cases. For example, let
a “negative sentence” be one with no, not, n’t, etc. and its variants; clearly these
are all related both morphologically and semantically.
But this is not sufficient, for it says nothing about the contrasts in (13)–(16):
(13) a. Few semantics students have ever climbed Mt Lafayette.
b. *Many semantics students have ever climbed Mt Lafayette.

(14) a. Few students turned in any correct answers on the exam.

b. *Many students turned in any correct answers on the exam.

(15) a. At most five students have ever managed to climb Mt Lafayette.

b. *At least five students have ever managed to climb Mt Lafayette.

(16) a. At most five students turned in any correct answers.

b. *At least five students turned in any correct answers.

It is much less clear what it means to call few and especially at most five
“negative.” One might have the intuition that these all place an upper bound on the
number of (in this case) students in question; that intuition would also capture the
cases with no, for those too place an upper bound (zero). Indeed, this intuition is
on the right track. But having moved away from the notion of a “negative” sentence
to something else, it is then not so obvious how this extends the cases with simple
negation, such as (1).
The notion of an upper bound (in the above sense) also does not account in any
obvious way for the following contrasts:
(17) a. Michael doubts that Kolya will ever learn to howl on command.
b. *Michael believes that Kolya will ever learn to howl on command.

(18) a. Michael doubts that any dog in the playgroup will turn out to be part wolf.
b. *Michael believes that any dog in the playgroup will turn out to be partwolf.

(18b) may be marginal for some speakers but only on the free-choice reading
(which easily emerges if will is changed to could). Note that NPIs are also allowed
in the antecedent clause of a conditional (the clause introduced by if), but not in
the consequent clause (the main clause); this also does not obviously fall under
the “upper bound on a number” hypothesis:
(19) a. If Kolya ever learns to howl on command, he’ll be able to appear in the dog show.
b. *If Kolya learns to howl on command, he’ll ever appear in the dog show.
 342 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES
16.2. THE DISTRIBUTION OF NPIs

(20) a. If any dog in the playgroup turns out to be part wolf, they’ll close the dogpark.
b. *If they open up a new playgroup, then any dog that goes will turn out to be part
wolf.

16.2.2. Second pass

The discussion so far has considered only the distribution of NPIs within VPs. The
plot thickens once one turns to their distribution within complex nouns that
combine with quantificational determiners. (The discussion here assumes the
analysis of the structure of NPs containing relative clauses that was discussed in
section 13.2.1.) As detailed in Chapter 10, quantificational determiners like some,
no, every, etc. denote Curry’ed relations between sets.
We have already documented the distribution of NPIs within the VPs of sentences
containing these determiners: no, at most three, and few all allow these; some,
many, at least three, and every do not. Consider now what happens within the N
portion of the relevant sentences. First, no, at most three, no more than three, and
few all allow NPIs in that environment too:

(21) No student who has ever climbed Mt Lafayette found clear and sunny weather at
the top.
(22) Few students who have ever climbed Mt Lafayette found clear and sunny weather
at the top.
(23) At most three students who have ever climbed Mt Lafayette found clear and sunny
weather at the top.
(24) No more than three students who have ever climbed Mt. Lafayette found clear and
sunny weather at the top.

In view of this, it is rather unsurprising that some, many, and at least three also do
not allow NPIs in the N portion, just as they do not allow them in the VP portion:

(25) *Some student who has ever climbed Mt Lafayette found clear and sunny weather
at the top.
(26) *Many students who have ever climbed Mt Lafayette found clear and sunny weather
at the top.
(27) *At least three students who have ever climbed Mt Lafayette found clear and sunny
weather at the top.
 343
(All of these facts can be repeated with any.) The discussion above seems a bit
mundane: the determiners treat NPIs within the N portion exactly the same as
within the VP portion.
But in fact not all determiners treat the N and the VP alike; every shows an
asymmetry. While an NPI is not allowed in the VP, it is fine within the N, and
similarly with any:

(28) *Every semantics student has ever climbed Mt Lafayette.

(29) Every student who has ever climbed Mt Lafayette said it was sleeting at the top.
(30) *Every student turned in any correct answer on the exam.
(31) Every student who turned in any correct answer on the exam passed the course.

16.3. The first breakthrough

16.3.1. Initial characterization of the environments

The first task, then, is to come up with a full characterization of the environments
which allow NPIs (or, to use the common terminology, the environments which
“license” NPIs). Early work on this within the history of generative grammar—
which was done in a period in which there was little attention to semantics and
more to syntax—assumed that the grammar contained some feature on “NPI
licensors” which said that an NPI was allowed if the NPI was in a certain syntactic
relation to the “licensor” (see, e.g., Klima 1964). Of course this leaves it as a purely
accidental and stipulative property as to what items bear this feature (as well as
what items are NPIs). Obviously there should be some explanation as to why
some items allow NPIs and some don’t, and clearly the best place to look would
be some facts about meaning. In any case, the asymmetry with every (and, for that
matter, with if) casts doubt on this solution; we need to ensure that every allows
an NPI within the N that it combines with, but not within the VP. One might think
that this is easily solved by taking the relevant syntactic relation to be c-
command: every c-commands the material within the N but not within the VP. (For
the definition of c-command, see Chapter 15, n. 4.) But then all other determiners
such as no should also allow an NPI in the N and not in the VP, since the syntax is
the same.
 344 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

Indeed, the asymmetry with every gives an indication that the story is about
meaning and not some arbitrary feature. For there is another relevant way in which
the meaning of every differs from that of the other quantificational determiners
discussed in the last section: the others are all symmetric (i.e., commutative) and
every is not. To clarify, we temporarily remove many and few from the mix (they
require further discussion, postponed until 16.3.2). Thus consider some, no, at
least three, and at most three.
These all denote Curry’ed relations between sets of individuals—relations which,
in all cases, are symmetric. For any two sets A and B, if [[some]](A)(B) is true then
[[some]](B)(A) is also true. One can convince oneself of this in two different ways.
The first is to note that the two sentences Some huskies howl and Some howlers
are huskies entail each other. But a note of caution: it can be dangerous to check
out these sorts of semantic properties by considering actual sentences because
two sentences with the same truth conditions can have different implicatures,
which can make this method misleading. (In this case these sentences do in fact
have different implicatures; the first implicates that there are also some non-
howling huskies but has no implicature one way or the other about other possible
howlers; the second implicates that there are other howlers but makes no
implicature about non-howling huskies.) A better method is to simply consider the
semantics of [[some]]; it applies to two sets A and B and returns true if there is a
nonempty intersection. It of course then makes no difference as to what order
these sets are taken. Similar facts hold for no, at least three, and at most three;
these are all symmetric. [[every]], on the other hand, is quite obviously not
symmetric: it is just the subset relation. The fact that it differs from the other
determiners in just this way—and is also the only one which exhibits the
asymmetry between the N argument and the VP argument—lends initial credence
to the supposition that the distribution of NPIs has nothing to do with a syntactic
restriction but follows in some way from the semantics.
The first major breakthrough on this came from insights in Fauconnier (1975)
and Ladusaw (1979); the latter in turn drawing on work of Barwise and Cooper
(1981). We initially give the insight informally, formalizing it in the next section.

(32) Definition of upward-entailing environment (to be refined in 16.4.1):

Take any sentence S containing an expression which denotes a set
B. Consider any expression ª whose denotation C is a superset of B, and consider a
sentence S’ in which is replaced by ª. Then if S entails S’, the environment in which
occurs is said to be an upward-entailing environment.
 16.3. THE FIRST BREAKTHROUGH 345

The definition here is awkward as it is phrased in terms of the substitution of

linguistic expressions rather than being given purely in terms of semantic
properties. In fact, as noted above, using linguistic expression substitution can be
misleading since, in constructing sentences where one substitutes in one
expression for another, one needs to be careful to sort out the implicatures and to
make sure that these are not disguising the actual entailments. But we define it
this way initially for expository convenience. The idea is that the environment is
upward-entailing if substituting in a superset is destined to preserve truth. As one
example, consider the VP in a simple sentence such as Sally danced. The set
[[dance]] is a subset of the set [[move]], and Sally danced entails Sally moved. An
ordinary VP then (one not containing negation) is an upward-entailing
environment. We can check this in terms of the sentences themselves (as we just
did) but we can also simply think about the semantics. The truth conditions for
any sentence of the form [NP VP] are (in set terms) that [[NP]] is a member of the
[[VP]] set. Obviously if this is true, that individual will also be a member of any
superset of the original [[VP]].
Now consider sentences with quantified determiners such as some N VP. Such
a sentence is true if and only if there is a non-empty intersection of
[[N]]S and [[VP]]S. Clearly if that is the case, then expanding the N set to something
bigger will preserve truth—if there was something in the intersection of the original
sets, it will still be there when we expand the N set. The same of course is true of
the VP set. So both the N and the VP positions are upward-entailing positions. This
will also be true for all determiners that set a lower bound on the size of the
intersection. So at least three N VP is also upward-entailing on both the N and the
VP position by the same logic. The same is true for three N VP as we are assuming
that three and at least three mean the same (and just have very different
implicatures; see the discussion at the end of section 10.4).

16.2. The point was made above by consulting pairs of actual sentences to
determine whether an environment is upward-entailing can be misleading
because it is easy to imagine relevant contexts for sentences and to thus draw
implicatures; if these are mistaken for the truth conditions the pattern can be
obscured. Assuming that three really does mean “at least three,” construct
examples to show this point, explaining exactly why folding in the implicatures
could lead one astray.
 346 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

We can also define the reverse notion of a downward-entailing environment:

(33) Definition of downward-entailing environment (to be refined in 16.4.1):

Take any sentence S containing an expression which denotes a set B. Consider any
expression Æ whose denotation A is a subset of B, and consider a sentence S’ in
which is replaced by Æ. Then if S entails S’, the environment in which occurs is said
to be a downward-entailing environment.

The idea here is that when the substitution of a subset is guaranteed to preserve
truth, the environment is downward-entailing. Take then the case of a sentence of
the form no N VP. This is true if the intersection of the N set and the VP set is empty.
Obviously, then, shrinking either set to some subset will not add anything more to
the intersection (it can only remove material from the intersection), and so
substituting in a subset in either position will preserve truth. Hence sentences of
this form are downward-entailing on both positions. The same holds true for
determiners like at most three. Again, if there are at most three things in the
intersection of an N set and a VP set, then shrinking either one will not change this
fact. This too, then, is downward-entailing on both the N position and the VP
position. Notice that an environment need not be either upward- or downward-
entailing. Consider exactly three N V. This says that the cardinality of the
intersection of the N set and the VP set is exactly three. Adding to the N set could
change this; if what the individual added happened also to be in the VP set then
the cardinality of the intersection is now greater than three. Subtracting something
from the N set could have the same bad consequence; we could have reduced
the cardinality of the intersection. So the N position is neither upward- nor
downward-entailing. The exact same logic holds for the VP position; it is neither.

16.3. Calculate whether the N position is upward-entailing,

downwardentailing, or neither in sentences of the form [most N VP]. Do the
same for the VP position. Once again it will be best to do this by thinking of the
meaning of most rather than constructing sentence pairs, to ensure that no
irrelevant factors are entering into the calculation of the truth conditions. The
meaning of most was discussed in sections 10.1.3 and 10.2.3.

A clear pattern is emerging: it appears that NPIs are allowed in the downward-
entailing positions—a proposal made originally in Ladusaw (1979). Thus no
allows NPIs in both the N and the VP, as does at most three. At least three, three,
and some all ban NPIs in both positions.
 16.3. THE FIRST BREAKTHROUGH 347

A caveat: on the basis of just this data, the requirement could instead be that NPIs
are allowed in non-upward-entailing positions. The difference concerns the
predictions regarding sentences like ?Exactly three students who have ever
climbed Mt Lafayette saw sun at the top, and ?Exactly three students have ever
climbed Mt Lafayette. As noted by the ? at the beginning of these, the judgments
actually seem to be intermediate; and speakers vary on these. We will bow here
to the common wisdom in the literature that the relevant environments are
downward-entailing environments rather than non-upward-entailing ones, and
will leave the intermediate status of the sentences above as an unsolved problem.
However, it will turn out that the solution to be discussed in section 16.4 can easily
be modified if these sentences are basically good, where their slight degradation
is a result of some other factor.

16.4. Sentences with most are also relevant here. In light of your answer to
(16.3) what are the different predictions with respect to most that are made by
the two hypotheses: (a) NPIs are allowed only in downwardentailing positions,
and (b) NPIs are allowed in non-upward-entailing positions? Construct the
relevant sentences and see what the facts actually are (they might be unclear).

What about the case of every? The beauty of the above hypothesis regarding
the distribution of NPIs is that the asymmetry with respect to every is completely
unsurprising. For every is in fact downward-entailing on the N position and
upward-entailing on the VP position. A sentence of the form every N VP is true if
the N set is a subset of the VP set. Obviously any set B which a subset of the N set
is still a subset of the VP set, so shrinking the N set to a subset is guaranteed to
preserve truth. By the same logic, expanding the VP set to a superset is also
guaranteed to preserve truth. And so the above hypothesis predicts that NPIs can
occur within the N but not within the VP, and this is just what we find.
Of course, the point of departure in this chapter was the observation that NPIs
occur in negative sentences—this is, after all, how they got their name.
And so one might wonder whether that observation is also accounted for. The
answer, happily, is yes. We have posited that [[didn’t]] combines with a set to
return its complement. Thus in a sentence of the form NP didn’t VP, is the VP a
downward-entailing environment? Indeed it is. Such a sentence is true if the
 348 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

individual denoted by the NP is in [[didn’t VP]] S which in turn means that that
individual is in the complement of [[VP]] S (equivalently, that individual is not in
[[VP]]S). If an individual is not in [[VP]] S then it is not in any subset of [[VP]]S either
(put differently, i.e., that individual is in the complement of any subset of [[VP]] S).
So the ability of NPIs to occur in ordinary negated VPs is unsurprising. Recall that
our treatment of negation, which involved taking didn’t as an unanalyzed whole, is
a placeholder for a more sophisticated treatment of not or n’t as combining with
auxiliaries, but the same point will hold with a more sophisticated treatment. 5
The pattern with conditionals falls into place as well, although readers should
be warned that there are aspects of the semantics of conditionals being
oversimplified here (see, e.g., Kratzer 1986). But the oversimplified outline is
nonetheless instructive. Hence consider any sentence of the form if S1 then S2. We
can assume that a successful account of the semantics of the
English conditional will predict the following. Let S 3 entail S1. Then if a sentence
such as if S1 then S2 is true, a sentence of the form if S3 then S2 is also true.6 Let us
recast this in set terms using the fact that the value of any S can be seen as a set
of possible worlds. Then, if S3 entails S1, that means that the set of worlds in which
S3 is true is a subset of the set of worlds in which S 1 is true. Hence, substituting in
a subset in the antecedent position of a conditional preserves truth; and so this is
a downward-entailing environment.

5
The discussion here surrounds only the case of NPIs within the VP. See Exercise
16.10 for discussion of NPIs in subject position.
6 Truth be told, we are oversimplifying and glossing over a full set of interesting questions

concerning the conditional in English. To give a flavor for this, note that You struck a match
and it was raining certainly entails You struck a match. But, as discussed in Stalnaker
(1968), the following entailment pattern does not hold:
(i) If you strike a match, it will light.
(ii) If you strike a match and it is raining, it will light.
(i) seems true and (ii) does not, and thus apparently (i) does not entail (ii). See Stalnaker
(1968) and many since for proposed solutions.
 16.3. THE FIRST BREAKTHROUGH 349

16.5. Putting this in set terms as above, show that the consequent clause of a
conditional (the then clause) is an upward-entailing environment.
16.6. Again, thinking of propositions as sets of worlds, show that (under a
certain set of assumptions about belief states) the complement of believe (as
in (17b)) is an upward-entailing position, while the complement of doubt (as in
(17a)) is a downward-entailing position. What are the assumptions one needs
to make about peoples’ belief states in terms of the rationality of the believer?
*16.7. Consider the following:

(i) a. Herman denied that he committed any acts of harassment.

b. Herman denied that he had ever harassed anyone.

Does the fact that NPIs occur here follow from the generalization regarding
downward entailment? In what way does a fine-grained analysis of the
meaning of deny bear on the answer?
16.8. Consider the following:

(i) Fewer than half of the bears have ever seen the other side of the
mountain.

Is the ability of ever to occur here surprising or predicted? What about the
occurrence of ever in (ii), where it seems (to this author at any rate) possible
but not perfect:

(ii) ?Fewer than half of the undergraduates who ever took a formal
semantics course went on graduate school in linguistics.
The story does not end here, of course. First, one would want to know why this
pattern holds. Second, so far we have only a descriptive generalization; the
question remains as to how to build something into the grammar from which this
generalization follows. These questions are the subject of 16.4. Moreover, there
are other places where NPIs occur; see the discussion in 16.5. But first we return
to many and few, discussion of which was postponed earlier.
16.3.2. A note on many and few

At first glance, many and few appear to be neither upward- nor downwardentailing
on the N position. Consider the pairs in (34) and (35):
(34) a. Many huskies dig holes in the snow to sleep.
 350 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

b. Many dogs dig holes in the snow to sleep.

(35) a. Few dogs like to pull sleds.

b. Few huskies like to pull sleds.

A common intuition is that (34) does not entail (35); huskies do in fact like to dig
holes in the snow to sleep, something which is probably not a general property of
dogs. So (34) appears true and (35) false. This would mean that many is not
upward-entailing on the N position (nor is it downward-entailing; Many dogs like
to play fetch certainly does not entail Many huskies like to play fetch, for it could
well be that all of the fetch-enjoyers are dogs of other breeds). Conversely, (35a)
appears not to entail (35b); dogs in general are not sled pullers but huskies are,
and so it would seem that (35a) is true and (35b) is not. By this reasoning, few is
not downward-entailing on the N argument (nor is it upward-entailing, as the
reader can easily verify). The reason why these appear to be neither upward- nor
downward-entailing is clear on reflection: they look like proportional determiners
(like most, more than half, etc.). That is, it appears that the number of individuals
who dig holes in the snow to sleep is—in (34a)—dependent on the number of
huskies, while in (34b) it is dependent on the number of dogs. But if it is true that
many and few are proportional determiners, the distribution of NPIs is puzzling.
For few happily allows these in the N position, while many does not:
(36) a. *Many students who have ever climbed Mt Lafayette stopped at the
Greenleaf hut.
b. Few students who have ever climbed Mt Lafayette stopped at the Greenleaf hut.

But if NPIs are restricted to downward-entailing environments, their ability to

occur in (36b) seems surprising. If the restriction is actually non-upwardentailing
environments then (36b) is unsurprising, but the (quite robust) badness of (36a) is
unaccounted for.
Once again though, the fact that these are apparent counterexamples is due to
the difficulty of sorting out the full set of truth conditions of a sentence from other
factors that enter in. The question at issue is simple: are many and few truly
proportional or is something else going on? As it turns out, the situation is similar
to the situation with apparent non-intersective adjective modifiers discussed in
section 8.4.4. Recall that the meaning of big seems to involve a “free variable” over
sets of individuals, so that—in variable-ful terms—[[big]] means something like
“big for a Dn,” where the value of Dn is set by context. (In variable-free terms, [[big]]
is ºD[big for a D] where the open argument slot is passed up.) A reasonable
 16.3. THE FIRST BREAKTHROUGH 351

hypothesis is that something similar is going on with many: what counts as “many”
depends on setting some contextual parameter, and this is often but not always
given by the size of the N set. In other words, the claim here is that many means
(informally) “greater than nn,” where nn is a “free variable” over numbers, and its
value is determined by a variety of contextual factors. (Again the requisite
translation of this into the variable-free setting is straightforward. We will continue
to use the terminology of a “free variable” simply because this allows short cuts
in the exposition.) If we hold nn constant, then (34a) does in fact entail (34b) and
similarly (35a) entails (35b). But the actual sentences are misleading, because
the use of dog vs husky as the noun that many combines with pressures us to set
a very different value for n in the two sentences. This is analogous to the pressure
to take big quite differently in big elephant and big flea; there is a strong preference
to use the head noun to set the value of Dn in “big for a Dn,” but this is not required.
Is there independent evidence that this line of reasoning is correct? Indeed
there is, and it comes from the following clever case discussed in Bennett (1974)
and Partee (1989a) (although Partee’s conclusion in that paper is actually
somewhat different from that discussed here). The relevant context is a
discussion of, for example, the dating habits of students at Brown (with respect to
other Brown students), and someone says:

(37) (At Brown), many women date many men.

The interest in this example is that the cardinality of the women set of Brown
students is roughly the same as the cardinality of the men set of Brown students.
Yet what we take to be the number that would satisfy this for the women case is
quite different than for the men case. Of course many
 352 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

is also vague—just exactly how to zero in on a precise value for n n is unclear even
once we fix the context. But this does not obscure the basic point here; one might
have the intuition that, for example, (37) requires about 300 women to be multi-
men daters. But surely, for each such woman, she does not need to date 300 men
in order for the sentence to be true!
The conclusion, then, is that many is not really a proportional determiner
(in the sense that, e.g., most is).22 The size needed to count as many for the subject
NP in (37) is different from that of the size needed for the object NP, and so its
value can but need not be set at some proportion of the size of the N set. The value
of n depends on a variety of contextual factors, one of which could happen to be
the size of the N set. As noted above, once we keep n constant, then if (34a) is
true, so is (34b). Or, put differently, in any world in which (38a) holds then so does
(38b), holding nn constant:
(38) a. | husky’ \ digs-holes-in-snow’| > nn

b. | dog’ \ digs-holes-in-snow’| > nn

Surely if (a) is true, then any way to expand the husky set into a superset will still
be true. Exactly the same logic holds in the few case. These, then, turn out to be
perfectly well-behaved, and the downward-entailing generalization holds for
many and few.

16.4. But why? And how?

Still, there are two unsatisfactory aspects with what has been accomplished so
far. First, we have stated a descriptive generalization to the effect that NPIs are
allowed only in downward-entailing environments, but have formulated no actual
principle to ensure this. It is interesting to note that sometimes the descriptive
generalization itself is treated as the actual principle in grammar: one often sees

22 The argument can be reproduced with few. But strictly speaking, the above shows
only that many and few have a non-proportional (cardinal) reading but not that they don’t
also have proportional readings. It is possible that both are ambiguous (see Partee 1989a)
but this does not affect the point here. As long as there is a cardinal reading for few it
follows that NPIs will be licensed by few; and even if there is a proportional reading for
many, then there will be no licensing of NPIs if NPIs exist only in downward-entailing
environments.
 16.4. BUT WHY? AND HOW? 353

work that assumes (implicitly or explicitly) that the grammar contains a principle
like the following:

(39) An NPI is allowed in a sentence only if c-commanded by a downwardentailing

operator.

(See Chapter 15, n. 4 for the definition of c-command.) Notice that no constraint
like this could be incorporated into the sort of Direct Compositional architecture
assumed here. (39) is a filter on trees (i.e., on syntactic representations), and we
have speculated that the grammar itself does not “see” trees and therefore cannot
contain principles that refer to them. And even in a theory which does
countenance filters of this sort, (39) is decidedly odd. For it mixes syntax and
semantics: downward-entailing is a semantic notion, c-command is a syntactic
notion. (39), then, is no ordinary filter: it is not simply a constraint on syntactic
representations. Rather it refers also to the meaning of certain bits of the
representation, and thus must be seen as a filter that blocks
representation/meaning pairs. It is thus not only stipulative, but if fully fleshed out
would be quite complex. And so the question remains as to how it is ensured that
NPIs have the distribution they do.8
Second, even considering (39) (or something like it) to be a correct descriptive
generalization, one asks why it should be so. For we have defined downward-
entailing in terms of sets and subsets, and one might well wonder why these
notions should have any relevance whatsoever for the distribution of some item.
And what is the meaning of the NPIs? Ultimately one should hope that the
constraint on their distribution follows from something about their meaning
(possibly combined with other principles). The next task, then, is to recast the
notion of downward-entailing in a more illuminating way. We then turn to the
meaning of (some) NPIs, and explore one influential proposal that makes progress
towards answering the question of why NPIs have the distribution that they do.

8
An argument that is sometimes given for (39), which refers to the syntax, is the fact
that NPIs (apparently) do not occur in subject position. As should be clear from Exercise
16.10, a purely semantic description of the licensing environments does not seem account
for this fact. However, it is not entirely clear that NPIs really cannot appear in subject
position, and there are theories that the badness of the simple cases (such as *Any mouse
didn’t chase Calico) have a pragmatic explanation. See de Swart (1998) for such an
account.
 354 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

16.4.1. The insight from semantic strength

The key intuition (already in Fauconnier 1975) is that downward-entailing

environments are environments where semantic strength is reversed. Semantic
strength itself is a notion about the amount of information conveyed. Thus we
begin by giving a recursive definition of the notion of semantic strength. The base
step (40a) concerns the strength of propositions. Recall that a proposition is a set
of worlds, and so a proposition p is at least as strong as a proposition q if p is a
subset of q. The intuition is that a speaker uttering a proposition p is (assuming
normal communication) telling a hearer that p is true in the actual world, thus
narrowing the space of worlds we might be living in. A stronger proposition gives
more information; it narrows the space of worlds more. Put differently, it is
“harder” for a stronger proposition to be true. We thus give a recursive definition
of strength in (40b) (for convenience, we give the definition of “at least as strong
as” rather than “stronger than” although it could easily be done either way):
(40) a. A proposition p is at least as strong as a proposition q if p q. (Thus p gives at least

as much information.) We will notate this as p q.

b. For any two functions f and g (whose ultimate result is 1 or 0), f g iff for all x, f(x)

g(x).

Note then that [[dances]] is at least as strong as [[moves]]. In set terms, for any
world w, [[dances]]wS is a subset of [[moves]]WS. Or, to put this in terms of the
functions: [[dances]] [[moves]] because for all x, [[dances]](x) [[moves]](x), and
the latter holds because for all w, [[dances]](x)(w) [[moves]](x)(w). (From now on
we will generally omit the world argument, as has been done throughout most of
this text.) The intuition can again be recast in terms of information. If one VP
characterizes a subset of another, the first is “harder” to be true of an individual,
and so to know that, for example, Sabrina dances is to know more than to know
that Sabrina moves.
From here, we can give a general definition of an upward monotone or
downward monotone function (the informal definition of downward- and upward-
entailing earlier are just special cases of this).
(41) Consider any function F whose ultimate result is t, and which takes as argument a
function (of any type whose ultimate result is t).
a. F is upward monotone iff for all f and g, if f g then F(f) F(g).

b. F is downward monotone iff for all f and g, if f g then F(g) F(f).

 16.4. BUT WHY? AND HOW? 355

(Note that both F above and the functions which are its argument must ultimately
have a truth value as their ultimate result; this is because is defined only for such
functions.)
We illustrate with every, which was the interesting case. In the earlier informal
discussion we spoke about the downward entailingness of the VP position in a
sentence of the form every N VP. Technically, though, it makes no sense to talk
about the VP “argument” of every; in a sentence like every ballerina danced, the
VP is actually an argument of every ballerina. (Recall that [[every]] is the Curry’ed
relation between sets, and is of type <<e,t>,<<e,t>,t>>.) So now consider [[every
ballerina]]. The definition in terms of sets and supersets showed that this is
upward-entailing (if some set S is in [[every ballerina]] then so is any superset of
S). But the intuition can also be phrased in terms of semantic strength. As shown
above, [[dance]] is stronger than [[moves]]. And in fact, it is “harder” for every
ballerina danced to be true than for every ballerina moved to be true. So [[every
N]] (for any N) is upward monotone; it preserves the strength direction. But every
itself reverses the strength direction (this is what it means to say that it is
downward monotone on its N argument). [[husky]] is stronger than [[dog]], but
[[every dog]] is stronger than [[every husky]]. It is easy to see this: for any set S, it
is “harder” for S to be in [[every dog]] than to be in [[every husky]]. To show this
more formally using the recursive definition: [[every]] is downward monotone
because for any two functions f such that f g, [[every(g)]] [[every(f)]]. This follows
from the recursive definition of . Given two f and g such that f g, [[every]](g)
[[every]](f) because for all functions h (of type <e,t>), [[every]](g)(h) [[every]](f)(h).
16.9. (i) Show that for any N, [[some N]] is upward monotone. Show that
[[some]] is upward monotone.
(ii) Show that for any N, [[no N]] is downward monotone. Show that
[[no]] is downward monotone.
(iii) We took the lexical meaning of honest to be a function of type <e,t>,
and we will use honest’1 as a notation for that function. However, the
lexical item honest can also shift to the prenominal version, of
category N/RN and with meaning ºP<e,t>[ºx[P(x) & honest1’(x)]]. Call
this meaning [[honest2]]. Is [[honest2]] upward monotone,
downward monotone, or neither? Show exactly how you arrive at
the conclusion you arrive at. (Ignore the possibility that honest—like
big—might be relativized to
some contextually salient comparison class.)
 356 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

*16.10. There is a complex set of questions surrounding the ability of any within
a subject that combines with a negated VP. Common wisdom is that they don’t;
cases like (ia) are bad (compared to (ib)):
(i) a. *Any mouse didn’t chase Calico.
b. Calico didn’t chase any mouse.

We have yet to give any semantics for expressions of the form any N (this will
be done in 16.4.2.2) but assume for the purposes of this exercise that these are
indeed GQs. That is, in the CG syntax they are of category S/(S/NP) with
meanings of type <e,<e,t>>.
One might then think at first glance that the badness of (ia) follows because
these are not strictly speaking in downward monotone environments. For the
VP is an argument of the subject, not vice versa, thus the definition above does
not include the subject. Unfortunately, this line of reasoning is not sufficient to
explain why (ia) is bad. Why not?
Incidentally, the oft-cited generalization that NPIs do not occur in subject
position is not really correct; there are well-known good cases like (ii) where
the NPI is embedded within the subject.
(ii) A book with any torn pages will not be found in this library.

See de Swart (1998) for a pragmatic explanation for the contrast between (ia)
and (ii).

It is now reasonable to speculate that what is important about

downwardentailing or downward monotone environments is that they reverse
semantic strength (or informativeness). Perhaps, then, there is something about
the meaning of NPIs which makes them odd or nonsensical in environments
where the strength is not reversed.

16.4.2. The second breakthrough: Kadmon and

Landman on any

And indeed this hypothesis is explored both in Kadmon and Landman (1993) and
in Krifka (1995). The two accounts differ, but both aim to derive the distribution
from the strength reversal generalization combined with the meanings of NPIs
(and with possibly other pragmatic principles). Here we explore only the former
proposal; the interested reader will be able to consult Krifka (1995) on their own
 16.4. BUT WHY? AND HOW? 357

after reading Chapter 18. Although we will not follow Kadmon and Landman in all
details, the discussion below is based on their general idea.
Recall the discussion in section 10.6 regarding the possibility that determiners
come with a slot for a domain restriction. (Readers who initially skipped that
section should consult it now.) This forms the background for the Kadmon and
Landman proposal. They suggest that any actually has almost the same meaning
as some, but contains as part of its meaning the notion that it is a domain widener.
That is, its meaning ensures that the relevant individual can be drawn from a
domain wider than the normal contextually salient domain. While Kadmon and
Landman do not actually formally write out the meaning for any and do not discuss
the issues surrounding domain restriction, one conceivable way to formalize their
proposal is to suppose that the meaning of some is as in (42a) (in variableful
terms) or (42a’) (in variable-free terms) and any is as in (42b or b’), where D is a
variable over sets of individuals:
(42) a. [[some]] = ºQ[ºP[QS \ D \ PS 6¼ ]] (using a free variable) a0. [[some]] =

ºD[ºQ[ºP[ QS \ D \ PS ¼6 ]]] (in variable-free terms)

b. [[any]] = ºQ[ºP[∃D’[D D’ & QS \ D’ \ PS 6¼ ]]] (using a free variable) b 0. [[any]] =

ºD[ºQ[ºP[∃D’[D D’ & QS \ D’ \ PS ¼6 ]]]] (in variable-free terms)

Thus (42b) gives any a meaning such that any student, for example, is ultimately
interpreted with respect to some contextually salient domain (exactly as is the
case for some student). But any says more: it says that the student in question
which ultimately satisfies the three-way intersection can be drawn from a wider
domain than the obvious one. It is also necessary to assume that the contextually
salient domain (the one which is widened) is never set to the widest possible
domain.23
Now consider the meaning of any pig vs the meaning of some pig. To illustrate
take the following two sentences, whose truth conditions are given below each
sentence (we give these using the free variable D rather than in variable-free terms
simply because this is easier to read and is fine for informal exposition):
(43) a. Some pig ate this morning. a 0. pig’S \ D \ ate-this-morning’ 6¼

23 This is because of the way we have formalized the meaning. If we assume that D can
be the widest possible domain, then a sentence using any could be false simply because
there is no way to widen D. This result is not intended as part of the Kadmon and Landman
proposal.
 358 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

b. *Any pig ate this morning. b 0 ∃D’[D D’ & pig’S \ D’ \ ate-

this-morning’ 6¼ ]

(43b) is of course bad, but—assuming that [[any]] is as given above—we know

that it should have the meaning shown in b’. This says that there is a D’ wider than
the normal D such that a pig drawn from D’ ate. Crucially, this is a weaker
statement than (43b). The claim that one can find a wider domain than the
contextually salient one is trivially true (given the assumption made above that D
is never the widest possible domain) and so this part has no effect on the truth
conditions. But (43b) also says that there’s a pig drawn from this wider domain
who ate. Thus if (43b) is true then (43a) is true, and so (b) is a weaker statement
than (a).
When would widening the domain strengthen a statement? In just those
environments where the strength directions are reversed: i.e., in monotone
decreasing environments. Let us check this with a couple of examples.
Consider the pair in (44):
(44) a. Every farmer who fed some pig decided not to then eat bacon for breakfast.
b. Every farmer who fed any pig decided not to then eat bacon for breakfast.

As discussed in Chapter 13, relative clauses denote sets; [[who fed some/any
pig]] is the set of feeders of a (domain-restricted) pig. Given the semantics of
relative clauses and of GQs in object position, [[who fed some pig]] is the set of
individuals x such that there is a pig out of some domain D that x fed. By the same
token, [[who fed any pig]] is the set of individuals x such that there is a pig out of a
wider domain D’ that x fed. In both cases this intersects with the set of farmers,
giving the following (informally shown) meanings to the two complex Ns at issue:
(45) a. [[farmer who fed some pig]] S = {x| x fed a pig in D}
b. [[farmer who fed any pig]] S = {x| x fed a pig in D’, where D’ is wider than the
contextually salient D}

Note that it is “easier” to be in the (b) set than in the (a) set; (a) is thus stronger
than (b). But every reverses strength direction, and so (44b) is a stronger
statement than (44a). Put differently, [[every farmer who fed any pig]] is a stronger
function than [[every farmer who fed some pig]]. If, however, we were to take
farmer who fed some/any pig as the N argument of some, we can see that the
version with any is weaker than the version with some.

16.11. Show that [[no farmer who fed some pig]] is weaker than [[no farmer
who fed any pig]].
 16.4. BUT WHY? AND HOW? 359

And so this brings us to the second key part of the Kadmon and Landman
proposal, which is that there is some principle as follows: any is allowed only if it
strengthens a statement. This is simply a stipulation here: what remains to be
explained is both why this should be true and from what this requirement follows.
But before elaborating on these points, we turn to ever to show that the basic
proposal can be generalized.

16.4.3. ever

Although Kadmon and Landman did not deal with ever, the proposal extends
rather naturally to ever. In fact, while some authors have had the intuition that
domain widening is not quite the correct analysis for any (see, e.g., Krifka 1995),
the parallel story with ever seems very plausible. To consider the semantics of
ever, something needs to be said about the semantics of tense in English. As a
first pass, past-tense sentences seem to existentially quantify over points in time.
Thus, (46) means something that can be represented as (47), where t is a variable
over moments in time, “now” is of course the speech time, and < means “earlier
than”:

(46) Strawbuilder went to market.

(47) ∃t[t < now & Strawbuilder go to market at t].

But in general the understanding of sentences is restricted to some

contextually salient window of times. This once again is illustrated by an example
due to Partee (1984) (although Partee actually used this example for a somewhat
different point). Suppose my husband and I are in the car beginning a four-day
vacation, and I turn to him and say (48):

(48) Uh oh. I didn’t turn off the stove.

Here there is a past tense (which we are assuming involves existential

quantification over times in the past) and a negation. The question thus arises as
to whether the meaning is appropriately represented in (49a) (with the negation
scoping over tense) or (49b) in which the scopes are reversed (note that we are
not concerned here with how the compositional semantics puts things together,
so these formulas are just intended as a way to notate the final meaning):

(49) a. ~∃t[t < now & I turn off the stove at t]

b. ∃t [t < now & ~[I turn off the stove at t]]
 360 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

Clearly (49b) is not correct. Of course there is some time at which I didn’t turn off
the stove, but this is hardly the point of the comment in (48). (49a) is closer: it’s
saying there is no time at which I turned off the stove. But that can’t quite be right
either; I could have turned off the stove many times in the past. We care only about
the fact that within some relevant interval, there is no time at which I turned off
the stove. So (49a) is fine, provided that the domain of quantification over times is
domain-restricted to some salient set of adjacent times (i.e., some interval). Thus,
as with quantifiers in general, quantification over times is arguably domain-
restricted.
What is the meaning of ever? It seems fairly clear that ever broadens the
interval of times at which the truth of a sentence is being evaluated. Consider, for
example, the contrast in (50):
(50) a. I haven’t climbed Mt Chocorua.
b. I haven’t ever climbed Mt Chocorua.

(50) is interpreted with respect to a much narrower interval. We could be

discussing my hiking activity this summer, and I say (50a) where it is clearly
understood as my not having climbed Mt Chocorua during this last hiking season.
But (50b) broadens that to a much greater temporal window (in fact one might
argue that ever broadens to the widest domain possible; we leave it open as to
whether or not that is correct).
The Kadmon and Landman account of any thus does extend naturally to ever:
ever widens the temporal domain at which a sentence is evaluated and we can
assume that there is some restriction here too that widening is not allowed unless
it also strengthens. As with the case of any, the domainwidened interval results in
a stronger sentence exactly in those environments in which the general strength
is reversed—i.e., in downward monotone environments.
 16.4. BUT WHY? AND HOW? 361

16.12. Show that (ia) is stronger than (ib), and that (iia) is weaker than (iib).
(Note that of course (iia) is bad, but assuming that it were good and that ever
only widens the temporal domain, one can show the point.) Do the same with
the pairs in (iii) and (iv) (this is relatively straightforward). To show this for (i) and
(ii) it is best to construct contexts in which the sentences without ever have a
clear restriction to some temporal interval, as was done in the text with respect
to (50).

i. a. Everyone who has ever climbed Mt Chocorua enjoyed the view.

b. Everyone who has climbed Mt Chocorua enjoyed the view. ii. a.
*Someone who has ever climbed Mt Chocorua enjoyed the view.
b. Someone who has climbed Mt Chocorua enjoyed the view.
iii. a. I have not ever climbed Mt Chocorua.
b. I have not climbed Mt Chocorua. iv.
a. *I have ever climbed Mt Chocorua.
b. I have climbed Mt Chocorua.

16.5. Open questions and further discussion

While this account goes quite far, there are many details left incomplete in the
Kadmon and Landman proposal, and which will also be left incomplete here. It is
still an open question as to just what the strengthening requirement on any and
ever follow from, or how to even exactly state this requirement. Is it encoded into
the lexical entry for any? Kadmon and Landman propose that indeed it is, but they
do not spell out just what the lexical entry actually looks like so as to enforce this
requirement.10 Note that

10
Notice that the requirement would have to be encoded in such a way that it can be
satisfied not directly when any combines with its sister, but much later in the combinatorics
(or, to use representational terms to make the point, “higher in the tree”). In other words,
consider a sentence like (i):
(i) Every farmer who fed any pig decided not to eat bacon.
Here the requirement of downward monotonicity is satisfied in virtue of the complex
N farmer who fed any pig occurring as argument of every. But any does not “know” when it
combines with pig that it will ultimately end up in such an environment. Thus if this
ultimately is encoded into the meaning of any, it has to stay “active” so that it projects to
larger expressions—in this case to the expression farmer who fed any pig.
 362 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

16.5. OPEN QUESTIONS AND FURTHER DISCUSSION

the meaning spelled out in (42b–b0) gives widening, but says nothing about
strengthening. Does this requirement follow from some pragmatic principle governing
competing utterances? One might think that the latter is quite reasonable—one can
imagine a principle to the effect that a speaker should not make a weaker statement
when a stronger statement is possible. But this is not obviously correct: it could well
be that the weaker statement is the only one that the listener is sure of. In other words,
it is not obvious that a pure pragmatic principle will explain the anomaly of Any pig ate
this morning. If
I am sure that some pig ate, but am not sure that that pig comes from the obvious
contextually salient domain (say, our farm), I might be able to say nothing more than
the sentence above. Yet it is still unacceptable.11
So let us recap the story told in this chapter. It was pointed out that early accounts
of the distribution of NPIs do nothing more than posit a syntactic feature on some
items with a stipulation that NPIs must occur in some syntactic relationship to items
with this feature. This was discarded for a semantic characterization of the
environments. We briefly entertained the possibility that the grammar instead
contains a stipulation like (39), and this too was discarded in favor of an analysis that
relies in part on the meanings of any and ever. Yet there remains a final (and
unformalized) stipulation to the effect that any and ever are allowed only if they
strengthen the utterance. One might, then, argue that we have simply cashed in one
stipulation for another. But in the opinion of at least this author this is not really true,
for each step along the way seems to be zeroing in on the full solution.
Before leaving this, though, we should also note that there is a vast amount of
literature on this domain and even plenty of controversy, and we have just scratched
the surface. Even within the broad picture here, we have had nothing to say about the
full distribution of NPIs. They are allowed in questions (both yes/no questions and wh-
questions) which are obviously beyond the scope here. There are, moreover, a number
of environments discussed as early as Klima (1964) which have not been covered

11
Kadmon and Landman (1993) also question the viability of a pragmatic account by
claiming that such sentences are “ungrammatical” rather than simply odd. However, we do not
really have a priori intuitions about what sorts of anomalies are due to the grammar (rendering
something “ungrammatical”) rather than due to other factors, and so this objection does not
seem valid. It is, however, true that the oddness cannot be cancelled, suggesting that
something more is at work than a pure pragmatic principle.
 363
here. NPIs also interact interestingly with presupposition; for discussion, see von
Fintel (1999) which should be reasonably accessible at this point. Moreover, it is well
known that NPIs can be licensed by implicatures (see Linebarger 1987 and much
subsequent discussion) and see Giannakidou (1997) for an alternative to the strength
reversal generalization presented in this chapter.

16.6. Semantic strength and scalar implicatures

The fact that semantic strength is reversed in certain environments is also relevant to
the distribution of scalar implicatures. And not surprisingly. After all, the Gricean
program discussed in section 10.4 posits that scalar implicatures result from
competition between a potentially stronger statement and a weaker one. Should a
speaker utter a weaker statement where a stronger statement directly competes with
the weaker one, then the hearer will conclude that the speaker did not have grounds
to utter the stronger one. This—combined with the assumption that the speaker is in
a position to know the truth or falsity of the stronger statement (the “knowledgeability
assumption”)—leads to the conclusion that the stronger statement is false.
Obviously, then, when strength is reversed one should see a different behavior. A
sentence with some should not implicate that the corresponding sentence with all is
false in those cases where the former is actually stronger than the latter. And in fact
this is exactly the pattern that we find.
Thus recall the discussion in section 10.4, where it was shown that some cannot
have not all as part of its meaning, because the not all suggestion disappears in
certain contexts. Among these was the linguistics department prerequisites scenario:
the department manual reads Everyone who has taken some Phonetics course is
allowed to enroll in Phonology I,and no one else is. As discussed there, an instructor
is not entitled (at least not by just what the manual says) to ban a student who has
taken all of the Phonetics courses.
(27) in section 10.4 showed that the exclusive or implicature disappears in the similar
sentence Everyone whohas taken (either)Phonetics I or Phonology I is allowed to
enroll in Phon/Phon II (and no one else is). These implicatures also disappear in the
antecedent of a conditional, as in (51):

(51) If you have taken some Phonetics course you may enroll in Phonology I.
16.6. SEMANTIC STRENGTH AND SCALAR IMPLICATURES
 364 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

This carries with it no implicature that the student who has taken all of the Phonetics
courses is not covered under (51). Similar facts can be constructed with or.
So exactly where does the some to not all implicature disappear? In downward
monotone environments!12 But this is exactly what we would expect. Consider the
pairs in (52):
(52) a. Every student who handed in some of the homework assignments got a B or better.
b. Everyone who handed in all of the homework assignments got a B or better.

(52a) is a stronger statement than (52b); it entails (52b) which is a reversal of the
usual situation with respect to some and all. For in normal (non-downward
monotone) environments a sentence with all entails the corresponding sentence with
some. But because this is a strength-reversing environment, the entailment direction
reverses. It is thus not surprising that (52a) does not implicate that (52b) is not true.
(Indeed, (52a) entails that (52b) is in fact true.)
It gets even more interesting. For in fact the implicature direction reverses; a
sentence containing all (in a downward-entailing environment) implicates that the
speaker is not in a position to say the corresponding sentence with some. Add to this
the speaker knowledgeability assumption, and it follows that a sentence all
implicates that the corresponding sentence with some is false. (The observation that
the normal implicatures disappear in downward monotone environments is
discussed in Horn 1989; the extended observation that the implicatures in fact
reverse in these environments can be found in Levinson 2000; see also Horn 1972.)13
Thus the general Gricean view predicts that on hearing (52b), a listener will conclude
(assuming speaker knowledgeability) that (52a) is false. That in turn means that there
is at least one student who handed in some of the homework assignments who did
not get a B or better. Of course the falsity of (52a) by itself is compatible with someone
who handed in all of the homework assignments not getting a B (since some does
include all). But the listener assumes that (52b) is true, and so the truth of (52b) (the
assertion) combined 12
This is not the only place they disappear, nor would one expect it to be since implicatures
arise from a complex interaction of factors.
13 The interaction of implicatures and downward-entailing environments is also discussed in
detail in Chierchia (2004) (and other more recent works by Chierchia on this subject), although
the conclusion that he draws from this fact is somewhat different from that drawn here.
with the falsity of (52a) (the implicature) leads the listener to conclude that there is at
least one student who handed in some but not all of the homework assignments who
didn’t get a B or better. (The discussion here assumes that (52b) presupposes that
there are students who handed in all of the homework assignments.) And indeed this
is exactly the inference that (in a normal conversational setting) a listener will draw
 365
on hearing (52b); i.e., that a B was not guaranteed if a student handed in only some
but not all of the homeworks.
Consider also the following:

(53) Not every student likes formal semantics.

As noted as early as Horn (1972), this implicates that there is some student who likes
formal semantics. That not every student does not entail some student can be shown
by our usual embedding this itself in (another) downward monotone context:

(54) If not every student in the school passes the statewide exam, the school will lose its
funding. (Otherwise, the funding will be guaranteed to continue.)

Surely the principal at the school hoping to receive continued funding—and knowing
that some students will fail—will not then tell all students that they should fail. But if
not every student was literally the same as “not every student but at least one
student,” then it would be quite a rational move on the part of the principal to make
sure that no student passes—for then the school would be guaranteed funding. Once
again, then, the not every to some inference disappears, showing that this is an
implicature, not part of the meaning of not every.
The reason for the implicature in (53) is exactly parallel to the discussion
surrounding (52). (55) is stronger than (53) and presumably competes with it; a
listener hearing (53) will thus conclude that the speaker was not in a position to utter
(55) and combining this with the knowledgeability assumption we arrive at the
conclusion that (55) is false:

(55) No student likes formal semantics.

This is part of the more general phenomenon of implicature reversal in downward

monotone environments. To simplify the discussion, think of not here as “scoping
over” the whole sentence, and so we recast this as (56):
16.6. SEMANTIC STRENGTH AND SCALAR IMPLICATURES

(56) It is not the case that every student likes formal semantics.

(This move is made simply to avoid giving a semantics for the version of not that
directly combines with a determiner like every. To do so is straightforward, as the
interested reader can verify, but it involves expository space and details that are
irrelevant to the point at hand.) It is not the case that is ordinary propositional negation
(thus of type <<s,t>,<s,t>>) and so of course it is a downward monotone function.
 366 16. NEGATIVE POLARITY ITEMS AND IMPLICATURES

Thus we would expect that (56) would implicate that the corresponding sentence in
(57) is false (note that (57) is equivalent to (55)):

(57) It is not the case that some student likes formal semantics.

And for this to be false, there must be at least one student who does like formal
semantics, exactly the implicature that we find here.

16.13. Consider the following:

(i) No student read every book on the reading list.

This implicates (given the usual caveats about speaker knowledgeability) that at
least one student did read at least one of the books on the reading list. Work
through exactly why this implicature arises.
 17 More binding phenomena
17.1. Functional questions 364 17.3. “Paycheck” pronouns
(and “donkey” pronouns?)
371 17.2. Binding and coordination
(in variable-free semantics) 370 17.4. “Hidden variables” 378

This chapter explores several (related) domains that involve “binding”

phenomena although not in the obvious sense, for many of them involve cases
where what is being bound is not an overt pronoun but rather what is sometimes
thought of as a hidden variable. An especially interesting aspect of these cases is
that they provide striking support for the variable-free view, as their analysis is
quite effortless under that view. But independent of the issues surrounding
variables, these domains form a rich set of (interrelated) phenomena which are
quite interesting in their own right.

17.1. Functional questions

17.1.1. The phenomenon

The first observation of interest centers on a surprising interpretation of

whquestions (and their possible answers). Some preliminary remarks are in order
concerning ordinary wh-questions (and answers) as in (1):
(1) Q. Who does Romeo admire (the most)? A.
Mercutio.

A serious account of the semantics of questions is beyond the scope of this book
(see Hamblin 1973, Karttunen 1977, and Groenendijk and Stokhof 1984 for
seminal proposals which have formed the basis of most subsequent work). Here
two points will suffice. The first is that the question asks for the identity of an
individual; one can paraphrase this as “what is the xe such that Romeo loves x (the
most)”? This is not meant as an actual analysis of the semantics, for this simply
368 17. MORE BINDING PHENOMENA

uses one question to paraphrase another. But the paraphrase is meant to stress
the fact that the question is about the identity of an individual; hence Mercutio is
an appropriate answer.
Second, some rudimentary account of the compositional semantics is needed.
As is well known, the interior portion of questions (e.g., Romeo admire in (1)) has
much in common with the interior of a relative clause. For example, in the LF
theory with movement it is assumed that who has raised from the object position
of love, leaving a trace and yielding (roughly) the structure in (2) which is also the
LF:

(2) Who [8 [SRomeo admire t8 (the most)]]

We restrict the discussion here just to the composition of the node labeled 8; this
is exactly the same as the composition of the corresponding material in the
relative clause (8) in Chapter 13. Hence, [ Romeo admire t8] is (on any g) that
8

function of type <e,t> which characterizes the set of individuals that Romeo loves.
We will not develop the rest of the composition of (2) as it requires a full theory of
the meaning of questions, and will simply assume [[ 8]] combines with the
meaning of who to give the full question. Note incidentally that although the who
is homophonous with the relative pronoun who they are not the same lexical item;
question words and relative pronouns are a distinct set. (For example, what is only
a question word.) In the Direct Compositional and variable-free approach
sketched in section 13.4, the meaning of Romeo admire (the most) is similar, but
is done without use of traces or LF. Recall that here Romeo lifts, undergoes g-sl,
and takes admire as argument. (For short, lifted Romeo function composes with
admire). No use of variables and assignments was needed for this, and so Romeo
admire is the characteristic function of the set of individuals that Romeo admires.
Again assume that this combines with [[who]].
Now consider (3):
(3) Who does every man admire (the most)?

This too can be a run-of-the-mill question asking for the identity of some
individual—here an appropriate answer might be Hillary Clinton. But it has
another interpretation, brought out by embedding this in the discourse in (4):
(4) Q. Who does every mani admire (the most)? A. Hisi mother.

How is it that the question here can have a reading in which his mother is an
appropriate answer? The initial temptation is to think of this as basically a
question about an individual, but where every man “scopes over” the question. In
other words, one can try to paraphrase this informally as “for every man y, what is
 17.1. FUNCTIONAL QUESTIONS 369
the identity of the individual x such that y admires x (the most)”? This assumes
that there is some way to combine the meaning of a question with the meaning of
a GQ, which depends in part on just what is the type of meaning for questions. But
even if those two can combine, the scope solution breaks down when one
switches to a GQ like no man, as was noted in Groenendijk and Stokhof (1984):
(5) Q. Who would no mani invite (to his wedding)? A. Hisi ex-wife.

Clearly it makes no sense to try to paraphrase this as “for no man y, what is the
identity of the x such that y loves x?” (See also Engdahl 1986 for detailed
discussion of problems with the scope strategy.)

17.1.2. An analysis with complex traces

In light of this, Groenendijk and Stokhof (1984) and Engdahl (1986) analyzed these
as functional questions. They ask not for the identity of an individual but rather for
the identity of a function from individuals to individuals. (5-Q) can be informally
paraphrased as: “what is the identity of the function f of type <e,e> such that for
no man, x, x invited f(x)?” By way of elaboration, we first embed the basic idea into
the LF theory using traces.1 The idea is that the movement of who can leave a
complex trace: one which corresponds to a variable over functions of type <e,e>
applied to a variable of type e. Mechanically this can be accomplished by giving
the trace two indices (see, e.g., Groenendijk and Stokhof 1984 and Chierchia
1991). As the two indices need to be kept distinct we notate the trace as tf-8/x-7.
The assignments have as their domain not only variables over individuals (for

1
The actual mechanics in Engdahl (1986) are slightly different as she was not using
traces but a different device, and Groenendijk and Stokhof (1984) also used a slightly
different notation. But these differences are inconsequential.
which we have been using the positive integers) but also an infinite number of
variables over functions of type <e,e>. (Indeed, as discussed in section 15.5, the
assignments have as their domain variables over all sorts of objects.) Call f 8 the
8th variable of type <e,e>, and x7 the 7th variable of type e. Then the interpretation
of tf-8/x-7 on any g is g(f8)(g(x7)). The relevant portion of the meaning of the question
is put together by assuming an LF as follows, where who has left a complex trace
and where no man has undergone QR and so the individual argument of the
complex trace is bound in the normal way that binding takes place:
370 17. MORE BINDING PHENOMENA

(6)
Q

who Λ f-8
S
GQ Λ x-7
no man S

NP VP

tx-7 V NP

invite tf-8/x-7

The value of the lowest S-node on any g is the proposition that g(x7) invited
g(f8)(g(x7)). The interpretation of x-7 closes off the individual variable, and so the
value of this node differs only on assignments which differ on the value of f 8 (since
that is the only open variable here). Thus on any g, [[ x-7]] is the set of individuals
who invited the value of g(f8) applied to themselves. This is then taken as argument
of [[no man]] (on g); so the value of the next S up the tree is—on any g—the
proposition that the intersection of the set just named and the man set is empty.
Then f8 is closed off by the º-abstraction step which is used to interpret f-8. The
result of that step is that the value of f-8 is (on any g) a function of type <<e,e>,t>. It
characterizes the set of functions f which are such that no man invited the value
of f applied to self. Put differently, it characterizes the set of functions f such that
the set of individuals x who invited f(x) has an empty intersection with the man
set. And this then presumably combines with the meaning of the question word
who (note that who would need to have a generalized meaning as it can combine
with functions of type <e,t> or—in this case— functions of type <<e,e>,t>). Notice,
incidentally, that each node labeled needs a full index such as f-7 or x-8. For
example, it is not sufficient to label the higher -node with just the index 7, because
the semantic interpretation of this node needs to be sensitive to whether or not
abstraction is happening over the function variable f-7 or the individual variable x-
7.
The analysis succeeds in giving the right meaning to a sentence which at first
glance appears rather intractable to paraphrase. But this particular
implementation of the basic analysis has costs (aside from the fact that here we
are abandoning Direct Compositionality). For there is no reason to expect the
existence of the functional reading for questions: this comes as quite a surprise.
 17.1. FUNCTIONAL QUESTIONS 371
This phenomenon is accounted for only by the hypothesis that traces can be
complex in the requisite way, but nothing in the theory of variables and traces so
far predicts that there should be complex as well as simple traces. In fact, Engdahl
pointed out that the traces need to be of arbitrary complexity; there are similar
questions asking for the identity of functions of type <e,<e,e>>. So, for example,
consider (7) in the context of a linguistics department with all male phonology
students, all female semantics professors, and a wonderful curriculum in which
the phonology students are required to take courses each year from every
semantics professor:
(7) Q. Which assignment does every phonology i student (every year) hope that every
semantics professorj will ignore? A. The first one that hei hands in to herj.

This requires the trace (after ignore) to be of the form tR-8/x-7/x-9 for R the 8th variable
of type <e,<e,e>> (where this is applied to two individual variables).

17.1.3. A variable-free analysis

Strikingly, the functional meaning for questions comes for free in the variable-free
account sketched in section 15.3. It does not require positing extra structure in
the “gap” position, and the existence of the functional reading follows
automatically from the mechanisms for binding in general. 24 For the sake of
discussion, we collapse the two steps of g-sl plus application into function
composition, which is a harmless expository simplification. The functional
reading for (4) is just a consequence of the fact that invite can undergo z; if no man
composes with z(invite) the result is exactly the meaning of the lambda node in
the LF in (6). It is the set of functions f of type <e,e> such that no-man z-invites f.
In other words, it is the set of functions f (of type <e,e>) such that the intersection
of the man set with the set of individuals who z-love f is empty. That set in turn is
the set of individuals x who love f(x). And so ultimately the question asks for the
identity of that function. Of course, z does not have to apply; when it doesn’t the
result is the ordinary individual reading for (4). So the existence of functional

24 There is a potential caveat: the meaning of who must be such that it takes as
argument both functions of type <e,t> (as in the ordinary individual reading of questions)
and one of type <<e,e>,t> for the functional questions. Whether a single item can be listed
with the more complex meanings derived from this is something we leave open, as it
cannot be answered without a more detailed account of the
372 17. MORE BINDING PHENOMENA

readings for questions is simply a consequence of the mechanisms for bound

pronouns in general.

17.1. Show the details of the syntactic and semantic composition of the
functional reading for no man invite in the variable-free view. (Note that the
syntactic category of this expression is different from that of the same string on
the individual reading.)

Engdahl’s observation that these can be asking for the identity of functions of
type <e,<e,e>>—or even more complex ones—follows as well, since there can be
various applications of z and g. Readers who worked through the details of
Exercise 15.23 should be able to see how to get the relevant reading for (7); those
who did not can content themselves with the intuition that the apparatus for
multiple binders and multiple pronouns extends automatically to multiply
complex functional questions.
Finally, we note that functional readings interact with WCO in just the ways one
would expect (under either the complex trace analysis or the variable-free
analysis). The relevant facts were first observed in May (1985) (although he did
not hook these into the analysis of functional questions). Thus (8a) has a
functional reading but (8b) does not:

meaning of question words. However, we can note that the variable-ful account of
functional questions also needs a way to predict these extra meanings for question words.
(8) a. Which woman would no man invite (to his wedding)? His ex-wife.
b. *Which woman would invite no man (to her wedding)? His ex-wife.

The fact that this contrast reduces to WCO was noted in Engdahl (1988) and in
Chierchia (1991). The point here is neutral between the LF-with-traces view and
the view here. In the LF view, the only way (9b) could have the functional reading
is for there to be a complex trace in subject position of invite (where whichwoman
moves, even though the movement here does not affect the word order), of the
general form tf-n/x-m, where noman is indexed m and hence raises and leaves a
trace indexed m in object position. While the WCO filter (9) in Chapter 15 does
not literally block this (as the interested reader can verify), it can easily be revised
to do so. (The term “pronoun” in that formulation would have to be expanded to
 17.1. FUNCTIONAL QUESTIONS 373
include both pronouns and the individual index portion of a complex trace. That
some revision is needed here may well be just an artifact of the particular
formulation that was chosen.) In the Direct Compositional and variable-free view,
the only way to arrive at the functional reading would be by application of s on
invite, which is impossible.

17.2. Binding and coordination (in variable-free semantics)

A persistent theme throughout this book has been that the Direct Compositional
tools allow for simple treatments of cases of coordination without having to
maintain some version of a conjunction reduction analysis. Chapter 5 motivated
Curry’ing transitive verbs in part to give a smooth account of sentences with VP
coordination; Chapters 10 and 11 showed how VP coordination interacts with the
analysis of GQs to give the appropriate readings of sentences containing both;
Chapter 14 showed some interesting interactions of coordination of Ss with GQs
in object position, and Chapter 13 showed that the adoption of the g-sl (or
function composition) allows for a simple analysis of sentences like (10) (the
basic point originally from Dowty 1988):
(9) Calista loves and Samantha hates model-theoretic semantics.

In light of these results, consider (10) in which one pronoun is apparently bound
by two different subject positions:
 374 17. MORE BINDING PHENOMENA

(10) a. Every mani loves but no mani wants to marry hisi mother.
b. Every third grade boyi loves and every fourth grade boyi hates hisi math teacher.
c. Every mani loves and no mani marries hisi mother.

We won’t dwell on how these may or may not be handled in the LF theory since
the issues are somewhat complex; for discussion see von Stechow (1990) and
Jacobson (1999). Suffice it to point out here that these are unproblematic, given
the variable-free fragment developed so far. (10c) will be used to illustrate. Every
man loves is put together as shown above in the discussion of functional
questions: the every man function composes with z-love to give an expression of
category S/RNPNP; this expression denotes the characteristic function of the set of
functions f such that every man z-loves f. no man marries is similar and is {f|f such
that no man z-marries f}. These two sets intersect (by the generalized definition
for and), and the expression everymanlovesandnomanmarries denotes (in set
terms) the set of functions f (of type <e,e>) such that every man z-loves f and no
man z-marries f. The analysis of his mother is also straightforward. It denotes the-
mother-of function and is of category NPNP. Both its syntax and semantics are thus
right to be the argument of every man loves and no man marries; and the truth
conditions thus require that the-mother-of function is in the set of functions
described above. (Note too that the fact that his mother is an appropriate answer
to a functional question follows in the same way: it denotes a function of type
<e,e> which is what the question asks for.)25

17.3. “Paycheck” pronouns (and “donkey” pronouns?)

17.3.1. Paychecks

Both theories of pronouns developed in Chapter 15 handle both free and bound
pronouns. But there is another use of pronouns which does not fall under these
categories, at least not in the obvious sense. These are called paycheck pronouns,

25 The fact that his mother is an appropriate answer to a functional question is not,
however, necessarily problematic for the variable-ful approach. For many proponents of
this approach would also argue that answers themselves are elliptical for full sentences,
and so this answer would be taken to hiddenly be the full sentence Every man 8 loves his8
mother (the most), where his8 is bound in the normal way.
 17.3 “PAYCHECK” AND “DONKEY” PRONOUNS 375

so-named because of their introduction into the literature in Karttunen (1969)

whose example was along the lines of (11) (modified from Karttunen’s actual
example):
(11) The woman who put her paycheck in the bank was far wiser than the one who put it in
the Brown University Employees Credit Union.

A more manageable example for the purposes of discussion is (12):

(12) Every third grade boyi loves hisi mother. But every fourth grade boy hates her.

Of interest here is the reading where her can be paraphrased as “his mother” (and
where his in the paraphrase is bound by the subject position of hates). On this
reading, her is neither a free nor a bound pronoun in the usual sense. Clearly it is
not free: in the reading of interest there is no single individual that it picks up. But
neither is it bound: the only thing it could be bound by is no fourth grade boy and
obviously this is not the reading we are after (and in any case has a gender clash).
Broadly speaking, there are two main analyses of paycheck pronouns.
One assumes the pronoun is—at some level of representation—represented as
fuller material; this is much like the ellipsis account of VP Ellipsis. Call this the full
NP analysis. This was suggested (for a related case) in Geach (1962) under the
rubric pronouns of laziness. The full NP analysis was argued for (for cases of this
sort) explicitly in Karttunen (1969, 1971), Partee (1975b), Jacobson (1977), and
for a slightly different but related group of cases in Evans (1980). It is revived in
more recent literature such as Heim (1990) and most notably Elbourne (2005)
(who takes the view— also espoused in much traditional transformational work—
that all pronouns are short for full linguistic material). If paycheck pronouns are
represented as full definite NPs at some level, there is no problem getting the
appropriate reading. At the level relevant for interpretation (LF, perhaps) her in (12)
is actually his mother, where his is bound by the subject position of hates in the
way that the binding of any ordinary pronoun takes place (one can take one’s pick
of theories for that part). This removes the mystery of how it is that the paycheck
reading is available. But it comes at a cost: some device is needed to effect the
mapping between the fuller representation and the simple pronoun (perhaps a
rule deleting linguistic material as in most of the accounts above). Incidentally,
one will often see these kinds of pronouns (and some other cases) discussed
under the rubric of e-type pronouns, a terminology from Evans (1980). Along with
this goes a common slogan that “e-type pronouns go proxy for definite
descriptions.” Presumably this is meant to claim that such pronouns are actually
 376 17. MORE BINDING PHENOMENA

full definite NPs at the level that inputs the semantics. Here, however, we will
continue to use the term paycheck pronouns rather than e-type pronouns, both
because the general points (and analysis) predate Evans’s analysis, and because
the term e-type pronouns is used by different authors in slightly different ways.
(Moreover, since the term e-type pronouns is also frequently coupled with the
above slogan, those using this term are often making a commitment to the full NP
analysis. The term paycheck pronouns is intended here in an analysis-neutral
way.)
The second broad tradition begins with seminal work of Cooper (1979),
subsequently slightly modified in Engdahl (1986), who observes an interesting
commonality between functional questions and paycheck pronouns. We exposit
this initially under the LF view. Like the trace of a functional question, Engdahl
posits (following Cooper) that a pronoun can correspond not just to a simple
variable over individuals, but also a complex variable, i.e., a variable over functions
of type <e,e> applied to an individual variable.4 Thus in (12b) the pronoun her can
be indexed as herf-8/x-7, where—just like the corresponding trace [[herf-8/x-7]]—on
any g is g(f8)(g(x7)). The rest of the composition is unsurprising: the individual
variable is bound in the way that binding of ordinary pronouns happens in general.
For example, in the LF theory, no fourth grade boy is, on the reading of relevance,
indexed 7 and undergoes QR, leaving tx-7 in subject position, and the composition
proceeds in the familiar way.
Key to the analysis is the fact that f8 is never bound, so the second sentence in
(12) contains a free variable over functions of type <e,e>. For any g, it asserts that
the set of lovers of the value of g(f8) applied to self has an empty intersection with
the set of fourth grade boys. The final assignment-dependent proposition is
applied by the listener to some contextually salient assignment. Although we have
left it open as to how an assignment becomes contextually salient, we assume
that there is a salient assignment in which f8 is mapped to the-mother-of function;
(which in turn is a salient function in view of the meaning of the first sentence).
Thus, using the

4
If traces are taken to be simply unpronounced pronouns, then this commonality is not
surprising. But it is not clear that it makes sense to consider traces as a species of
pronouns, as it is not obvious that traces should be listed in the lexicon with an index. After
all, the index is inherited from that of the moved expression.
informal terminology that we have often used in this text, the free variable f 8 picks
up the contextually salient mother-of function.
 17.3 “PAYCHECK” AND “DONKEY” PRONOUNS 377

Engdahl further points out that—just like the case of functional traces—
pronouns can have meanings that are indefinitely complex. Consider (13):

(13) The womani who told Searsj that the bill shei owed themj was in the mail was wiser than
the woman who told Filenes that it had not yet been mailed.

Here it is, in this account, a function of type <e,<e,e>> applied to two individual
variables; in the end it picks up the function that maps people and stores to the
bill that the former owes the latter.

17.3.2. A variable-free analysis of paycheck pronouns

While again quite elegant in some respects, the Cooper/Engdahl theory— like the
full NP theory—comes at a cost. For one must posit additional pronouns in the
lexicon. Not only are there (an infinite number of) pronouns with simple indices,
but there are also pronouns with complex indices, where the indices themselves
can be arbitrarily complex. But it turns out that if the basic insight in this analysis
is translated directly into the variable-free view here, the existence of paycheck
readings on pronouns is completely unsurprising, and requires no extra lexical
entries for pronouns. With only one small (and quite natural) modification to the
system above, these readings come as an automatic consequence of the system.
To show this, we translate the Cooper/Engdahl account into the variablefree
system in three steps. For the first step, we temporarily pretend that there are
variables and retain the free function variable from the Cooper/Engdahl proposal.
The individual variable, however, can immediately be eliminated. The relevant
reading is obtained just like the functional question case in (4) and (5); hates in the
second clause undergoes z. In other words, the meaning of the second sentence
can be represented as
[[no fourth grade boy]](z([[loves]])(f)).
Of course, this is cheating since it makes use of a free function variable f. So the
second step is to eliminate this and treat the paycheck reading of her in a way
analogous to any other free pronoun. Let the actual meaning of paycheck her be
not only the identity function over individuals, but also the identity function over
functions of type <e,e>. Thus it can have the meaning ºf<e,e>[f], which is of type
<<e,e>,<e,e>>. The category for this pronoun is (unsurprisingly) (NP NP)(NP ). ThisNP

category looks complex but its existence follows from the recursive definition of
categories. The superscript feature NPNP passes all the way up by the normal g
 378 17. MORE BINDING PHENOMENA

rules. So the second clause is composed as shown in (14) which is a schematic

representation of both the syntax and the semantics:
(14) g<e,e>(no-fourth-grade-boy) (g<e,e>(z (hates)) (paycheck-her))

Recasting this in perhaps somewhat more transparent notation, the final meaning
is ºf<e,e>[no-fourth-grade-boy(z-hates f)], or, equivalently, ºf<e,e> [no-fourth-grade-
boy(ºx[x hates f(x)])].

*17.2. Show the full syntactic and semantic computation which is sketched
above, using the lambda notation to show that this does indeed mean what is
claimed above.

But there still remains one unsatisfying aspect, for the analysis so far relies on
the assumption that her has two meanings: the identity function on individuals
and the identity function on functions of type <e,e>. This has little advantage over
the variable-ful analogue if these two meanings need to be listed in the lexicon. It
would be an accident that pronouns are homophonous in this way (and here too
would be needed an infinite number of pronoun meanings as they can be
indefinitely complex). But the final step relies on the observation that the
paycheck meaning of her is nothing more than the g operation applied to the
ordinary meaning of her.5 To predict the existence of this (and all the other extra
meanings), we simply need to generalize the g-sup operation to apply not only to
ordinary slashes but also superscripts: let an expression of category A B map to
one of category (AC)B with the normal g semantics. This generalization was
C

anticipated in the final fragment given at the end of Part III, so readers can consult
the gsup rule (R-7) given there to see this formalized. (15) shows that g applied to
the ordinary meaning:

(15) g<e,e>(ºf[f]) = ºh<e,e>[ºye[ºf[f](h(y))]] = ºh<e,e>[ºye[h(y)]] = ºh<e,e>[h]

All more complex paycheck meanings are just further applications of the gsup
rule. Finally, Jacobson (1977) and (2000) detail that paycheck pronouns

5
I owe this important observation to Mark Hepple.
 17.3 “PAYCHECK” AND “DONKEY” PRONOUNS 379

also show WCO effects in just the way that one would expect under either the full
NP theory or any variant of the Cooper/Engdahl theory (including the variable-free
variant above).

17.3.3. Donkeys

No discussion of pronouns would be complete without at least a brief mention of

another unexpected reading for pronouns: the so called donkey pronouns as in
(16) (Geach 1962):
(16) Every farmer who owns a donkey beats it.

The actual scope of what goes under the rubric of donkey pronouns varies in
different works; here we use the term to refer to cases where the apparent
antecedent or binder is indefinite, is in a position which normally cannot bind the
relevant pronoun, and where the interpretation of the pronoun covaries with
something else (in this case, with the farmers). Notice that these cannot (in an
obvious way) be analyzed as a case where a donkey “scopes out” and binds it
because this gives the wrong meaning. In other words, this cannot be paraphrased
as There exists a donkey such that every farmer who owns it beats it.
Many authors have argued that these are just another instance of paycheck
pronouns. Take, for example, the Cooper/Engdahl analysis sketched above.
Indeed Cooper’s (1979) proposal was originally designed specifically for the case
of donkey pronouns. Thus, using Engdahl’s modification of this, the idea is that
here too it is simply a complex variable such as f-8 (x-7) where the individual
variable x-7 is bound in the normal way by the subject slot, and the f variable
remains free and picks up the contextually salient function mapping each farmer
to the donkey that he owns. The translation into the variable-free setting is obvious
and proceeds exactly as for the case of any other paycheck pronoun, so for the
sake of saving space we will not make that translation here. And the same idea
has an obvious analogue under the full NP view; here it would simply be short for
the donkey that he owns.
The well-known difficulty with assimilating donkey pronouns to paycheck
pronouns is that there need not be a unique donkey owned by each farmer. (16) is
(according to many speakers) felicitous in a context where Farmer MacDonald is
a mule breeder and hence owns eight donkeys. If it is analyzed as the
donkey that he owns, then (where he is understood
 380 17. MORE BINDING PHENOMENA

MacDonald) this should be infelicitous, but many speakers find this fine. And note
that while uniqueness with the is generally relativized to context (possibly by a
domain restriction as part of the meaning of the), this fact is not of any help. For
(16) can (for many speakers) be felicitously uttered in a context in which no one
particular donkey of the mule breeder is salient. An analysis based on the
Cooper/Engdahl tradition (whether in variable-free or variable-ful terms) has the
same problem. A function, by definition, maps each member of its domain to a
unique member of the co-domain, so analyzing this as involving a free function
variable or argument slot runs into the same problem: the mapping between
farmers and donkeys in (16) need not be function. And for those speakers
(including the present author) who actually do find (16) very odd in the above
scenario (where a farmer may own multiple donkeys), the problem persists, since
other examples make the same point. Thus Heim (1982) constructs an example
which illustrates the basic problem but seems impeccable:
(17) Everyone who bought a sage plant bought nine others to go along with it.

The literature abounds with possible solutions; a number of authors have

argued that if one relativizes the interpretation of these sentences to small enough
situations, one can indeed maintain that donkey pronouns reduce to paycheck
pronouns by finding a unique donkey in (16) and sage plant in (17). (See Heim
1990, Elbourne 2005, and others.) On the other hand, a very different kind of
analysis of donkey sentences is found in the related traditions of Discourse
Representation Theory (Kamp and Reyle 1993), File Change Semantics (Heim
1982), and Dynamic Semantics (Groenendijk and Stokhof 1991 and many since),
all of which enrich the semantics to give an account of the way in which sentences
change the context as the discourse proceeds. (See also Barker and Shan 2008
for a different recent account.) And it should be noted that there are a number of
other well-known questions with respect to the interpretation of donkey
pronouns. For example, in (16) most speakers have the intuition that our mule
breeder MacDonald must be a mass donkey abuser and beat every one of his eight
donkeys (thus this is making some claim about every farmer–donkey owning pair).
Yet clearly this requirement is not always there, as shown by the fact that (18)
(from Schubert and Pelletier 1989) surely does not have the analogous
requirement:
(18) Everyone who had a quarter put it in the parking meter.
 17.3 “PAYCHECK” AND “DONKEY” PRONOUNS 381

For more discussion, see among others, Kadmon (1987) and Kanazawa (1994). In
any case we will leave it open here as to the best analysis of donkey pronouns,
noting that the hope would be—with Cooper (1979), Heim (1990), Elbourne
(2005), and others—that these can reduce to paycheck pronouns.

17.4. “Hidden variables”

17.4.1. Hidden variables can always be complex

There are many lexical items that have been argued to contain hidden variables as
part of their meaning. (Initially we phrase this in variable-ful terms as this once
again is expositorily convenient.) This has come up in various places throughout
this book: the domain restriction variable with respect to quantificational
determiners, the comparison class variable found in adjectives like tall, and the
numerical standard variable n in the meaning of many. There is a striking (although
perhaps not entirely surprising) generalization that emerges about these: these
can always have the kind of complex readings that we have seen for pronouns and
functional gaps. Call this the paycheck generalization.
Thus, sticking again for the moment to the terminology of a theory with
variables, all of these hidden variables can not only be simple, but can also be
variables over functions of some sort applied to an individual variable. Thus call
them paycheck variables: a paycheck variable consists of an indexed variable F i
of type <e,z> (where z is any type) applied to an indexed individual variable. Then
the generalization that appears to emerge is that any time there is a variable of
type z as part of the meaning of some expression, it can also be a paycheck
variable of the form Fi(xj). (One might suspect that it can actually be even more
complex—this suspicion seems to be correct and will be discussed below.)
Consider first the case of quantifier domain restrictions. Recall from section
10.6 that there are two conceivable explanations for the fact that a GQ like every
student is interpreted with respect to a given domain. One is that it contains “free
variable” of type <e,t> as part of its meaning: [[every student]] is ºP[student’ \ DS
PS]. (Recall that, as detailed in
 382 17. MORE BINDING PHENOMENA

section 15.5.1, there must also be some index on each of the free variables, but
for the sake of exposition we will consistently suppress the index in the discussion
here.) An alternative explanation (see, e.g., Kratzer 2004) rests on the observation
that sentences are generally interpreted not with respect to full worlds but smaller
situations, and the domain restriction might just be a consequence of this.
While the latter is undoubtedly true, there is an interesting argument that the
domain restriction is in any case a part of the meaning of the GQ. (This argument
is originally due to von Fintel 1994 and is further discussed in Stanley and Szabo
2000.) This is that the domain restriction can be complex, in the sense discussed
above. Consider, for example, (20) in the context set up in (19): 6

(19) It is well known that the ceramics courses at Brown are quite difficult. One
requirement for a student to pass is that she/he must be able to crack a pot with
only a nutcracker. And this requirement turns out to be the downfall of many a
student. This can be especially worrisome for seniors who might be in danger of not
meeting the graduation requirement if they don’t pass the class. The Dean of the
College always sends a note to all faculty during the year, asking if they have any
worries about the (hopefully) graduating seniors in their class. And this year it is a
complete mess, for it turns out that:

(20) Every ceramics professor is pretty sure that no senior will successfully complete the
pot-cracking requirement.

How is no senior interpreted in this context? It of course is not “no senior in the
world.” Nor is it “no senior at Brown.” It is not even “no senior who is taking a
ceramics class.” For suppose that Tina (a senior) is in Professor Carberry’s
ceramics class, and another senior Zorro is in Professor Glazie’s ceramic class.
Professor Carberry might be quite sure that Tina will fail at pot-cracking, but (20)
does not require him to have any opinion about how Zorro will do. (Of course, in
the scenario at hand, Professor Glazie must feel that Zorro will be incapable of
completing the requirement, but Glazie need not have shared this information
with Carberry.) Rather, this is interpreted as “no senior in the course of x,” where x
is bound by every ceramics professor. Thus we can represent the meaning of no
senior as

6 Notice incidentally that positing a “free” variable in these cases would lead us to expect

to find cases where it is also bound; it is however difficult to construct convincing cases
along these lines probably because there happens not to be the right convergence of
properties to find “binding” of something of type <e,t>.
 17.4. “HIDDEN VARIABLES” 383

ºP[senior’S \ F<e,<e,t>>(xe)S PS]. As is similar to the case of paycheck pronouns on the

Cooper/Engdahl analysis, x ultimately is bound in the normal way that binding
happens by the position occupied by every ceramics professor, F remains free and
picks up the contextually salient function mapping each professor into the set of
students in her/his course.
Similar facts can be constructed with the. In section 8.3 it was posited that the
meaning of the is a function of type <<e,t>,e>, but where an NP like the
psychologist is also interpreted with respect to some domain. Hence the meaning
of the might be fleshed out as a function from P (of type <e,t>) to the unique (or
most salient) member of PS \ DS, where D is a free variable of type <e,t> whose
value is fixed by context. Interestingly, here too the restriction can be complex and
hence best represented as a paycheck variable. Suppose we are at an
interdisciplinary conference of linguists and psychologists, with equal numbers of
each. In order to achieve maximal cross-talk between the two groups the
organizers pair each linguist with a psychologist to talk together at dinner. But it
was not successful, for I report this as:
(21) Unfortunately, no linguist felt that the psychologist
knew much about language.

Here the meaning of the psychologist can be represented

as

Øx[x psychologist’s\ R<e,<e,t>>(y)S] where y is bound by no linguist, and R remains

free and picks up the salient relation of being paired with.

Following some observations of Breheny (2003) (although he was arguing for a

different point), one can show that these paycheck domain variables can be even
more complex. Let’s expand our scenario to a multidisciplinary conference of
linguists, philosophers, and psychologists, with an equal number of participants
from each field. To maximize interaction among the participants, each linguist is
assigned to pour a drink for one and only one philosopher at the party, and also
one and only one psychologist. And each philosopher will pour a drink to their
assigned linguist and to their assigned psychologist, and similarly for the
psychologists. It turns out to be a particularly bad year for jobs for recent PhDs in
philosophy. And so the philosophers at the conference—all being philosophers of
language—hatch a plot: they decide kill off the linguists, thereby creating some
new openings. And so each philosopher serves a poisoned drink to her/his
 384 17. MORE BINDING PHENOMENA

appointed linguist (a foolproof plot, for they can always blame the psychologists
for the mass die-off of linguists). But the plot failed: the linguistics figured it out,
and alas:
(22) No philosopher managed to get any linguist to actually accept the drink.

What is the domain restriction on the drink? Obviously it is not unique in the world,
nor the unique drink at the party. Nor can it even be the unique drink relative to
each philosopher, for the philosopher also served a drink to a psychologist. Nor
can it be the unique drink relative to the linguist, for she also gets a drink from a
psychologist (after all, this is what the philosophers are banking on to put the
blame on the psychologists). So it must be the drink unique to the pairing of the
philosopher and his/her assigned linguist. In other words, the semantics of the
drink can be represented as

Øx[x drink’S \ Q<e,<e,<e,t>>>(y)(z)S]. z is bound by the subject of accept; y is bound by

the subject of manage; and Q remains free and picks up the salient relation
between three individuals and drinks such that for all x,y,z, Q(y)(z)(x) holds if z
serves x to y. (The order of placement of these variables in the prose description
is arbitrary.)

*17.3. Construct a case using a quantificational determiner like every or no to

show that there can be a domain restriction which is complex in the same way
as the domain restriction on the drink above (i.e., where the domain restriction
must be a domain in <e,<e,<e,t>>>).

The point can also be made with other cases of “hidden” variables. Take the
comparison class for adjectives like tall. As noted originally in Kennedy (1997),
this too can be complex where it is bound into by the subject slot, as in (23):
(23) All of John’s children are tall.

What is the comparison class for tall (on the obvious interpretation)? It cannot be
everyone in the world, of course. But neither is it all children. Rather it is, for each
child, the set of children approximately their own age.
Hence in the simple cases we can think of tall as meaning “tall for a P i” (for
Pi the ith variable of type <e,t>), but it can also mean “tall for a F i(xj)” for
 17.4. “HIDDEN VARIABLES” 385

F the ith variable of type <e,<e,t>>. In (23), the individual variable x is bound by the
subject position in the way binding normally takes place, while the F variable
remains free, and picks up the contextually salient function mapping any child to
the set of individuals of the same age.
17.4. Construct a case that the “hidden” variable n as part of the meaning of
many can also be a complex variable of the form Ni(xj) where N is the ith
variable that is a function from individuals to numbers. (We leave it open as to
the status of numbers in the theory; we have not formally introduced these as
part of the model-theoretic apparatus.)

17.4.2. Translating into variable-free

Strikingly, the paycheck generalization follows immediately once the notion of a

hidden variable is translated into the variable-free setting. First as discussed in
section 15.5.2, the notion of a “hidden variable” is recast as an argument slot.
Take, for example, the case of domain restriction. If it is correct that the meaning
of every senior encodes the expectation that this will be limited to some domain,
this means that [[every senior]] contains an extra argument slot; it is
ºQ<e,t>[ºP<e,t>[senior’S \ QS PS]]. Hence, [[every senior]] is of type <<e,t>,<<e,t>,t>>
(a function into GQ meanings rather than a pure GQ). And that slot will of course
be passed throughout the combinatorics by the normal g process. In the end, a
sentence like every senior will fail is a function from domains to propositions, and
the listener supplies this to a contextually salient domain. One additional aspect
that must be resolved here concerns the syntax: what is the appropriate
superscript feature on the category of every senior? Unlike the case of slots that
come from the meaning of pronouns, this is less straightforward, since we have
no overt pro-form of the appropriate type. For convenience, assume that it is N, so
every senior actually has category (S/(S/NP)) N. (We assume that ultimately the
argument slot comes from the meaning of every as was shown in section 15.5.2.)

17.5. Show the details of the syntactic and semantic composition of a simple
case like Every student failed in the context where the domain is understood
as something simple, such as the set of students at Brown.
 386 17. MORE BINDING PHENOMENA

As noted above, the paycheck generalization is automatic, for this slot can be
made more complex by application of g-sup and can be “bound into” by z. (Recall
that g-sup was generalized in the fragment at the end of Part III to apply to
superscript argument positions as well as slash argument positions.) Consider,
for example, a slightly trimmed-down version of (20) Every ceramics professor
thinks that no senior will complete the requirement in the
Seniors-Fail-Ceramics scenario. As was worked out (for a parallel case) in
Exercise 17.5, no senior will complete the requirement is of category S N with
meaning of type <<e,t>,t>—it is waiting to find a domain to give a proposition. But
it can also map to a fancier meaning and category by g-sup. The relevant mapping
yields an expression category (SNP)(N ). Semantically, this wants a function of type
NP

<e,<e,t>> (a two-place relation) to give something that is a function of type <e,t>.

(Computing the actual meaning is left for the exercise below.) Now consider think.
Its lexical category is (S/NP)/S, but it can map by z to (S/NP)/S NP. This is the run-
of-the-mill binding rule. But in the case at hand it does not combine with just an
SNP but with the more complex category above. This is unproblematic: application
of g-sup maps this to (S/NP)(N )/(SNP) (N ) and so can take no senior will complete
NP NP

the requirement as argument. Lest this seem a mysterious mechanical exercise,

the intuition can be summarized as follow. Normally no senior will complete the
requirement has a domain restriction argument slot
(the analogue of a free variable in a variable-ful theory), and this passes all the way
up, never being bound, and so in the end is a free slot. But in the scenario at hand,
more happens. This maps by g-sup to the more complex slot of type <e,<e,t>>.
The e-position created in that step is ultimately bound by z on think, and the two-
place relation slot just continues to pass up by g-sup. So this is quite analogous
to the variable-ful treatment where the complex variable has a part which remains
free and a part which is bound by the normal binding process. The only difference
is that here the complex meaning of the domain restriction is unsurprising, for it
follows that it can be made more complex by application of g-sup. Fancier
examples such as (22) (the philosophers-kill-linguists scenario) are also
automatic and simply involve additional applications of the rules.

*17.6. Show the full details of the above derivation.

*17.7. Show the syntactic and semantic composition of (46) (in the case where
tall is restricted to a contextually salient function of type <e,<e,t>>). Since
 17.4. “HIDDEN VARIABLES” 387

expressions like all of John’s children are not in the fragment, recast this as
every child of John.
One final case that was treated earlier as a case of a free variable encoded as
part of the meaning of an item was the analysis of VPE. Recall that it was posited
in section 14.4.2 that an auxiliary like will of category VP/VP and of type
<<e,t>,<e,t>> could map to a second item which is just a VP and has a meaning
will’(P). In the variable-free view this can’t be quite right, since it contains crucial
use of a free variable. But the necessary revision is obvious; the category of the
“ellipsis” will is actually VPVP with meaning ºP[will’(P)]. Does the paycheck
generalization extend to this argument slot as well? Indeed it does: a relevant
example is a case like (24):7

(24) Every fourth grade boyi hopes that the principal won’t call his i mother; but every third
grade boy hopes that the principal will.

This is analogous to the sloppy identity cases discussed in section 15.4.1.3,

except that here the missing VP is embedded further down. So here we have the
elliptical version of will, but (as in the case of ACD) it picks up not a function of
type <e,t> but rather the two-place relation [[call his mother]] of type <e,<e,t>>.
(That [[call his mother]] is of this type follows in the variable-free view because it
contains a pronoun; it is the two-place relation ºx[ºy[y call the mother of x]]. Thus
ellipsis-will maps to something of category (VPNP)(VP ) which is of type
NP

<<e,<e,t>>,<e,<e,t>>>. The VPNP superscript passes all the way up by g-sup; it is

“bound into” by z on hopes. The reader can fill in the full details.

7
Actually examples of this general sort were discussed in Sag (1976) who thought they
were bad, but it has since been shown that the particular examples used by Sag had other
problems (see Evans 1988 for discussion).
 18 Additional semantic
dimensions: The semantics of focus
18.1. Computing alternatives 18.4. Further compositional issues 393
compositionally 386
395
18.5. Focus and binding
18.2. The meaning of only 390
18.3. Additional focus-sensitive words 18.6. More generally: 397
Multidimensional semantics
392
It is easy to get the idea that the enterprise of formal compositional semantics has
little to say about the many subtle nuances expressed linguistically in actual
conversational contexts. The propositional content of a sentence— i.e., its truth
conditions—is one dimension and certainly a central one. For it is the proposition
expressed by a sentence that adds to our understanding of the facts of the world.
But speakers use linguistic forms to express many additional aspects of meaning
(broadly construed). For example, a speaker can express her attitude towards the
proposition (as in Oh shoot, the puppy just chewed up my students’ exams!)
(items like “oh shoot” are sometimes called expressives). Another non-truth-
conditional kind of meaning is exemplified by a sentence like: He is a phonologist
but he loves model-theoretic semantics. This suggests that phonologists
generally don’t love model-theoretic semantics. But is not a matter of truth
conditions, for if it turns out that all phonologists love model-theoretic semantics
this sentence could still be true. Or, in the case that will be under consideration in
this chapter, a sentence like LARRY loves eels (with heavy stress on LARRY )
conveys that there is something at issue as to whether other people love eels.
Incidentally, while all three of the examples above involve inferences that are not
part of the truth conditions, they also are not conversational implicatures. For here
it is the particular linguistic forms—oh shoot, but, and special stress on LARRY—
that give rise to the inferences. Hence the role of these forms is part of English
speakers’ knowledge, and not something that follows just from principles of
cooperative conversation. Gricean conversational implicatures, on the other
hand, refer to those inferences that follow from general principles of reasoning
and/or principles governing cooperative conversation (combined, of course, with
the literal meaning of a sentence) and hence not generally taken to be part of the
grammatical knowledge of a speaker. 1 The inferences here are crucially tied in to
 389
the form, and are sometimes referred to under the rubric of conventional
implicature: something which is conventionalized but not literally part of the truth
conditions.
What this shows is that indeed sentences have more than one dimension as
part of their meaning (using “meaning” in a broad sense here). But for at least
some of these dimensions, the basic compositional tools developed for the truth-
conditional dimension can help elucidate the compositional contribution of the
relevant linguistic forms. The contribution of “focal” or “contrastive” stress is the
poster child for this point, and so it is to this that we now turn. The discussion will
necessarily be brief and incomplete, but hopefully sufficient as to give a flavor of
the enterprise.

18.1. Computing alternatives compositionally

The analysis here is based on the seminal work of Rooth (1985), although it
departs from Rooth in some detail. Rooth’s contribution often goes under the
rubric of alternative semantics. This does not mean an alternative to the usual
view of semantics. Rather, Rooth showed that the role of focal stress is to conjure
up a set of alternatives, and he provided a way for these to be computed
compositionally.

1
But, as discussed in Chapter 10, n. 4, there are recent works that claim that
conversational implicatures are also part of the grammatical computation. See that
footnote for references, as well as references to defenses of the Gricean view.
18.1. COMPUTING ALTERNATIVES COMPOSITIONALLY

As noted above, focal stress in certain cases contributes nothing to the actual
truth conditions of a sentence, yet it does create this additional dimension of
meaning. Thus while (1a) and (1b) have the same truth conditions, they are
appropriate in different contexts:
(1) a. COYOTES eat chipmunks.
b. Coyotes eat CHIPMUNKS.

(In these sentences, coyotes and chipmunks are not plural common nouns, but
are what is known as bare plural NPs; see Carlson 1980.) (1a) is most likely
uttered in a context where what is at issue is the question of what sort of animals
 390 18. THE SEMANTICS OF FOCUS
are chipmunk-eaters; (1b) is more appropriate in a context where what is at issue
is the feeding habits of coyotes.
Especially interesting is the fact that focus actually does contribute to the truth
conditions in some cases. This emerges when sentences like these are
embedded in constructions involving so-called focus-sensitive words like only.
Thus take the pairs in (2) and (3) ((3) is considerably less natural but its analysis is
simpler and so we will often use this kind of frame in this chapter):
(2) a. I only believe that COYOTES eat chipmunks.
b. I only believe that coyotes eat CHIPMUNKS.

(3) a. It is only the case that COYOTES eat chipmunks.

b. It is only the case that coyotes eat CHIPMUNKS.

(3a) denies that other relevant animals eat chipmunks while (3b) is entirely neutral
on that point. Take, for example, the case of an environmentalist charged with
figuring out what is responsible for the declining population of chipmunks in her
region. A group of overly zealous homeowners who all love the local chipmunks
are thinking that the solution is to thin out the population of wolves and mountain
lions. “Not so,” says the environmentalist, “It’s only the case that COYOTES eat
chipmunks.” (Of course she probably utters a more colloquial version of (3a).) And
so here (3a) entails that the wolves and mountain lions are not preying on the
chipmunks. But (3b) has no such entailment. It might instead be uttered by a
coyote defender (concerned with the fact that humans seem to feel that the
coyote population needs to be thinned) in a context where chipmunks are quite
plentiful. The coyote defender can argue against thinning the coyote population by
uttering (3b)—here she is committed to the fact that coyotes are not bothering any
other species. Thus the meaning of only (or, it is only the case) is such that it is
sensitive to the role of stress. In general, then, the role of this kind of stress is to
conjure up a set of alternatives (an intuition which is reflected by the common use
of the term contrastive stress). Moreover, when embedded under a focus-
sensitive word like only, those alternatives become relevant to the truth
conditions.
In light of this, Rooth (1985) proposed that the meaning of expressions has a
second dimension of meaning. In addition to what is known as its regular semantic
value, it has a focus value which is also compositionally computed. From now on,
when we wish to distinguish the focus value of Æ from its regular value, we will
use the notation [[Æ]]REG vs [[Æ]]FOC; we also continue to use [[Æ]] (and, for that
matter, Æ’) interchangeably with [[Æ]]REG whenever there is no need to distinguish
 391
this from the focus value. The focus value of an expression is a set of things of the
same kind as the regular value of the expression (and the context will restrict this
to relevant alternatives). For a full sentence like COYOTES eat chipmunks it is
therefore a set of alternative propositions. That the focus value must be
compositionally computed is obvious from the fact that the location of the stress
is on a single word, but what is at issue is alternative propositions.
Hence, let every linguistic expression (even those containing no focal stress)
have these two values. In the case of items with no focal stress, the focus value is
simply the singleton set of its regular value. (Having all items have a focus value
allows for a smooth compositional computation.) Focused items, on the other
hand, have as their value a set of alternatives, where the actual members of the
set of alternatives is determined by context. Moreover, one member of the set of
alternatives which comprise the focus value will be the regular value of the
expression (this is a technical move to make the compositional computation
simpler).
Consider then a simple case like (4):
(4) a. Sally SWIMS.
b. SALLY swam.

In (4a), the regular value for Sally is s, and the focus value is {s}. The regular value
for SWIMS is swim’ and the focus value is a set such as {swim’, run’, walk’, jog’}.
(Imagine a context in which what is at issue is how Sally gets her exercise.) In (4b)
the regular value of SALLY is again s, but here the focus value is {s,b,r, . . . }, while
[[swim]]REG is swim’, but [[swim]]FOC is just {swim’}.
18.1. COMPUTING ALTERNATIVES COMPOSITIONALLY

The above gives just the regular and focus values of the lexical items; the
interesting step is in how the compositional computation of focus values works.
This is quite straightforward: the computation of the focus value always parallels
that of the regular value. Take (4a). In the computation of the regular value,
[[SWIM]]REG applies to [[Sally]]REG. For the focus computation, each member of the
focus value of the function applies (one by one) to each member of the focus value
of the argument. Thus R-1 can be expanded to include the focus computation as
follows (we again use boldfaced curly brackets to indicate the ordered pair of
regular and focus values):
R-1-amended: Let Æ be an expression of the form <[Æ], A/B, {[[Æ]] REG, [[Æ]]FOC}> and
 392 18. THE SEMANTICS OF FOCUS
be an expression of the form <[], B, REG
, [[]]FOC}>. Then there is an
expression ª of the form <[FCAT(Æ) ]), A,
FOCFOC
&& X = f(y)]}}>.

Notice that there can be more than one item in a sentence with focal stress.
Take for example (5) a context where Farmer MacDonald asks which of his
animals will manage to find which of the edible plants in his field:
(5) PORKY will find the TRUFFLES.

Here the main value is find’(the truffles’)(porky’) and the focus value is a more
complex set of alternative propositions which differ both on the subject and
object position. For example, let our relevant set of individuals be {p, el, ey} (Porky
the pig, Elsie the cow, and Eeyore the donkey), and our relevant set of edible
plants be {t, c, a} (truffles, clover, and arugula). Then the focus value of (5) has nine
propositions: {find’(t)(p), find’(t),(el),
find’(t),(ey), find’(c)(p), etc.}.2

18.1. Show in detail the compositional computation of the focus value for (5)
(using the focus value for PORKY and for TRUFFLES as given above). As usual,
ignore will.

2
Actually, (5) is ambiguous between a reading in which the effect of the focal stress is
to give alternatives to truffles and a reading (called broad focus) in which the alternatives
are alternatives to the entire VP; thus the focus value of the VP would be something like
(using informal prose): {find the truffles, roll in the mud, grunt, . . . }. This is discussed in
section 18.4.
18.2. The meaning of only

What happens with this combines with a focus-sensitive word like only? The
lexical meaning of only is such that it accesses not just the regular value but also
the focus value—and so the truth conditions for something like It is only the case
that Sally SWIMS makes crucial reference to the alternatives.
 393
To flesh this out, there are some issues regarding both the syntax and the
semantics of only that need to be (briefly) dealt with. First, its syntax is complex.
Only can combine directly with the focused material itself, as in (6):
(6) a. Only SALLY swims.
b. Tom saw only SALLY.

But it need not, it can also combine with a VP which contains focused material
within it, as in (7):
(7) a. I only believe that Sally SWIMS.
b. Coyotes only eat CHIPMUNKS.

The end result of the contribution of only in, say, (7b) is (among other things) to
assert that none of the alternative propositions are true (i.e., it entails that coyotes
don’t eat mice or cats or rabbits). This means that its contribution is, in some
sense, a sentence-level contribution. The same is true in (6): (6a) denies that
Martha swims and that Tom swims. Yet only in the syntax does not combine
directly with sentences.
By now the Direct Compositional techniques for handling this apparent
mismatch should be quite familiar; one can simply “fancy up” the packaging of the
meaning of only to respect its syntax. But to save space, we will not do this here.
Rather, we will take the expositorily convenient tack of pretending that we are
dealing with a single word “it is only the case that” which attaches itself to a
sentence. (This is analogous to the move made for negation in section 3.4.) Hence
we recast (7b) as (3b) given above.
The second point concerns the semantics. Does (3b) (and (7b)) assert that
coyotes eat chipmunks, or do these sentences just presuppose this? There is
debate in the literature about this, but nothing in the remarks here hinge on the
outcome of this debate. The discussion is simplest if we assume that this is part
of the assertion rather than a presupposition, but the meaning given below could
be revised. In light of this, we now somewhat informally give a “lexical entry” for it
is only the case:
 394 18. THE SEMANTICS OF FOCUS

18.2. THE MEANING OF ONLY

(8) it is only the case, S/RS, [[it is only the case]] takes both the regular and the focus value of the
S that it combines with, and returns true (at a world w and time t) iff [[S]] REG = 1, and 8p[p
[[S]]FOC and p 6¼ [[S]]REG [ p is false (at w and t)]].

Thus prefixing this to a sentence with meaning p (for some proposition) has the effect
of asserting that p is true and that all propositions in the focus value which are not p
itself are false.26 Note that (8) is incomplete as it leaves open what is the focus value
of a sentence like (3b). Assume here that it is only the case also “swallows up” the
alternatives, so that the resulting focus value is—as in the case for material with no
focused constituents within it— just the singleton set containing the regular value.
Thus the focus value for (1b) coyotes eat CHIPMUNKS is the set of propositions of
the form (stating these propositions in ordinary English) {coyotes eat chipmunks,
coyotes eat beavers, coyotes eat deer, etc.} and the environmentalist asserting (3b) is
saying that the first of those propositions is true, but none of the others are.

*18.2. Take the occurrence of only in (6a). Assume that only combines here with
SALLY, and so is of category GQ/ RNP. Give it a meaning such that (6a) is
synonymous with It is only the case that SALLY swims (under the meaning given for
the latter above). In other words, the task is to “fancy up” the meaning of it is only
the case shown in (8) to allow this to combine directly with the NP, even though its
contribution is ultimately to say something about alternative propositions.

Consider now the case of a sentence with more than one focused item, as in (5). As
noted above, the alternatives are the set of propositions varying on the set of
contextually salient animals (in this case Elsie, Eeyore, and Porky) and contextually
salient edible plants (clover, arugula, and truffles).

26 This is a slight oversimplification. A better approximation of the meaning of only is such

that it negates not all members of the focus value distinct from the main proposition, but
rather all members that are not entailed by the main proposition. For example, if it is correct
that four entails three (as would follow if three means “at least 3”) then we surely don’t want
to say that It’s only the case that he has taken FOUR linguistics courses entails that he has not
taken three linguistics courses.
 395
Suppose that Farmer MacDonald is concerned about the resourcefulness of his
animals: would they be able to survive on their own should he die? He knows that there
are good edible plants on his farm, and so to test their survival ability he lets them
loose to see what they find. We report the sad outcome of this test:
(9) Unfortunately, it is only the case that PORKY found the TRUFFLES.

This negates all of the alternative propositions in the focus value of (5), indicating to
our altruistic farmer that only Porky will survive and then only if the land is filled with
enough4 truffles.

*18.3. Consider (i) in the same scenario:

(i) Unfortunately, only PORKY found TRUFFLES.

This can (at least for some speakers) have the same value as (9), and can be used
to report that no other animal found any of the other edible plants. Does this follow
from your answer to Exercise 18.2? If not, revise that so that it does follow. 4 You may
want to try to do this informally.

18.3. Additional focus-sensitive words

Only is not the only lexical item sensitive to the focus value of the expression with
which it combines. Among the other focus-sensitive words are also (and its variants
such as too) and even. Consider the pair in (10):
(10) a. It’s also the case that DENISE snowboards.
b. It’s also the case that Denise SNOWBOARDS.

(10a) presupposes that someone else snowboards. If presupposition failure is a

matter of a sentence having no truth value, then the “lexical entry” for it is also the case
can be written informally as follows:

4
This judgment was disputed by an anonymous referee, but has been checked with many
speakers so perhaps there is some speaker variation here. Another instance of this with a
different scenario can be found in Jacobson (2012).
18.4. FURTHER COMPOSITIONAL ISSUES
 396 18. THE SEMANTICS OF FOCUS

(11) it is also the case that, S/RS,

[[it-is-also-the-case-that-S]] = 1 if [[S]]REG =1 and ∃p[p 6¼ [[S]]REG & p [[S]]FOC; [[it-is-

also-the-case-that-S]] = 0 if [[S]]REG = 0 and ∃p[p ¼6 [[S]]REG & p [[S]]FOC; [[it-is-also-the-

case-that-s]] is undefined otherwise.

(Of course a different treatment of presupposition would require modifying

(11) in the appropriate way.)
Even is even more interesting (pun intended). A sentence like It’s even the case
that BALTIMORE beat the Red Sox requires some way to rank the alternative
propositions in the focus value of BALTIMORE beat the Red
Sox on a scale—perhaps a scale of likelihood. Moreover, it presupposes or at least
implies that all of the propositions which are lower-ranked are true, and it asserts
that the ordinary value of BALTIMORE beat the Red Sox is true. The exact nature
of the types of scales required by even has been the subject of considerable
research; see, e.g., Kay (1990) for interesting work on this.

18.4. Further compositional issues

There are two further issues about the compositional semantics of focus values
that we mention here without solution, as the goal is just to whet the reader’s
appetite to pursue this further. First, the computation of the alternatives in the
focus value need not start with the actual item that is stressed. In a context in
which I am at my vacation house in New Hampshire after a snowstorm and
someone asks me what I did this morning I can answer (12): (12) I spent the morning
shoveling SNOW.

Most likely in this context what is at issue is surely not alternative substances that
I might have been shoveling. Rather, the focus value in the relevant context is a set
like {I spent the morning writing my textbook; I spent the morning cross-country
skiing; I spent the morning shoveling snow}. In other words, the focus is really on
the entire VP shoveling snow, but the rules of English prosody are such that the
actual stress needs to land on one word, and so it lands on snow. This is called
broad focus. The set-up in this chapter does not account for the broad-focus
interpretation and so additional principles effecting a mapping between the actual
prosody and the computation of alternatives are needed. For some discussion,
see Selkirk (1984), but note that this is not embedded in a kind of Direct
Compositional architecture and allows the grammar to crucially refer to trees.
 397
Thus a detailed account of the mapping between the prosody and the contribution
of focus is an interesting project for the Direct Compositional program.
The second issue concerns what is known as second occurrence focus. (5) (the
Porky/truffle case) is a case where more than one item has focal stress, and (9) shows
what can happen when such a sentence combines with it is only the case. Here we
set up the Concerned Farmer MacDonald scenario, which supported a reading in
which the effect of it is only the case was to negate all of the alternatives. But only (or
its more verbose counterpart it is only the case that) can also “swallow up” just one of
the items in focus, while the other remains active and is thus available for additional
focus-sensitive items. The sentences and scenarios are sufficiently complex that
using the circumlocutious it is only/also the case only makes things worse, so we
illustrate with the more natural uses of only and also. Thus let us return to Farmer
MacDonald. This time his concern is that all of the edible plants except the arugula on
his farm will be killed in a severe drought brought on by global warming. As a result, he
is training all of his animals to get used to eating only arugula. Then we can say:
(13) a. Farmer MacDonald has begun to feed only ARUGULA to Elsie.
b. He’s also begun to feed only ARUGULA to EEYORE.

ARUGULA in
(13b) is in small capital letters to indicate that it does not receive the same
phonological prominence as EEYORE, but it does receive some (see Rooth 1996 for
discussion of the actual acoustics here).
Note first that ARUGULA is what gives rise to the alternatives that only in (13b) is
sensitive to. Assume that also has scope over only (as both the semantics and the
word order would suggest). Then here only “swallows up” the alternative propositions
that are introduced by the prosodic focus on ARUGULA but it cannot have the effect of
saying that the only plant-fed-toanimal pair is arugula to Eeyore. After all, (13a) asserts
(or presupposes) that MacDonald feeds arugula to Elsie as well. (And (13b) itself
presupposes that MacDonald feeds arugula to some other animal.) Hence here the
compositional contribution of only is to negate alternatives such as {MacDonald feeds
truffles to Eeyore, MacDonald feeds clover to Eeyore, . . . }, saying nothing about the
other animals. Then role of also is to introduce the presupposition (satisfied in virtue
of (13a)) that indeed there is a true proposition of the form “Mac feeds only arugula to
x” for some x besides Eeyore. The key point, then, is that the alternatives stemming
from focal
 398 18. THE SEMANTICS OF FOCUS

18.5. FOCUS AND BINDING

stress on EEYORE have to still be active after only has made its contribution, and these
are then relevant to the contribution of also.
Another example to bring this out can be constructed with even. Here we add the
fact that donkeys are well known to be quite adverse to arugula. We can then follow
(13a) with (13c), where the role of even is parallel to that discussed above for also:
(13) c. He has even begun to feed only ARUGULA to EEYORE.

This means that the compositional computation of a sentence with two foci must
sometimes make a kind of nested focus where one focus-sensitive word operates on
one set of alternatives, and another on the alternatives to those alternatives. We leave
it open here as to how to accomplish this. Incidentally, the same remarks hold if we
replace also, even, and only with their more verbose counterparts it is only/also/even
the case that. This means that the phenomenon demonstrated is not tied to the fact
that only in (13b) and (13c) forms a constituent with ARUGULA; the reader can construct
the relevant examples.

18.5. Focus and binding

Section 15.4.1 showed that the pronoun in (14) can have the understanding shown
there either by being a free pronoun or a bound pronoun.

(14) Jacki called hisi mother.

One piece of evidence for this was the existence of both the strict and the sloppy
reading with VP Ellipsis, as in (15) (see the discussion in section 15.4.1.3):
(15) Jacki called hisi mother and Bill did too.

Similar facts emerge with focus.5 Thus note that focal stress on JACK in (16) conjures
up two different kinds of alternative sets:
(16) JACKi called hisi mother.

One possibility is that what is at issue is what other boys called their own mothers (call
this the bound, sloppy, or co-varying reading). This emerges

5
 399
It is well known that there are a number of similarities between ellipsis and focus; see
Rooth (1992) for some relevant discussion.
naturally in the Mothers’ Day Scenario of section 15.4.1.3. The other is the “free”
reading—i.e., what is at issue is who else called Jack’s mother; this emerges in the
Hospitalized Mother scenario. As one might expect, one can even more clearly bring
out the two readings by embedding (16) it is only the case that, where the two readings
are truth-conditionally distinct:

(17) It is only the case that JACKi called hisi mother.

It is worth considering how the accounts of binding given in Chapter 15 can be folded
in with the account of focus to explicitly predict this. We will work this through in the
variable-free account, leaving it as an exercise for the reader (in 18.4) to work through
the details in the LF account.
Before considering the bound or co-varying reading, a word is necessary about the
focus values of expressions containing free pronouns in the variable-free setting. In
other words, consider first a simple case like (18) where him is simply a free pronoun:
(18) NORA called him.

The details of the compositional computation are beyond the scope of this work, but
a reasonable guess is that the focus value of (18) will be a function from individuals to
alternative propositions. In other words, the focus value of (18) is a function of the form
ºx[{nora called x, sally called x, macdonald called x, . . . }]. Like with the regular value
of an expression with a free pronoun, this will be applied by the listener to some
contextually salient individual which ultimately gives the set of alternative
propositions of relevance. Thus if the pronoun in (16) is not bound, the focus value of
(16) is the set of alternatives ºx[{Jack Sprat called x’s mother, Billy called x’s mother,
Tom Thumb called x’s mother . . . }]. Applying this to the contextually salient Jack gives
the relevant free reading.6
The bound reading is also straightforward (under any theory of binding), as long as
we continue to assume that the computation of the focus value of any sentence
always parallels the computation of its regular value. Thus one way to arrive at the
relevant understanding (of the regular value) of (16) is via application of z to call. The
result of this is that z-call his mother is the set of self-mother-callers, which then takes
[[Jack]] as argument. The focus

6
More needs to be said as something is needed to ensure that the individual taken as
argument of the regular value is the same as the one taken as argument of the focus value. See
Jacobson (2012) for details.
 400 18. THE SEMANTICS OF FOCUS

18.6. MULTIDIMENSIONAL SEMANTICS

value of the lexical item call is just {call’}. But since the focus computation exactly
parallels the regular computation, this means that when this maps to z(call’) in the
computation of the regular value, exactly the same will happen in the focus value. his
mother is the mother-of function; and so the focus value of call his mother will be
{z(call’)(the-mother-of’)}, still a singleton set. But because [[JACK]] FOC is a set like {j, b,
s, . . . }, the focus value of JACK called his mother will be the set {j z-called the-mother-
of, b zcalled the-mother-of, s z-called the-mother-of, . . . }. And this is just what we
want for the bound or co-varying reading of (16). (17) is straightforward; it is only the
case combines with this, and hence the regular value of (17) asserts that Jack is a self
’s mother-caller, and says that none of the other propositions in the focus value of (16)
is true.

18.4. Work through the analysis of (16) on the bound reading using the LF theory of
pronominal binding (section 15.2), while still maintaining the basic set-up for focus
computation given here. (Note that that set-up is neutral with respect to the issue
of how to do binding.)

18.6. More generally: Multidimensional semantics

The computation of alternatives and the way in which they interact with focus-
sensitive words like only is, as noted earlier, the poster child for the notion that the
compositional rules can compute more than one dimension simultaneously. But the
focus value is not the only dimension that has been explored in this regard. In a classic
paper, Karttunen and Peters (1979) proposed that presuppositions and what are
known as conventional implicatures can also be computed compositionally in parallel
with the “main semantic” value of a sentence. (A conventional implicature is an
inference that is tied to linguistic form—such as the implicatures associated with
but— and hence not derived by Gricean principles, but which does not appear to be
part of the truth conditions.) For some discussion of multidimensional semantics, see
Potts (2005).
 19 Intensionality and the
syntax/semantics interface
19.1. Intensionality-sensitive words 399 19.2. Revising the fragment:
Type-sensitive rules 403

A chapter on intensionality could either be very long or very short. The long chapter
(which should really be at least a book) would include a rich discussion both of
lexical semantics and of issues about how natural languages carve up the
dimensions of times and the space of possible worlds. For the lexical meanings
of words which crucially combine with intensions give us a window into the latter
domain. For example, a sentence like It must be raining appears to involve some
sort of universal quantification over worlds, but surely it is not saying that it is
raining in all possible worlds (that would make it a logical necessity that it is
raining). Rather it says that it is raining in all worlds in some subset of worlds—
perhaps all those compatible with the speakers’ beliefs. Other modal auxiliaries
access other sets of worlds. (One can think about this for any of the following
cases: She can lift 150 pounds; It can sometimes snow in May; He should take
five courses this semester, etc. See Kratzer 2012 for rich discussion of the
semantics of some modal auxiliaries.) Other words that make crucial reference to
some set of worlds as part of their meaning are those verbs known as
propositional attitude verbs. As the name implies, these indicate the subject’s
attitude towards some proposition and include think, believe, hope, etc. As noted
in Exercise 4.2, even such ordinary words as bachelor arguably encode a
19.1. INTENSIONALITY-SENSITIVE WORDS

modal dimension (i.e., refer to other possible worlds), and as pointed out in
section 2.3.4 ordinary grammatical devices as the progressive construction also
make reference to possible worlds. And there are many words and morphemes
whose meanings crucially combine with items of type <i,x> for some x—i.e. words
402 19. INTENSIONALITY
whose meanings include something about times. The pasttense morpheme is one
such obvious example.
But this will be a short chapter. We give just a few examples of words which
crucially are sensitive to the intension of expressions with which they combine,
and then comment on the ways in which the fragment needs revision to take this
into account.

19.1. Intensionality-sensitive words

First consider a verb like think, as in (1):

(1) Sarah thinks that the world’s largest bell weighs 800 pounds.

For convenience, we have been treating think’ as of type <t,<e,t>> (with frequent
reminders that this is oversimplified). The reason that think’ cannot really be just
of type <t,<e,t>> is that this would make it a function from {1,0} to (the
characteristic function of) a set of individuals. But Sarah is not having some
thought about 1 or 0, but rather a proposition. So in reality, think’ is (ignoring times)
of type <<w,t>,<e,t>>. To assert (1) is to assert that the set of worlds consistent
with Sarah’s beliefs is a subset of the worlds in which the proposition “the world’s
largest bell weighs 800 pounds” is true. Note incidentally that this particular way
of putting it assumes that Sarah need not have a single world in her belief state.
Presumably there are many propositions about which she has no opinion and so
the object of someone’s beliefs is a set of possible worlds. Thus think’ combines
with a proposition p (a set of worlds), and returns the set of individuals x such that
for each x, the set of worlds in their belief state is a subset of p. That the set of
worlds compatible with the relevant individuals’ beliefs is only a subset of p is
easy to see, for p includes, for example, worlds in which the bell sentence is true
and Porky found Farmer MacDonald’s arugula as well as other worlds in which the
bell sentence is true and Porky failed the arugula test. Sarah might believe that
Porky failed and that the bell sentence is true, and thus her belief worlds are only
a subset of those where the bell sentence is true.
What this means is that the fragment must be revised, for the full app rule
(R-1) was formulated to have the function combine with the extension of the
expression that it takes as argument. Before making the necessary revision, let’s
give the meaning of a few other words that are sensitive to intensions. First, take
 403
one which is sensitive to times—a classic example is former (as in former
senator). This is of category N/ RN. Notice thought that (unlike ordinary adjectives
such as tall and red) it has no predicative counterpart:
(2) *That senator is former.

This is hardly surprising; the semantics for predicative adjectives in section 8.1
takes the individual denoted by the subject and says it is in the set characterized
by the adjective. But it makes no sense to speak of the set of former things; one is
former only with respect to some other set. For this reason, the prenominal
version is obviously not intersective in the sense that red or even tall is (recall the
semantics of R-2 (mod) which is responsible for the prenominal versions of tall
and red, and recall the discussion in section 8.4 justifying the claim that tall is
intersective. To show that former is not intersective, we can note not only that
there is no set of things that are “former,” but also that a former senator is not a
senator (unlike a tall senator). Hence prenominal former is not derived by R-2 from
a predicative version; it is just listed in the lexicon as N/N.

*19.1. This fact actually introduces a wrinkle into the fragment so far. Single-
word nominal modifiers are exceptional in their directional feature, and this
was handled by a clause regarding the output of R-2 (mod) which treats single-
word outputs in an exceptional way. Comment on why the behavior of former
creates a problem for this account, and on a possible way or ways to solve this.

What about the meaning of former? As it combines with a time-dependent set

to give a time-dependent set, its meaning is as follows:

(3) [[former]] = ºP<i,<e,t>>[ºi[ºx[∃i’[i’ < i & P(x)(i’) & ~P(x)(i)]]]]

In prose: this takes a time-dependent set P and returns a function that maps an
individual x to true at a time i if there is an earlier time i’ at which x was in P and x
is no longer in P.
404 19. INTENSIONALITY
19.1. INTENSIONALITY-SENSITIVE WORDS

19.2. If one describes Tim as now being a former former senator, it must be the
case that Tim is now a senator. Show exactly why this inference follows, given
the definition in (3).

Former is sensitive to intensions in the sense of combining with timedependent

sets. Other prenominal modifiers of this sort include future (a future astronaut),
perhaps past (though the range of nouns that past can combine with seems rather
limited), and a few others whose meaning is more complex and arguably contains
a hidden variable or argument slot as part of their meaning; see the following
exercise.

*19.3. Thus note the following sentence, and comment on its implications for
the analysis of the meaning of previous (and for the issues discussed in section
17.4).

(i) Every class begins with a review of the previous class.

There are also a few prenominal adjectives sensitive not necessarily to times
but to worlds. A nice example is the word wannabe which might be unfamiliar to
some readers as it is a reasonably recent coinage (1980s according to some
online sources) and may be best-known in American English. (But to a reader
unfamiliar with this word, do not worry, for we will actually define it.) Examples
are in (4):
(4) a. The wannabe drill sergeant spent hours and hours in the gym doing pushups.
b. Every syntactician is a wannabe semanticist.

While there might be some additional implicatures associated with this word at
least a rough approximation to its meaning is in (5) (we ignore the temporal
parameter here):
(5) ºP<w,<e,t>>[ºw[ºx[8w’ [w’ in set of want-worlds of x in w, P(w’)(x)]]]]
 405
Not surprisingly, wannabe cannot be used predicatively:
(6) *The semanticist/drill sergeant is wannabe.

(Although wannabe has by now also shifted to a noun such that a wannabe is now
a predicative noun, with perhaps an understood noun argument slot as in “You’ve
become a wannabe,” from <http://www.copyblogger.com/areyou-a-wannabe/>).

*19.4. Most of the clearly non-intersective adjectives (such as former, future,

wannabe, etc.) do not have predicative forms, as we would expect. But there is
at least one whose behavior is puzzling. Thus a fake gun is necessarily not an
actual gun, yet (i) is perfectly fine:

(i) That gun is fake.

But notice also the existence of expressions like toy gun, water gun, and— most
interestingly—real gun (along with That gun is real). While the first two are noun
compounds and might have frozen meanings, it is still the case that these—and
in particular real gun—suggests that part of the solution might lie in there being
an extended meaning of gun. (Note that this is not limited to gun; cf. fake
Christmas tree.) Elaborate on this, and suggest a possible meaning for fake if
this is correct. Most interestingly, discuss the semantic composition of (ii)
(recalling relevant discussion from Chapter 17):

(ii) Everything in that store is fake.

This question is meant to be entirely open-ended and is an invitation to

speculation; don’t expect to necessarily find a tidy answer.
19.5. Write out a meaning for the auxiliary will, as in (i):

(i) The Red Sox will beat the Yankees.

The task here is to think of this purely as a “future operator” (there might well
be additional aspects to the meaning of will brought out in other contexts, but
these can be ignored here). Note though that one does need some notion of the
structure of time. Assume that time is a set of discrete points ordered by a
relation < (earlier than).

19.2. REVISING THE FRAGMENT

406 19. INTENSIONALITY

19.2. Revising the fragment:

Type-sensitive rules

Once intensionality is folded in, the combinatory and unary rules given in
Appendices to Parts I–III are in need of some revision. Before delving into this, we
will for simplicity fold the worlds and times in together as has been done in various
sections in this book, and continue to use s for the set of world/time pairs. Now
consider the app rule. Most of the cases considered in this book involved the
application of a function to the extension of its argument. Hence recall that R-1—
the app rule—was formulated for this. In the fragment as spelled out in the
Appendices it is given extensionally, but the full version that is needed for most of
the material up until this chapter would be INT-R-1 as follows:
INT-R-1. If Æ is an expression of the form <[Æ], A/B, Æ’> and is an expression of the form
<[], B, ’> then there is an expression ª of the form
<FCAT(Æ)([Æ])([]), A, ºw[Æ’(w)(’(w))]>.

This is a perfectly straightforward revision. The problem is that it is not always

correct—for all of the cases considered above involve cases where a function
applies to the intension and not the extension of the expression with which it
combines. The syntax looks exactly the same either way. Thus former and the
prenominal modifier tall are both of category N/RN, but the first applies to the
intension of senator in former senator and the latter to the extension (in tall
senator). So the semantic composition is not always a transparent reflection of
the syntactic composition.
There are (at least) two conceivable solutions. The first is the one taken in
Montague (1973) which a reader who has carefully followed the various
Montagovian solutions mentioned throughout might be able to guess. Keep the
semantic composition uniform and simple, and put the complexity in the lexical
entries. In other words, this strategy is to revise R-1 so that a function always
applies to an intension. But those words that “really” want only extensions do so
by building this fact into their lexical meaning. Take, for example, grunt, whose
meaning we have been assuming is of type <s<e,t>>. Montague’s suggestion to
keep the combinatorics completely uniform is to revise this so that its meaning is
of type <s,<<s,e>,t>>. Call grunt*’ the “ordinary” grunt function—i.e., the function
 407
that maps all grunters to 1 and non-grunters to 0. Then one can recast [[grunt]] as
what we will call grunt’-
MONT as follows: ºs[ºX<s,e>[grunt*’(s)(X(s))]]. This can then apply to the intension of
the truffle-loving pig and the lexical meaning ensures that ultimately the actual
grunt function is applied to the extension of that NP at the world in question.
This is a possible strategy, and it will mean that all of the other rules in our
fragment need only minor modification. But it is a trade-off: while keeping the
rules simple and elegant it seems (to many modern researchers) to be a rather
heavy-handed approach to lexical meaning. The second solution, which is more
in keeping with the more “modern” thinking on such things, is to keep the lexical
meanings simpler but simply make R-1 slightly more complex. Indeed, the
strategy is simply to use a version of the notion of typedriven interpretation. Recall
(see the discussion in section 6.5) that this was originally proposed in Klein and
Sag as a means of allowing the semantic side of the syntactic combinatory rules
to be predictable from the semantic types of the combining expressions. For the
most part this was not needed in the CG fragment here because the rules
themselves are stated extremely generally. Later this idea was imported into non-
Direct Compositional frameworks as a way of predicting just what semantic
operation will interpret a mother node in a tree on the basis of the types of its
daughter(s). (See especially Heim and Kratzer 1998.) But here we will reintroduce
this notion as a way to predict when intensions are crucial. The basic idea is
simple and obvious: if a function asks for an extension, it gets an extension; if it
asks for an intension, it gets an intension. To make this precise, we simply define
the semantic operation app as follows:

(7) Type-sensitive app:

i. If Æ’ is of type <s,<a,b>> (for a 6¼ <s,x> for some x), and ’ is of type <s,a>, then

app(Æ’(’)) = ºs[Æ’(s)(’(s))].

ii. If Æ’ is of type <s,<<s,a>,b>> and ’ is of type <s,a>, then app(Æ’(’)) = ºs[Æ’(s)(’)].

This, however, is not quite enough, as we need to revise some of the other rules
to also take into account intensions and operate in a type-driven way. Note first
that all of the rules need minor “tweaking” in that they apply to expressions whose
actual values are always intensional objects; i.e., their values are of type <s,a> (for
some a), while the rules were formulated purely extensionally. For the most part
these revisions are straightforward and the reader can supply the details. (The
408 19. INTENSIONALITY
strategy is quite similar to the strategy seen earlier of having rules ignore argument
positions in the semantics that correspond to superscript positions in the
19.2. REVISING THE FRAGMENT

syntax.) But some of the rules also need to be stated in a type-sensitive way, as
they will need to be sensitive to the distinction between extensions and
intensions. We give just one example, which is the lift rule (l). If senator is lifted
from its basic meaning to combine with a modifier like former it must do so in a
different way than when it combines with tall. Hence we restate l as follows:

(8)Type-sensitive l:
i. Take any expression Æ of type <s,a>. Then ext-l(Æ’) is a function of type <s,<a,b>,b>
such that ext-l(Æ’) = ºs[ºf<a,b>[f(Æ’(s))]]. ii Take any expression Æ of type <s,a>. Then
int-l(Æ’) is a function of type <s,<<s,a>,b>,b> such that int-l(Æ’) = ºs[ºf<a,b>[f(Æ’)]].

Either version of l can apply to senator, but only the intensional lift in (ii) will yield
an output that can take former as argument.
Ultimately one would hope to avoid separate statements for each of the rules
that need to be recast and to have very general statements governing the
intensional versions of all the rules, but this will not be attempted here. We will,
however, note one other possible way to revise the earlier fragment. Recall that
the strategy of having the closed-class items denote constant functions was
motivated by giving a uniform type and combinatorics for everything. The
Montagovian strategy noted just above of course takes this to the extreme. But if
it is correct to depart in any case from that strategy, one might rethink this
treatment of the closed-class items, and have them merely have extensions. One
might, then, add a further type-sensitive version of app as follows:

iii. If Æ’ is of type <a,b> (for a 6¼ s), and ’ is of type <s,a>, then app(Æ’(’)) =

ºs[Æ’(’(s))].

So, for example, a case-marking preposition like of can be taken to be just the
identity function over individuals (i.e., of type <e,e> and not of type <s,<e,e>>); if it
combines with the fastest skier (of type <s,e>) it does so by (iii) above (the result
is that of the fastest skier will have the same meaning as the fastest skier). Of
course if this strategy is correct, additional principles will be needed to fill out the
full system. In the end, the trade-off between simplifying the lexical entries and
 409
potentially complicating the various rules will depend on just exactly how the
latter are stated and how generally this can be done.
This will not be attempted here. We leave this as a project for any interested
reader, along with many of the other suggestions and open questions that have
emerged throughout this text. And of course there are vast areas of natural
language semantics (including cross-linguistic work) not even touched in this
book. But we hope to have given the reader the background, tools, and—most
importantly—the appetite to delve into this endlessly rich territory.
 References

AJDUKIEWICZ, KAZIMERZ. 1935. Die syntaktische Konnexität. Studia Philosophica 1: 1–27.

ANDERSON, ALAN. 1951. A note on subjunctive and counterfactual conditionals. Analysis 12:
35–8.
BACH, EMMON. 1968. Nouns and noun phrases. In E. Bach and R. Harms (eds), Universals in
Linguistic Theory. New York: Holt, Rinehart and Winston, 90–122.
BACH, EMMON. 1979. Control in Montague grammar. Linguistic Inquiry 10: 515–31.
BACH, EMMON. 1980. In defense of passive. Linguistics and Philosophy 3: 297–341.
BACH, EMMON and ROBIN COOPER. 1978. The NP-S analysis of relative clauses and
compositional semantics. Linguistics and Philosophy 2: 145–50.
BACH, EMMON, ELOISE JELINEK, ANGELIKA KRATZER, and BARBARA PARTEE (eds). 1995. Quantification
in Natural Languages. Dordrecht: Kluwer.
BARKER, CHRIS. 1995. Possessive Descriptions. Stanford: CSLI.
BARKER, CHRIS. 2001. Introducing continuations. In R. Hastings, B. Jackson, and Z.
Zvolenszky (eds), SALT XI: Proceedings of the Eleventh Conference on Semantics and
Linguistic Theory. Ithaca: CLC, 20–35.
BARKER, CHRIS and CHUNG-CHIEH SHAN. 2008. Donkey anaphora is in-scope binding.
Semantics and Pragmatics 1: 1–42.
BARSS, ANDREW and HOWARD LASNIK. 1986. A note on anaphora and double objects. Linguistic
Inquiry 17: 347–54.
BARWISE, JON and ROBIN COOPER. 1981. Generalized quantifiers and natural language.
Linguistics and Philosophy 4: 159–219.
BAYER, SAMUEL. 1996. The coordination of unlike categories. Language 72: 579–616.
BENNETT, MICHAEL. 1974. Some extensions of a Montague fragment of English. PhD
dissertation, UCLA.
BOLINGER, DWIGHT. 1967. Adjectives in English: Attribution and predication. Lingua 18: 1–34.
BOUTON, LAWRENCE. 1970. Antecedent-contained pro-forms. In M. Campbell,
J. Lindholm, A. Davison, W. Fisher, L. Furbee, J. Lovins, E. Maxwell, J. Reighard, and S.
Straight (eds), Papers from the Sixth Regional Meeting of the Chicago Linguistic Society.
Chicago: Chicago Linguistic Society, 154–67.
BREHENY, RICHARD. 2003. A lexical account of implicit (bound) contextual dependence. In R.
Young and Y. Zhou (eds), SALT XIII: Proceedings of the Thirteenth Conference on
Semantics and Linguistic Theory. Ithaca: CLC, 37–54.
BRESNAN, JOAN and JANE GRIMSHAW. 1978. The syntax of free relatives in English. Linguistic
Inquiry 9: 331–91.
CARLSON, GREGORY. 1980. Reference to Kinds in English. New York: Garland.
 REFERENCES 411
CARLSON, GREGORY. 1981. The distribution of free-choice any. In R. Hendrick, C. Masek, and
M. Miller (eds), Papers from the Seventeenth Regional Meeting of the Chicago Linguistic
Society. Chicago: Chicago Linguistic Society, 8–23.
CARPENTER, BOB. 1998. Type-Logical Semantics. Cambridge, MA: MIT Press.
CHIERCHIA, GENNARO. 1991. Functional WH and weak crossover. In D. Bates (ed.),
Proceedings of the Tenth West Coast Conference on Formal Linguistics. Stanford:
CSLI, 75–90.
CHIERCHIA, GENNARO. 2004. Scalar implicatures, polarity phenomena, and the syntax/
pragmatics interface. In A. Belletti (ed.), Structures and Beyond, Volume 3: The
Cartography of Syntactic Structures. Oxford: Oxford University Press, 39–103.
CHIERCHIA, GENNARO, DANNY FOX, and BENJAMIN SPECTOR. 2012. The grammatical view of scalar
implicatures and the relationship between semantics and pragmatics. In P. Portner, C.
Malenborn, and K. von Heusinger (eds), Semantics: An International Handbook of
Natural Language Meaning, Volume 3. Berlin:
Mouton de Gruyter, 2297–332.
CHIERCHIA, GENNARO and SALLY MCCONNELL-GINET. 1990. Meaning and Grammar:
An Introduction to Semantics. Cambridge, MA: MIT Press.
CHOMSKY, NOAM. 1957. Syntactic Structures. The Hague: Mouton.
CHOMSKY, NOAM. 1965. Aspects of the Theory of Syntax. Cambridge, MA: MIT
Press.
CHOMSKY, NOAM. 1970. Deep structure, surface structure and semantic interpretation. In D.
Steinberg and L. Jakobovits (eds), Semantics. Cambridge: Cambridge University Press,
183–214.
CHOMSKY, NOAM. 1976. Conditions on rules of grammar. Linguistic Analysis 2:
303–51.
CHOMSKY, NOAM. 1977. On wh-movement. In P. W. Culicover, T. Wasow, and A. Akmajian
(eds), Formal Syntax. New York: Academic Press, 71–132. CHOMSKY, NOAM. 1981. Lectures
on Government and Binding. Holland: Foris
Publications.
COOPER, ROBIN. 1979. The interpretation of pronouns. In F. Heny and H. Schnelle (eds),
Syntax and Semantics, Volume 10: Selections from the Third Groningen Round Table.
New York: Academic Press, 61–92.
CORMACK, ANNABEL. 1984. VP anaphora: Variables and scope. In F. Landman and F. Veltman
(eds), Varieties of Formal Semantics. Dordrecht: Foris Publications,
81–102.
CURRY, HASKELL AND ROBERT FEYS. 1958. Combinatory Logic, Volume 1. Amsterdam: North
Holland.
DAYAL, VENEETA. 1995. Licensing any in non-negative/non-modal context. In M. Simons and
T. Galloway (eds), SALT V: Proceedings of the Fifth Conference on Semantics and
Linguistic Theory. Ithaca: CLC, 72–93.
DOWNING, PAMELA. 1977. On the creation and use of English compound nouns. Language 53:
810–42.
 412 REFERENCES
DOWTY, DAVID. 1979. Word Meaning and Montague Grammar. Dordrecht: Kluwer.
DOWTY, DAVID. 1982. Grammatical relations and Montague grammar. In P. Jacobson and G.
Pullum (eds), The Nature of Syntactic Representation. Dordrecht: Reidel, 79–130.
DOWTY, DAVID. 1988. Type-raising, functional composition, and non-constituent
coordination. In R. Oehrle, E. Bach, and D. Wheeler (eds), Categorial Grammars and
Natural Language Structures. Dordrecht: Kluwer, 153–98.
DOWTY, DAVID. 1992. Variable-free syntax, variable-binding syntax, the natural deduction
Lambek calculus, and the crossover constraint. In J. Mead (ed.), Proceedings of the
Eleventh West Coast Conference on Formal Linguistics.
Stanford: CSLI, 161–76.
DOWTY, DAVID, ROBERT WALL, and STANLEY PETERS. 1981. Introduction to Montague Semantics.
Dordrecht: D. Reidel.
ELBOURNE, PAUL. 2005. Situations and Individuals. Cambridge, MA: MIT Press.
EMONDS, JOSEPH. 1976. A Transformational Approach to Syntax: Root, StructurePreserving,
and Local Transformations. New York: Academic Press.
ENGDAHL, ELISABET. 1986. Constituent Questions: The Syntax and Semantics of Questions with
Special Reference to Swedish. Dordrecht: Reidel.
ENGDAHL, ELISABET. 1988. Relational interpretation. In R. Kempson (ed.), Mental
Representations: The Interface between Language and Reality. Cambridge:
Cambridge University Press, 63–82.
EVANS, FREDERIC. 1988. Binding into anaphoric verb phrases. In J. Powers and K. de Jong
(eds), Proceedings of the Fifth Annual Eastern States Conference on
Linguistics. Columbus: Ohio State University, 122–9.
EVANS, GARETH. 1980. Pronouns and variables. Linguistic Inquiry 11: 337–62.
FAUCONNIER, GILLES. 1975. Pragmatic scales and logical structure. Linguistic Inquiry 6: 353–
75.
VON FINTEL, KAI. 1994. Restrictions on quantifier domains. PhD dissertation, University of
Massachusetts.
VON FINTEL, KAI. 1999. NPI licensing, Strawson entailment, and context dependency. Journal
of Semantics 16: 97–148.
FOX, DANNY. 2002. Antecedent-contained deletion and the copy theory of movement.
Linguistic Inquiry 33: 63–96.
FOX, DANNY. 2003. On logical form. In R. Hendrick (ed.), Minimalist Syntax.
Oxford: Blackwell, 82–123.
FOX, DANNY. 2007. Free choice disjunction and the theory of scalar implicatures. In U.
Sauerland and P. Stateva (eds), Presupposition and Implicature in Compositional
Semantics. New York: Palgrave Macmillan, 71–120.
FOX, DANNY and JON NISSENBAUM. 1999. Extraposition and scope: A case for overt qr. In S. Bird,
A. Carnie, J. Haugen, and P. Norquest (eds), Proceedings of the Eighteenth West Coast
Conference on Formal Linguistics. Somerville, MA: Cascadilla Press, 132–44.
FRAZIER, LYN. 1978. On comprehending sentences: Syntactic parsing strategies. PhD
dissertation, University of Connecticut.
 REFERENCES 413
GAZDAR, GERALD. 1982. Phrase structure grammar. In P. Jacobson and G. Pullum (eds), The
Nature of Syntactic Representation. Dordrecht: Kluwer, 131–86.
GAZDAR, GERALD, EWAN KLEIN, GEOFFREY PULLUM, and IVAN SAG. 1985. Generalized Phrase
Structure Grammar. Oxford: Blackwell and Cambridge, MA: Harvard University Press.
GAZDAR, GERALD and IVAN SAG. 1981. Passive and reflexives in phrase structure grammar. In J.
Groenendijk, T. Janssen, and M. Stokhof (eds), Formal Methods in the Study of Language.
Amsterdam: Mathematical Centre Tracts, 131–52.
GEACH, PETER T. 1962. Reference and Generality: An Examination of Some Medieval and
Modern Theories. Ithaca: Cornell University Press.
GEACH, PETER T. 1970. A program for syntax. Synthese 22: 3–17.
GEURTS, BART. 2009. Scalar implicature and local pragmatics. Mind and Language 24: 51–79.
GIANNAKIDOU, ANASTASIA. 1997. The landscape of polarity items. PhD dissertation, University of
Groningen.
GRICE, H. PAUL. 1975. Logic and conversation. In P. Cole and J. Morgan (eds), Syntax and
Semantics, Volume 3: Speech Acts. New York: Academic Press,
41–58.
GRIMSHAW, JANE. 1982. Subcategorization and grammatical relations. In A. Zaenen (ed.),
Subjects and Other Subjects. Bloomington, IN: Indiana University Linguistics Club, 35–
55.
GROENENDIJK, JEROEN and MARTIN STOKHOF. 1984. Studies on the semantics of questions and
the pragmatics of answers. PhD dissertation, University of Amsterdam.
GROENENDIJK, JEROEN and MARTIN STOKHOF. 1991. Dynamic predicate logic. Linguistics and
Philosophy, 14: 39–100.
HAMBLIN, C. L. 1973. Questions in Montague English. Foundations of Language 10:
41–53.
HANKAMER, JORGE and IVAN SAG. 1984. Toward a theory of anaphoric processing. Linguistics
and Philosophy 7: 325–45.
HAYES, BRUCE and MAY ABAD. 1989. Reduplication and syllabification in Ilokano. Lingua 30:
331–74.
HEIM, IRENE. 1982. The semantics of definite and indefinite noun phrases. PhD dissertation,
University of Massachusetts.
HEIM, IRENE. 1990. E-type pronouns and donkey anaphora. Linguistics and Philosophy 13:
137–77.
HEIM, IRENE and ANGELIKA KRATZER. 1998. Semantics in Generative Grammar.
Oxford: Blackwell.
HENDRIKS, HERMAN. 1993. Studied flexibility: Categories and types in syntax and semantics.
PhD dissertation, University of Amsterdam.
HEPPLE, MARK. 1990. The grammar and processing of order and dependency: A categorial
approach. PhD dissertation, University of Edinburgh.
HORN, LAURENCE. 1972. On the semantic properties of logical operators in English.
PhD dissertation, UCLA.
HORN, LAURENCE. 1989. A Natural History of Negation. Chicago: University of Chicago Press.
 414 REFERENCES
HORN, LAURENCE. 2004. Implicature. In L. Horn and G. Ward (eds), The Handbook of
Pragmatics. Oxford: Blackwell, 3–28.
HUYBREGTS, RINY. 1976. Overlapping dependencies in Dutch. Utrecht Working Papers in
Linguistics 1: 24–65.
JACKENDOFF, RAY. 1972. SemanticInterpretation in Generative Grammar. Cambridge,
MA: MIT Press.
JACOBS, RODERICK and PETER ROSENBAUM. 1968. English Transformational Grammar. Waltham,
MA: Xerox College Publishing.
JACOBSON, PAULINE. 1977. The syntax of crossing coreference sentences. PhD dissertation,
University of California at Berkeley (published by Garland Press,
New York, 1980).
JACOBSON, PAULINE. 1983. On the Syntax and Semantics of Multiple Relatives in English.
Bloomington, IN: Indiana University Linguistics Club.
JACOBSON, PAULINE. 1987. Phrase structure, grammatical relations, and discontinuous
constituents. In A. Ojeda and G. Huck (eds), Syntax and Semantics, Volume 20:
Discontinuous Constituency. New York: Academic Press, 27–69.
JACOBSON, PAULINE. 1992. Antecedent contained deletion in a variable-free semantics. In C.
Barker and D. Dowty (eds), SALT II: Proceedings of the Second Conference on Semantics
and Linguistic Theory. Columbus: Ohio State University, 193–213.
JACOBSON, PAULINE. 1993. i-within-i effects in a variable-free semantics and a categorial
syntax. In P. Dekker and M. Stokhof (eds), Proceedings of the Ninth Amsterdam
Colloquium. Amsterdam: ILLC, University of Amsterdam, 349–68.
JACOBSON, PAULINE. 1998. Antecedent-contained deletion and pied-piping: Evidence for a
variable-free semantics. In D. Strolovitch and A. Lawson (eds), SALT VIII: Proceedings
of the Eighth Conference on Semantics and Linguistic Theory. Ithaca:
CLC, 74–91.
JACOBSON, P AULINE. 1999. Towards a variable-free semantics. Linguistics and Philosophy 22:
117–85.
JACOBSON, PAULINE. 2000. Paycheck pronouns, Bach-Peters sentences, and variablefree
semantics. Natural Language Semantics 8: 77–155.
JACOBSON, PAULINE. 2003. Binding without pronouns (and pronouns without binding). In G. J.
Kruiff and R. Oehrle (eds), Binding and Resource Sensitivity.
Dordrecht: Kluwer, 57–96.
JACOBSON, PAULINE. 2007. Direct compositionality and variable-free semantics: The case of
“Principle B” effects. In P. Jacobson and C. Barker (eds), Direct Compositionality. Oxford:
Oxford University Press, 191–236.
JACOBSON, PAULINE. 2008. Direct compositionality and variable-free semantics: The case of
antecedent-contained deletion. In K. Johnson (ed.), Topics in Ellipsis. Cambridge:
Cambridge University Press, 30–68.
JACOBSON, PAULINE. 2012. Direct compositionality and “uninterpretability”: The case of
(sometimes) “uninterpretable” features on pronouns. Journal of Semantics
29: 305–43.
 REFERENCES 415
JÄGER, GERHARD. 1996. Topics in Dynamic Semantics. PhD dissertation, Humboldt
University.
KADMON, NIRIT. 1987. On unique and non-unique reference and asymmetric quantification.
PhD dissertation, University of Massachusetts. (Published by Garland Press, New York,
1993.)
KADMON, NIRIT and FRED LANDMAN. 1993. Any. Linguistics and Philosophy 16:
353–422.
KAMP, HANS and UWE REYLE. 1993. From Discourse to Logic: Introduction to Model-Theoretic
Semantics of Natural Language, Formal Logic and Discourse Representation Theory.
Dordrecht: Kluwer.
KANAZAWA, MAKOTO. 1994. Weak vs. strong readings of donkey sentences and monotonicity
inferences in a dynamic setting. Linguistics and Philosophy 17: 109–58.
KARTTUNEN, LAURI. 1969. Pronouns and variables. In R. Binnick, A. Davison, G. Green, and J.
Morgan (eds), Proceedings of the Fifth Regional Meeting of the Chicago Linguistic
Society. Chicago: Chicago Linguistic Society, 108–16.
KARTTUNEN, LAURI. 1971. Definite descriptions with crossing coreference. Foundations of
Language 7: 157–82.
KARTTUNEN, LAURI. 1977. Syntax and semantics of questions. Linguistics and Philosophy 1: 3–
44.
KARTTUNEN, LAURI and STANLEY PETERS. 1979. Conventional implicature. In C.-K. Oh and D. A.
Dinneen (eds), Syntax and Semantics 11: Presupposition.
New York: Academic Press, 1–56.
KATZ, JERROLD and JERRY FODOR. 1963. The structure of a semantic theory. Language 39: 170–
210.
KAY, PAUL. 1990. Even. Linguistics and Philosophy 13: 59–111.
KAYNE, RICHARD. 1984. Connectedness and Binary Branching. Dordrecht: Foris
Publications.
KEENAN, EDWARD. 1971. Names, quantifiers, and a solution to the sloppy identity problem.
Papers in Linguistics 4: 1–22.
KEENAN, EDWARD and LEONARD FALTZ. 1985. Boolean Semantics for Natural Language. Dordrecht:
Reidel.
KEENAN, EDWARD and JONATHAN STAVI. 1986. A semantic characterization of natural language
determiners. Linguistics and Philosophy 9: 253–326.
KENNEDY, CHRISTOPHER. 1997. Projecting the adjective: The syntax and semantics of gradability
and comparison. PhD dissertation, University of California at Santa Cruz. (Published by
Garland Press, New York, 1999.)
KENNEDY, CHRISTOPHER. 2007. Vagueness and grammar: The semantics of relative and
absolute gradable adjectives. Linguistics and Philosophy 30: 1–45.
KLEIN, EWAN and IVAN SAG. 1985. Type-driven translation. Linguistics and Philosophy 8: 163–
201.
KLIMA, EDWARD. 1964. Negation in English. In J. FODOR and J. KATZ (eds), The Structure of
Language. Englewood Cliffs, NJ: Prentice Hall, 246–323.
 416 REFERENCES
KLUENDER, ROBERT. 1998. On the distinction between strong and weak islands: A processing
perspective. In P. Culicover and L. McNally (eds), Syntax and Semantics, Volume 29: The
Limits of Syntax. New York: Academic Press, 241–80.
KRATZER, ANGELIKA. 1986. Conditionals. In A. Farley, P. Farley, and K. McCollough (eds),
Papers from the Parasession on Pragmatics and Grammatical Theory. Chicago: Chicago
Linguistic Society, 115–35.
KRATZER, ANGELIKA. 1998. More structural analogies between pronouns and tenses. In D.
Strolovitch and A. Lawson (eds), SALT VIII: Proceedings of the Eighth Conference on
Semantics and Linguistic Theory. Ithaca: CLC, 92–109.
KRATZER, ANGELIKA. 2004. Covert quantifier restrictions in natural language. Presentation
accessible at: <semanticsarchive.net/Archive/mIzMGUyZ/Covert%
20Quantifier%20Domain%20Restrictions.pdf>.
KRATZER, ANGELIKA. 2012. Modals and Conditionals. Oxford: Oxford University
Press.
KRIFKA, MANFRED. 1995. The semantics and pragmatics of polarity items. Linguistic Analysis
25: 209–57.
KRIPKE, SAUL. 1980. Naming and Necessity. Cambridge, MA: Harvard University
Press.
KUNO, SUSUMU. 1975. Three perspectives in the functional approach to syntax. In R.
Grossman, L. J. San, and T. Vance (eds), Papers from the Parasession on
Functionalism. Chicago: Chicago Linguistic Society, 276–336.
LADUSAW, WILLIAM. 1979. Polarity sensitivity as inherent scope relations. PhD dissertation,
University of Texas, Austin.
LAHIRI, UTPAL. 1998. Focus and negative polarity in Hindi. Natural Language Semantics 6: 57–
123.
LAKOFF, GEORGE. 1971. On generative semantics. In D. Steinberg and L. Jakobovits (eds),
Semantics. Cambridge: Cambridge University Press, 232–96.
LAKOFF, GEORGE and STANLEY PETERS. 1969. Phrasal conjunction and symmetric predicates. In
D. Reibel and S. Schane (eds), Modern Studies in English: Readings in Transformational
Grammar. Englewood Cliffs, NJ: Prentice-Hall, 113–42.
LAMBEK, JOACHIM. 1958. The mathematics of sentence structure. The American
Mathematical Monthly 65: 154–70.
LANGACKER, RONALD. 1969. On pronominalization and the chain of command. In D. Reibel
and S. Schane (eds), Modern Studies in English. Englewood Cliffs, NJ:
Prentice-Hall, 160–86.
LARSON, RICHARD. 1988. On the double object construction. Linguistic Inquiry 19:
335–91.
LASNIK, HOWARD. 1976. Remarks on coreference. Linguistic Analysis 2: 1–22.
LEVINSON, STEPHEN. 1983. Pragmatics. Cambridge: Cambridge University Press.
LEVINSON, STEPHEN. 2000. Presumptive Meanings: The Theory of Generalized Conversational
Implicature. Cambridge, MA: MIT Press.
LEWIS, DAVID. 1970. General semantics. Synthese 22: 18–67.
 REFERENCES 417
LINEBARGER, MARCIA. 1987. Negative polarity and grammatical representation. Linguistics
and Philosophy 10: 325–87.
LINK, GODEHARD. 1983. The logical analysis of plurals and mass terms: A latticetheoretical
approach. In R. Bauerle, C. Schwartze, and A. von Stechow (eds), Meaning, Use, and the
Interpretation of Language. Berlin: de Gruyter, 302–23.
MAY, ROBERT. 1977. The grammar of quantification. PhD dissertation, MIT.
MAY, ROBERT. 1985. Logical Form: Its Structure and Derivation. Cambridge, MA: MIT Press.
MCCAWLEY, JAMES. 1971. Where do noun phrases come from? In D. Steinberg and L.
Jakobovits (eds), Semantics. Cambridge: Cambridge University Press, 217–31.
MCCAWLEY, JAMES. 1987. Some additional evidence for discontinuity. In G. Huck and A. Ojeda
(eds), Syntax and Semantics, Volume 20: Discontinuous Constituency. New York:
Academic Press, 185–200.
MERCHANT, JASON. 2001. The Syntax of Silence: Sluicing, Islands, and the Theory of Ellipsis.
Oxford: Oxford University Press.
MITCHELL, JONATHAN. 1986. The formal semantics of point of view. PhD dissertation, University
of Massachusetts.
MONTAGUE, RICHARD. 1970. English as a formal language. In B. Visentini et al.
(eds), Linguaggi Nella Societá e Nella Tecnica. Milan: Edizioni di Communita,
188–221.
MONTAGUE, RICHARD. 1973. The proper treatment of quantification in ordinary
English. In K. J. J. Hintikka, J. M. E. Moravcsik, and P. Suppes (eds), Approaches to Natural
Language: Proceedings of the 1970 Stanford Workshop on Grammar and Semantics.
Dordrecht: Reidel, 221–42.
MOORTGAT, MICHAEL. 1997. Categorial type logics. In J. van Benthem and A. ter Meulen (eds),
Handbook of Logic and Language. Amsterdam: Elsevier, 95–179.
MORRILL, GLYN. 1994. Type-Logical Grammar: Categorial Logic of Signs.
Dordrecht: Kluwer Academic.
MUNN, ALAN. 1993. Topics in the syntax and semantics of coordinate structures. PhD
dissertation, University of Maryland.
OEHRLE, RICHARD. 1976. The grammatical status of the English dative alternation.
PhD dissertation, MIT.
OEHRLE, RICHARD. 1990. Categorial frameworks, coordination, and extraction. In A. Halpern
(ed.), Proceedings of the Ninth West Coast Conference on Formal
Linguistics. Stanford: CSLI, 411–25.
PARSONS, TERENCE. 1990. Events in the Semantics of English. Cambridge, MA: MIT
Press.
PARTEE, BARBARA. 1973. Some transformational extensions of Montague grammar. Journal of
Philosophical Logic 2: 509–34.
PARTEE, BARBARA. 1975a. Montague grammar and transformational grammar. Linguistic
Inquiry 6: 203–300.
PARTEE, BARBARA. 1975b. Deletion and variable binding. In E. Keenan (ed.), Formal Semantics
of Natural Languages. Cambridge: Cambridge University Press, 16–34.
 418 REFERENCES
PARTEE, BARBARA. 1984. Nominal and temporal anaphora. Linguistics and Philosophy 7: 243–
86.
PARTEE, BARBARA. 1989a. Many quantifiers. In J. Powers and K. De Jong (eds), Proceedings of
the Fifth Eastern States Conference on Linguistics. Columbus: Ohio
State University, 383–402.
PARTEE, BARBARA. 1989b. Binding implicit variables in quantified contexts. In C. Wiltshire, R.
Graczyk, and B. Music (eds), Proceedings of the Twenty-Fifth Regional Meeting of the
Chicago Linguistic Society. Chicago: Chicago Linguistic
Society, 342–65.
PARTEE, BARBARA. forthcoming. A History of Formal Semantics. Oxford: Oxford University
Press.
PARTEE, BARBARA and VLADIMIR BORSCHEV. 2003. Genitives, relational nouns, and argument-
modifier ambiguity. In E. Lang, C. Maienborn, and C. FabriciusHansen (eds), Modifying
Adjuncts. Berlin: de Gruyter, 67–112.
PARTEE, BARBARA and MATS ROOTH. 1983. Generalized conjunction and type ambiguity. In R.
Bäuerle, C. Schwarze, and A. von Stechow (eds), Meaning, Use and Interpretation of
Language. Berlin: Walter de Gruyter, 361–83.
POLLARD, CARL. 1984. Generalized phrase structure grammars, head grammars, and natural
language. PhD dissertation, Stanford University.
POLLARD, CARL and IVAN SAG. 1994. Head-Driven Phrase Structure Grammar.
Chicago: University of Chicago Press.
POPE, EMILY. 1971. Answers to yes–no questions. Linguistic Inquiry 2: 69–82.
POSTAL, PAUL. 1966. On so-called “pronouns” in English. In F. Dinneen (ed.), Report of the
Seventeenth Annual Round Table Meeting on Linguistics and Language Studies.
Washington, DC: Georgetown University Press, 177–206.
POSTAL, PAUL. 1970. On coreferential complement subject deletion. Linguistic Inquiry 1:
439–500.
POSTAL, PAUL. 1971. Cross-over Phenomena. New York: Holt,
Rinehart and Winston.
POSTAL, PAUL. 1974. On certain ambiguities. Linguistic Inquiry 5: 367–424.
POTTS, CHRISTOPHER. 2005. The Logic of Conventional Implicature. Oxford: Oxford
University Press.
QUINE, WILLARD. 1960. Variables explained away. Proceedings of the American Philosophical
Society 104: 343–7.
REINHART, TANYA. 1983. Anaphora and Semantic Interpretation. London: Croom
Helm.
ROOTH, MATS. 1985. Association with focus. PhD dissertation,
University of Massachusetts.
ROOTH, MATS. 1992. Ellipsis redundancy and reduction redundancy. In S. Berman and A.
Hestik (eds), Proceedings of the Stuttgart Ellipsis Workshop. Stuttgart:
University of Stuttgart, 1–26.
 REFERENCES 419
ROOTH, MATS. 1996. On the interface principles for intonational focus. In T. Galloway and J.
Spence (eds), SALT VI: Proceedings of Sixth Conference on Semantics and Linguistic
Theory. Ithaca: CLC, 202–26.
ROSS, JOHN. 1967. Constraints on variables in syntax. Cambridge, MA: MIT Press.
ROSS, JOHN. 1969. On the cyclic nature of English pronominalization. In D. Reibel and S.
Schane (eds), Modern Studies in English. Englewood Cliffs, NJ: PrenticeHall, 187–200.
RUSSELL, BENJAMIN. 2006. Against grammatical computation of scalar implicatures. Journal
of Semantics 23: 361–82.
RUSSELL, BERTRAND. 1905. On denoting. Mind 14: 479–93.
SAG, IVAN. 1976. Deletion and logical form. PhD dissertation, MIT.
SCHÖNFINKEL, MOSES. 1924. Über die bausteine der mathematischen logik. Mathematische
Annalen 92: 305–16.
SCHUBERT, LENHART and FRANCIS JEFFRY PELLETIER. 1989. Generically speaking, or using
discourse representation theory to interpret generics. In G. Chierchia, B. Partee, and R.
Turner (eds), Properties, Types, and Meaning, Volume II:
Semantic Issues. Dordrecht: Kluwer, 193–268.
SELKIRK, ELISABETH. 1984. Phonology and Syntax: The Relation between Sound and Structure.
Cambridge: MIT Press.
SHAN, CHUNG-CHIEH and CHRIS BARKER. 2006. Explaining crossover and superiority as left-to-
right evaluation. Linguistics and Philosophy 29: 91–134.
SHARVIT, YAEL. 1998. Possessive wh-expressions and reconstruction. In P. Tamanji and K.
Kusumoto (eds), Proceedings of the Twenty-Eighth Meeting of the North Eastern
Linguistic Society. Amherst: GLSA, 409–23.
SHIEBER, STUART. 1985. Evidence against the context-freeness of natural language.
Linguistics and Philosophy 8: 333–43.
SMITH, CARLOTA. 1969. Ambiguous sentences with “and.” In D. Reibel and S. Schane (eds),
Modern Studies in English: Readings in Transformational Grammar. Englewood Cliffs,
NJ: Prentice Hall, 75–9.
STALNAKER, ROBERT. 1968. A theory of conditionals. In N. Rescher (ed.), Studies in Logical
Theory. Oxford: Oxford University Press, 98–112.
STALNAKER, ROBERT and RICHMOND THOMASON. 1973. A semantic theory of adverbs. Linguistic
Inquiry 4: 195–220.
STANLEY, JASON and ZOLTAN SZABO. 2000. On quantifier domain restriction. Mind and Language
15: 219–61.
VON STECHOW, ARNIM. 1990. Layered traces. Paper presented at the 40th Conference on
Language and Logic. Revfülöp, Hungary.
STEEDMAN, MARK. 1987. Combinatory grammars and parasitic gaps. Natural Language and
Linguistic Theory 5: 403–39.
STEEDMAN, MARK. 1996. Surface Structure and Interpretation. Cambridge, MA: MIT Press.
STOCKWELL, ROBERT, PAUL SCHACHTER, and BARBARA PARTEE. 1972. The Major Syntactic
Structures of English. New York: Holt, Rinehart, and Winston.
 420 REFERENCES
STORTO, GIANLUCA. 2000. On the structure of indefinite possessives. In B. Jackson and T.
Matthews (eds), SALT X: Proceedings of the Tenth Conference on Semantics and
Linguistic Theory. Ithaca: CLC, 203–20.
STOWELL, TIM. 1981. Origins of phrase structure. PhD dissertation, MIT.
STOWELL, TIM. 1983. Subjects across categories. The Linguistic Review 2: 258–312.
STRAWSON, PETER F. 1950. On referring. Mind 59: 320–44.
DE SWART, HENRIETTE. 1998. Licensing of negative polarity items under inverse scope. Lingua
105: 175–200.
SZABOLCSI, ANNA. 1987. Bound variables in syntax: Are there any? In J. Groenendijk et al.
(eds), Proceedings of the Sixth Amsterdam Colloquium. Amsterdam: ITLI,
University of Amsterdam, 331–53.
SZABOLCSI, ANNA. 1992. Combinatory categorial grammar and projection from the lexicon. In
I. Sag and A. Szabolcsi (eds), Lexical Matters. Stanford: CSLI, 241–68.
WASOW, THOMAS. 1972. Anaphoric relations in English. PhD dissertation, MIT.
WILLIAMS, EDWIN. 1978. Across-the-board rule application. Linguistic Inquiry 9:
31–43.
WINTER, YOAD. 1995. Syncategorematic conjunction and structured meanings. In M. Simons
and T. Galloway (eds), SALT V: Proceedings of the Fifth Conference on Semantics and
Linguistic Theory. Ithaca: CLC, 387–404.
ZWARTS, FRANS. 1998. Three types of polarity. In F. Hamm and E. Hinrichs (eds), Plural
Quantification. Dordrecht: Kluwer, 177–238.
 Index
Abad, May 122 Bolinger, Dwight 120
Adjectives complex 90, 113 Borschev, Vladimir 131
intersective 126–8, 400–2 Bouton, Lawrence 266–8
postnominal 120 predicative 113– Breheny, Richard 380
16, 127–8, 400–2 prenominal 90, Bresnan, Joan 104
120–8, 401
Adjuncts 230 Carlson, Gregory 55, 335, 387
Agreement 96–7, 111 Carpenter, Bob 88
Ajdukiewicz, Kazimerz 88 Cartesian product 22
Alphabetic variants 150 Case marking 84, 96–7, 103–4, 111, 288
Also 392–3 accusative 97 dative 97 genitive 130–1,
Alternative semantics 386–97 278–80, 306 lexical case 97 nominative
And (see conjunction) 97 syntactic case 97
Anderson, Alan 38 Case selection 96
Antecedent-Contained Ellipsis (see also Categorial Grammar 8, 55, 88–106
Ellipsis) 265–72 C-Command 281, 284, 299, 311, 316,
Any 334–6 340, 350
free-choice 335, 338 Cell (see partition)
argslot-l 258–63, 272–3, 328 Argument Cellmates (see partition)
Slot Lift (see argslot-l) Characteristic function (of a set) 35
Chiercjia, Gennaro 31, 173, 361,
Assignments/Assignment functions
135–43, 145–50, 151, 160–1, 214–20, 366, 370
229, 274–8, 278–81, Chomsky, Noam 9–11, 16, 84–5, 108,
286–7, 293–4, 317–18, 366–7 209, 210, 249, 281
Autonomy (of syntax) 107–12 Circumfix 91
Auxiliary (verbs) 37, 50–1, 115, 228, Closed-class items (see also open-class
235–6, 266–9, 271, 317, 326, 384 items) 102, 324–5, 405
Coindexation 17, 281, 283–4, 285, 291,
Bach, Emmon 4, 85–6, 213–14, 248 299, 306, 311, 316
Barker, Chris 130, 256, 284, 287, Combinatory Categorial Grammar 88
316, 377 Combinatory Logic 275, 286
Barss, Andrew 284 Comparison class 30–1, 126–7, 179,
Barwise, Jon 341 318–20, 378–81
Bayer, Samuel 100, 181 Compositionality 42–51
Bennett, Michael 185, 348 Compounding 122
Binding (see also variables) 11, 264, Conditionals (see also counterfactual
conditionals) 338, 345–6, 360
275–311, 364–84, 395–7
 INDEX 423
Conjunction (see also coordination) 69, CPs 84, 89–90, 95, 229–31
82, 100–4, 158–62, 184, Cross-categorial 82, 103, 162, 190, 199,
225, 276 cross-categorial 226, 228, 239
82 generalized 198–203 of Cross-linguistic Semantics 4, 168, 406
generalized quantifier 173–8 of Crossover
nouns 201–3 of Ss 69, 73–5, 80– Strong Crossover 283
3, 162 of transitive verbs 200 of Weak Crossover 143, 281–4,
VPs 72, 80–3, 200 298–301, 305–7, 369–70, 376
Conjunction Reduction 73, 191, 255, Cross-serial dependencies
370
(see also Dutch) 87
Constituency (tests for) 68, 160, 224–6,
Curry, Haskell 76, 275, 286, 300
240–2
Curry’ed functions 76–83, 101, 124,
Context (of discourse) 5–6, 30–1,
154, 167, 198–9, 234–8
117–19, 164, 170–3, 178–9,
189–91, 202, 214, 241, 246–7,
Dative alternation 153
266–71, 277, 285–6, 294, 307–10,
Dayal, Veneeta 335
313–14, 317–19, 348–9, 354–5,
Determiner Phrases (DP) (see also
373–4, 376–7, 379–81, 387–8
determiners) 16, 94, 250
Context-free phrase structure grammar
Determiners 53–4, 94, 178–9 complex
44, 70, 86–7
166–7 proportional 347–9
Contextually salient domain (see also quantificational 53, 118, 157–62,
domain restriction) 164, 354–5,
167–8, 178, 195–6, 211–12,
359, 382
317–20, 339–43, 378
Contrastive stress 310, 386–8
the 54, 118–19, 179, 183
Cooper, Robin 202, 213–14, 341,
Direct Compositionality 42–3
373–8, 380
Directional feature 93–5, 103, 165,
Coordinate Structure Constraint
237–9, 400
233, 254–5
Discontinuous constituent 85
Coordination (see also conjunction)
Discourse context (see context)
100, 181, 224–6, 254, 262–3,
Discourse Representation Theory 377
370–1 non-constituent (see also
Disjunction 46, 74, 82, 100–4
Right-Node
Division (see Geach rule)
Raising) 100, 226–7
Domain restriction 119, 177, 178–80,
Copy theory of movement 209
210, 213–14, 318–20, 354–6,
Coreference accidental
377–83
307 asserted 18
Domain widening 354–8
presupposed 18
Donkey pronouns 371, 376–8
Cormack, Annabel 270
Downing, Pamela 130
Counterfactual conditionals (see also
Downward monotone 351–3, 357–8,
conditionals) 37–40, 53–6
361–3
Counterfactual reasoning 61
Dowty, David 12, 39–40, 84–5, 100,
 226, 287, 370 Dutch Former 400–2
87, 247 Fox, Danny 173, 209, 249
Dynamic Semantics 377 Fragment 12–13
Frazier, Lyn 189
Elbourne, Paul 285, 372, 377–8 Frege’s principle 42–3
Ellipsis 307–12, 384 of VPs 265–72, 285, Function xvii, 19, 24–6 constant function
314, 317, 330, 56–8, 102–4, 136,
372, 384, 395 142, 146, 151, 216–19, 286,
Embedded sentences 87, 226, 231, 247, 287, 405 function composition
272–3, 291, 303–4 221–4, 226–7,
Empty operators 123, 267 232–7, 369–71 partial function
Engdahl, Elisabet 366–70, 373–80 25, 34–6, 58, 62–3,
Entailment 32 downward-entailment 80–1, 103, 109, 143, 183, 217
343–6, Functional questions 364–74 answers
347–51, 361 upward- to 371
entailment 341–7, 351–2 Equivalence Future operator 402
relation 21–4
Evans, Frederic 270, 384 Gazdar, Gerald 8, 44, 84, 89, 96–7, 104,
Evans, Gareth 372–3 Even 221, 231–2
392–3 Geach, Peter T. 233, 272, 376
Events 38–40, 46, 58–9, 66–7, 80 Geach rule 233–40
Ever 334–6 Gender 96, 288–9, 294, 372
Expressives 385 Generalized Phrase Structure Grammar
Faltz, Leonard 185 Fauconnier, Gilles (see also Gazdar, Klein, Pullum,
341, 351 feature bundles 96–8, 114–15 and Sag) 8, 44, 70, 89, 221, 237
Features (syntactic) 59, 93, 94–5, 96–8, Generative Semantics 10, 108, 248–50,
103–4, 122, 135, 165, 194–6, 229, 264
232, 237–9, 261, 267, 288–9, German 87, 91, 97–8, 247, 258
340–1, 359 Geurts, Bart 173
Few 346–9 Giannakidou, Anastasia 334, 360
Feys, Robert 76, 275, 286, 300 Government-Binding Theory 85, 89,
File Change Semantics 377 von 101, 108, 215
Fintel, Kai 178–9, 288, 360, 379 Gradable properties 114, 127 Grice, H.
First-order logic 49–50, 67, Paul 171
133–44, 160–1, 163–4, 199, Grimshaw, Jane 90, 104, 111
234, 238 Groenendijk, Jeroen 364–6, 377 g-
Focus 310, 385–97 broad rule (see also Geach rule) g-slash
393 second occurrence 236–7, 328 generalized 237–9
394 g-sup 289, 329 generalized
303–4
Focus-sensitive operators 387–97
Fodor, Jerry 61
Hamblin, C. L. 364
 INDEX 425
Hankamer, Jorge 314 Hayes, 295, 342–3
Bruce 122 Intersection categories 100
Head 86, 94, 126, 348 Intransitive verbs (see verbs)
Head-Driven Phrase Structure Iota operator 154, 183 Island
Grammar (see also Pollard and constraints 209, 232, 264 i-
Sag) 89, 237 within-i condition 16, 320–2
Head-first generalization 94
Heim, Irene 71, 76, 99, 111, 121, Jabberwocky examples 109–10
143, 216–17, 266, 372, Jackendoff, Ray 10
377–8, 404 Jacobs, Roderick 73
Hendriks, Herman 247, 258, 2 Jacobson, Pauline 16, 85, 101, 126, 209,
64, 272 214, 232, 243, 270–1, 282, 284,
Hepple, Mark 287, 375 287, 302, 312, 314, 326, 371,
Hindi 335 372–5, 392, 396
Hittite 213 Jäger, Gerhard 287
Horn, Laurence 171, 236, 335, Join 199
361–2 Kadmon, Nirit 335, 353–9, 378
Huybregts, Riny 87 Kamp, Hans 378
Identity function 115, 219, 230, 261 on Kanazawa, Makoto 378
individuals 104, 129, 152, 261, Karttunen, Lauri 364, 372, 397
288, 374–5, 405 on Katz, Jerrold 61
properties 115 Kay, Paul 393
Ilokano (see also reduplication) 122 Kayne, Richard 84
Imperatives 28, 335–6 Keenan, Edward 340, 359
Implicatures 46–8, 75, 341–2 Kennedy, Christopher 114, 127, 381
conventional 386 conversational 386, Klein, Ewan 8, 44, 84, 89, 96–7, 98–9,
397 grammatical computation of 173 104, 111, 221, 232, 404 Klima,
Scalar implicatures 48, 168–73, 177, Edward 340, 359
360–3 Kluender, Robert 232
Incorporation (of a noun) 122 Kratzer, Angelika 71, 76, 99, 111, 121,
Indexicals 29, 119, 313–15 143, 178, 216–17, 266, 314, 345,
Indices 17–18,217,248,277,279–81,374 379, 398, 404
Infixation (see also Wrap) 85–6, 91, 106, Krifka, Manfred 337, 353–6
195, 329–30 Kripke, Saul 56–7, 324 Kuno,
Inflectional Phrase (IP) 44 Susumu 232
Information structure 232–3
Intensionality 36, 93, 398–406 Ladusaw, William 334, 341–4
Interpretive Semantics 10, 111–12, 276 Lahiri, Utpal 335
Intersection 7, 21, 24, 81–2, 101–2, Lakoff, George 10, 264
120–1, 124–8, 158, 166–8, 174–5, Lambda abstraction 144–7
182–4, 198–202, 210, 216, 230, Lambda calculus xvii, 133–54
Lambda conversion 147–50 Modifiers (modification) 90, 95, 103,
Lambek, Joachim 91 104, 120–1, 126–8, 209–14,
Landman, Fred 335, 353–9 230–2, 401, 403
Langacker, Ronald 285 Montague Grammar 8–12
Langendoen’s Constraint 16 Montague, Richard 8–12, 36, 88–9, 118,
Larson, Richard 85, 315 163, 182–6, 249, 256, 403
Lasnik, Howard 284–5 Moortgat, Michael 256
Last resort strategy 123, 192–3 morphological rules 115, 121–2
Levinson, Stephen 119, 173, 361 Morphology 55, 59, 91, 112, 129
Lewis, David 11, 28, 70 Morrill, Glyn 100
Lexeme 95, 115, 121–2, 288 Most 343–4
Lexical rules 122–3 Movement 214–20, 239–42, 250–1
Lexical semantics 60–1 covert 250, 263, 283 movement
Lexicon 5, 43, 95, 122–4 rule 85–6, 152,
LF (Logical Form) 27, 50, 74–5, 107–8, 249, 315–16 overt
160–3, 190–2, 210, 234, 245, 248– 250
54, 263–5, 267, 269–70, Multidimensional semantics 397 Munn,
274–6, 278–81, 283–4, 366–8 Alan 82, 101
lift function 183 syntax of 194–6 type-
sensitive 405 Value Lift 272 Names 54–6, 89, 102, 182–5
Lifting 181–97, 328 Negation 11, 49–51, 102–3, 234–6, 265,
Linebarger, Marcia 360 334, 338, 345, 363 wide
Link, Godehard 53 scope 190–1
Logical Form (see LF) Negative polarity items 177, 333–60 in
Lower 183 subject position 350, 353
Lowest types hypothesis 189–93, 203 Neo-Davidsonian event semantics
(see also events) 66
Many 346–9 Nissenbaum, Jon 249
Markerese 28 NMOD rule (see mod) 295–6, 327
Maxim of quantity 171 Noun Phrases (NP) 16, 52–8, 130
May, Robert 249, 272, 369 Definite NPs 183, 305–15
McCawley, James 10, 225, 248 Nouns basic-level 116, 202
McConnell-Ginet, Sally 31 common 54–5, 61, 116–17
Meet 199 compounds 130
Merchant, Jason 268 conjunction (see conjunction of
Metalanguage 45, 134, 248 nouns)
Minimal Attachment hypothesis 189 proper 54, 305 relational 95–6, 128–
Minimalism 44, 89, 108, 215, 249 31, 202, 280,
Minimizers 336–7 Mitchell, 321–2
Jonathan 288 mod 327, 400 Number 96, 122
Model-theoretic semantics 27–8
 INDEX 427
Object language 45 Possible worlds (see also intensionality)
Oehrle, Richard 153, 232 31–40, 345, 398–9
Only 387–8, 390–2 Postal, Paul 18, 264, 283–4, 385
Open-class items (see also closed-class Potts, Christopher 397
items) 102, 324–5 Pragmatics (see also implicatures) 34,
Open proposition 290 48, 164, 169–71, 201–2
Operator (O) 149–50 Or Prepositions 54, 94, 100–4, 128–9, 261
(see disjunction) Presupposition 6, 30–5, 58, 80, 103,
Ordered n-tuple 22, 68, 118–19, 360, 397 projection
76, 79 119
Ordered pair xvii, 21–6, 67–8, 76–9, 136 Principle A 281
Principle B 281–2
Parsons, Terence 66, 80 Principle C 281
Partee, Barbara 10–11, 131, 165, 185–6, Principle of compositionality
189–93, 199, 202–3, 210, 257–9, (see Frege’s principle)
264, 314, 348–9, 356, 372 Principles and Parameters 215
Partition 21–4, 140–2 Processing 127, 189–93, 232, 254, 263
cell in 24, 140–1 Pro-drop languages 294
cellmates in 140–2 Progressive 40, 399
Parts of speech 54–5, 96, 115 Pronouns 17, 274–322, 370 bound 143,
adjective 54 adverb 54 article (see 277–82, 287–93, 395–7 e-type 372–3
Determiners) 54 determiner (see free (see also variables free) 281,
Determiners) 54 noun (see Nouns) 285–6, 293–4, 300, 311, 395–7 of
54 preposition 54 verb 54 laziness 372 relative 208, 214, 220,
Passive 33, 152, 282 227–8, 264, 283
Past-tense 41, 59, 356 Pullum, Geoffrey 8, 44, 84, 89, 96–7,
Paycheck generalization 378 104, 221, 232
Paycheck pronouns 371–8
Pelletier, Francis Jeffry 377 QR (see quantifier raising)
Peters, Stanley 12, 397 Quantificational determiner
Phantom category 231 (see determiners)
Phonologically null elements (see also Quantifier 53, 83, 133, 138–43, 264, 286,
empty operators) 216 357, 398 generalized 156–
Phrase structure rules 43–5, 68, 70, 73, 79, 163 in object position 244–73
84–7, 248 lowering 248–55 raising 75, 248–
Pied-piping 208, 232 55 scope ambiguity 246, 286
Plurals 53–5, 95–6, 122 bare plurals vacuous quantification 139–40
387 plural individuals 53, 96, 122, Questions (see also functional questions)
182 28, 227–8, 364–5
Pollard, Carl 86–7, 89
Pope, Emily 120
 question words 367–9 wh- Set theory (Cont.)
questions 227–8, 359 yes/no proper subset 20
questions 359 proper superset 20
Quine, Willard 287 singleton 19 subset
20 superset 20
Reciprocals 281, 317
Shan, Chung-Chieh 284, 287,
Reconstruction 209
316, 377
Reduplication (see also Ilokano) 122
Sharvit, Yael 209
Reflexives 23–4, 143, 281–2
Shieber, Stuart 86
Reinhart, Tanya 284
Silent lexical item 123–5
Relation 22
Sloppy identity 307–12, 314, 395
Relative clause 5, 207–43 noun
Small clause 114–16, 321–2
modification argument
S operation 300–1
209–12
Specifier 94
Reyle, Uwe 377
Spector, Benjamin 173
Right-Node Raising 100, 225
Stalnaker, Robert 11, 345
Rigid designators 56, 183, 324
Stanley, Jason 178, 288, 379 Stavi,
Root clauses 114–15
Jonathan 168 von Stechow, Arnim 371
Rooth, Mats 165, 185–6, 189–93, 199,
Steedman, Mark 88–91, 221, 231–2, 282
202–3, 257–9, 264, 310, 386–8,
Stockwell, Robert 210
394, 395
Stokhof, Martin 364–6, 377
Rosenbaum, Peter 73
Storto, Gianluca 131
Ross, John 208–10, 232–3, 254, 264, 285,
Stowell, Tim 114
308
Strawson, Peter F. 31, 118–19
Russell, Benjamin 173
Strict identity 308–15, 395
Russell, Bertrand 30–1, 58, 118–19
Suffixation 50, 55, 59, 91, 98, 115, 122
Superlatives 118, 128 de Swart, Henriette
Sag, Ivan 8, 44, 84, 89, 96–104, 111,
350, 353
221, 231–2, 266, 314, 384, 404
Syncategorematic 45, 100
Scalar implicature (see implicatures) Synonymy 33
Schachter, Paul 210 Szabo, Zoltan 178, 288, 379
Schönfinkel, Moses 76 Szabolcsi, Anna 281, 287
Schönfinkelization (see Curry’ed functions)
Schubert, Lenhart 377 Tautology 34, 143, 149
Scope 11, 190–1, 235–6, 246–7, 252–4, Tense Phrase (TP) 44
257–65, 272, 282, 316, 366 The (see Determiners, the)
Selkirk, Elisabeth 393 Thomason, Richmond 11
Semantic strength 351–63 Three-place verbs 68, 84–7, 247, 261,
Set theory 19–26 cardinality 20, 25, 315–17, 329–30
166, 343, 348 element 19–20
Times 28, 41, 48–9, 53, 356–7, 398–401,
member 19 null set 20 power set 20
402
 INDEX 429
Traces 214–20, 251, 279, 373 complex Verb phrases (see also intransitive verbs)
366–8 50, 58–60, 84
Transformational Grammar 11, 69, 73, 153, VP Anaphora (see ellipsis of VPs)
285, 372 VP-internal subject hypothesis 254–5 Wall,
Trees (as representations) 70–1, 85, 99, Robert 12
111, 216–18, 263, 274, 284, Wasow, Thomas 283
350, 394 Williams, Edwin 225
Truth conditional semantics 28–31 Word order rules 93–7, 124, 165, 195, 228,
Type-driven 99, 111, 216 326, 330
Type-driven interpretation 98, 404 Wrap 86–7, 91, 94, 106, 315–17, 329–30
Type-driven principles 99, 219, 254
Type Lifting (see Lifting) X-bar theory 43, 91, 94, 230
Type-Logical Grammar (see also Carpenter
and Categorial Zero derivation 122
Grammar) 8, 88 Z-rule 290–4, 298, 316, 321, 329
Type mismatch (repair) 192–4 Zwarts, Frans 337
Type-sensitive rules 403–6
Type shift (unary) rule 121–3, 186,
192–3, 264

Underspecification 97, 115, 122

underspecified categories 95, 194
Union category 100
Upward monotone 351–3

Vague sentences 30
Vagueness 31, 161, 246, 349
Variable free (semantics) 221, 232, 271,
276, 281, 286–305, 307–17,
319–22, 348, 354, 364–65,
368–76, 382–84, 396–97
Variables 16, 274–322 binding of
138 bound 142–3 free 127, 143,
317–20, 348, 378 hidden 378–84
paycheck 378–84
Verbs intransitive (see also verb
phrases) 58–62, 80, 90, 94–5,
113–5 transitive 23–4, 59, 66–
87, 90,
198–200, 224, 256–7, 270–2