Intensional semantics

K A I VO N F I N T E L & I R E N E H E I M
Kai von Fintel & Irene Heim

Intensional semantics

Massachusetts Institute of Technology • Spring 2021

About these lecture notes

These lecture notes have been evolving for many years now, starting with
some notes from the early 1990s by Angelika Kratzer, Irene Heim, and
Kai von Fintel, which have since been modified and expanded many times
by Irene and/or Kai, with feedback and contributions from colleagues and
students.

We encourage the use of these notes in courses at other institutions. Of

course, you need to give full credit to the authors and you may not use
the notes for any commercial purposes. If you use the notes, we would
like to be notified and we would very much appreciate any comments,
criticism, and advice on these materials.

Cite as follows:
von Fintel, Kai & Irene Heim. 1997–2021. Intensional semantics. MIT.

GitHub repository with the current & up-to-date development version:

https://github.com/fintelkai/fintel-heim-intensional-notes
Acknowledgments
Thanks to generations of students in our classes. Thanks to colleagues
who have used these notes in their classes. Here’s an almost certainly in-
complete list of people who have helped us in some way or other with
these notes (telling us about typos, asking questions, pointing out issues,
etc.): Paul Elbourne, Nick Fleisher, Graeme Forbes, Adam Goodkind,
Aaron Jacobs, Magda Kaufmann, Ezra Keshet, Angelika Kratzer, Manfred
Kupffer, Eric Mack, Yuliya Manyakina, Liz Smeets, Vesela Tihomirova
Simeonova, Ede Zimmermann.

Some advice
1. These notes presuppose familiarity with the material, concepts, and
notation of the Heim & Kratzer textbook.
2. There are numerous exercises throughout the notes. It is highly
recommended to do all of them and it is certainly necessary to do so
if you at all anticipate doing semantics-related work in the future.
3. The notes are designed to go along with explanatory lectures. You
should ask questions and make comments as you work through the
notes.
4. While most of the object language examples are from English,
semantics is a cross-linguistic enterprise. Students (and teachers)
should bring in a cross-linguistic perspective throughout.
5. Students with semantic ambitions should also at an early point start
reading supplementary material (as for example listed at the end of
each chapter of these notes).
6. Prospective semanticists may start thinking about how they would
teach this material.
7. For more advice, see http://kaivonfintel.org/prerequisites/.

Not a spectator sport

“As someone once said, Mathematics is not a spectator sport. To learn Harold Simmons, An introduction to
and understand Mathematics you have to get stuck in and get your hands category theory, 2011, Cambridge Uni-
versity Press
dirty. You have to do the calculations, manipulations, and proofs yourself,
not just read the stuff and pretend you understand it.”

The same goes for formal semantics.

Contents

PA RT I THE FI FTH DIMENSION

1 Beginnings 3

2 Attitudes, conditionals, modals 21

3 Restricting and ordering 47

P A R T II E S C A P I N G O PAC I T Y

4 Specificity and Transparency 71

5 The Third Reading 89

P A R T III TIME

6 Beginnings of tense and aspect 103

7 Embedded tenses 121

P A R T IV QUESTIONS

8 Interrogative clauses 149

BI BLIOGRAPHY
 PA RT I

The Fifth Dimension

Hypotheticals, ‘imaginaries’, conditionals, the syntax of counter-factuality and
contingency may well be the generative centres of human speech. […] Language
is the main instrument of man’s refusal to accept the world as it is.

George Steiner, After Babel (1975), p.226 & p.228

1 Beginnings

1.1 Displacement . . . . . . . . . . . . . . . . . . . . . . 3
1.2 An intensional semantics in 10 easy steps . . . . . . . 6
1.2.1 Laying the foundations . . . . . . . . . . . . 6
1.2.2 Intensional operators . . . . . . . . . . . . . 10
1.3 Comments and complications . . . . . . . . . . . . . 14
1.3.1 Intensions all the way? . . . . . . . . . . . . . 14
1.3.2 Why talk about other worlds? . . . . . . . . 15
1.3.3 The worlds of Sherlock Holmes . . . . . . . 16
1.3.4 What’s next and a general pattern . . . . . . 17
1.4 *Issues with an informal meta-language . . . . . . . . 18
1.5 Further readings . . . . . . . . . . . . . . . . . . . . . 19

1.1 Displacement

Hockett 1960 presented a list of DESIGN FEATURES OF HUMAN LANGUAGE,

which continues to play a role in current discussions of animal commu-
nication and the evolution of language. One of the design features is DIS-
PLACEMENT: human language is not restricted to discourse about the actual
here and now. We use language to speculate about how things might have Hockett cites this passage from Childe
been different, about what would happen if we don’t find dinner soon; we 1936: p.30 to illustrate the utility of dis-
placement:
wonder what our friend thinks the world is like; we may want to tell our
boss what it would take for us not to resign. “[…] parents can, with the aid of lan-
How does natural language untie us from the actual here and now? guage, instruct their offspring how to deal
One degree of freedom is given by the ability to name entities and refer with situations which cannot conveniently
be illustrated by actual concrete exam-
to them even if they are not where we are when we speak: ples. The child need not wait till a bear
(1) Rahel is in Hamburg. attacks the family to learn how to avoid
it. Instruction by example alone in such
This kind of displacement is not something we will explore here. We’ll a case is liable to be fatal to some of the
pupils. Language, however, enables the el-
take it for granted.
ders to forewarn the young of the danger
Consider a sentence with no names of absent entities in it: while it is absent, and then demonstrate
the appropriate course of action.”
(2) It is snowing. [said in Cambridge]

On its own, (2) makes a claim about what is happening right now here in
Cambridge. But there are devices at our disposal that can be added to (2),
4 THE FI FTH DIMENSION

resulting in claims about snow in displaced situations. Displacement can

occur in the TEMPORAL dimension:
(3) At noon yesterday, it was snowing.

This sentence makes a claim not about snow now but about snow at noon
yesterday, a different time from now. We will look at temporal semantics
in Part 2 of this book.
In Part 1 of this book, we will focus on what might be called the MODAL The terms MODAL and MODALITY descend
dimension. Here’s an example of modal displacement: from the Latin modus, “way”, and are an-
cient terms pertaining to the way a propo-
(4) COUNTERFACTUAL CONDITIONAL sition holds, necessarily, contingently, etc.
If the storm system hadn’t been deflected by the jet stream, it would For more on the history, see Auwera &
Aguilar 2015.
have been snowing.
If we wanted to be more fanciful, we
This sentence makes a claim not about snow in the actual world but about could call this the Fifth Dimension (after
snow in the world as it would have been if the storm system hadn’t been the four dimensions of space and time).
deflected by the jet stream, a world distinct from the actual one (where Take a look at the original intro to The
the system did not hit us), a merely POSSIBLE WORLD. Twilight Zone: https://www.youtube.com/
watch?v=vB1Ot9MEOOs
Natural language abounds in modal constructions (see Kratzer 1981).
Here are some other examples:
(5) MODAL AUXILIARIES
It may be snowing.

(6) MODAL ADVERBS

Possibly, it will snow tomorrow.

(7) PROPOSITIONAL ATTITUDES

Miriam believes that it is snowing.

(8) EVIDENTIALS
It appears that it is snowing.

(9) HABITUALS
Ellen smokes.
(10) GENERICS
Bears like honey.

(11) IMPERATIVES
Get your snow shovels ready!

(12) SUFFICIENCY AND EXCESS

Linda is old enough to watch The X-Files.
Klara is too expensive to hire.

(13) INFINITIVAL
{ RELATIVES
}
an
I know expert to talk to.
the
 BEGINNINGS 5

Exercise 1.1 Collect some examples of modal displacement in a language other

than English (whether spoken by you or someone else you know). These can
serve as a touchstone of understanding as we proceed. □

In this chapter, we will put in place the basic framework of INTEN-

SIONAL SEMANTICS, the kind of semantics that models displacement of the
point of evaluation in temporal and modal dimensions. To do this, we will
start with one rather special example of modal displacement:
(14) In the world of Sherlock Holmes, a detective lives at 221B Baker
Street.
(14) doesn’t claim that a detective lives at 221B Baker Street in the actual
world (presumably a false claim), but that in the world as it is described
in the Sherlock Holmes stories of Sir Arthur Conan Doyle, a detective
lives at 221B Baker Street (a true claim, of course). We choose this exam-
ple rather than one of the more well-studied displacement constructions
because we want to focus on conceptual and technical matters for now
before we get distracted by many interesting complications.
The questions we want to answer are: How does natural language
achieve this feat of modal displacement? How do we manage to make
claims about other possible worlds? And why would we want to?
To make displacement possible and compositionally tractable, we need
meanings of natural language expressions, and of sentences in particu-
lar, to be displaceable in the first place. They need to be “portable”, so to
speak, able to make claims about more than just the actual here and now.
And we need other natural language expressions that take that portable
meaning and apply it to some situation other than the actual here and
now. That is what intensionality is all about.
The basic idea of the account we’ll develop is this:

• expressions are assigned their semantic values relative to a possible

world;

• in particular, sentences have truth-values in possible worlds;

• in the absence of modal displacement, we evaluate sentences with

respect to the “actual” world, the world in which we are speaking;

• modal displacement changes the world of evaluation;

• displacement is effected by special operators, whose semantics is our

primary concern here.

A terminological note: we will call the sister of the intensional operator its
PREJACENT, a useful term introduced by our medieval colleagues.
6 THE FI FTH DIMENSION

1.2 An intensional semantics in 10 easy steps

1.2.1 Laying the foundations

Step 1: Possible worlds

Our first step is to introduce possible worlds. This is not the place to dis-
cuss the metaphysics of possible worlds in any depth. Instead, we will
just start working with them and see what they can do for us. Basically,
a possible world is a way that things might have been. In the actual world,
there are two coffee mugs on my desk, but there could have been more
or less. So, there is a possible world — albeit a rather bizarre one — where
there are 17 coffee mugs on my desk. We join Heim & Kratzer in adduc-
ing this quote from Lewis (1986: 1f.):

The world we live in is a very inclusive thing. Every stick and

every stone you have ever seen is part of it. And so are you
and I. And so are the planet Earth, the solar system, the entire
Milky Way, the remote galaxies we see through telescopes,
and (if there are such things) all the bits of empty space be-
tween the stars and galaxies. There is nothing so far away
from us as not to be part of our world. Anything at any dis-
tance at all is to be included. Likewise the world is inclusive
in time. No long-gone ancient Romans, no long-gone ptero-
dactyls, no long-gone primordial clouds of plasma are too far
in the past, nor are the dead dark stars too far in the future, to
be part of the same world. …
The way things are, at its most inclusive, means the way
the entire world is. But things might have been different, in
ever so many ways. This book of mine might have been fin-
ished on schedule. Or, had I not been such a commonsensical
chap, I might be defending not only a plurality of possible
worlds, but also a plurality of impossible worlds, whereof you
speak truly by contradicting yourself. Or I might not have
existed at all — neither myself, nor any counterparts of me.
Or there might never have been any people. Or the physical
constants might have had somewhat different values, incom-
patible with the emergence of life. Or there might have been
altogether different laws of nature; and instead of electrons
and quarks, there might have been alien particles, without
charge or mass or spin but with alien physical properties that
nothing in this world shares. There are ever so many ways
that a world might be: and one of these many ways is the way
that this world is.
 BEGINNINGS 7

Previously, our “metaphysical inventory” included a domain of entities It’s possible that your previous inventory
and a set of two truth-values and increasingly complex functions between also included pluralities, events, and/or de-
grees. We’re just adding to the menagerie
entities, truth-values, and functions thereof. Now, we will add possible
now. Questions arise about what the limits
worlds to the inventory. Let’s assume we are given a set 𝑊 , the set of all are and whether the inventory is universal.
possible worlds, which is a vast space since there are so many ways that For some discussion, see the manuscript
things might have been different from the way they are. Each world has “A typology of semantic entities” (Rett
2019).
as among its parts entities like you and me and these coffee mugs. Some
of them may not exist in other possible worlds. So, strictly speaking each
possible worlds has its own, possibly distinctive, domain of entities. What
we will use in our system, however, will be the grand union of all these
world-specific domains of entities. We will use 𝐷 to stand for the set of all
possible individuals.
Among the many possible worlds that there are — according to Lewis,
there is a veritable plenitude of them — is the world as it is described in
the Sherlock Holmes stories by Sir Arthur Conan Doyle. In that world,
there is a famous detective Sherlock Holmes, who lives at 221B Baker
Street in London and has a trusted sidekick named Dr. Watson. Our sen-
tence In the world of Sherlock Holmes, a detective lives at 221B Baker Street
displaces the claim that a famous detective lives at 221B Baker Street from
the actual world to the world as described in the Sherlock Holmes stories.
In other words, the following holds (until we revise it):
(15) The sentence In the world of Sherlock Holmes, a detective lives at 221B
Baker Street is true in a world w iff the sentence a detective lives at
221B Baker Street is true in the world as it is described in the Sher-
lock Holmes stories.
What this suggests is that we need to make space in our system for
having devices that control in what world a claim is evaluated. This is
what we will do now.

Step 2: The evaluation world parameter

Recall from H&K that we were working with a semantic interpretation

function that was relativized to an assignment function 𝑔, which was
needed to take care of pronouns, traces, variables, etc. From now on, we
will relativize the semantic values in our system to possible worlds as well.
What this means is that from now on, our interpretation function will
have two superscripts: a world 𝑤 and an assignment 𝑔: ⟦·⟧𝑤,𝑔 . For a given
expression 𝜙, we call ⟦𝜙⟧𝑤,𝑔 the EXTENSION of 𝜙 at 𝑤 (relative to 𝑔). The
extension of an expression at a world is the central notion of an inten-
sional semantics.
8 THE FI FTH DIMENSION

So, the prejacent embedded in (14) will have its truth-conditions de- Recall from H&K, pp.22f, that what’s
scribed as follows: inside the interpretation brackets is a men-
tion of an object language expression.
(16) For any world 𝑤 and assignment function 𝑔: They make this clear by bold-facing all
⟦a famous detective lives at 221B Baker Street⟧𝑤,𝑔 = 1 object language expressions inside inter-
pretation brackets. In this book, we will
iff a famous detective lives at 221B Baker Street in world 𝑤.
follow common practice in the field and
It is customary to refer to the world for which we are calculating the not use a special typographic distinction,
but let it be understood that what is inter-
extension of a given expression as the EVALUATION WORLD. In the absence
preted are object language expressions.
of any shifting devices, we would normally evaluate a sentence in the ac-
tual world. But then there are shifting devices such as our in the world of
Sherlock Holmes. We will soon see how they work. But first some more
pedestrian steps: adding lexical entries and composition principles that are
formulated relative to a possible world. This will allow us to derive the
truth-conditions as stated in (16) in a compositional manner.

Step 3: Lexical entries

Among our lexical items, we can distinguish between items which have a
WORLD-DEPENDENT semantic value and those that are world-independent.
Let’s start with the entry for famous:
(17) For any 𝑤 ∈ 𝑊 and any assignment function 𝑔:
⟦famous⟧𝑤,𝑔 = 𝜆𝑥 ∈ 𝐷. 𝑥 is famous in 𝑤.

Of course, “𝜆𝑥 ∈ 𝐷. . . . ” is short for “𝜆𝑥 : 𝑥 ∈ 𝐷. . . . ”. Get used to seman-

ticists condensing their notation whenever convenient! A further step of
condensation is taken below: “𝜆𝑥 : 𝑥 ∈ 𝐷𝑒 . . . . ” becomes “𝜆𝑥𝑒 . . . . ”.
Always make sure that you actually understand what the notation
means. Here, for example, we are saying that the semantic value of the
word famous with respect to a given possible world 𝑤 and a variable as-
signment 𝑔 is that function that is defined for an argument 𝑥 only if 𝑥 is
a member of the domain of individuals and that, if it is defined, yields the
truth-value 1 if and only if 𝑥 is famous in 𝑤. (17) does not mean that the
function maps 𝑥 to “𝑥 is famous in 𝑤”, which would be very weird: map-
ping an individual to a meta-language statement!
A couple more predicates: (18) shows another customary conden-
sation of notation: the universal quantifi-
(18) a. ⟦detective⟧𝑤,𝑔 = 𝜆𝑥 ∈ 𝐷. 𝑥 is a detective in 𝑤. cation over evaluation worlds and assign-
b. ⟦lives-at⟧𝑤,𝑔 = 𝜆𝑥 ∈ 𝐷. 𝜆𝑦 ∈ 𝐷. 𝑦 lives-at 𝑥 in 𝑤. ment functions is left implicit and on the
face of it, the parameters seem unbound.
The set of detectives will obviously differ from world to world, and so But they are.
will the set of famous individuals and the set of pairs where the first ele-
ment lives at the second element.
 BEGINNINGS 9

Other items have semantic values which do not differ from world to
world. The most important such items are certain “logical” expressions,
such as truth-functional connectives and determiners: Again, note the ruthless condensation of
the notation in (c) and (d): variables are
(19) a. ⟦and⟧𝑤,𝑔 = 𝜆𝑢 ∈ 𝐷𝑡 . 𝜆𝑣 ∈ 𝐷𝑡 . 𝑢 = 𝑣 = 1. subscripted with the type of the domain
b. ⟦the⟧𝑤,𝑔 = 𝜆𝑓 ∈ 𝐷 ⟨𝑒,𝑡 ⟩ : ∃!𝑥 [𝑓 (𝑥) = 1]. the 𝑦 such that 𝑓 (𝑦) = 1. that their values are constrained to come
from.
c. ⟦every⟧𝑤,𝑔 = 𝜆𝑓 ⟨𝑒,𝑡 ⟩ . 𝜆ℎ ⟨𝑒,𝑡 ⟩ . ∀𝑥𝑒 : 𝑓 (𝑥) = 1 → ℎ(𝑥) = 1.
d. ⟦a/some⟧𝑤,𝑔 = 𝜆𝑓 ⟨𝑒,𝑡 ⟩ . 𝜆ℎ ⟨𝑒,𝑡 ⟩ . ∃𝑥𝑒 : 𝑓 (𝑥) = 1 & ℎ(𝑥) = 1.

Note that there is no occurrence of 𝑤 on the right-hand side of the en-

tries in (19). That’s the tell-tale sign of the world-independence of the
semantics of these items.
We will also assume that proper names have world-independent se-
mantic values, that is, they refer to the same individual in any possible
world.
(20) a. ⟦Noam Chomsky⟧𝑤,𝑔 = Noam Chomsky.
b. ⟦Sherlock Holmes⟧𝑤,𝑔 = Sherlock Holmes.
c. ⟦221B Baker Street⟧𝑤,𝑔 = 221B Baker Street.

Step 4: Composition principles

The old rules of Functional Application, Predicate Modification, and 𝜆-

Abstraction can be retained almost intact. We just need to modify them
by adding world-superscripts to the interpretation function. For example:
(21) FUNCTIONAL APPLICATION (FA)
If 𝛼 is a branching node and {𝛽, 𝛾 } the set of its daughters, then,
for any world 𝑤 and assignment 𝑔: if ⟦𝛽⟧𝑤,𝑔 is a function whose
domain contains ⟦𝛾⟧𝑤,𝑔 , then ⟦𝛼⟧𝑤,𝑔 = ⟦𝛽⟧𝑤,𝑔 (⟦𝛾⟧𝑤,𝑔 ).

The rule simply passes the world parameter down.

Exercise 1.2 Formulate the new versions of Predicate Modification and 𝜆-

Abstraction. □

Step 5: Truth

We will want to connect our semantic system to the notion of the TRUTH
OF AN UTTERANCE. We first adopt the “Appropriateness Condition” from
Heim & Kratzer (p.243):
(22) APPROPRIATENESS CONDITION
A context 𝑐 is appropriate for an LF 𝜙 only if 𝑐 determines a variable
assignment 𝑔𝑐 whose domain includes every index which has a free
occurrence in 𝜙.
10 THE FI FTH DIMENSION

We then intensionalize Heim & Kratzer’s definition of truth and falsity

of utterances:
(23) TRUTH AND FALSITY CONDITIONS FOR UTTERANCES
An utterance of a sentence 𝜙 in a context 𝑐 in a possible world 𝑤 is
true iff ⟦𝜙⟧𝑤,𝑔𝑐 = 1 and false if ⟦𝜙⟧𝑤,𝑔𝑐 = 0.

Exercise 1.3 Compute under what conditions an utterance in possible world 𝑤 7

(which may or may not be the one we are all living in) of the sentence a famous
detective lives at 221B Baker Street is true. □

1.2.2 Intensional operators

So far we have merely “redecorated” the system inherited from Heim Here’s an illustration of the basic idea
& Kratzer. We have introduced possible worlds into our inventory, our we’re trying to implement:

lexical entries and our old composition principles. But with the tools we
⟦𝑂𝑝 𝜙⟧𝑤
have now, all we can do so far is to keep track of the world in which we
=
evaluate the semantic value of an expression, complex or lexical. We will
controls
get real mileage once we introduce INTENSIONAL OPERATORS which are
⟦𝑂𝑝⟧𝑤 ⟦𝜙⟧
capable of shifting the world parameter. We mentioned that there are a
number of devices for modal displacement. As advertised, for now, we
will just focus on a very particular one: the expression in the world of Sher-
lock Holmes. We will assume, as seems reasonable, that this expression is a
sentence-modifier both syntactically and semantically.

Step 6: A syncategorematic entry

We begin with a heuristic step. We want to derive something like the

following truth-conditions for our sentence:
(24) ⟦in the world of Sherlock Holmes,
a famous detective lives at 221B Baker Street⟧𝑤,𝑔 = 1
iff the world 𝑤 ′ as it is described in the Sherlock Holmes stories is
such that there exists a famous detective in 𝑤 ′ who lives at 221B
Baker Street in 𝑤 ′ .
We would get this if in general we had this rule for in the world of Sher-
lock Holmes:
(25) For any sentence 𝜙, any world 𝑤, and any assignment 𝑔:
⟦in the world of Sherlock Holmes, 𝜙⟧𝑤,𝑔 = 1
iff the world 𝑤 ′ as it is described in the Sherlock Holmes stories is
′
such that ⟦𝜙⟧𝑤 ,𝑔 = 1.

This is a so-called SYNCATEGOREMATIC treatment of the meaning of

this expression. Instead of giving an explicit semantic value to the expres-
sion, we specify what effect it has on the meaning of a complex expression
that contains it. In (25), we do not compute the meaning for in the world
of Sherlock Holmes, 𝜙 from the combination of the meanings of its parts,
 BEGINNINGS 11

since in the world of Sherlock Holmes is not given a separate meaning, but
in effect triggers a special composition principle. This format is very com- The diamond ^ symbol for possibility
mon in modal logic systems, which usually give a syncategorematic se- is due to C.I. Lewis, first introduced in
Lewis & Langford 1932, but he made no
mantics for the two classic modal operators (the necessity operator □ and
use of a symbol for the dual combination
the possibility operator ^). When one only has a few closed class expres- ¬^¬. The dual symbol □ (“Box”) was later
sions to deal with that may shift the world parameter, employing syncat- devised by F.B. Fitch and first appeared
egorematic entries is a reasonable strategy. But we are facing a multitude in print in 1946 in a paper by his doctoral
student Barcan (1946). See footnote 425
of displacement devices. We will therefore need to make our system more of Hughes & Cresswell 1968. Another
modular. notation one finds is 𝐿 for necessity and 𝑀
We want to give in the world of Sherlock Holmes its own meaning and for possibility, the latter from the German
möglich ‘possible’.
combine that meaning with that of its prejacent by a general composi-
tion principle. The Fregean slogan we adopted says that all composition is
function application (modulo the need for 𝜆-abstraction and the possible
need for predicate modification). See Heim & Kratzer, Section 4.3, pp.
What we will want to do is to make (24) be the result of functional ap- 63–72 for a reminder about the status of
predicate modification.
plication. But we can immediately see that it cannot be the result of our
usual rule of functional application, since that would feed to in the world
of Sherlock Holmes the semantic value of a famous detective lives in 221B
Baker Street in 𝑤, which would be a particular truth-value, 1 if a famous
detective lives at 221B Baker Street in 𝑤 and 0 if there doesn’t. And what-
ever the semantics of in the world of Sherlock Holmes is, it is certainly not a
truth-functional operator.
We need to feed something else to in the world of Sherlock Holmes. At
the same time, we want the operator to be able to shift the evaluation
world of its prejacent. Can we do this?

Exercise 1.4 How would you show that in the world of Sherlock Holmes is
not a truth-functional operator? □

Step 7: Intensions

We will define a richer notion of semantic value, the INTENSION of an ex-

pression. This will be a function from possible worlds to the extension of
the expression in that world. The intension of a sentence can be applied
to any world and give the truth-value of the sentence in that world. In-
tensional operators take the intension of their prejacent as their argument,
that is we will feed the intension of the embedded sentence to the shifting
operator. The operator will use that intension and apply it to the world it
wants the evaluation to happen in. Voilà.
Now let’s spell that account out. Our system actually provides us with Just like H&K, we make no claim that the
two kinds of meanings. For any expression 𝛼, we have ⟦𝛼⟧𝑤,𝑔 , the seman- semantic values that are attributed to ex-
pressions in our framework fully capture
tic value of 𝛼 in 𝑤, also known as the EXTENSION of 𝛼 in 𝑤. But we can
what is informally meant by “meaning”.
also calculate 𝜆𝑤 .⟦𝛼⟧𝑤,𝑔 , the function that assigns to any world 𝑤 the ex- But certainly, intensions come closer to
tension of 𝛼 in that world. This is usually called the INTENSION of 𝛼. We “meaning” than extensions.
will sometimes use an abbreviatory notation for the intension of 𝛼:
12 THE FI FTH DIMENSION

(26) ⟦𝛼⟧𝑔¢ := 𝜆𝑤 .⟦𝛼⟧𝑤,𝑔 The notation with the subscripted cent-

sign comes from Montague Grammar. See
It should be immediately obvious that since the definition of intension Dowty, Wall & Peters 1981: p. 147.
abstracts over the evaluation world, intensions are not world-dependent.
The definition here is simplified, in that it
Note that strictly speaking, it now makes no sense anymore to speak glosses over the fact that some expres-
of “the semantic value” of an expression 𝛼. What we have is a semantic sions, in particular those that contain
system that allows us to calculate extensions (for a given possible world PRESUPPOSITION TRIGGERS, may fail to
have an extension in certain worlds. In
𝑤) as well as intensions for all (interpretable) expressions. We will see that
such a case, the intension has no exten-
when 𝛼 occurs in a particular bigger tree, it will always be determinate sion to map such a world to. Therefore,
which of the two “semantic values” of 𝛼 is the one that enters into the the intension will have to be a partial
compositional semantics. So, that one — whichever one it is, the extension function. So, the official, more “pedan-
tic”, definition will have to be as follows:
or the intension of 𝛼 — might then be called “the semantic value of 𝛼 in 𝑔
⟦𝛼⟧¢ := 𝜆𝑤 : 𝛼 ∈ dom(⟦⟧𝑤,𝑔 ).⟦𝛼⟧𝑤,𝑔 .
the tree 𝛽”.
It should be noted that the terminology of EXTENSION vs. INTENSION is The Port-Royal logicians distinguished
time-honored but that the possible worlds interpretation thereof is more EXTENSION from COMPREHENSION. Leib-
niz preferred the term INTENSION rather
recent. The technical notion we are using is certainly less rich a notion
than COMPREHENSION. The notion prob-
of meaning than traditionally assumed. (For example, Frege’s “modes of ably goes back even further. See Spencer
presentation” are not obviously captured by this possible worlds imple- 1971 for some notes on this. The possible
mentation of extension/intension.) worlds interpretation is due to Carnap
1947.

Step 8: Semantic types and semantic domains

If we want to be able to feed the intensions to lexical items like in the

world of Sherlock Holmes, we need to have the appropriate types in our
system.
Recall that 𝑊 is the set of all possible worlds. And recall that 𝐷 is the
set of all POSSIBLE INDIVIDUALS and thus contains all individuals existing in
the actual world plus all individuals existing in any of the merely possible
worlds.
We now expand the set of semantic types, to add intensions. Intensions
are functions from possible worlds to all kinds of extensions. So, basically
we want to add for any kind of extension we have in our system, a corre-
sponding kind of intension, a function from possible worlds to that kind
of extension.
We add a new clause, (27c), to the definition of semantic types: This is as good a place as any to add a ser-
mon about the notation for types. Func-
(27) SEMANTIC TYPES tional types like ⟨𝑒, 𝑡 ⟩ are written with
a. 𝑒 and 𝑡 are semantic types. angled brackets because of the mathemat-
ical connection between functions and
b. If 𝜎 and 𝜏 are semantic types, then ⟨𝜎, 𝜏⟩ is a semantic type. ordered pairs. The basic types like 𝑒 and
c. If 𝜎 is a semantic type, then ⟨𝑠, 𝜎⟩ is a semantic type. 𝑡 , however are not written with angled
brackets, since they are not of the type of
d. Nothing else is a semantic type. functions. Please do not make this mis-
take.
 BEGINNINGS 13

We also add a fourth clause to the previous definition of semantic do-

mains:
(28) SEMANTIC DOMAINS
a. 𝐷𝑒 = 𝐷, the set of all possible individuals
b. 𝐷𝑡 = {0, 1}, the set of truth-values
c. If 𝜎 and 𝜏 are semantic types, then 𝐷 ⟨𝜎,𝜏 ⟩ is the set of all func-
tions from 𝐷𝜎 to 𝐷𝜏 .
d. INTENSIONS: If 𝜎 is a type, then 𝐷 ⟨𝑠,𝜎 ⟩ is the set of all functions
from 𝑊 to 𝐷𝜎 .

Clause (d) is the addition to our previous system of types. The func- Note a curious feature of this set-up: there
tions of the schematic type ⟨𝑠, . . .⟩ are intensions. Here are some examples is no type 𝑠 and no associated domain.
This corresponds to the assumption that
of intensions:
there are no expressions of English that
take as their extension a possible world,
• The intensions of sentences are of type ⟨𝑠, 𝑡⟩, functions from pos- that is, there are no pronouns or names
sible worlds to truth values. These are usually called PROPOSITIONS. referring to possible worlds. We will ac-
Note that if the function is total, then we can see the sentence as tually question this assumption in a later
chapter. For now, we will stay with this
picking out a set of possible worlds, those in which the sentence is
more conventional set-up.
true. More often than not, however, propositions will be PARTIAL
functions from worlds to truth-values, that is functions that fail to
map certain possible worlds into either truth-value. This will be the
case when the sentence contains a presupposition trigger, such as
the. The famous sentence The King of France is bald has an intension
that (at least in the analysis sketched in Heim & Kratzer) is unde-
fined for any world where there fails to be a unique King of France.
• The intensions of one-place predicates are of type ⟨𝑠, ⟨𝑒, 𝑡⟩⟩, func-
tions from worlds to set of individuals. These are usually called
PROPERTIES.
• The intensions of expressions of type 𝑒 are of type ⟨𝑠, 𝑒⟩, functions
from worlds to individuals. These are usually called INDIVIDUAL
CONCEPTS.

Step 9: A lexical entry for a shifter

We are ready to formulate the lexical entry for in the world of Sherlock
Holmes: This is not yet the final semantics, see
Section 1.3 for complications. One com-
(29) ⟦in the world of Sherlock Holmes⟧𝑤,𝑔 = plication we will not even start to discuss
𝜆𝑝 ⟨𝑠,𝑡 ⟩ . the world 𝑤 ′ as it is described in the Sherlock Holmes stories is that obviously it is not a necessity that
is such that 𝑝 (𝑤 ′ ) = 1. there are Sherlock Holmes stories in the
first place and that the use of this opera-
That is, in the world of Sherlock Holmes expects as its argument a func- tor presupposes that they exist; so a more
fully explicit semantics would need to
tion of type ⟨𝑠, 𝑡⟩, a proposition. It yields the truth-value 1 iff the proposi-
build in that presuppositional component.
tion is true in the world as it is described in the Sherlock Holmes stories. Also, note again the condensed notation:
All that’s left to do now is to provide in the world of Sherlock Holmes “𝜆𝑝 ⟨𝑠,𝑡 ⟩ . . . . ” stands for the fully official
with a proposition as its argument. This is the job of a new composition “𝜆𝑝 : 𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . . . . ”.
principle.
14 THE FI FTH DIMENSION

Step 10: Intensional Functional Application

We add the new rule of Intensional Functional Application.

(30) INTENSIONAL FUNCTIONAL APPLICATION (IFA)
If 𝛼 is a branching node and {𝛽, 𝛾 } the set of its daughters, then,
for any world 𝑤 and assignment 𝑔: if ⟦𝛽⟧𝑤,𝑔 is a function whose
domain contains ⟦𝛾⟧𝑔¢ , then ⟦𝛼⟧𝑤,𝑔 = ⟦𝛽⟧𝑤,𝑔 (⟦𝛾⟧𝑔¢ ).

This is the crucial move. It makes space for expressions that want to
take the intension of their sister as their argument and do stuff to it. Now,
everything is in place. Given (29), the semantic argument of in the world
of Sherlock Holmes will not be a truth-value but a proposition. And thus,
in the world of Sherlock Holmes will be able to check the truth-value of its
prejacent in various possible worlds. To see in practice that we have all we
need, please do the following exercise.

Exercise 1.5 Calculate the conditions under which an utterance in a given pos-
sible world 𝑤 7 of the sentence in the world of the Sherlock Holmes stories, a
famous detective lives at 221B Baker Street is true. □

Exercise 1.6 What in our system prevents us from computing the extension of
Watson is slow, for example, by applying the intension of slow to the extension
of Watson? What in our system prevents us from computing the extension of
Watson is slow by applying the intension of slow to the intension of Watson?
□

Exercise 1.7 What is wrong with the following equation: Please think about this exercise before
looking at Section 1.4, which explores this
(31) (𝜆𝑥 . 𝑥 is slow in 𝑤) (Watson) = Watson is slow in 𝑤 ? issue.
[ Hint: there is nothing wrong with the following:
(32) (𝜆𝑥 . 𝑥 is slow in 𝑤) (Watson) = 1 iff Watson is slow in 𝑤. ] □

1.3 Comments and complications

1.3.1 Intensions all the way?

We have seen that to adequately deal with expressions like in the world of
Sherlock Holmes, we need an intensional semantics, one that gives us access
to the extensions of expressions across the multitude of possible worlds.
At the same time, we have kept the semantics for items like and, every, and
a unchanged and extensional. This is not the only way one can set up an
intensional semantics. The following exercise demonstrates this.

Exercise 1.8 Consider the following “intensional” meaning for and:

(33) ⟦and⟧𝑤,𝑔 = 𝜆𝑝 ⟨𝑠,𝑡 ⟩ . 𝜆𝑞 ⟨𝑠,𝑡 ⟩ . 𝑝 (𝑤) = 𝑞(𝑤) = 1.
 BEGINNINGS 15

With this semantics, the conjunction and would operate on the intensions of the
two conjoined sentences. In any possible world 𝑤, the complex sentence will be
true iff the component propositions are both true of that world.
Compute the truth-conditions of the sentence In the world of Sherlock
Holmes, Holmes is quick and Watson is slow both with the extensional
meaning for and given earlier and the intensional meaning given here. Is there
any difference in the results? □

There are then at least two ways one could develop an intensional system.
(i) We could “generalize to the worst case” and make the semantics de-
liver intensions as the semantic value of an expression. Such systems
are common in the literature (see Lewis 1970, Cresswell 1973).
(ii) We could maintain much of the extensional semantics we have de-
veloped so far and extend it conservatively so as to account for non-
extensional contexts.
We have chosen to pursue (ii) over (i), because it allows us to keep the
semantics of extensional expressions simpler. The philosophy we follow is
that we will only move to the intensional sub-machinery when triggered
by an expression that creates a non-extensional context. As the exercise
just showed, this might be more a matter of taste than a deep scientific
decision. We will turn to questions of expressive power later in this book.

1.3.2 Why talk about other worlds?

Why would natural language bother having such elaborate mechanisms
to talk about other possible worlds? While having devices for spatial and
temporal displacement (talking about Hamburg or what happened yes-
terday) seems eminently reasonable, talking about worlds other than the
actual world seems only suitable for poets and the like. So, why?
The solution to this puzzle lies in a fact that our current semantics of
the shifter in the world of Sherlock Holmes does not yet accurately capture:
modal sentences have empirical content, they make CONTINGENT claims,
claims that are true or false depending on the circumstances in the actual
world.
Our example sentence In the world of Sherlock Holmes, a famous detective
lives at 221B Baker Street is true in this world but it could easily have been
false. There is no reason why Sir Arthur Conan Doyle could not have
decided to locate Holmes’ abode on Abbey Road.
To see that our semantics does not yet capture this fact, notice that in
the semantics we gave for in the world of Sherlock Holmes:
(34) ⟦in the world of Sherlock Holmes⟧𝑤,𝑔 =
𝜆𝑝 ⟨𝑠,𝑡 ⟩ . the world 𝑤 ′ as it is described in the Sherlock Holmes stories
is such that 𝑝 (𝑤 ′ ) = 1.

there is no occurrence of 𝑤 on the right hand side. This means that the
truth-conditions for sentences with this shifter would be world-independent.
16 THE FI FTH DIMENSION

In other words, they are predicted to make non-contingent claims that

are either true no-matter-what or false no-matter-what. This needs to be
fixed.
The fix is obvious: what matters to the truth of our sentence is the
content of the Sherlock Holmes stories as they are in the evaluation world.
So, we actually need the following semantics for our shifter:
(35) ⟦in the world of Sherlock Holmes⟧𝑤,𝑔 =
𝜆𝑝 ⟨𝑠,𝑡 ⟩ . the world 𝑤 ′ as it is described in the Sherlock Holmes stories
in 𝑤 is such that 𝑝 (𝑤 ′ ) = 1.

We see now that sentences with this shifter do make a claim about the
evaluation world: namely, that the Sherlock Holmes stories as they are
in the evaluation world describe a world in which such-and-such is true.
So, what is happening is that although it appears at first as if modal state-
ments concern other possible worlds and thus couldn’t really be very in-
formative, they actually only talk about certain possible worlds, those that
stand in some relation to what is going on at the ground level in the ac-
tual world. As a crude analogy, consider:
(36) My grandmother is sick.

At one level this is a claim about my grandmother. But it is also a claim

about me: namely that I have a grandmother who is sick. Thus it is with
modal statements. They talk about possible worlds that stand in a certain
relation to the actual world and thus they make claims about the actual
world, albeit slightly indirectly.

1.3.3 The worlds of Sherlock Holmes

So far, we have played along with colloquial usage in talking of the world
of Sherlock Holmes. But it is important to realize that this is sloppy talk.
Lewis (1978) writes:
[I]t will not do to follow ordinary language to the extent of
supposing that we can somehow single out a single one of
the worlds [as the one described by the stories]. Is the world
of Sherlock Holmes a world where Holmes has an even or
an odd number of hairs on his head at the moment when he
first meets Watson? What is Inspector Lestrade’s blood type?
It is absurd to suppose that these questions about the world
of Sherlock Holmes have answers. The best explanation of
that is that the worlds of Sherlock Holmes are plural, and the
questions have different answers at different ones.
The usual move at this point is to talk about the set of worlds “COMPATIBLE An equivalent way to phrase this is that we
WITH the (content of ) Sherlock Holmes stories in 𝑤”. We imagine that we are looking at the worlds not excluded by
the (content of the) stories.
ask of each possible world whether what is going on in it is compatible
with the stories as they were written in our world. Worlds where Holmes
 BEGINNINGS 17

lives on Abbey Road are not compatible. Some worlds where he lives at
221B Baker Street are compatible (again not all, because in some such
worlds he is not a famous detective but an obscure violinist). Among the
worlds compatible with the stories are ones where he has an even number
of hairs on his head at the moment when he first meets Watson and there
are others where he has an odd number of hairs at that moment.
What the operator in the world of Sherlock Holmes expresses is that its
complement is true throughout the worlds compatible with the stories. In
other words, the operator universally quantifies over the compatible worlds.
Our next iteration of the semantics for the operator is therefore this:
(37) ⟦in the world of Sherlock Holmes⟧𝑤,𝑔 =
𝜆𝑝 ⟨𝑠,𝑡 ⟩ . ∀𝑤 ′ compatible with the Sherlock Holmes stories in 𝑤 :
𝑝 (𝑤 ′ ) = 1.

At a very abstract level, the way we parse sentences of the form in the
world of Sherlock Holmes, 𝜙 is that both components, the in-phrase and the
prejacent, determine sets of possible worlds and that the set of possible
worlds representing the content of the fiction mentioned in the in-phrase
is a subset of the set of possible worlds determined by the prejacent. We
will encounter the same rough structure of relating sets of possible worlds
in other intensional constructions.
This is where we will leave things. There is more to be said about fic-
tion operators like in the world of Sherlock Holmes, but we will just refer
to you to the relevant literature. In particular, one might want to make
sense of Lewis’ idea that a special treatment is needed for cases where
the sentence makes a claim about things that are left open by the fiction
(no truth-value, perhaps?). One also needs to figure out how to deal with
cases where the fiction is internally inconsistent. In any case, for our pur-
poses we’re done with this kind of operator.

1.3.4 What’s next and a general pattern

With the basic framework of intensional semantics in place, we can now

look at a succession of intensional operators. In particular, we will explore
the semantics of propositional attitude predicates such as believe or want,
modal auxiliaries such as must or might, and conditional sentences.
We will see a general pattern at work. In our Sherlock Holmes exam- 𝑀 [ 𝑓 (𝑎) ] (𝜙 )
ple, an intensional operator can be seen as taking an anchor 𝑎 (the Sher- 𝑀: a quantificational/modal relation be-
lock Holmes stories as they are in actual world) and project from it a set of tween two sets of worlds (proposi-
tions)
compatible worlds, this can be encapsulated in a “flavor” function 𝑓 from
𝑎: the anchor of the modal claim
anchors to sets of possible worlds; the operator then makes a claim with a
certain quantificational force 𝑀 about the relation between the projected 𝑓 : the flavor function that projects a set of
worlds from the anchor
set of worlds and the prejacent worlds 𝜙. Indirectly then, a claim is thereby
𝜙: the prejacent set of worlds (proposi-
made about the anchor.
tion)
18 THE FI FTH DIMENSION

We will look at the non-trivial issues that arise when several inten-
sional operators interact (modals under attitudes, modals in the conse-
quent of a conditional, etc.). We will also see that constituents of the pre-
jacent can sometimes be evaluated with respect to a world that is not the
world that the intensional operator is taking us to (so-called de re read-
ings). Further, we will move from worlds to times and explore the seman-
tics of tense and aspect. And, for the intrepid, this can all come together
by exploring how tense and aspect interact with attitudes, modality, and
conditionals.

1.4 *Issues with an informal meta-language

Exercise 1.7 asks what is wrong with writing something like Thanks to Magda Kaufmann, Angelika
Kratzer, and Ede Zimmermann for discus-
(38) (𝜆𝑥 . 𝑥 is slow in 𝑤) (Watson) = Watson is slow in 𝑤. sions of the issues explored in this section,
which is optional on a first pass, as indi-
Think about it. On the left hand side of the “=” sign is a meta-language cated by the star on the section title.
expression consisting of a 𝜆-expression (so some kind of function) applied
to an individual (contributed by the meta-language name “Watson”). The
function is a function from individuals to truth-values that will deliver
the truth-value 1 iff the individual is slow in world 𝑤. So, what we have
on the left hand side is the result of a function from individuals to truth-
values applied to an individual. In other words, on the left hand side we
have a truth-value, namely the truth-value 1 if Watson is slow in 𝑤 and
the truth-value 0 if Watson is not slow in 𝑤.
Now, what do we have on the right hand side of the “=”? We have the
meta-language sentence “Watson is slow in 𝑤”. That is not nor does it
contribute a truth-value. It is a statement of fact. Truth-values are not the
same as statements of fact.
The proper thing to do is to write
(39) (𝜆𝑥 . 𝑥 is slow in 𝑤) (Watson) = 1 iff Watson is slow in 𝑤.

There are actually two ways to parse the statement in (39), both legiti-
mate it appears.
On one parse, the major connective is the meta-language expression
“iff”. On its left hand side is a meta-language statement (that applying
the function to the individual Watson gives the truth-value 1) and on
the right hand side of the “iff” we have another meta-language statement
(that Watson is slow in 𝑤). So, the whole thing says that these two state-
ments are equivalent: (i) that function applied to that individual gives us
the truth-value 1, and (ii) that Watson is slow in 𝑤.
The other parse is perhaps more conspicuously represented as follows: Is this weird? It turns out that natural
{ language, not just our semi-formal meta-
1 if Watson is slow in 𝑤 language, has conditionals that seem very
(40) (𝜆𝑥 . 𝑥 is slow in 𝑤) (Watson) =
0 if Watson is not slow in 𝑤 similar: I fear [the consequences if we fail].
See Lasersohn 1996, Frana 2017, and
Blümel 2019 for some discussion.
 BEGINNINGS 19

Here, the “=” sign is the major connective. The left hand side is a meta-
language expression that resolves to a truth-value and the right hand
side as well contributes a truth-value: 1 if such and such and 0 if such and
such.
H&K, of course, introduced a convention that allowed meta-language
statements to be used in a place where a truth-value was expected (p.37,
(9)):
Read “[𝜆𝛼 : 𝜙 . 𝛾]” as either (i) or (ii), whichever makes sense.
(i) “the function which maps every 𝛼 such that 𝜙 to 𝛾”
(ii) “the function which maps every 𝛼 such that 𝜙 to 1, if 𝛾,
and to 0 otherwise”
Since it never makes sense to map anything to a meta-language state-
ment, no ambiguity will ever arise.
So, one might want to extend this leeway and use it in the case of (38) This is the approach of von Stechow 1991.
as well. We could say that in general, meta-language statements supply
truth-values wherever that makes sense. In that case, (38) is just shorthand
for (39).
Alternatively, one can introduce a new notation that indicates that a This is the approach Ede Zimmermann
meta-language statement is being used to contribute a truth-value: (pc) advocates and has been using in his
{ classes.
1 if 𝛼
(41) ⊢ 𝛼 ⊣ =
0 if otherwise

Lastly, one could abandon the H&K informal meta-language approach

altogether and introduce a rigidly formalized meta-language.
These lecture notes will proceed to follow H&K’s approach and will
not introduce any further innovations. So, (38) is illicit and only (39) is
acceptable.

1.5 Further readings

There is considerable overlap between this chapter and Chapter 12 of
Heim & Kratzer 1998. Here, we approach intensional semantics from a
slightly different angle. It would probably be beneficial if you read H&K’s
Chapter 12 in addition to this chapter and if you did the exercises in there.
Come to think of it, some other ancillary reading is also recommended.
You may want to look at relevant chapters in other textbooks (for exam-
ple: Dowty, Wall & Peters 1981: Chapters 5&6, Gamut 1991: Volume II:
Intensional Logic and Logical Grammar, Chierchia & McConnell-Ginet
2000: Chapter 5: Intensionality, and Zimmermann & Sternefeld 2013).
The Stanford encyclopedia of philosophy is always a good resource. Men-
zel 2016 is the entry on possible worlds. A couple of influential philosoph-
ical works on the metaphysics and uses of possible worlds are Kripke 1980
and Lewis 1986. An interesting paper on the origins of the modern possi-
ble worlds semantics for modal logic is Copeland 2002.
20 THE FI FTH DIMENSION

A must read for students who plan to go on to becoming specialists

in semantics, together with a handbook article putting it in perspective:
Montague 1973 and Partee & Hendriks 1997.
To learn more about discourse about fiction, read Lewis 1978. Recent
reconsiderations: Bonomi & Zucchi 2003, Hanley 2004, Proudfoot 2006.
An interesting paper that explores the meaning of fictional texts: Bauer
& Beck 2014. Inconsistencies in fictions and elsewhere are discussed in:
Varzi 1997 and Lewis 1982.
Some other interesting work on stories and pictures and their content:
Ross 1997, Zucchi 2001, Blumson 2009. More recently, there’s been quite
a bit of work on pictorial semantics, see for example Abusch & Rooth
2017, Greenberg 2018, Maier & Bimpikou 2019.
Hockett & Altmann 1968 is a follow-up to Hockett’s original article on
design features. See Emonds 2011 for a recent re-appraisal. If you’re in-
terested in whether displacement really is an exclusive feature of human
language and cognition, there are some recent survey articles: Cheke &
Clayton 2010, Redshaw 2014. Two very recent articles about displace-
ment in humans and other animals are Leahy & Carey 2020 and Redshaw
& Suddendorf 2020.
Astonishingly, Lewis’ doctrine of the reality of the plurality of pos-
sible worlds is being paralleled (pun absolutely intended) by theoretical
physicists in a number of ways. There is a controversial “many worlds”
interpretation of quantum mechanics, for example. Other terms found are
the “multiverse” and “parallel universes”. See for starters, Kai’s blog en-
try on a popular book on the issue, http://kaivonfintel.org/many-worlds/,
MIT physics professor Max Tegmark’s page on the topic, http://space.mit.
edu/home/tegmark/crazy.html, and a Fresh Air interview with physicist
Brian Greene about his book The Hidden Reality: Parallel Universes and
the Deep Laws of the Cosmos: http://www.npr.org/2011/01/24/132932268/
a-physicist-explains-why-parallel-universes-may-exist.
2 Attitudes, conditionals, modals

2.1 Attitudes . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.1.1 Hintikka’s idea . . . . . . . . . . . . . . . . . 22
2.1.2 Iterated attitudes . . . . . . . . . . . . . . . . 24
2.2 Conditionals . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 Modals . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3.1 Syntactic assumptions . . . . . . . . . . . . . 29
2.3.2 Quantification over possible worlds . . . . . 29
2.3.3 Contingency, flavors, context-dependency . 31
2.3.4 Epistemic vs. Circumstantial . . . . . . . . . 33
2.3.5 Toward an analysis . . . . . . . . . . . . . . . 34
2.4 Explorations and variations . . . . . . . . . . . . . . . 37
2.4.1 Accessibility relations . . . . . . . . . . . . . 37
2.4.2 Conversational backgrounds . . . . . . . . . 42

Introduction

Towards the end of the first chapter, we identified a general schema for 𝑀 [ 𝑓 (𝑎) ] (𝜙 )
modal displacement operators. It begins with a “flavor function” that 𝑀: a quantificational/modal relation be-
“projects” a set of relevant worlds from an “anchor”, and then a quantifi- tween two sets of worlds (proposi-
tions)
cational claim is made about those worlds and their relation to the preja-
𝑎: the anchor of the modal claim
cent. We will now see this pattern at work in three kinds of constructions:
propositional attitudes, conditionals, and modals. These look superficially 𝑓 : the flavor function that projects a set of
worlds from the anchor
quite dissimilar:
𝜙: the prejacent set of worlds (proposi-
(1) a. Charlotte believes that Lucy is smart. tion)
b. If Lucy is smart, she will cancel the meeting.
c. Lucy might cancel the meeting.

The intensional operator in (1a) is the lexical verb believe, in (1b) it is the
subordinating complementizer if, and in (1c) it is the auxiliary verb might.
Despite this surface variety, the core semantic contributions are very sim-
ilar.
Propositional attitude predicates like believe project a set of worlds from
the mental state of their subject and relate those worlds to the worlds de-
scribed by the prejacent proposition. Conditionals select a relevant subset
22 THE FI FTH DIMENSION

of the worlds described by their antecedent and relate them to the worlds
described by their consequent. And the modal might says that some worlds
in a relevant set make the prejacent proposition true.
We will now look at these ideas in more detail and build some nimble-
ness in deploying the technical notions here.

2.1 Attitudes

We move from the contents of works of fiction to the contents of the

minds of people. Instead of worlds compatible with the Sherlock Holmes
stories, we look at the worlds compatible with the beliefs or other mental
states of a person.

2.1.1 Hintikka’s idea

Expressions like believe, know, doubt, expect, want, regret, and so on are usu- According to Hintikka 1969, the term
ally said to describe PROPOSITIONAL ATTITUDES, expressing relations be- PROPOSITIONAL ATTITUDE goes back to
Russell 1940.
tween individuals (the attitude holder) and propositions (intensions of
sentences).
The simple idea is that Amandine believes that Letícia is a spy claims that The possible worlds semantics for propo-
Amandine believes of the proposition that Letícia is a spy that it is true. sitional attitudes was in place long before
the extension to fiction contexts was pro-
Note that for the attitude ascription to be true it does not have to hold
posed. Our discussion here has inverted
that Letícia is actually a spy. But where — in which world(s) — does Letí- the historical sequence for pedagogical
cia have to be a spy for it be true that Amandine believes that Letícia is a purposes.
spy?
We might want to be inspired by the colloquial phrase “the world ac-
cording to …” and say that Amandine believes that Letícia is a spy is true
iff in the world according to Amandine, Letícia is a spy. We immediately
recall from Chapter 1 that we need to fix this idea up by making space
for multiple worlds compatible with Amandine’s beliefs and by tying the
truth-conditions to contingent facts about the evaluation world. That is,
what Amandine believes is different in different possible worlds.
The following lexical entry thus offers itself:
(2) ⟦believe⟧𝑤,𝑔 =
𝜆𝑝 ⟨𝑠,𝑡 ⟩ . 𝜆𝑥 . ∀𝑤 ′ compatible with 𝑥 ′𝑠 beliefs in 𝑤 : 𝑝 (𝑤 ′ ) = 1.

What is going on in this semantics? We conceive of Amandine’s be-

liefs as a state of her mind about whose internal structure we will remain
agnostic, a matter left to other cognitive scientists. What we require of It is important to realize the modesty of
it is that it embody opinions about what the world she is located in looks this semantics: we are not trying to figure
like. In other words, if her beliefs are confronted with a particular possible out what belief systems are and partic-
world 𝑤 ′ , they will determine whether that world may or may not be the ularly not what their internal workings
are. That is the job of psychologists (and
world as they think it is. What we are asking of Amandine’s mental state philosophers of mind, perhaps). For our
is whether any state of affairs, any event, anything in 𝑤 ′ is in contradic- semantics, we treat the belief system as a
tion with anything that Amandine believes. If not, then 𝑤 ′ is compatible black box that determines for each possible
world whether it considers it possible that
it is the world it is located in.
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 23

with Amandine’s beliefs. For all Amandine believes, 𝑤 ′ may well be the
world where she lives. Many worlds will pass this criterion, just consider
as one factor that Amandine is unlikely to have any precise opinions about
the number of leaves on the tree in front of my house. Amandine’s be-
lief system determines a set of worlds compatible with her beliefs: those
worlds that are viable candidates for being the actual world, as far as her
belief system is concerned.
Now, Amandine believes a proposition iff that proposition is true in all
of the worlds compatible with her beliefs. If there is just one world com-
patible with her beliefs where the proposition is not true, that means that
she considers it possible that the proposition is not true. In such a case, we
can’t say that she believes the proposition.
Here is the same story in the words of Hintikka (1969), the source for
this semantics for propositional attitudes:

My basic assumption (slightly simplified) is that an attri-

bution of any propositional attitude to the person in ques-
tion involves a division of all the possible worlds (…) into two
classes: into those possible worlds which are in accordance
with the attitude in question and into those which are in-
compatible with it. The meaning of the division in the case
of such attitudes as knowledge, belief, memory, perception,
hope, wish, striving, desire, etc. is clear enough. For instance,
if what we are speaking of are (say) 𝑎’s memories, then these
possible worlds are all the possible worlds compatible with ev-
erything he remembers. […]
How are these informal observations to be incorporated
into a more explicit semantical theory? According to what I
have said, understanding attributions of the propositional at-
titude in question (…) means being able to make a distinction
between two kinds of possible worlds, according to whether
they are compatible with the relevant attitudes of the person
in question. The semantical counterpart to this is of course a
function which to a given individual person assigns a set of
possible worlds.
However, a minor complication is in order here. Of course,
the person in question may himself have different attitudes in
the different worlds we are considering. Hence this function We recognize our schema: the anchor is
in effect becomes a relation which to a given individual and the pair of an individual and a world, the
flavor function projects from the anchor a
to a given possible world 𝜇 associates a number of possible
set of worlds compatible with the relevant
worlds which we shall call the ALTERNATIVES to 𝜇. The rela- attitude of the anchor individual in the
tion will be called the alternativeness relation. (For different anchor world.
propositional attitudes, we have to consider different alterna- Note that what Hintikka calls “alternative-
tiveness relations.) ness relations” are now more commonly
known as ACCESSIBILITY RELATIONS. We
will encounter such relations again and
again, and we will explore their formal
properties later on.
24 THE FI FTH DIMENSION

Exercise 2.1 Let’s adopt Hintikka’s idea that we can use a function that maps
𝑥 and 𝑤 into the set of worlds 𝑤 ′ compatible with what 𝑥 believes in 𝑤. Call this
function B. That is,
(3) B = 𝜆𝑥 . 𝜆𝑤 . {𝑤 ′ : 𝑤 ′ is compatible with 𝑥’s beliefs in 𝑤 }.

Using this notation, our lexical entry for believe could look as follows:
(4) ⟦believe⟧𝑤,𝑔 = 𝜆𝑝 ⟨𝑠,𝑡 ⟩ . 𝜆𝑥 . B (𝑥)(𝑤) ⊆ 𝑝.

We are here indulging in the usual sloppiness in treating 𝑝 both as a function

from worlds to truth-values and as the set characterized by that function.
Here now are two “alternatives” for the semantics of believe:
(5) Aᴛᴛᴇᴍᴘᴛ 1 (ᴠᴇRY ᴡRᴏNG)
[ ]
⟦believe⟧𝑤,𝑔 = 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . 𝜆𝑥 ∈ 𝐷. 𝑝 = B (𝑥)(𝑤) .

(6) Aᴛᴛᴇᴍᴘᴛ 2 (ᴀᴌSᴏ ᴠᴇRY ᴡRᴏNG)

[ ]
⟦believe⟧𝑤,𝑔 = 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . 𝜆𝑥 ∈ 𝐷. 𝑝 ∩ B (𝑥)(𝑤) ≠ ∅ .

Explain why these do not adequately capture the meaning of believe. □

Exercise 2.2 Follow-up: The semantics in (6) would have made believe into
an existential quantifier of sorts: it would say that some of the worlds compati-
ble with what the subject believes are such-and-such. You have argued (success-
fully, of course) that such an analysis is wrong for believe. But are there attitude If you can’t find any candidates that sur-
predicates with such an “existential” meaning? Discuss some candidates. □ vive scrutiny, can you speculate why there
might be no existential attitude predicates?
[Warning: this is underexplored territory!]
Exercise 2.3 Propose a semantics for the adjective alleged as in Vera is an
alleged kleptomaniac. Do not assume any hidden structure. Try to relate your After this little exercise, you might be
interested in some really tough questions
semantics to the verb allege as in Romelu alleged that Vera is a kleptomaniac.
about intensionality inside noun phrases:
□ Bogal-Allbritten 2013, Bogal-Allbritten
& Weir 2017 and Hirsch 2017: especially
Chapter 4.
2.1.2 Iterated attitudes
Semantics is a lab science in several ways. The most crucial way is that we
learn a lot about our objects of study when we put them together and see
how they react to each other.
We expect attitudes to be able to iterate: an attitude claim is a sentence
with contingent truth-conditions and thus provides a proposition that
in turn could be the complement of another attitude claim. In fact, one
suspects that much of the fabric of human life involves iterated attitudes:
we wonder whether Emma realizes that Caroline believes that Janet has
invited Preston for dinner without telling Abby.

Exercise 2.4 Derive the truth-conditions of Emma believes that Caroline

believes that Preston is from Texas. [Follow-up question: if Emma’s belief is
correct, does that mean that Preston is from Texas?] □
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 25

2.2 Conditionals

In many ways, conditionals are the archetypal construction of displace-

ment: the consequent is evaluated not against the actual here and now but
against the scenario conjured up by the antecedent. Consider a few con-
ditional sentences:
(7) a. If Kim left before 6am, she got there in time.
b. If there’s an earthquake tomorrow, this house will collapse.
c. If there had been a massive snowstorm last night, Kai would
have stayed home.

These represent the three main subtypes of conditionals (there are more):
(7a) is an “indicative” conditional about the past, (7b) is an indicative con-
ditional about the future, and (7c) is a “subjunctive” conditional. For the
moment, the differences will be left aside.
The basic idea of how conditionals work is this: the if -clause whisks
us away to a particular possible world (or maybe a set thereof ) and the
consequent clause is asserted to be true of that world (or those worlds).
But what world(s) are we being taken to? The most obvious requirement
is that the antecedent of the conditional needs to be true of the world(s).
But there’s more.
Given our discussions of how the semantics of fiction operators anchors Lewis used a rather whimsical example to
them in facts about the actual world (the content of the relevant body of start off his seminal 1973 book on coun-
terfactuals: “If kangaroos had not tails,
fiction) and how the semantics of attitude predicates is anchored in the
they would topple over”. For another ex-
mental states of an individual in the actual world, it shouldn’t come as a ample of counterfactual whimsy, consider
surprise that conditionals are similarly anchored. So, look at the examples this scene from the TV show “Big Bang
in (7): what in the actual world are they about? Theory”: https://www.youtube.com/
watch?v=0lpY0Kt4bn8. As the examples
Here’s a first attempt of an answer: (7a) is about the local transporta- in the text make clear, conditionals are
tion system, the weather, the traffic, and so on. (7b) is about the sturdiness actually very down-to-earth in real life.
of this house, facts of geology, laws of physics, and so on. (7c) is about
Kai’s proclivities (such as avoiding traffic snarls), the local climate, and so
on. Since the conditionals are anchored in real world facts, they are no
mere flights of fancy and whether they are true depends on those facts. If
today’s traffic was particularly bad, it may be false that Kim’s leaving be-
fore 6am would have got her there in time. If the architects went to great
lengths to make the house earthquake-safe, (7b) may well be false. And if
there was an attendance-mandatory faculty meeting, Kai may well have
come in in spite of a massive snowstorm.
So, the outlines of the semantics of conditionals are clear: if takes us
to worlds where the antecedent is true but that match the actual world in
certain relevant features. And the consequent then is evaluated in those
worlds. There are many details to work out and we’ll keep returning to
that task. But for now, we put forward a placeholder analysis.
26 THE FI FTH DIMENSION

We will treat if as a higher-order operator that together with the an-

tecedent creates an intensional operator with a semantics very similar to
the final analysis we gave to in the world of Sherlock Holmes. But where
the fiction operator directly encoded what features of the actual world it’s
sensitive to (the Sherlock Holmes fiction), conditionals rely on context for
this job. Here’s a first draft of the proposal:
(8) ⟦if⟧𝑤,𝑔 = 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . 𝜆𝑞 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ .
∀𝑤 ′ : 𝑝 (𝑤 ′ ) = 1 & 𝑤 ′ is relevantly like 𝑤 → 𝑞(𝑤 ′ ) = 1.

The contextual anchoring to features of the evaluation world 𝑤 is here

effected by the placeholder “relevantly like 𝑤”. This is crucial because
otherwise the conditional would talk about any world whatsoever where
the antecedent is true. This would make the truth-conditions not just not
contingent on the actual world but also far too strong to allow most sensi-
ble conditionals to be true ever.
Think about the earthquake conditional (7b): we would derive the ab-
surdly strong truth-conditions that the conditional is true iff all of the
worlds where there is a major earthquake in Cambridge tomorrow are
worlds where my house collapses.
There are some obvious and immediate problems with this analysis.
For one, while it’s easy to imagine circumstances where the conditional
(7b) is judged to be true, there surely are possible worlds where there’s
an earthquake but my house does not collapse: perhaps, the builders in
that world used all the recommended best practices to make the build-
ing earthquake-safe, perhaps it’s a world where I’m simply unreasonably
lucky, or the house is immediately adjacent to much sturdier neighboring
buildings which keep it propped up, or Harry Potter flies by and protects
the house at the last minute (he owes me a favor, after all). This problem
(that the house doesn’t in fact collapse in all possible worlds where there’s
an earthquake but that the conditional can still be judged true in some
worlds) is accompanied with another problem: whether the conditional
is true depends on what the world is like. Was the house built to exacting
standards? Is it propped up by its neighbors? Does Harry Potter owe me a
favor? That is the problem solved by restricting the quantifier over worlds
to world “relevantly like 𝑤”.
Obviously, this is a semantics with a “placeholder”, because what does
“relevantly like” mean precisely? Now, just because the semantics is there-
fore rather vague and context-dependent doesn’t mean it is wrong. As
Lewis 1973: p.1 writes:

Counterfactuals are notoriously vague. That does not mean

that we cannot give a clear account of their truth conditions.
It does mean that such an account must either be stated in
vague terms — which does not mean ill-understood terms —
or be made relative to some parameter that is fixed only within
rough limits on any given occasion of language use.
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 27

The insight articulated by Lewis here is very important. Applying Strawson 1950 famously wrote: “Neither
mathematical or logical methods to analyzing natural language mean- Aristotelian nor Russellian rales give the
exact logic of any expression of ordinary
ing often arouses severe skepticism, precisely because natural language
language; for ordinary language has no
is often vague and context-dependent. But that just means that an ade- exact logic.”
quate analysis needs to not ignore vagueness and context-dependence and
rather be clear about where they enter.
All the more reason to refine our initial draft of the proposal. We put a
placeholder for context-dependence in the meta-language (“worlds rel-
evantly like the evaluation world”) but that is not really sufficient. We
would like to embed the analysis in a general framework for how context
enters the semantics. For the purposes of this book, we will adopt an ap-
proach that generalizes from the analysis of “free” pronouns in the Heim
& Kratzer textbook.
In H&K, chapters 9–11, a technical implementation of context-dependency
is developed for pronouns and their referential (and E-Type) readings.
Referential pronouns are analyzed there as free variables, appealing to
a general principle that free variables in an LF need to be supplied with
values from the utterance context. If we want to describe the context-
dependency of conditionals (and as we’ll soon see, modals) in a technically
analogous fashion, we can think of their LF-representations as incorpo-
rating or subcategorizing for a kind of invisible pronoun, a free variable
that effects the anchoring of the conditional claim to relevant features of
the evaluation world.
Concretely, we posit LF-structures where if doesn’t just take two We are using the notation for variables of
propositions as its arguments but also an object language variable of type types other than 𝑒 introduced by Heim
& Kratzer, p. 213. An index on a variable
⟨𝑠, ⟨𝑠, 𝑡⟩⟩:
now is an ordered pair of a natural num-
(9) ber and a type. The variable assignments
relative to which we calculate semantic
𝜓 consequent values now are functions from ordered
pairs of a natural number and a type to
elements of the domain of objects of that
𝜙 antecedent
type.

if 𝑓 ⟨ 3,⟨𝑠,𝑠𝑡 ⟩⟩

We have written the silent pronoun as “𝑓 ” to remind you of “flavor”, the

evocative term we used to describe the anchoring of intensional operators.
So, we have a variable over flavor functions that will return a set of worlds
when applied to a given world. We will give if the job of telling 𝑓 that
what we need is the set of worlds that 𝑓 assigns to the evaluation world:
(10) ⟦if⟧𝑤,𝑔 = 𝜆𝑓 ∈ 𝐷 ⟨𝑠,⟨𝑠,𝑡 ⟩ ⟩ . 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . 𝜆𝑞 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ .
∀𝑤 ′ : 𝑓 (𝑤) (𝑤 ′ ) = 1 & 𝑝 (𝑤 ′ ) = 1 → 𝑞(𝑤 ′ ) = 1.

Together this means that a conditional says about the evaluation world
𝑤 0 that among the worlds that are 𝑓 -related to 𝑤 0 , the ones where the
antecedent is true are all worlds where the consequent is also true.
28 THE FI FTH DIMENSION

We get different flavors of conditionals from different contextual res-

olutions of 𝑓 . Consider for example the earthquake conditional in (7b).
Context might assign to 𝑓 the function that when given an evaluation
world 𝑤 returns the set of worlds that “agree” with 𝑤 on how sturdy
this house is, what the local geology is like, and what the laws of physics
are. Then, the conditional claims that the actual world is such that all
the worlds that agree with it via 𝑓 and where there is an earthquake are
worlds where this house collapses.
With context-dependency comes the spectre of indeterminacy. A fa-
mous example is a pair attributed by Quine 1960: p. 221 to Nelson Good-
man (imagine these being said while the Korean War was going on):

(11) a. If Caesar were in command, he would use the atom bomb.

b. If Caesar were in command, he would use catapults.

Both versions seem possible: saying (11a) would talk about worlds where
Caesar with all his ruthlessness is in command, while (11b) would talk
about worlds where Caesar’s own arsenal comes with him.
We will return to conditionals in the next chapter when some addi-
tional complications become necessary.

Exercise 2.5 We continue our experiments and embed a conditional under an

attitude. Compute the truth-conditions of the following sentence:
(12) Allie believes that if there’s an earthquake, this house will collapse.

Exercise 2.6 Here’s a potentially simpler analysis of conditionals. What if the

silent pronoun were of type ⟨𝑠, 𝑡⟩, a set of worlds? If would then take the contex-
tually assigned set of worlds and say that all the worlds in that set in which the
antecedent is true are worlds where the consequent is true. Spell out the details of
this idea. How would it fare when confronted with the embedded conditional (12)
from the previous exercise? □
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 29

2.3 Modals

The final empirical addition of this chapter are modal auxiliaries like may,
must, can, have to, etc. Most of what we say here should carry over straight-
forwardly to modal adverbs like maybe, possibly, certainly, etc. We will
make certain syntactic assumptions, which make our work easier but
which leave aside many questions that at some point deserve to be ad-
dressed.

2.3.1 Syntactic assumptions

We will assume, at least for the time being, that a modal like may is a
RAISING predicate (rather than a CONTROL predicate), i.e., its subject is
not its own argument, but has been moved from the subject-position of
its infinitival complement. So, we are dealing with the following kind of
structure:
(13) a. Ann may be smart.
b. [ Ann [ 𝜆1 [ may [ t1 be smart ]]]]

Actually, we will be working here with the even simpler structure be-
low, in which the subject has been reconstructed to its lowest trace posi-
tion. (E.g., these could be generated by deleting all but the lowest copy in
the movement chain.) We will be able to prove that movement of a name We will talk about reconstruction in more
or pronoun never affects truth-conditions, so at any rate the interpreta- detail later.

tion of the structure in (13b) would be the same as that of (14). As a mat-
ter of convenience, then, we will take the reconstructed structures, which We will assume that even though Ann be
allow us to abstract away from the (here irrelevant) mechanics of variable smart is a non-finite sentence, this will
not have any effect on its semantic type,
binding.
which is that of a sentence, which in turn
(14) may [ Ann be smart ] means that its semantic value is a truth-
value. This is hopefully independent of
So, for now at least, we are assuming that modals are expressions that the (interesting) fact that Ann be smart on
its own cannot be used to make a truth-
take a full sentence as their semantic argument. Now then, what do modals
evaluable assertion.
mean?

2.3.2 Quantification over possible worlds

The basic idea of the possible worlds semantics for modal expressions is This idea goes back a long time. It was
that they are quantifiers over possible worlds. Toy lexical entries for must famously held by Leibniz, but there are
precedents in the medieval literature; see
and may, for example, would look like this:
Knuuttila 2003. See Copeland 2002 for
(15) ⟦must⟧𝑤,𝑔 = 𝜆𝑝 ⟨𝑠,𝑡 ⟩ . ∀𝑤 ′ : 𝑝 (𝑤 ′ ) = 1. the modern history of the possible worlds
analysis of modal expressions.
(16) ⟦may⟧𝑤,𝑔 = 𝜆𝑝 ⟨𝑠,𝑡 ⟩ . ∃𝑤 ′ : 𝑝 (𝑤 ′ ) = 1.

A necessity modal like must says that all worlds make its prejacent true, Sometimes, people call necessity modals
while a possibility modal like may says that some worlds make its preja- “universal modals” and possibility modals
“existential modals”, which obviously
cent true. Note that our previous intensional operators were all universal
presupposes this quantificational analysis.
30 THE FI FTH DIMENSION

quantifiers (unless you found some existential attitudes in Exercise 2.2), so

the existential force of may is a new frontier for us.
The analysis in (15)/(16) is too crude (in particular, notice that it would
make modal sentences non-contingent — there is no occurrence of the
evaluation world on the right hand side!). But it does already have some
desirable consequences that we will seek to preserve through all subse-
quent refinements. It correctly predicts a number of intuitive judgments
about the logical relations between must and may and among various
combinations of these items and negations. To start with some elemen-
tary facts, we feel that must 𝜙 entails may 𝜙, but not vice versa:
(17) You must stay.
Therefore, you may stay. VALID

(18) You may stay.

Therefore, you must stay. INVALID

The somewhat stilted it is not the case-

(19) a. You may stay, but it is not the case that you must stay. construction is used in (19a) to make cer-
tain that negation takes scope over must.
b. You may stay, but you don’t have to stay. CONSISTENT When modal auxiliaries and negation are
together in the auxiliary complex of the
We judge must 𝜙 incompatible with its “inner negation” must [not 𝜙 ], same clause, their relative scope seems not
but find may 𝜙 and may [not 𝜙 ] entirely compatible: to be transparently encoded in the surface
order; specifically, the scope order is not
(20) You must stay, and/but also, you must leave. (leave = not stay). reliably negation ≻ modal. (Think about
CONTRADICTORY examples with mustn’t, can’t, shouldn’t, may
not etc. What’s going on here? This is an
(21) You may stay, but also, you may leave. CONSISTENT interesting topic which we must set aside
for now. See the references at the end
We also judge that in each pair below, the (a)-sentence and the (b)- of the chapter for relevant work.) With
sentences say the same thing. modal main verbs (such as have to), this
complication doesn’t arise; they are con-
(22) a. You must stay. sistently inside the scope of clause-mate
b. It is not the case that you may leave. auxiliary negation. Therefore we can use
You aren’t allowed to leave. (19b) to (unambiguously) express the same
scope order as (19a), without having to
(You may not leave.) resort to a biclausal structure.
(You can’t leave.)
The parenthesized variants of the (b)-
(23) a. You may stay. sentences in (54) are pertinent here only
to the extent that we can be certain that
b. It is not the case that you must leave.
negation scopes over the modal. In these
You don’t have to leave. examples, apparently it does, but as we
You don’t need to leave. remarked above, this cannot be taken for
(You needn’t leave.) granted in all structures of this form.

Given that stay and leave are each other’s negations (i.e. ⟦leave⟧𝑤,𝑔 = In logicians’ jargon, must and may behave
𝑤,𝑔 𝑤,𝑔 𝑤,𝑔
⟦not stay⟧ , and ⟦stay⟧ = ⟦not leave⟧ ), the LF-structures of these as DUALS of each other. For definitions of
“dual”, see Barwise & Cooper 1981: p. 197
equivalent pairs of sentences can be seen to instantiate the following schemata:
or Gamut 1991: vol.2, 238.
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 31

(24) a. must 𝜙 ≡ not [may [not 𝜙]]

b. must [not 𝜓 ] ≡ not [may 𝜓 ]

(25) a. may 𝜙 ≡ not [must [not 𝜙]]

b. may [not 𝜓 ] ≡ not [must 𝜓 ]

Our present analysis of must, have-to, … as universal quantifiers and of More linguistic data regarding the “paral-
may, can, … as existential quantifiers straightforwardly predicts all of the lel logic” of modals and quantifiers can be
found in Horn’s dissertation (Horn 1972).
above judgments, as you can easily prove.

2.3.3 Contingency, flavors, context-dependency

We already said that the semantics we started with is too simple-minded.
In particular, we have no dependency on the evaluation world, which
would make modal statements non-contingent. This is not correct.
If one says It may be snowing in Cambridge, that may well be part of use-
ful, practical advice about what to wear on your upcoming trip to Cam-
bridge. It may be true or it may be false. The sentence seems true if said
in the dead of winter when we have already heard about a Nor’Easter that
is sweeping across New England. The sentence seems false if said by a
clueless Australian acquaintance of ours in July.
The contingency of modal claims is not captured by our current se-
mantics. All the may-sentence would claim under that semantics is that
there is some possible world where it is snowing in Cambridge. And
surely, once you have read Lewis’ quote in Chapter 1, where he asserts
the existence of possible worlds with different physical constants than we
enjoy here, you must admit that there have to be such worlds even if it is
July. The problem is that in our semantics, repeated here:
(26) ⟦may⟧𝑤,𝑔 = 𝜆𝑝 ⟨𝑠,𝑡 ⟩ . ∃𝑤 ′ : 𝑝 (𝑤 ′ ) = 1,

there is no occurrence of 𝑤 on the right hand side. This means that the Conversely, the plenitude of possible
truth-conditions for may-sentences are world-independent. In other words, worlds would make must-claims very
likely false if they are not reigned in or
they make non-contingent claims that are either true whatever or false
anchored somehow.
whatever, and because of the plenitude of possible worlds they are more
likely to be true than false. This needs to be fixed. But how?
Well, what makes it may be snowing in Cambridge seem true when we
know about a Nor’Easter over New England? What makes it seem false
when we know that it is summer in New England? The idea is that we
only consider possible worlds COMPATIBLE WITH THE EVIDENCE AVAILABLE
TO US. And since what evidence is available to us differs from world to
world, so will the truth of a may-statement.
(27) ⟦may⟧𝑤,𝑔 = 𝜆𝑝. ∃𝑤 ′ compatible w/ the evidence in 𝑤 : 𝑝 (𝑤 ′ ) = 1. From now on, we will leave off type-
specifications such as that 𝑝 has to be of
(28) ⟦must⟧𝑤,𝑔 = 𝜆𝑝. ∀𝑤 ′ compatible w/ the evidence in 𝑤 : 𝑝 (𝑤 ′ ) = 1. type ⟨𝑠, 𝑡 ⟩, whenever it is obvious what
they should be and when saving space is
aesthetically called for.
32 THE FI FTH DIMENSION

Let us consider a different example:

(29) You have to be quiet.

Imagine this sentence being said based on the house rules of the particu-
lar dormitory you live in. Again, this is a sentence that could be true or
could be false. Why do we feel that this is a contingent assertion? Well,
the house rules can be different from one world to the next, and so we
might be unsure or mistaken about what they are. In one possible world,
they say that all noise must stop at 11pm, in another world they say that
all noise must stop at 10pm. Suppose we know that it is 10:30 now, and
that the dorm we are in has either one or the other of these two rules, but
we have forgotten which. Then, for all we know, you have to be quiet may
be true or it may be false. This suggests a lexical entry along these lines:
(30) ⟦have-to⟧𝑤,𝑔 = 𝜆𝑝. ∀𝑤 ′ compatible with the rules in 𝑤 : 𝑝 (𝑤 ′ ) = 1.

Again, we are tying the modal statement about other worlds down to
certain worlds that stand in a certain relation to the actual world: those
worlds where the rules as they are here are obeyed.
A note of caution: it is very important to realize that the worlds com-
patible with the rules as they are in 𝑤 are those worlds where nothing
happens that violates any of the 𝑤-rules. This is not at all the same as
saying that the worlds compatible with the rules in 𝑤 are those worlds
where the same rules are in force. Usually, the rules do not care what the
rules are, unless the rules contain some kind of meta-statement to the ef-
fect that the rules have to be the way they are, i.e. that the rules cannot
be changed. So, in fact, a world 𝑤 ′ in which nothing happens that vio-
lates the rules as they are in 𝑤 but where the rules are quite different and
in fact what happens violates the rules as they are in 𝑤 ′ is nevertheless
a world compatible with the rules in 𝑤. For example, imagine that the
only relevant rule in 𝑤 is that students go to bed before midnight. Take
a world 𝑤 ′ where a particular student goes to bed at 11:30 pm but where
the rules are different and say that students have to go to bed before 11
pm. Such a world 𝑤 ′ is compatible with the rules in 𝑤 (but of course not
with the rules in 𝑤 ′ ).
Apparently, there are different flavors of modality, varying in what
kind of facts in the evaluation world they are sensitive to. The semantics
we gave for must and may above makes them talk about evidence, while
the semantics we gave for have-to made it talk about rules. But that was
just because the examples were hand-picked. In fact, in the dorm scenario
we could just as well have said You must be quiet. And, vice versa, there is
nothing wrong with using it has to be snowing in Cambridge based on the
evidence we have. In fact, many modal expressions seem to be multiply
ambiguous. The English modal have to is probably the world champion in
this regard:
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 33

(31) a. It has to be raining.

b. Visitors have to leave by six pm.
c. You have to go to bed in ten minutes.
d. I have to sneeze.
e. To get home in time, you have to take a taxi.

Traditional descriptions of modals often distinguish a number of “read- Beyond “epistemic” and “deontic,” there is
ings”: EPISTEMIC, DEONTIC, ABILITY, CIRCUMSTANTIAL, DYNAMIC, …. Here a great deal of terminological exuberance.
Sometimes all non-epistemic readings are
are some initial illustrations.
grouped together under the term ROOT
(32) EPISTEMIC MODALITY MODALITY (nobody knows why).

A: Where is John?
B: I don’t know. He may be at home.

(33) DEONTIC MODALITY

A: Am I allowed to stay over at Janet’s house?

B: No, but you may bring her here for dinner.

(34) CIRCUMSTANTIAL/DYNAMIC MODALITY

A: I will plant the rhododendron here.

B: That’s not a good idea. It can grow very tall.

How are may and can interpreted in each of these examples? What do
the interpretations have in common, and where do they differ?
In all three examples, the modal makes an existentially quantified claim
about possible worlds. This is usually called the MODAL FORCE of the claim.
What differs is what worlds are quantified over, sometimes called the
MODAL FLAVOR. In EPISTEMIC modal sentences, we quantify over worlds
compatible with the available evidence. In DEONTIC modal sentences, we
quantify over worlds compatible with the rules and/or regulations. And in
the CIRCUMSTANTIAL modal sentence, we quantify over the set of worlds
which conform to the laws of nature (in particular, plant biology). What
speaker B in (34) is saying, then, is that there are some worlds conforming
to the laws of nature in which this rhododendron grows very tall.

2.3.4 Epistemic vs. Circumstantial

Do flavors of modality correspond to some sorts of signals in the struc- In the earlier Kratzer 1981, the hy-
ture of sentences? Read the following famous passage from Kratzer 1991 drangeas were Zwetschgenbäume ‘plum
trees’. The German word Zwetschge, by
and think about how the two sentences with their very different modal
the way, is etymologically derived from
meanings differ in structure: the name of the city Damascus (Syria), the
center of the ancient plum trade.
Consider sentences (35) and (36):
(35) Hydrangeas can grow here.
34 THE FI FTH DIMENSION

(36) There might be hydrangeas growing here.

The two sentences differ in meaning in a way which is illus-

trated by the following scenario.

Suppose I acquire a piece of land in a far away country

and discover that soil and climate are very much like at home,
where hydrangeas prosper everywhere. Since hydrangeas are
my favorite plants, I wonder whether they would grow in
this place and inquire about it. The answer is (35). In such a
situation, the proposition expressed by (35) is true. It is true
regardless of whether it is or isn’t likely that there are already
hydrangeas in the country we are considering. All that mat-
ters is climate, soil, the special properties of hydrangeas, and
the like. Suppose now that the country we are in has never
had any contacts whatsoever with Asia or America, and the
vegetation is altogether different from ours. Given this evi-
dence, my utterance of (36) would express a false proposition.
What counts here is the complete evidence available. And this
evidence is not compatible with the existence of hydrangeas.

(35) together with our scenario illustrates the pure CIR-

CUMSTANTIAL reading of the modal can. […]. (36) together
with our scenario illustrates the epistemic reading of modals.
[…] circumstantial and epistemic conversational backgrounds
involve different kinds of facts. In using an epistemic modal,
we are interested in what else may or must be the case in our
world given all the evidence available. Using a circumstan-
tial modal, we are interested in the necessities implied by
or the possibilities opened up by certain sorts of facts. Epis-
temic modality is the modality of curious people like histo-
rians, detectives, and futurologists. Circumstantial modality
is the modality of rational agents like gardeners, architects,
and engineers. A historian asks what might have been the
case, given all the available facts. An engineer asks what can
be done given certain relevant facts.

Consider also the very different prominent meanings of the following

two sentences, taken from Kratzer as well:
(37) a. Cathy can make a pound of cheese out of this can of milk.
b. Cathy might make a pound of cheese out of this can of milk.

2.3.5 Toward an analysis

How can we account for this variety of readings? One way would be to
write a host of lexical entries, basically treating this as a kind of (more or
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 35

less principled) ambiguity/polysemy. Another way, which is preferred by

many people, is to treat this as a case of context-dependency, as argued in
seminal work by Kratzer (1977, 1978, 1981, 1991).
According to Kratzer, what a modal brings with it intrinsically is just It is well-known that natural language
a modal force, that is, whether it is an existential (possibility) modal or a quantification is in general subject to con-
textual restriction. See for example von
universal (necessity) modal. What worlds it quantifies over is determined
Fintel 1994: Ch.2 and Stanley & Szabó
by context. In essence, the context has to supply a restriction to the quan- 2000.
tifier. How can we implement this idea? Well, we just have to transpose
the setup we put in place for conditionals.
We give modals a flavor argument, a silent pronoun of type ⟨𝑠, 𝑠𝑡⟩, just
like we did with if. So we posit LF-structures like this:
(38) [ [must 𝑓 ⟨𝑛,⟨𝑠,𝑠𝑡 ⟩⟩ ] [you quiet] ]

𝑓 ⟨𝑛,⟨𝑠,𝑠𝑡 ⟩ ⟩ , again, is a variable over functions from worlds to (characteristic

functions of ) sets of worlds, which — like all free variables — needs to re-
ceive a value from the utterance context. For example, it may be supplied
with the function which, for any world 𝑤, yields the set {𝑤 ′ : the house
rules that are in force in 𝑤 are obeyed in 𝑤 ′ }. If we apply this function to
a world 𝑤 1 , in which the house rules read “no noise after 10 pm”, it will
yield a set of worlds in which nobody makes noise after 10 pm. If we ap-
ply the same function to a world 𝑤 2 , in which the house rules read “no
noise after 11 pm”, it will yield a set of worlds in which nobody makes
noise after 11 pm.
The new lexical entries for must and may that will fit this new structure Compare these entries for the modals
are these: to the entry for if in (10). What are the
differences?
(39) a. ⟦must⟧𝑤,𝑔 =
𝜆𝑓 ∈ 𝐷 ⟨𝑠,𝑠𝑡 ⟩ . 𝜆𝑞 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . ∀𝑤 ′ ∈ 𝑊 [𝑓 (𝑤) (𝑤 ′ ) = 1 → 𝑞(𝑤 ′ ) = 1] in set talk: 𝑓 (𝑤 ) ⊆ 𝑞

b. ⟦may⟧𝑤,𝑔 =
𝜆𝑓 ∈ 𝐷 ⟨𝑠,𝑠𝑡 ⟩ . 𝜆𝑞 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . ∃𝑤 ′ ∈ 𝑊 [𝑓 (𝑤)(𝑤 ′ ) = 1 & 𝑞(𝑤 ′ ) = 1] in set talk: 𝑓 (𝑤 ) ∩ 𝑞 ≠ ∅

Exercise 2.7 Let 𝑤 be a world, and assume that the context supplies an assign-
ment 𝑔 such that:
(40) 𝑔(𝑓 ⟨ 17,⟨𝑠,𝑠𝑡 ⟩⟩ ) = 𝜆𝑤 . 𝜆𝑤 ′ . the rules in force in 𝑤 are obeyed in 𝑤 ′ .

Compute the truth-conditions for the LF in (41):

(41) [ must 𝑓 ⟨ 17,⟨𝑠,𝑠𝑡 ⟩ ⟩ [you quiet] ]

Does truth-value of (41) correctly depend on the evaluation world 𝑤? □

On this approach, the epistemic, deontic, etc. “readings” of individ-

ual occurrences of modal verbs come about by a combination of two
separate things. The lexical semantics of the modal itself encodes just a
quantificational force, a relation between sets of worlds. This is either the
subset-relation (universal quantification; necessity) or the relation of non-
36 THE FI FTH DIMENSION

disjointness (existential quantification; possibility). The covert variable

next to the modal when applied to the evaluation world yields a set of
worlds, and this functions as the quantifier’s restrictor. The labels “epis-
temic”, “deontic”, “circumstantial” etc. group together certain conceptu-
ally natural classes of possible values for this covert restrictor.
Notice that, strictly speaking, there is not just one deontic reading (for
example), but many. A speaker who utters
(42) You have to be quiet.

might mean: ‘I want you to be quiet,’ (i.e., you are quiet in all those worlds
that conform to my preferences). Or she might mean: ‘unless you are
quiet, you won’t succeed in what you are trying to do,’ (i.e., you are quiet
in all those worlds in which you succeed at your current task). Or she
might mean: ‘the house rules of this dormitory here demand that you
be quiet,’ (i.e., you are quiet in all those worlds in which the house rules
aren’t violated). And so on. So the label “deontic” appears to cover a whole Proponents of polysemy accounts of the
open-ended set of imaginable “readings”, and which one is intended and variety of modal flavors will presumably
have to tackle the apparent limitlessness
understood on a particular utterance occasion may depend on all sorts of
of variation in some principle way. See
things in the interlocutors’ previous conversation and tacit shared assump- Viebahn & Vetter 2016 for a polysemy
tions. (And the same goes for the other traditional labels.) account.
A disappointing feature of our analysis in (39) is that a lot of the work
is being done by the modals: they don’t just take a restriction as their ar-
gument but they have to enforce that this restriction is evaluated in the
evaluation world. This is a departure from the ideal that modals are simply
quantifiers over possible worlds. It would be preferable to merely present
them with a set of worlds to quantify over rather than giving them the
responsibility of obtaining this set by applying an accessibility relation to
the evaluation world. We will later put in place a different framework (for
unrelated reasons) that will make good on this vision.

Exercise 2.8 Describe two worlds 𝑤 1 and 𝑤 2 so that

⟦must 𝑓 ⟨ 17,⟨𝑠,𝑠𝑡 ⟩⟩ you quiet⟧𝑤1,𝑔 = 1
and ⟦must 𝑓 ⟨ 17,⟨𝑠,𝑠𝑡 ⟩⟩ you quiet⟧𝑤2,𝑔 = 0. □

Exercise 2.9 In analogy to the deontic relation 𝑔(𝑓 ⟨ 17,⟨𝑠,𝑠𝑡 ⟩⟩ ) defined in (40),
define an appropriate relation that yields an epistemic reading for a sentence like
You may be quiet. □

Exercise 2.10 Describe the set of worlds that constitutes the understood restric-
tor of have to in each of the examples in (31). □

Exercise 2.11 Describe values for the covert ⟨𝑠, 𝑠𝑡⟩-variable that are intuitively
suitable for the interpretation of the modals in the following sentences:
(43) As far as John’s preferences are concerned, you may stay with us.
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 37

(44) According to the guidelines of the graduate school, every PhD candidate
must take 9 credit hours outside his/her department.

(45) John can run a mile in 5 minutes.

(46) This has to be the White House.

(47) This elevator can carry up to 3000 pounds.

For some of the sentences, different interpretations are conceivable depending on

the circumstances in which they are uttered. You may therefore have to sketch the
utterance context you have in mind before describing the accessibility relation. □

Exercise 2.12 Collect two naturally occurring examples of modalized sen-

tences (e.g., sentences that you overhear in conversation, or read in a newspaper or
novel – not ones that are being used as examples in a linguistics or philosophy
paper!), and give definitions of values for the covert ⟨𝑠, 𝑠𝑡⟩-variable which account
for the way in which you actually understood these sentences when you encoun-
tered them. (If the appropriate interpretation is not salient for the sentence out of
context, include information about the relevant preceding text or non-linguistic
background.) □

Exercise 2.13 Modals can be iterated like other intensional operators:

(48) You might have to walk.

Describe a context in which (48) would be an appropriate utterance. State plau-

sible values for the ⟨𝑠, 𝑠𝑡⟩ flavor functions restricting the two modals. Posit an LF
for (48). Derive the truth-conditions compositionally. □

2.4 Explorations and variations

With the basic analyses of attitudes, conditionals, and modals in place, we

turn to some technical explorations.

2.4.1 Accessibility relations

In all of the cases we have looked at the intensional operator ranges over
worlds that are relevant related to the evaluation world (via some kind
of anchor). In the case of conditional and modals, in fact, we made the
operator take a relation between worlds as its contextually provided first
argument:
(49) ⟦if⟧𝑤,𝑔 = 𝜆𝑓 ⟨𝑠,𝑠𝑡 ⟩ . 𝜆𝑝 ⟨𝑠,𝑡 ⟩ . 𝜆𝑞 ⟨𝑠,𝑡 ⟩ .
∀𝑤 ′ : 𝑓 (𝑤) (𝑤 ′ ) = 1 & 𝑝 (𝑤 ′ ) = 1 → 𝑞(𝑤 ′ ) = 1.

(50) ⟦must⟧𝑤,𝑔 = 𝜆𝑓 ⟨𝑠,𝑠𝑡 ⟩ . 𝜆𝑞 ⟨𝑠,𝑡 ⟩ . ∀𝑤 ′ : 𝑓 (𝑤) (𝑤 ′ ) = 1 → 𝑞(𝑤 ′ ) = 1.

38 THE FI FTH DIMENSION

In the Hintikka-semantics for attitudes as well, we think of the worlds

quantified over as those that stand in a particular relation to the evaluation
world. Which relation that is is lexically specified by the attitude predi-
cate. So, believe ranges over worlds compatible with the subject’s beliefs in
the evaluation world.
It might be useful to rehearse some mathematical basics. A relation is
something that two elements stand in. A relation might not be symmet-
rical (or as we say outside of math, it might not be reciprocated), so the
order of elements is part of the notion. The formal model of a relation R
therefore is a set of ordered pairs drawn from two sets 𝐴 and 𝐵:

R = {⟨𝑥, 𝑦⟩ : 𝑥 ∈ 𝐴 & 𝑦 ∈ 𝐵}

In the function-centric and Schönfinkeled world we inherit from Heim

& Kratzer, this notion of a set of ordered pairs becomes a staggered func-
tion that takes first an element from 𝐴, then an element from 𝐵, and then
yields a truth-value:

𝑓 R = 𝜆𝑥 : 𝑥 ∈ 𝐴. (𝜆𝑦 : 𝑦 ∈ 𝐵. ⟨𝑥, 𝑦⟩ ∈ R)

Since we are used to thinking of a function from some type to truth-

values as characterizing a set of things of that type, we can reconceive
a relation as a function from elements of the first set to a set of elements
from the second set (the ones they stand in the relevant relation to).
Note that in the case of accessibility relations, we are dealing with a One might call this an “endorelation”.
relation between worlds. That is, the elements that stand in the relation
come from the same set.
Recall now (for example, from Section 6.6 of H&K) that the linguis- Why do we call reflexivity, transitiv-
tic study of determiners benefitted quite a bit from an investigation of the ity, and symmetry “formal” properties of
relations? The idea is that certain prop-
formal properties of the relations between sets of individuals that deter-
erties are “formal” or “logical”, while
miners express. We can do the same thing here and ask about the formal others are more substantial. So, the fact
properties of the accessibility relation associated with belief versus the one that the relation “have the same birth-
associated with knowledge, etc. The obvious properties to think about are day as” is symmetric seems a more formal
fact about it than the fact that the re-
reflexivity, transitivity, and symmetry. lation holds between my daughter and
A relation is REFLEXIVE iff for any object in the domain of the relation my brother-in-law. Nevertheless, one of
we know that the relation holds between that object and itself. Which ac- the most common ways of characteriz-
ing formal/logical notions (permutation-
cessibility relations are reflexive? Take a knowledge-induced accessibility
invariance, if you’re curious) does not in
relation (let’s call it K𝑥 ): fact make symmetry etc. a formal/logical
(51) 𝑤K𝑥 𝑤 ′ iff 𝑤 ′ is compatible with what 𝑥 knows in 𝑤. notion. So, while intuitively these do seem
to be formal/logical properties, we do not
We are asking whether for any given possible world 𝑤, we know that know how to substantiate that intuition.
See MacFarlane 2005 for discussion.
K𝑥 holds between 𝑤 and 𝑤 itself. It will hold if 𝑤 is a world that is com-
patible with what we know in 𝑤. And clearly that must be so. Take our We talk here about knowledge entailing
(or even presupposing) truth but we do
body of knowledge in 𝑤. The concept of knowledge crucially contains
not mean to say that knowledge simply
the concept of truth: what we know must be true. So if in 𝑤, we know equals true belief. For initial overview of
that something is the case then it must be the case in 𝑤. So, 𝑤 must be the issues, see the Stanford Encyclopedia
compatible with all we know in 𝑤. K𝑥 is reflexive. entry: Ichikawa & Steup 2018.
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 39

Now, if an attitude 𝑋 corresponds to a reflexive accessibility relation,

then we can conclude from 𝑎 𝑋𝑠 that 𝑝 being true in 𝑤 that 𝑝 is true in 𝑤.
This property of an attitude predicate is often called VERIDICALITY. It is to In modal logic notation: □𝑝 → 𝑝. This
be distinguished from FACTIVITY, which is a property of attitudes which pattern is sometimes called T or M, as is
the corresponding system of modal logic.
presuppose – rather than (merely) entail – the truth of their complement.
If we consider the relation B𝑥 pairing with a world 𝑤 those worlds
′
𝑤 which are compatible with what 𝑥 believes in 𝑤, we no longer have
reflexivity: belief is not a veridical attitude. It is easy to have false beliefs, The difference between believe and know
which means that the actual world is not in fact compatible with one’s in natural discourse is quite delicate, espe-
cially when one considers first person uses
beliefs, which contradicts reflexivity. And many other attitudes as well
(I believe the earth is flat vs. I know the earth
do not involve veridicality/reflexivity: what we hope may not come true, is flat).
what we remember may not be what actually happened, etc.
In modal logic, the correspondence between formal properties of the
accessibility relation and the validity of inference patterns is well-studied.
What we have just seen is that reflexivity of the accessibility relation cor-
responds to the validity of □𝑝 → 𝑝. Other properties correspond to other
characteristic patterns. Let’s see this for transitivity and symmetry.
TRANSITIVITY of the accessibility relation corresponds to the inference In the literature on epistemic modal logic,
□𝑝 → □□𝑝. The pattern seems not obviously wrong for knowledge: if the pattern is known as the KK THESIS
or POSITIVE INTROSPECTION. In general
one knows that 𝑝, doesn’t one thereby know that one knows that 𝑝? But
modal logic, it is the characteristic axiom
before we comment on that, let’s establish the formal correspondence be- 4 of the modal logic system S4, which is a
tween transitivity and that inference pattern. This needs to go in both system that adds 4 to the previous axiom
directions. M/T. Thus, S4 is the logic of accessibil-
ity relations that are both reflexive and
What does it take for the pattern to be valid? Assume that □𝑝 holds for transitive.
an arbitrary world 𝑤, i.e. that 𝑝 is true in all worlds 𝑤 ′ accessible from 𝑤.
Now, the inference is to the fact that 𝑝 again holds in any world 𝑤 ′′ ac- not p
cessible from any of those worlds 𝑤 ′ accessible from 𝑤. But what would
prevent 𝑝 from being false in some 𝑤 ′′ accessible from some 𝑤 ′ accessi- p
w3
ble from 𝑤? That could only be prevented from happening if we knew
that 𝑤 ′′ itself is accessible from 𝑤 as well, because then we would know w2
from the premiss that 𝑝 is true in it (since 𝑝 is true in all worlds accessible
from 𝑤). Ah, but 𝑤 ′′ (some world accessible from a world 𝑤 ′ accessible
from 𝑤) is only guaranteed to be accessible from 𝑤 if the accessibility re-
w1
lation is transitive (if 𝑤 ′ is accessible from 𝑤 and 𝑤 ′′ is accessible from 𝑤 ′ ,
then transitivity ensures that 𝑤 ′′ is accessible from 𝑤). This reasoning has
shown that validity of the pattern requires transitivity. The other half of
proving the correspondence is to show that transitivity entails that the
pattern is valid.
The proof proceeds by reductio. Assume that the accessibility relation
is transitive. Assume that (i) □𝑝 holds for some world 𝑤 but that (ii) □□𝑝
doesn’t hold in 𝑤. We will show that this situation cannot obtain. By (i),
𝑝 is true in all worlds 𝑤 ′ accessible from 𝑤. By (ii), there is some non-𝑝
world 𝑤 ′′ accessible from some world 𝑤 ′ accessible from 𝑤. But by transi-
tivity of the accessibility relation, that non-𝑝 world 𝑤 ′′ must be accessible
40 THE FI FTH DIMENSION

from 𝑤. And since all worlds accessible from 𝑤 are 𝑝 worlds, 𝑤 ′′ must be
a 𝑝 world, in contradiction to (ii). So, as soon as we assume transitivity,
there is no way for the inference not to go through.
Now, do any of the attitudes have the transitivity property? It seems
rather obvious that as soon as you believe something, you thereby be-
lieve that you believe it (and so it seems that belief involves a transitive
accessibility relation). And in fact, as soon as you believe something, you
believe that you know it. But one might shy away from saying that know-
ing something automatically amounts to knowing that you know it. For
example, many are attracted to the idea that to know something requires
that (i) that it is true, (ii) that you believe it, and (iii) that you are justi-
fied in believing it: the justified true belief analysis of knowledge. So, now
couldn’t it be that you know something, and thus (?) that you believe you
know it, and thus that you believe that you are justified in believing it,
but that you are not justified in believing that you are justified in believing
it? After all, one’s source of knowledge, one’s reliable means of acquiring
knowledge, might be a mechanism that one has no insight into. So, while
one can implicitly trust (believe) in its reliability, and while it is in fact re-
liable, one might not have any means to have trustworthy beliefs about it.
[Further worries about the KK Thesis are discussed in Williamson 2000.]
What would the consequences be if the accessibility relation were SYM- In modal logic notation, (52) looks as
METRIC? Symmetry of the accessibility relation R corresponds to the va- follows: 𝑝 → □^𝑝, known simply as B
in modal logic. The system that combines
lidity of the following principle:
T/M with B is often called Brouwer’s
(52) Brouwer’s Axiom [: System (B), after the mathematician L.E.J.
[ ]] Brouwer, not because he proposed it but
∀𝑝∀𝑤 : 𝑤 ∈ 𝑝 → ∀𝑤 ′ 𝑤 R𝑤 ′ → ∃𝑤 ′′ [𝑤 ′ R𝑤 ′′ & 𝑤 ′′ ∈ 𝑝] because it was thought that it had some
connections to his doctrines. Brouwer got
Here’s the reasoning: Assume that 𝑅 is in fact symmetric. Pick a world his own commemorative stamp from the
𝑤 in which 𝑝 is true. Now, could it be that the right hand side of the in- Netherlands:
ference fails to hold in 𝑤? Assume that it does fail. Then, there must be
some world 𝑤 ′ accessible from 𝑤 in which ^𝑝 is false. In other words,
from that world 𝑤 ′ there is no accessible world 𝑤 ′′ in which 𝑝 is true. But
since 𝑅 is assumed to be symmetric, one of the worlds accessible from 𝑤 ′
is 𝑤 and in 𝑤, 𝑝 is true, which contradicts the assumption that the infer-
ence doesn’t go through. So, symmetry ensures the validity of the infer-
ence.
The other way (validity of the inference requires symmetry): the in-
ference says that from any 𝑝-world (“𝑝-world” is a very common way of not p
saying “world of which 𝑝 is true”), we can only access worlds from which,
in turn, there is at least one accessible 𝑝-world. But imagine that 𝑝 is true w2 w3
in 𝑤 but not true in any other world. So, the only way for the conclusion
of the inference to hold automatically is to have a guarantee that 𝑤 (the
only 𝑝-world) is accessible from any world accessible from it. That is, we
need to have symmetry. QED. p
w1
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 41

To see whether a particular kind of attitude is based on a symmetric

accessibility relation, we can ask whether Brouwer’s Axiom is intuitively
valid with respect to this attitude. If it is not valid, this shows that the ac-
cessibility relation can’t be symmetric. In the case of a knowledge-based
accessibility relation (epistemic accessibility), one can argue that symmetry
does not hold (thanks to Bob Stalnaker (pc to Kai von Fintel) for help with
the following reasoning):
The symmetry condition would imply that if something hap-
pens to be true in the actual world, then you know that it is
compatible with your knowledge (Brouwer’s Axiom). This
will be violated by any case in which your beliefs are con-
sistent, but mistaken. Suppose that while 𝑝 is in fact true,
you feel certain that it is false, and so think that you know
that it is false. Since you think you know this, it is compat-
ible with your knowledge that you know it. (Since we are
assuming you are consistent, you can’t both believe that you
know it, and know that you do not). So it is compatible with
your knowledge that you know that not 𝑝. Equivalently: you This and the following step rely on the
don’t know that you don’t know that not 𝑝. Equivalently: you duality of necessity and possibility: 𝑞 is
compatible with your knowledge iff you
don’t know that it’s compatible with your knowledge that 𝑝.
don’t know that not 𝑞.
But by Brouwer’s Axiom, since 𝑝 is true, you would have to
know that it’s compatible with your knowledge that 𝑝. So if
Brouwer’s Axiom held, there would be a contradiction. So
Brouwer’s Axiom doesn’t hold here, which shows that epis-
temic accessibility is not symmetric.
Game theorists and theoretical computer scientists who traffic in logics All one really needs to make NI valid is
of knowledge often assume that the accessibility relation for knowledge to have a EUCLIDEAN accessibility rela-
tion: any two worlds accessible from the
is an equivalence relation (reflexive, symmetric, and transitive). But this is
same world are accessible from each other.
appropriate only if one abstracts away from any error, in effect assuming It is a nice little exercise to prove this, if
that belief and knowledge coincide. One striking consequence of working you have become interested in this sort
with an equivalence relation as the accessibility relation for knowledge is of thing. Note that all reflexive and Eu-
clidean accessibility relations are transitive
that one predicts the principle of NEGATIVE INTROSPECTION to hold: and symmetric as well — another nice little
(53) NEGATIVE INTROSPECTION (NI) thing to prove.

If one doesn’t know that 𝑝, then one knows that one doesn’t know
that 𝑝. (¬□𝑝 → □¬□𝑝).

This surely seems rather dubious: imagine that one strongly believes
that 𝑝 but that nevertheless 𝑝 is false, then one doesn’t know that 𝑝, but
one doesn’t seem to believe that one doesn’t know that 𝑝, in fact one be-
lieves that one does know that 𝑝.
42 THE FI FTH DIMENSION

2.4.2 Conversational backgrounds

There is a well-known technical variant on relativizing the semantics

of intensional operators to accessibility relations. The basic idea, due to
Kratzer, is to make intensional operators work with a set of proposi-
tions rather than a set of worlds. For example, must would say that a cer-
tain set of propositions together entail the prejacent. A conditional if 𝑝,
𝑞 would say that a certain set of propositions when augmented with the
antecedent proposition 𝑝 will entail the consequent proposition 𝑞. And
one could conceive of a belief state as providing a set of propositions (each
of which is a belief of the agent), so that the intensional operator believe
would claim that this set of propositions entails the prejacent.
We need to add the usual layer of contingency, which in this case
means that the relevant ingredient is actually a function from evaluation
worlds to sets of propositions.
At this point, we will assume that the resulting set of propositions is
always consistent, by which we mean that there’s at least one world where
all of the propositions in the set are true:
(54) A set of propositions P is CONSISTENT relative to a set of worlds 𝑊
iff ∃𝑤 ∈ 𝑊 : ∀𝑝 ∈ P : 𝑝 (𝑤) = 1).

From any consistent set of propositions, we can retrieve the set of worlds
characterized by it: those worlds such that each proposition in the set is
true in them. If we think of propositions as sets of worlds, this corresponds
to the grand intersection of the set of propositions:
(55) For any consistent set of propositions P,
∩
𝑐ℎ𝑎𝑟 (P) = {𝑤 : ∀𝑝 ∈ P : 𝑝 (𝑤) = 1} = (P)

The semantics of the modal must, for example, can now be rewritten:
∩
(56) ⟦must⟧𝑤,𝑔 = 𝜆M ⟨𝑠,⟨⟨𝑠𝑡,𝑡 ⟩⟩⟩ . 𝜆𝑝. (M (𝑤)) ⊆ 𝑝

The context supplies a function M from worlds to sets of propositions and

must claims that when applied to the evaluation world, M yields a set of
propositions that jointly entail the prejacent 𝑝, or in other words: all the
worlds where all the propositions in M (𝑤) are true are worlds where the
prejacent 𝑝 is true.
Kratzer called functions of type ⟨𝑠, ⟨⟨𝑠𝑡, 𝑡⟩⟩⟩ CONVERSATIONAL BACK-
GROUNDS and used the term MODAL BASE for the conversational back-
grounds that restrict modal operators. She also sometimes calls the result-
ing sets of propositions PREMISE SETS.
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 43

Note that we can retrieve accessibility relations from conversational

backgrounds:
(57) For any conversational background M of type ⟨𝑠, ⟨𝑠𝑡, 𝑡⟩⟩, we define
the corresponding accessibility relation 𝑓 M of type ⟨𝑠, 𝑠𝑡⟩ as follows:
𝑓 M := 𝜆𝑤 . 𝜆𝑤 ′ . ∀𝑝 [M (𝑤) (𝑝) = 1 → 𝑝 (𝑤 ′ ) = 1].

In words, 𝑤 ′ is 𝑓 M -accessible from 𝑤 iff all propositions 𝑝 that are as-

signed by M to 𝑤 are true in 𝑤 ′ .
What motivates the complication from accessibility relations to con-
versational backgrounds? One consideration that we will turn to in the
next chapter concerns inconsistent premise sets. But there is also another,
perhaps more intuitive, motivation. A conversational background is the
sort of thing that is identified by phrases like what the law provides, what we
know, etc. Take the phrase what the law provides. What the law provides is
different from one possible world to another. And what the law provides
in a particular world is a set of propositions. Likewise, what we know differs
from world to world. And what we know in a particular world is a set of
propositions. The intension of what the law provides is then that function
which assigns to every possible world the set of propositions 𝑝 such that
the law provides in that world that 𝑝. Of course, that doesn’t mean that
𝑝 holds in that world itself: the law can be broken. And the intension of
what we know will be that function which assigns to every possible world
the set of propositions we know in that world. Now, consider:
(58) (In view of what we know,) Brown must have murdered Smith.

The in view of -phrase may explicitly signal the intended conversational

background. Or, if the phrase is omitted, we can just infer from other
clues in the discourse that such an epistemic conversational background is
intended.
What follows are some (increasingly technical exercises) on conversa-
tional backgrounds.

Exercise 2.14 Imagine that we model individual 𝑥’s belief state with a set of “B S” stands for “belief state”.
propositions BS𝑥 . Now, when 𝑥 forms a new opinion, we could model this by
adding a new proposition 𝑝 to BS𝑥 . So, BS𝑥 now contains one further element.
There are now more opinions. What happens to the set of worlds compatible with
𝑥’s beliefs? Does it get bigger or smaller? Is the new set a subset or superset of the
previous set of compatible worlds? □

Exercise 2.15 Kratzer calls a conversational background (modal base) RᴇᴀᴌIS-

ᴛIᴄ iff it assigns to any world a set of propositions that are all true in that world.
The modal base what we know is realistic, the modal bases what we believe
and what we want are not.
Show that a conversational background M is realistic iff the corresponding
accessibility relation 𝑓 M (defined as in (57)) is reflexive. □
44 THE FI FTH DIMENSION

Exercise 2.16 Let us call an accessibility relation ᴛRIᴠIᴀᴌ if it makes every

world accessible from every world. 𝑓 is ᴛRIᴠIᴀᴌ iff ∀𝑤∀𝑤 ′ : 𝑤 ′ ∈ 𝑓 (𝑤). What
would the conversational background M have to be like for the accessibility rela-
tion 𝑓 M to be trivial in this sense? □

Exercise 2.17 [For the intrepid only!] The definition in (57) specifies, in effect,
a function from 𝐷 ⟨𝑠,⟨𝑠𝑡,𝑡 ⟩⟩ to 𝐷 ⟨𝑠,𝑠𝑡 ⟩ . It maps each function M of type ⟨𝑠, ⟨𝑠𝑡, 𝑡⟩⟩
to a unique function 𝑓 M of type ⟨𝑠, 𝑠𝑡⟩. This mapping is not one-to-one, how-
ever. Different elements of 𝐷 ⟨𝑠,⟨𝑠𝑡,𝑡 ⟩⟩ may be mapped to the same value in 𝐷 ⟨𝑠,𝑠𝑡 ⟩ .
Prove this claim. I.e., give an example of two functions M and M ′ in 𝐷 ⟨𝑠,⟨𝑠𝑡,𝑡 ⟩⟩
for which (57) determines 𝑓 M = 𝑓 M ′ .
As you have just proved, if every function of type ⟨𝑠, ⟨𝑠𝑡, 𝑡⟩⟩ qualifies as a
‘conversational background’, then two different conversational backgrounds can
collapse into the same accessibility relation. Conceivably, however, if we imposed
further restrictions on conversational backgrounds (i.e., conditions by which only a
proper subset of the functions in 𝐷 ⟨𝑠,⟨𝑠𝑡,𝑡 ⟩⟩ would qualify as conversational back-
grounds), then the mapping between conversational backgrounds and accessibility
relations might become one-to-one after all. In this light, consider the following
potential restriction:
(59) Every conversational background M must be “closed under entailment”;
i.e., it must meet this condition:
∀𝑤 .∀𝑝 [∩M (𝑤) ⊆ 𝑝 → 𝑝 ∈ M (𝑤)].
(In words: if the propositions in M (𝑤) taken together entail 𝑝, then 𝑝
must itself be in M (𝑤).)

Show that this restriction would ensure that the mapping defined in (57) will be
one-to-one. □

Further readings
We have just put in place some of the basics of the analysis of attitudes,
conditionals, and modals. Much of the work in this area will become ac-
cessible to you after the following chapter. For now, we recommend just a
few further readings.
Swanson 2011 is a recent survey on attitudes.
Further connections between mathematical properties of accessibility
relations and logical properties of various notions of necessity and possi-
bility are studied extensively in modal logic, see Hughes & Cresswell 1996
and Garson 2018, especially section 7 and 8, “Modal Axioms and Condi-
tions on Frames”, “Map of the Relationships between Modal Logics”. An We encourage you to visit the ad-
open access textbook on modal logic is Zach 2019. mirable Open Logic Project (https:
//openlogicproject.org/), that the Zach
A thorough discussion of the possible worlds theory of attitudes, and
2019 textbook is part of.
some of its potential shortcomings, can be found in Bob Stalnaker’s work
(1984, 1999).
Two introductory readings on conditionals are von Fintel 2011, 2012.
 AT T I T U D E S , C O N D I T I O N A L S , M O D A L S 45

The most important background readings on modals are the two pa-
pers Kratzer 1981, 1991. There are updated versions of Kratzer’s classic
papers in her volume “Modals and conditionals” (Kratzer 2012). A major
resource on modality is Paul Portner’s book: Portner 2009. You might
also profit from some survey-ish type papers on modals and modality: von
Fintel 2005, von Fintel & Gillies 2007, Swanson 2008, Hacquard 2009a.
3 Restricting and ordering

3.1 Conditionals and modals . . . . . . . . . . . . . . . . 47

3.1.1 First steps . . . . . . . . . . . . . . . . . . . . 47
3.1.2 A serious problem . . . . . . . . . . . . . . . 49
3.1.3 If -clauses as restrictors . . . . . . . . . . . . 51
3.2 Ordering . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2.1 The best we can get . . . . . . . . . . . . . . 54
3.2.2 Mathematical interlude on orderings . . . . . 55
3.2.3 Ordering worlds . . . . . . . . . . . . . . . . 57
3.2.4 Kratzer’s way . . . . . . . . . . . . . . . . . . 58
3.2.5 More cases of modal ordering . . . . . . . . 59
3.2.6 Ordering in conditionals . . . . . . . . . . . 61
3.3 The interplay of ordering and restricting . . . . . . . 62

3.1 Conditionals and modals

We have a basic theory of conditionals in place as well as a basic theory Thus continues our “chemistry” exper-
of modals. Both are treated as intensional operators that move us from iments, combining expressions we have
an initial theory about and seeing what
the initial evaluation world to another set of worlds: in the case of con-
happens when they are put in contact.
ditionals, those worlds where the antecedent is true that are otherwise
relevantly like the evaluation world; in the case of modals, to whatever
worlds the contextually supplied accessibility relation assigns to the evalu-
ation world. This means that if a sentence contains both a conditional and
a modal, we expect them to work together to express nested intensional
shifting.

3.1.1 First steps

In many cases, the experiments seem to have the expected outcome. Con-
sider the following attempt to embed a modal in the antecedent of a con-
ditional:
(1) If one can get to that beach by bike, Iris did just that.

This seems to have the meaning we predict: the conditional take us to

worlds where a certain modal fact holds — that the beach is reachable by
48 THE FI FTH DIMENSION

bike — and then we say that in those worlds Iris went to the beach by
bike.
Or consider our friend Howard again:
(2) If Howard has to pay a heavy fine, he will be broke.

Note that by itself, being under the obligation to pay a fine doesn’t auto- In our system, this could be captured
matically mean that one does. So, (2) indirectly signals that if Howard has by saying that the context supplies an 𝑓
that assigns to the evaluation world only
to pay a heavy fine, he will comply and thus he will be broke.
worlds where Howard does what he is
One thing that lots of people think is not straightforwardly possible is obligated to, at least in as much as paying
epistemic modals in conditionals antecedent. Papafragou 2006, for exam- fines is concerned.
ple, finds the following examples problematic:
(3) a. ?If Max must be lonely, his wife will be worried.
b. ?If Max may be lonely, his wife will be worried.

One possible explanation is that conditionals signal that the antecedent What exactly we might mean by “iffy”
is “iffy”. An epistemic modal statement can only be “iffy” if the speaker and how exactly conditionals signal iffiness
is an interesting question. Any ideas?
is not certain about what “the evidence” is. That can only be if “the ev-
idence” is not the evidence that the speaker has full access to. Relevant
examples include the “cancer scenarios” of DeRose 1991 and the “Master-
mind” cases of von Fintel & Gillies 2007, 2011:
(4) a. If John might have cancer [the doctors haven’t told us], he will
have to see an expert in Boston.
b. If there have to be two reds, your next move is obvious.

These seem to have acceptable readings where the conditional takes us to

worlds where the embedded epistemic modal claim holds.
Consider next modals in the consequent of conditionals. On the face of
it, examples abound:
(5) a. If jaywalking is illegal here, then that guy has to pay a fine.
b. If Caspar vacuums on Saturday, then Chris has to cook on Sun-
day.
(5b) is from a term paper by Moss. (5c) is
c. If Britanny drinks Coke, she must drink Coke. due to Zvolenszky.
d. If Howard returns his book late, he has to pay a fine.
e. If Cosette has to be home by midnight, she ought to think
about leaving now.

We will see, however, that there are significant challenges for composi-
tional semantics hiding here. We begin with the interaction of epistemic
modals and conditionals.
 RE STRICTI NG AND ORDERI NG 49

3.1.2 A serious problem

Our friends have been driving in the Massachusetts hinterlands, inexplica-

bly without iPhones or GPS, and are relying entirely on an old-fashioned
map. They’ve just passed through a little town with an iconic New Eng-
land church and are looking on the map to try to figure out where they
are. They have concluded that they are either on Route 117 or on Route
62. There are two plausible candidate towns on Route 117 (Maynard and
Stow) and just one on Route 62 (Hudson).
So, given all the evidence available to them, there are three live possi-
bilities: that they are in Hudson (on Rte 62, and they don’t know it), that
they are in Stow (on Rte 117, and they don’t know it), and that they are
in Maynard (on Rte 117, and they don’t know it).
It’s true when they say:
(6) We might be in Maynard.

since there are worlds compatible with their evidence where they are in
Maynard.
It’s true when they say:
(7) If we’re on Route 62, we’re in Hudson.

because of the three towns that they know they might be in, only Hudson
is on Rte 62.
Our semantics for conditionals has the conditional take us to worlds
that are (i) in 𝑓 (𝑤), here in the set of worlds compatible with their evi-
dence and (ii) are antecedent worlds. Among the worlds compatible with
their evidence, all 𝑝-worlds (worlds where they are in Rte 62) are worlds
where they are in Hudson.
So far so good. But here are some problematic cases (with the intuitive
truth-value judgments in the given scenario):
(8) a. If we’re on Route 117, we might be in Stow. True
b. If we’re on Route 117, we might be in Hudson. False
c. If we’re on Route 62, we must be in Hudson. True

We will now see that these cannot be explained in our framework!

Take (8c) first. The conditional takes us to those worlds that are (i)
compatible with their evidence or with what they know (which includes
their knowledge that they don’t know in which of the three towns they
are) and (ii) where they are on Rte 62. In all of those worlds, they are in
Hudson, but in none of them do they know or have any additional evidence
that they are in Hudson. So, we expect (8c) to be false, contrary to fact.
Now, take (8b). We go to worlds compatible with their evidence where
they are on Route 117. All of them are worlds where they are on Route
117 without knowing that they are there, since they know that they don’t
know where they are. In those worlds, is it true that for all they know,
50 THE FI FTH DIMENSION

they are in Hudson? Yes. Therefore, we predict that (8b) is true, again
contrary to fact.
Finally, while we do predict that (8a) is true, the reason is simply that
because of their ignorance, any world compatible with their evidence is a
world where any of the relevant might-claims is true. So, the conditional
antecedent is predicted to make no difference to the truth of the epistemic
possibility claim in its consequent, contrary to fact (our intuitions are dif-
ferent for (8a) vs. (8b)).
Before we turn to one of the dominant ways of accounting for the
meanings of the examples in (8), let’s consider a tempting idea about (8c):
what if the epistemic necessity modal must actually scopes over the condi-
tional, even though it appears in its consequent? At first glance, the result-
ing meaning is not far off the target. The claim would be that in all the
epistemically accessible worlds the conditional (7) if we’re on Rte 62, we’re
in Hudson is true. We already saw that that conditional is true in the actual
world, and there’s no reason to say that it isn’t true in other epistemically
accessible worlds. So far, so good.
Unfortunately, there are several reasons for doubting that this LF with
wide-scope for the modal is a convincing solution to our troubles:

1. For the might-version, we can’t resort to wide-scoping the modal,

because the epistemic conditional if we’re on Rte 117, we’re in Stow
is false in the actual world and in fact in all epistemically accessible
worlds: they have evidence that entails that it doesn’t follow from
being on Rte 117 that they’re in Stow. So, (8a) would be predicted
to be false under the wide-scope LF. So, no matter how we scope
the modal, we either have trivial truth or falsity, and not what we
want: non-trivial truth-conditions.
2. Epistemic must carries with it an “evidential signal” (von Fintel &
Gillies 2010):
(9) It must be raining.

This signals that the speaker inferred the truth of it is raining indi-
rectly on the basis of other evidence. The signal is still perceptible
when the prejacent is a generalization or conditional:
(10) a. Elephants must dislike bees.
b. George must have to be home by midnight.
c. It must be that if they’re on Rte 62, they are in Clinton.

But (8c) is not felt to have such a signal, other than that the an-
tecedent would be an additional piece of information in the deduc-
tion that they’re in Hudson.
3. We also need to be able to analyze epistemic conditionals with two
modals in a conjunctive consequent (examples like this are discussed
in Gillies 2010):
 RE STRICTI NG AND ORDERI NG 51

(11) If we’re on Rte 62, we must be in Hudson and might be very

close to the Horseshoe pub.

A wide-scope analysis of the modals in (11) seems impossible.

Conclusion: our analysis really is in trouble and can’t be rescued by wide-

scoping.
In the next section, we present the dominant treatment of the interac-
tion of conditionals and modals in current linguistic semantics. There are
alternatives that deserve to be considered, but that would lead us beyond
the bounds of this textbook.

3.1.3 If-clauses as restrictors

The problem we have encountered here with the interaction of an if -
clause and the modal operator might is similar to others that have been
noted in the literature. Most influentially, Lewis in his paper “Adverbs of
Quantification” (Lewis 1975) showed how hard it is to find an adequate
analysis of the interaction of if -clauses and ADVERBS OF QUANTIFICATION
like never, rarely, sometimes, often, usually, always in sentences like these:

 Always 


 


 


 Usually 


 


 Often 

(12) if a driver sees a friend walking, she offers her a ride.
 Sometimes 
 

 


 Rarely 


 


 Never 

 
Lewis proposed that in these cases, the adverb is the only operator at work
and that the if -clause serves to restrict the adverb. Thus, it has much the
same function that a common noun phrase has in a determiner-quantification.

The if of our restrictive if -clauses should not be regarded

as a sentential connective. It has no meaning apart from the
adverb it restricts. The if in always if …, …, sometimes if …,
…, and the rest is on a par with the non-connective and in
between …and …, with the non-connective or in whether …or
…, or with the non-connective if in the probability that …if
…. It serves merely to mark an argument-place in a polyadic
construction. (Lewis 1975: p. 11)

Building on Lewis’ insight, Kratzer argued for a uniform treatment of

if -clauses as restrictors. She claimed that

the history of the conditional is the story of a syntactic mis-

take. There is no two-place if …then connective in the logical
forms of natural languages. If -clauses are devices for restrict-
ing the domains of various operators. (Kratzer 1986)
52 THE FI FTH DIMENSION

Let us repeat this:

(13) KRATZER’S THESIS
If -clauses are devices for restricting the domains of various opera-
tors.
Kratzer’s Thesis gives a unified picture of the semantics of conditional
clauses. Note that it is not meant to supplant previous accounts of the
meaning of conditionals. It just says that what those accounts are analyz-
ing is not the meaning of if itself but the meaning of the operators that
if -clauses restrict.
Let us see how this idea helps us with our most problematic case, (8c) If
we’re on Route 62, we must be in Hudson. The idea is to deny that there are
two quantifiers over worlds in this sentence. Instead, the if -clause merely
contributes a further restriction to the modal must. In effect, the modal is Note that this is not the same as worlds
not quantifying over all the worlds compatible with our friends’ knowl- where they are on Route 62 and know that
they are. This is made vivid by examples
edge but only over those where they are on Route 62. It then claims that
ultimately due to Thomason (as noted in
all of those worlds are worlds where they are in Hudson. Since that is in van Fraasen 1980): If my partner is embez-
fact true, we now correctly predict that (8c) is true. zling, I will never know (because I’m so bad at
What we don’t yet have is a compositional calculation. What does it accounting).

mean in structural terms for the if -clause to be restricting the domain of

the modal? We present here a particularly “flat-footed” implementation.
The idea is that if -clauses serve as modifiers of the modal flavor argument
of modals.
Recall the entry we had for must:
(14) ⟦must⟧𝑤,𝑔 = 𝜆𝑓𝑠,𝑠𝑡 .𝜆𝑝𝑠𝑡 . ∀𝑤 ′ : 𝑤 ′ ∈ 𝑓 (𝑤) −→ 𝑝 (𝑤 ′ ) = 1.

We supplied the modal with a covert “pronoun” of type ⟨𝑠, 𝑠𝑡⟩ as its first
argument:
(15)

must 𝑓3,⟨𝑠,𝑠𝑡 ⟩ we be in Hudson

We assume a structure where the if -clause is the sister of the 𝑓 -pronoun:

(16)

must we be in Hudson

𝑓3,⟨𝑠,𝑠𝑡 ⟩

if

we are on Rte 62
 RE STRICTI NG AND ORDERI NG 53

The idea now is that the two restrictive devices work together: we This implementation of the restrictor the-
just feed to the modal the intersection of (i) the set of worlds that are 𝑓3 - ory was considered (and, without much of
an argument, dismissed) in Section 3.2 of
accessible from the actual world, and (ii) the set of worlds where they are
von Fintel 1994 and is also found in lec-
on Route 62. ture notes from 2004 by von Stechow. In
Obviously, we don’t hear if -clauses where they are located in this tree. Section 3.3 of von Fintel 1994, a different
So, on the way to PF, the if -clause must be moved to the periphery. Note approach is developed, which is adopted
(at least for illustration) by Kratzer 2015.
that if -clauses can appear on the left periphery or on the right. The idea is that if -clauses are variable
modifiers, something very like variable
Exercise 3.1 Write an appropriate lexical entry for if to make the structure binders without entirely overwriting the
in (16) interpretable. The idea is that if takes its antecedent proposition and an current variable assignment. Further ref-
erences about the restrictor theory are
accessibility relation and returns a restricted accessibility relation that a world given at the end of this chapter.
stands in (relative to the evaluation world) iff it stands in the original accessibility
relation plus makes the antecedent true. □

Exercise 3.2 Consider the example with two modals in the consequent we gave
earlier:
(11) If we’re on Rte 62, we must be in Hudson and might be very close to the
Horseshoe pub.

How could this be analyzed? □

What about conditionals that do not contain a modal operator or any

other operator that the if -clause could be restricting? Kratzer proposed
that such bare conditionals contain implicit operators. There’s a case for The one-case vs. multi-case distinction is
at least two such operators: (i) a covert generic frequency operator (giv- discussed and so-named in Kadmon 1987.

ing rise to what are sometimes called “multi-case” conditionals), and (ii) a
covert (epistemic?) necessity operator akin to must (giving rise to ordinary
“one-case” conditionals):
(17) a. If Polly sees a husky, she pets it.
b. If Kim left before 6am, she got there in time.

3.2 Ordering

Our semantics treats modals as quantifiers over worlds. The set of worlds
they quantify over is supplied by context by way of assigning an acces-
sibility function to a covert pronoun, and is optionally subject to being
restricted by an if -clause, as developed just now.
We have distinguished (at least) epistemic and deontic accessibility
functions:
(18) a. The keys must be in the car. epistemic
b. Your guest can’t stay past midnight. deontic

For deontic modals, and perhaps more generally what Portner 2017
calls “priority” expressions, there is a strong argument that there is more
at play.
54 THE FI FTH DIMENSION

3.2.1 The best we can get

Consider our friend Howard:

(19) [We think that Howard forgot to return his library book.]
He has to pay a $5 fine.

According to our analysis, the deontic modal has to here claims that he
pays a $5 fine in all of the worlds compatible with what the relevant rules
(the library regulations, in this case) require. But wait: surely the rules
really require everyone to return their books on time! And so, the worlds
compatible with the rules are all worlds where all books are returned on
time, including Howard’s, and thus nobody pays a fine. How can (19)
actually be true?
It’s clear that (19) is naturally understood in such a way that its truth
depends both on facts about Howard’s actions and facts about the library
regulations. For instance, it will be judged true if (i) Howard did indeed
fail to return his book, and (ii) the regulations mandate a fine in such
cases. It may be false either because the regulations are different or be-
cause Howard did return the book. Our accessibility relation therefore
needs to be more complex, combining facts and regulations.
A second attempt at specifying the accessibility relation might thus go The proposal in (20) suggests that there’s
something like this: a temporal asymmetry here. But this is
not necessarily so. Prakken & Sergot 1996
(20) 𝜆𝑤 . 𝜆𝑤 ′ . [what happened in 𝑤 ′ up to now is the same as what hap- present the case of a set of formal and in-
pened in 𝑤 and 𝑤 ′ conforms to what the rules in 𝑤 demands]. formal regulations governing the appear-
ance and use of holiday cottages, which
The problem with (20) is that, unless there were no infractions of the say that there are not to be any fences
but that if there are fences, they must be
rules at all in 𝑤 up to now, no world 𝑤 ′ will be accessible from 𝑤. There-
white. In such a case, one could say: This
fore, (19) is predicted to follow logically from the premise that Howard fence should be white, expressing the kind of
broke some rule. This does not represent our intuition about its truth complex flavor we are dealing with.
conditions.
A better definition of the appropriate accessibility relation has to be
even more complicated:
(21) 𝜆𝑤 . 𝜆𝑤 ′ . [what happened in 𝑤 ′ up to now is the same as what hap-
pened in 𝑤 and 𝑤 ′ conforms at least as well to what the rules in 𝑤
demands as does any other world in which what happened up to
now is the same as in 𝑤].

(21) makes explicit that there is an important difference between the ways
in which facts about Howard’s actions on the one hand, and facts about
the rules on the other, enter into the truth conditions of sentences like
(19). Worlds in which Howard didn’t do what he did are simply excluded
from the domain of the modal here. Worlds in which the rules aren’t
obeyed are not absolutely excluded. Rather, we restrict the domain to
those worlds in which the rules are obeyed as well as they can be, consid-
ering what has happened. We exclude only those worlds in which there
 RE STRICTI NG AND ORDERI NG 55

are infractions above and beyond those that are shared by all the worlds in
which Howard has done what he has done. The analysis of (19) thus cru-
cially involves the notion of an ordering of worlds: here they are ordered
according to how well they conform to what the rules in 𝑤 demand.
The diagnosis then is that the modal here is not a pure deontic modal.
Rather, its flavor is complex. We take for granted the fact that Howard
did not return his book on time. Consider then just those worlds where
Howard did not return the book. None of those worlds are fully compati-
ble with the rules. But among those worlds, the ones where he pays a fine
satisfy the rules as best as possible. So, the flavor of the modal combines
some actual world circumstances (the book was not returned on time) and
what the rules require (late books incur a fine). And the flavor is essentially
complex. Imagine that Howard is a scofflaw who never pays fines. If this
fact were part of the flavor of the modal in (21), then we would expect
the sentence to be false, but intuitively it is true. And as we already saw, a
purely deontic reading would also predict the sentence to be false.
So, the flavor of the modal in (21) is best characterized as the following
mixture: it quantifies over worlds where (i) the same relevant things hap-
pened as in the evaluation world and (ii) apart from that, things develop as
best as possible according to the rules.
This is a very common pattern: intensional operators have complex
flavors that combine a set of circumstances taken for granted and some
way of identifying the best worlds within the set of worlds characterized
by those circumstances.
If we want to stick to our simple semantics, with its flavor function
(from evaluation worlds to sets of worlds quantified over), we have to lo-
cate the complexity in the pragmatics of determining a salient value to the
context-dependent flavor. For (21), the contextual variable assignment
would have to assign to the accessibility function variable a function like
the one in (21).
Kratzer famously proposed a different diagnosis: modals are doubly rel-
Modal base
ative, requiring two separate contextually supplied pieces of information.
In addition to the accessibility relation or MODAL BASE (a function that as-
Flavor
signs a set of accessible worlds to any evaluation world), modals are also
sensitive to an ORDERING of the accessible worlds.
Ordering

3.2.2 Mathematical interlude on orderings

All order relations are binary relations between the elements of a set. Here
we’re dealing with sets of worlds, but we can order soccer teams (by how
good they are), people (by how rich they are, or by how close to Rome
they were born), or race horses (by how they finished in the race). You
get the idea.
In some sense, the most general kind of order is a PREORDER, one that
you could read as “at least as highly ranked as”. It would be REFLEXIVE,
56 THE FI FTH DIMENSION

since any element is trivially at least as highly ranked as itself. And it would
be TRANSITIVE, since if 𝑥 is at least as highly ranked as 𝑦 and 𝑦 is at least as
highly ranked as 𝑧, then surely 𝑥 is at least as highly ranked as 𝑧.
When a preorder is also ANTI-SYMMETRIC (no distinct elements can
have the same rank), it is called a PARTIAL ORDER. An order is STRICT if
it actually doesn’t allow elements at the same rank, so it is transitive but
asymmetric. An order is TOTAL if it is a COMPLETE relation: any two dis-
tinct elements are related in one way or the other.
An order is WELL-FOUNDED if for any subset of the domain of the rela-
tion, there are elements in the subset that are “optimal”: there is no other
element in the subset that is strictly better by the ordering. Any order on
a finite set is well-founded, but not every order on an infinite set is.
We provide two handy charts, one list of ours and one from Amartya
Sen’s influential book Collective Choice and Social Welfare (Sen 1970).

𝑅 is irreflexive on 𝑆 ∀𝑥 ∈ 𝑆 : ¬𝑅(𝑥, 𝑥)
𝑅 is reflexive on 𝑆 ∀𝑥 ∈ 𝑆 : 𝑅(𝑥, 𝑥)
𝑅 is transitive on 𝑆 ∀𝑥, 𝑦, 𝑧 ∈ 𝑆 : 𝑅(𝑥, 𝑦) & 𝑅(𝑦, 𝑧) → 𝑅(𝑥, 𝑧)
𝑅 is symmetric on 𝑆 ∀𝑥, 𝑦 ∈ 𝑆 : 𝑅(𝑥, 𝑦) ↔ 𝑅(𝑦, 𝑥)
𝑅 is antisymmetric on 𝑆 ∀𝑥, 𝑦 ∈ 𝑆 : 𝑅(𝑥, 𝑦) & 𝑅(𝑦, 𝑥) → 𝑥 = 𝑦
𝑅 is asymmetric on 𝑆 ∀𝑥, 𝑦 ∈ 𝑆 : 𝑅(𝑥, 𝑦) → ¬𝑅(𝑦, 𝑥)
𝑅 is complete on 𝑆 ∀𝑥, 𝑦 ∈ 𝑆 : 𝑥 ≠ 𝑦 → 𝑅(𝑥, 𝑦) ∨ 𝑅(𝑦, 𝑥)
𝑅 is dense on 𝑆 ∀𝑥, 𝑦 ∈ 𝑆 : 𝑥 ≠ 𝑦 & 𝑅(𝑥, 𝑦) →
∃𝑧 ∈ 𝑆 : 𝑧 ≠ 𝑥 & 𝑧 ≠ 𝑦 & 𝑅(𝑥, 𝑧) & 𝑅(𝑧, 𝑦)
𝑅 is well-founded on 𝑆 ∀𝑆 ′ ⊆ 𝑆 ∃𝑥 ∈ 𝑆 ′ :
¬∃𝑦 ∈ 𝑆 ′ : 𝑅(𝑦, 𝑥) & ¬𝑅(𝑥, 𝑦)

Figure 3.1: Properties of CHOICE

relationsSETS
MAXIMAL ELEMENTS AND 9
We specify below the terminology to be used in this book and The first few chapters of Sen’s book are
also note a few alternative names used in the literature. 3 worth reading. A semantic introduction to
orders can be found in Landman 1991. If
Properties satisfied Name to be used Other names used you’re curious about the mathematical as-
in this work in the literature pects of this topic, we recommend Davey
& Priestley 2002.
1. reflexivity and transi- quasi-ordering pre-ordering
tivity
2. reflexivity, transitivity ordering complete pre-ordering;
and completeness complete quasi-order-
ing; weak ordering
3. reflexivity, transitivity partial ordering ordering
and anti-symmetry
4. reflexivity, transitivity, chain linear ordering; complete
completeness and ordering; simply
anti-symmetry ordering
5. transitivity and asym- strict partial
metry ordering
6. transitivity, asymmetry strong ordering ordering; strict ordering;
and completeness strict complete
ordering

1*2 Maximal Elements

Figure 3.2: andsummary
Sen 1970’s Choice Sets
of types of orders
Corresponding to the binary relation of "weak preference" R
("at least as good as"), we can define relations of "strict preference"
P and of "indifference" /.
DEFINITION 1*1. xPy<-+[xRy & ~(yRx)]
DEFINITION 1*2. xIy*->[xRy&yRx]
 RE STRICTI NG AND ORDERI NG 57

3.2.3 Ordering worlds

Armed with all this information we can return to the ordering of worlds.
When we think about worlds being ordered with respect to how well
they obey a law, for example, it’s clear that we need to allow different
worlds to be “tied”, to be at the same rank. So, our order will need to
be a preorder. It’s also wise not to assume that for two given worlds, we
can necessarily order them with respect to each other at all. So, our order
will not be total. It will be a useful idealization if we assume that our or- This is essentially the same as what is also
ders will be well-founded. Then, we can assume that we can always find known as the “Limit Assumption” in con-
ditional/modal semantics. For discussion,
worlds that are “as good as possible”. In sum, we assume that the context
see Lewis 1974 and now S. Kaufmann
will supply a function from evaluation worlds to a well-founded preorder 2017.
on worlds. So, now we can find for any set of worlds the contextually best
worlds among them.
Given a well-founded preorder ≤, we define a function from sets of Be aware that the literature is hopelessly
worlds to subsets thereof that yields the ≤-best worlds in the original set: inconsistent with respect to the direction
of the symbols for “at least as good as” or
(22) For any set of worlds 𝑝 and any well-founded preorder ≤ on 𝑊 : “better than”. Here, we use “𝑤1 ≤ 𝑤2 ”
O ≤ (𝑝) = {𝑤 ∈ 𝑝 : ¬∃𝑤 ′ (𝑤 ′ ≤ 𝑤 & ¬(𝑤 ≤ 𝑤 ′ ))} for “𝑤1 is at least as good 𝑤2 ” because
the intuition is that 𝑤1 is less far from the
We can now state the semantics for doubly-relative modals. Here’s the ideal than 𝑤2 .
new entry for must:
(23) ⟦must⟧𝑤,𝑔 = 𝜆𝑓𝑠,𝑠𝑡 . 𝜆𝑜𝑠,𝑠𝑠𝑡 : 𝑜 (𝑤) is a well-founded preorder on 𝑊 . 𝜆𝑝𝑠𝑡 .
∀𝑤 ′ ∈ O𝑜 (𝑤 ) (𝑓 (𝑤)) : 𝑝 (𝑤 ′ ) = 1

As before, the context supplies a flavor function 𝑓 that yields for any world
as set of 𝑓 -accessible worlds. The context now also supplies a function that
for any world yields a well-founded preorder. We will call this function
the ORDERING SOURCE (a term from Kratzer). The semantics of modals
uses the preorder to order the set of accessible worlds and to find the best
accessible worlds. The modal must, as a necessity modal, claims that all of
the best accessible worlds are worlds where its prejacent is true.

Exercise 3.3 Write a doubly-flavored lexical entry for the possibility modal
may. □

Let’s see how the analysis applies to (19): Howard has to pay a fine.

• The modal base will be a function that assigns to any evaluation

world the set of worlds where the same relevant circumstances hold.
Since in our stipulated evaluation world, Howard failed to return
his book, the modal base will assign to our world a set of worlds
that only contains worlds where Howard didn’t return the book.
• The ordering source will be a function that assigns to any evalua-
tion world an ordering that represents what the rules in that world
prefer. For our cases and for the given evaluation world, the or-
dering will prefer worlds where no book is returned late to worlds
58 THE FI FTH DIMENSION

where books are returned late. It further prefers worlds where fines
are paid for late books to worlds where no fines are paid.
• For our simple example then, any world in the modal base where
Howard pays a fine will count as better than an otherwise similar
world where he doesn’t. The very best worlds simpliciter are worlds
where there’s never any late books, but since there aren’t any such
worlds in the modal base, the ordering has to make do with what
it’s given.
• Modals then make quantificational claims about the best worlds in
the modal base (those for which there isn’t a world in the modal
base that is better than them).
• In our case, (19) claims that in the best worlds (among those where
Howard failed to return his book), he pays a fine.

3.2.4 Kratzer’s way

As we’ve seen in the previous chapter, Kratzer actually calculates the set
of accessible worlds from a function of type ⟨𝑠, ⟨𝑠𝑡, 𝑡⟩⟩ (what she calls a
conversational background). Recall that the construction is simple: find the
worlds where all the propositions in the set are true. The additional power
that comes with using sets of propositions remains essentially unused in
the case of modal bases. So, to simplify matters, we will stick with our
earlier assumption that the domain of modals is a set of worlds resulting
from applying a contextually supplied value of a pronoun of type ⟨𝑠, 𝑠𝑡⟩ to
the evaluation world.
For the ordering component of the doubly-relative semantics of modals, For our example about Howard’s fine,
Kratzer’s construction is quite intuitive. We assume that the ordering we might imagine that for our evaluation
world, the ordering source assigns a set of
source argument is a function from evaluation worlds to sets of propo-
propositions that contains (among others)
sitions. The idea is now that such a set P of propositions can be used to the following two propositions: (i) nobody
order the worlds in the modal base. returns books late, (ii) anybody who re-
For any pair of worlds 𝑤 1 and 𝑤 2 , we say that 𝑤 1 comes closer than 𝑤 2 turns a book late pays a fine. It’s crucial
here that the second proposition is (vac-
to the ideal set up by P (in symbols: 𝑤 1 < P 𝑤 2 ), iff the set of proposi- uously) true when nobody returns books
tions from P that are true in 𝑤 2 is a proper subset of the set of proposi- late.
tions from P that are true in 𝑤 1 .
Here’s a more precise derivation of the ordering procedure:
(24) For any set of propositions P, we define a strict partial order <𝑃 :
∀𝑤 ′, 𝑤 ′′ : 𝑤 ′ <𝑃 𝑤 ′′ iff
∀𝑝 ∈ P (𝑤 ′′ ∈ 𝑝 → 𝑤 ′ ∈ 𝑝) and ∃𝑝 ∈ P (𝑤 ′ ∈ 𝑝 ∧ 𝑤 ′′ ∉ 𝑝)
𝑤 is better than 𝑤 ′′ according to P iff all propositions in 𝑃 that
′

hold in 𝑤 ′′ also hold in 𝑤 ′ but some hold in 𝑤 ′ that do not also hold
in 𝑤 ′′ .
And of course, once we have such a strict partial order and we assume
well-foundedness, as before, we can define the selection function that
gives us the set of < P -best worlds from any set 𝑋 of worlds:
 RE STRICTI NG AND ORDERI NG 59

(25) O <P (𝑋 ) = {𝑤 ∈ 𝑋 : ¬∃𝑤 ′ (𝑤 ′ < P 𝑤)}

Finally, we need to rewrite the semantics of modals so that they are

sensitive both to a modal base/accessibility relation and an ordering source
of the Kratzerian kind:
(26) ⟦must⟧𝑤 = 𝜆𝑓𝑠,𝑠𝑡 .𝜆𝑜𝑠,𝑠𝑡𝑡 .𝜆𝑝. ∀𝑤 ′ ∈ O𝑜 (𝑤 ) (𝑓 (𝑤)) : 𝑝 (𝑤) = 1.

Exercise 3.4 Give an LF for the sentence Ashlyn must leave that will work
with the entry for must in (26). □

3.2.5 More cases of modal ordering

The combination of “facts on the ground” and “doing the best with what
we have” is pervasive in intensional constructions. We will discuss a suc-
cession of brief case-studies but there’s still much to explore.

How to get to Harvard Square. Our initial example involved deontic

modality. Next is the closely connected case of goal-oriented or teleolog-
ical modality. You want to get to Harvard Square from MIT. This is your
primary goal. You have secondary goals: to not spend a lot of money and
to stay reasonably safe. I say:
(27) You ought to take the Red Line.

A two-factor analysis of (27) might look like this: In von Fintel & Iatridou 2008, a three-
factor analysis is proposed, distinguishing
• The modal base is circumstantial. It assigns to the evaluation world technically between the primary goal
a set of propositions describing the relevant circumstances. Imagine and a secondary ordering source. There’s
further discussion in work by Rubinstein
that for our world, the following facts are relevant: (i) you are in
(2012, 2014).
Kendall Square, (ii) you can get to Harvard Square on the Red Line,
by taxi or Lyft, or on foot, (iii) the Red Line costs $1.25, takes 10
minutes, and is safe, (iv) the taxi/Lyft costs $10, is fast, and is safe, (v)
walking is free, slow, and possibly unsafe this time of the night.
• The ordering source describes your goals (often called a teleological
ordering source). Your relevant goals in our world are: (i) you get
to Harvard Square, (ii) you pay less than $2, (iii) you are safe.
• What are the best worlds in the modal base according to the given
ordering source? The worlds where you take the Red Line.
• That’s why it seems true to say (27).

The weakness of must. A more controversial application of the two-

factor analysis is found in Kratzer 1991: she argues that apparently strong
necessity modals in their epistemic construal are actually weakened by
ordering. Consider someone who observes people coming into a win-
dowless classroom with wet outer clothing. They might say:
(28) It must be raining.
60 THE FI FTH DIMENSION

Kratzer argues that what underlies (28) is a combination of an epistemic

modal base and an ordering source based on not completely exception-
less generalization and assumptions. In more recent work, von Fintel & The debate continues with Lassiter 2016,
Gillies 2010 have argued that any weakness in the meaning of (28) has Del Pinal & Waldon 2019, von Fintel &
Gillies 2020.
to do with an “evidential signal” (that the evidence for the prejacent is
merely indirect) rather than a lack of logical strength.
Wanting what’s best. Intriguingly, we can find the facts + order-
ing phenomenology in the semantics of propositional attitudes as well,
most clearly with bouletic attitudes like want. Heim 1992 tells the story of
someone who in her dream scenario would not teach next semester but
knows that she will have to. She does have preferences on what week days
to teach. So, she says:
(29) I want to teach Tuesdays and Thursdays next semester.

In terms of the two-factor analysis, the idea here is that the accessibility Heim herself gives a dynamic semantics
relation is “doxastic”: the modal base contains the worlds compatible with analysis for attitudes that incorporates
these insights. In later work, von Fin-
the agent’s beliefs. In the case at hand, all worlds in that modal base have
tel 1999 gives a static version of a two-
her teach. The ordering is based on her preferences: among the worlds factor analysis that is closely modeled after
in the modal base, she prefers the ones where she teaches Tuesdays and Kratzer’s analysis of modals. See also Ru-
Thursdays. binstein 2017 for a recent discussion.

The Good Samaritan. Prior 1958 introduced the following “Paradox

of the Good Samaritan”. Imagine that someone has been robbed and John
is walking by. It is easy to conceive of a code of ethics that would make
the following sentence true:
(30) John ought to help the person who was robbed.

In our previous one-factor semantics for modals, we would have said

that (30) says that in all of the deontically accessible worlds (those com-
patible with the code of ethics) John helps the person who was robbed.
Prior’s point was that under such a semantics, something rather unfortu-
nate holds. Notice that in all of the worlds where John helps the person
who was robbed, someone was robbed in the first place. Therefore, it
will be true that in all of the deontically accessible worlds, someone was
robbed. Thus, (30) will entail:
(31) It ought to be the case that someone was robbed.

It clearly would be good not make such a prediction.

The doubly-relative analysis of modality can successfully avoid this
unfortunate prediction. We conceive of (30) as being uttered with respect
to a circumstantial modal base that includes the fact that someone was
robbed. Among those already somewhat ethically deficient worlds, the
relatively best ones are all worlds where John helps the victim.
Note that we still have the problematic fact that among the worlds in
the modal base, all are worlds where someone was robbed, and we would
thus appear to still make the unfortunate prediction that (31) should be
 RE STRICTI NG AND ORDERI NG 61

true. But this can now be fixed. For example, we could say that ought 𝑝
is semantically defective if 𝑝 is true throughout the worlds in the modal
base. This could be a presupposition or some other ingredient of mean-
ing. So, with respect to a modal base which pre-determines that someone
was robbed, one couldn’t felicitously say (31).
Consequently, saying (31) would only be felicitous if a different modal
base is intended, one that contains both 𝑝 and non-𝑝 worlds. And given
a choice between worlds where someone was robbed and worlds where
nobody was robbed, most deontic ordering sources would probably choose
the no-robbery worlds, which would make (31) false, as desired.

3.2.6 Ordering in conditionals

Historically, the use of an ordering of possible worlds in the semantics of Even though Lewis’ work appeared five
intensional operatores was first proposed by Stalnaker and Lewis in their years after Stalnaker’s, it was indepen-
dently developed. See Arvan 2015 and
work on conditionals (Stalnaker 1968, Lewis 1973). In our attempt at an
Fn.29 in Starr 2019 for further details
analysis for conditionals in Section 2.2, we used a one-factor approach about the intellectual history of these
where conditionals quantify over all contextually accessible antecedent ideas.
worlds. Stalnaker and Lewis argue that the logic of conditionals, primarily
counterfactuals, reveals the effects of an ordering. Consider the case of a The match makes an early appearance in
perfectly normal match in front of us and the following two counterfac- Goodman 1947.

tual conditionals:
(32) a. If I had struck this match, it would have lit.
b. If I had dipped this match in water and struck it, it would have
lit.
Intuitively, it is clear that (32a) may well be true while (32b) is almost cer-
tainly false. But according to our previous analysis, which is a version of
what in the trade is called a “strict conditional” analysis, this cannot be. If
all of the accessible worlds where I strike this match are worlds where it
lights, then this must a fortiori also be true of that subset of the accessible
worlds where I strike this match after having dipped it in water.
In the Stalnaker-Lewis ordering semantics, (32b) doesn’t anymore fol-
low from (32a). The most highly ranked worlds where I strike the match
might well be worlds where I don’t dip it in water first, and so (32a) could
be true while (32b) is false.
Within this book’s framework, we would replace the meaning for if
from (10) in Section 2.2 with the following two-factor meaning:
(33) ⟦if⟧𝑤,𝑔 = 𝜆𝑓𝑠,𝑠𝑡 .𝜆𝑜𝑠,𝑠𝑡𝑡 .𝜆𝑝𝑠𝑡 .𝜆𝑞𝑠𝑡 . ∀𝑤 ′ ∈ O𝑜 (𝑤 ) (𝑓 (𝑤) ∩ 𝑝) : 𝑞(𝑤 ′ ) = 1.

The crucial move here is that the antecedent proposition 𝑝 is used to find
the 𝑝-worlds that are among the 𝑓 -accessible worlds. The resulting set of
accessible 𝑝-worlds is then ordered and the best accessible worlds 𝑝 are
what the universal quantifier makes a claim about.
62 THE FI FTH DIMENSION

But wait. In Section 3.1, we argued that if -clauses don’t actually do

their own modal quantification, so the analysis in (33) can’t be quite what
we should adopt here. In fact, from the restrictor point of view, the se-
mantics of conditionals and their reliance on an ordering on possible
worlds should be a compositional effect: the if -clause serves as a restric-
tor for a (possibly covert) modal and it is that modal that is sensitive to an
accessibility relation and an ordering. In the next section, we explore the
interplay of conditionals and modals in a theory that involves modal bases
and orderings. Things are not entirely straightforward.

3.3 The interplay of ordering and restricting

We return to Howard and the saga of the possibly late library book. Imag-
ine that we don’t know whether he returned it late but we’re confident
we understand the library regulations. So, we say:
(34) If Howard returned his book late, he has to pay a fine.

There is one analysis that will not work: the restrictor analysis plus the One might think that instead of starting
pragmatically sophisticated but semantically simple idea that modals are with the worlds that match the evalua-
tion world in whether the book was re-
just sensitive to an accessibility relation even when that relation incor-
turned late, we could start with the worlds
porates considerations of ordering. So, assume that the accessibility rela- compatible with our evidence. Then, we
tion is the one that delivers the best worlds among those that match the would include both worlds where the
evaluation world with respect to whether the book was late or not. Now, book was late and where it wasn’t. But
even among those, the not-late worlds
imagine a scenario where unbekownst to us, Howard actually returned are the best. So, then again, the if -clause
the book on time. It would seem that this would not threaten the intu- could not find any antecedent worlds
itive truth of (34). But according to the analysis under consideration, we among the accessible worlds.
predict something odd: the worlds that match the evaluation world are all
worlds where the book was not late and the best ones among them thus
are also not-late worlds (and of course they are worlds without a fine).
The if -clause would then try to find late worlds in that set of accessible
worlds, but since there are none, we would get an empty set of worlds to
quantify over, resulting either in vacuous truth or some kind of infelicity.
This is not correct.
The two-factor analysis (modal base 𝑓 , ordering 𝑜) can help with the
issue. The idea is that the if -clause does not restrict the 𝑜-best 𝑓 -worlds
but restricts 𝑓 before 𝑜 enters into the picture. In other words, we posit an
LF like this one:
 RE STRICTI NG AND ORDERI NG 63

(35)

𝑜 5,⟨𝑠,𝑠𝑡𝑡 ⟩ Howard pays a fine

have to

𝑓3,⟨𝑠,𝑠𝑡 ⟩

if

the book is late

To evaluate this LF, we can use the entry for if you developed in Exer-
cise 3.1 and the entry for modals in (26). Note that for the if -clause not to
be an idle restrictor, the worlds 𝑓 delivers need to include worlds where
the antecedent is true.

Exercise 3.5 In fact, calculate the truth-conditions for (35) with the ingredients
just mentioned. □

Exercise 3.6 At this point, it would be very instructive to read Kratzer 1991.
In particular, you should study her version of the Samaritan Paradox in her Sec-
tion 3.2 and her solution in her Section 8. □

It appears then that we have found vindication both for the two-factor While Zvolensky’s 2002 SALT paper was
semantics for modals and for the restrictor theory. There’s a problem instrumental in bringing the issue into
the natural language semantics literature,
though, illustrated by Zvolenszky 2002 with the following example:
it had been already identified by Frank
1996 and in fact, it had been well-known
as the “if p, ought p” problem in the more
(36) If Britney Spears drinks Coke in public, then she must drink Coke
logico-philosophical world. See Carr 2014
in public. for references and an important recent
contribution.
An intuitively plausible use case for (36) would be a situation where we
consider the possibility of seeing Spears drinking Coke in public. We As a matter of fact, Spears was paid to
endorse Pepsi.
conclude that if we saw her drinking Coke, we would be able to deduce
that she is somehow contractually obligated to drink Coke.
Unfortunately, our current analysis predicts that (36) should, if any-
thing, be vacuously true rather than making a contingent claim. The
if -clause would restrict whatever accessibility relation we’re using and
ensure that we’re only dealing with worlds where Spears drinks Coke in
public. Then, the ordering would apply to find the best such worlds. Fi-
nally, the modal would say that the best such worlds are worlds where she
drinks Coke in public. That is of course trivially true.
Consider what is needed to make the modal claim (that Spears must
drink Coke in public) non-trivial. The domain of quantification needs to
include worlds where she doesn’t drink Coke (perhaps she drinks Pepsi,
64 THE FI FTH DIMENSION

or Poland Spring Water). Then, the modal’s ordering source identifies

the best worlds and the modal says that all of those are Coke worlds. This
non-triviality can only arise if if -clause doesn’t in fact restrict the modal
to Coke worlds in the first place. So, is the restrictor theory in trouble?
At this point, we must recall that under the restrictor theory, if -clauses
can also restrict covert modals. This means that when we have an if -
clause and an overt modal, we can’t automatically conclude that the con-
ditional is restricting that modal. It might be restricting another modal, a
covert one, instead.
So, let’s look at (36) in that light. Imagine that the if -clause is restrict-
ing a higher covert epistemic modal. Then, we are taken from the evalu-
ation world to those worlds compatible with our evidence where Spears
drinks Coke. We are saying that in all of those worlds Spears is obligated
to drink Coke. That will be a non-trivial claim if the worlds the lower
modal quantifies over contains both Coke and non-Coke worlds.
More technically, we’d be working with an LF like the one sketched
in (37). We use □ for the covert modal. We need to distinguish four modal
parameter “pronouns”.
(37) □ (𝑓1 if Coke) (𝑜 1 ) [ must 𝑓2 𝑜 2 Coke ]

What values for the four parameter would yield the reading we want? 𝑓1
should give us the worlds compatible with our evidence. 𝑜 1 could be triv-
ial or the kind of not entirely reliable evidence that Kratzer 1991 argues
for. Note that the higher modal will pass down only Coke-worlds. Cru-
cially, 𝑓2 would give us both Coke and non-Coke worlds when applied
to the epistemically accessible worlds passed on by the higer operator. 𝑜 2 ,
finally, would encode Spear’s contractual obligations.
With this, we have rescued the restrictor theory. The existence of such See Geurts 2004 for a discussion of how
nested structures is predicted by the framework and now we see that they this kind of structure is expected.

are empirically attested.

Let’s return to the case of Howard and his book again. As it turns out,
we could get the correct reading for our sentence If Howard returned the
book late, he has to pay a fine without having if directly restrict the overt
modal. Assume that the if -clause restricts a higher covert modal, again
maybe an epistemic one. And now, in contrast to the Spears example, as-
sume that the lower 𝑓 -relation assigns to any of the worlds passed down
(all of which will be late book worlds) a set of worlds matching the facts
about whether the book was returned late. So, in effect, the conditional
antecedent will end up restricting both modals: the higher one directly
and the lower one indirectly since the lower accessibility relation will en-
sure matching of the relevant fact.
This possibility led Frank 1996 to propose that deontic modals in fact
never are directly restricted:
 RE STRICTI NG AND ORDERI NG 65

There are in fact no truly deontically modalized if -conditionals.

Instead, we assume conditionals with a deontic modal op-
erator in the consequent clause to be analyzed throughout
in terms of an implicit or explicit epistemically (or circum-
stantially) based modal operator. The deontic modal adverb is
then to be analyzed within the scope of the ‘higher’ epistemic
modal operator.

Now, if this is in fact feasible, why should we then still maintain the
restrictor analysis in the first place? Why not go back to the earlier idea
that if is a modal operator in its own right?
To evaluate this possibility, let’s return once more to our friends who
do not know precisely where they are. Imagine that they have reason to
think that among the two Rte 117 possibilities, Stow is much more likely
than Maynard, perhaps because they know Maynard a bit better and think
that if they were there, they would very likely recognize some build-
ing or other. They have no reason to think that Stow is more likely than
Hudson (which is on Rte 62, remember). So, they say:
(38) If we’re on Route 117, we ought to be in Stow.

Now, to be clear: they are still lost, so their evidence is compatible with
them being on either of the two routes and in any of the three towns.
So, for (38) to have a chance of being true, the modal ought here can’t
simply map the evaluation world to the worlds epistemically accessible
from it (and then potentially order the resulting worlds). We need to nar-
row things down to the Route 117 worlds, even though our friends don’t
know whether they’re there, even if they are. So, as things stand, we still
need the restrictor analysis. And we need the restriction to happen before
the ordering applies. So, for now, (38) provides the strongest argument
we have for both the restrictor theory and the two-factor semantics for
modals.
At this point, we leave this intriguing and complex inquiry and refer
you to the ever-growing literature, some we’ve already mentioned and
more we will list below.

Further readings

Conditionals. In-depth overviews of the logico-philosophical perspective

on conditionals include Nute 1984, Edgington 1995, Bennett 2003, and
Starr 2019.
Ideas pertaining to the implementation of the restrictor theory are
found in Reich 2009 and Ebert, Ebert & Hinterwimmer 2014. Schlenker
2004 presents an alternative that is at least plausibly in the same orbit. One
should really take into account what we know about the syntax of condi-
tionals, some of which is discussed in Chapter 3 of Kai’s thesis but much
66 THE FI FTH DIMENSION

more is found in Bhatt & Pancheva 2006. Rothschild 2011 explains the
restrictor theory to philosophers, might be useful as an additional reading.
Gillies 2010 proposes an ingenious alternative to the restrictor theory.
Khoo 2011 files a complaint about the coverage of Gillies’ proposal. The
interaction of conditionals with probability operators is very tricky, see
Egré & Cozic 2011 and von Fintel & Gillies 2015 for some discussion.
Higginbotham 1986 identified the interaction of negative quantifiers
like no student with conditionals as a compositionality puzzle. The restric-
tor theory has sometimes been seen as a solution to the puzzle (von Fintel
1998) and sometimes not (von Fintel & Iatridou 2002). See Dekker 2001,
Higginbotham 2003, Leslie 2009, Huitink 2010, Klinedinst 2011, Kratzer
2015, Lauer & Nadathur 2016 among others.
To understand only if -conditionals, von Fintel 1997 needs to work
hard to neutralize the universal force of bare conditionals stemming from
Kratzer’s implicit necessity modal.
Herburger 2015, 2016, and Bassi & Bar-Lev 2017 consider the possi-
bility that bare conditionals are actually (sometimes or always) existen-
tially quantified.
Attitudes. Linguistic work on attitudes has often been concerned with
various co-occurrence patterns, particularly which moods (indicative or
subjunctive or infinitive) occur in the complement and whether negative
polarity items are licensed in the complement. Mood licensing: Portner
1997. NPI-Licensing: Kadmon & Landman 1993, von Fintel 1999, Gian-
nakidou 1999.
Tamina Stephenson in her MIT dissertation and related work explores
the way attitude predicates interact with epistemic modals and taste predi-
cates in their complements: Stephenson 2007a,b.
Jon Gajewski in his MIT dissertation and subsequent work explores
the distribution of the NEG-RAISING property among attitude predicates
and traces it back to presuppositional components of the meaning of the
predicates: Gajewski 2005, 2007.
Interesting work has also been done on presupposition projection in
attitude contexts: Asher 1987, Heim 1992, Geurts 1998.
Modals. On the syntax of modals, there are only a few papers of un-
even quality: Bhatt 1997, Wurmbrand 1999, Cormack & Smith 2002,
Butler 2003. Follow up on older references from the bibliographies in
these papers. The following paper explores some issues in the LF-syntax
of epistemic modals: von Fintel & Iatridou 2003,
Valentine Hacquard’s MIT dissertation is a rich source of cross-linguistic
issues in modality, as is Fabrice Nauze’s Amsterdam dissertation: Hac-
quard 2006, Nauze 2008.
Some more recent work by Hacquard deals with deriving and corre-
lating modal flavors with syntactic position of the modal auxiliaries: Hac-
quard 2010, Hacquard 2013. A recent handbook article by Hacquard on
 RE STRICTI NG AND ORDERI NG 67

actuality entailments (involving the interaction of modality with aspect;

we’ll discuss aspect later in these notes): Hacquard 2016.
The semantics of epistemic modals has become a hot topic recently.
Here are some of the main references: Hacking 1967, Teller 1972, DeRose
1991, Egan, Hawthorne & Weatherson 2005, Egan 2007, MacFarlane
2006. Stephenson 2007a, Hawthorne 2007, von Fintel & Gillies 2008a,
von Fintel & Gillies 2008b.
A paper by Pranav Anand and Valentine Hacquard tackles what hap-
pens to epistemic modals under attitude predicates: Anand & Hacquard
2013.
Evidentiality is a topic closely related to epistemic modality. Some ref-
erences: Willett 1988, Aikhenvald 2004, Drubig 2001, Blain & Déchaine
2007, McCready & Ogata 2007, Speas 2008, von Fintel & Gillies 2010.
Modals interact with disjunction and indefinites to generate so-called
FREE CHOICE-readings, which are a perennial puzzle. Here is just a very
small set of initial references: Kamp 1973, Zimmermann 2000, Schulz
2005, Aloni 2007, Alonso-Ovalle 2006, Fox 2007, van Rooij 2006.
Restricting and ordering. Very important for the problem we ended
the chapter on are these papers: Kaufmann & Schwager 2009 and Condo-
ravdi & Lauer 2016.
 P A R T II

Escaping opacity
4 Specificity and Transparency

4.1 Predictions of our framework . . . . . . . . . . . . . 71

4.2 Raised subjects . . . . . . . . . . . . . . . . . . . . . . 76
4.2.1 Non-specific readings for raised subjects . . . 76
4.2.2 Syntactic “Reconstruction” . . . . . . . . . . 80
4.2.3 Some Alternatives to Syntactic Reconstruction 83
4.3 Further readings . . . . . . . . . . . . . . . . . . . . . 88

4.1 Predictions of our framework

When a DP occurs in the scope of an intensional operator, our frame-

work makes clear predictions. Consider, for example:
(1) Emma wanted Taylor to buy a book about soccer.

Imagine that we give the following simplified meaning to want: We know from the previous chapter ex-
actly in what sense this meaning is sim-
(2) ⟦want⟧𝑤,𝑔 = 𝜆𝑝.𝜆𝑥 . ∀𝑤 ′ : 𝑥’s wants in 𝑤 are satisfied in 𝑤 ′ plified: it doesn’t capture the interplay of
→ 𝑝 (𝑤 ′ ) = 1. beliefs and preferences.

In other words, 𝑥 wants 𝑝 is true iff 𝑝 is true in all worlds where 𝑥’s wants
are satisfied. Further, assume that the DP a book about soccer is interpreted
within the embedded clause. Then, we claim, (1) will be true iff in all of
the worlds that satisfy all of Emma’s wants, there is a book about soccer
that Taylor buys. (You prove this claim in the following exercise.)

Exercise 4.1 Draw the obvious, if simplified, LF for (1) and calculate its truth-
conditions. □

Now, consider what happens if the object of the lower verb QRs and ad-
joins to the matrix clause:
(3) [a book about soccer] (1 [Emma wanted Taylor to buy 𝑡 1 ])

When you calculate the truth-conditions of (3) [please do so], you will
get a result that is very different from the previous exercise. Now what is
claimed is that there is a book about soccer, call it 𝑥, such that in all of the
worlds satisfying all of Emma’s wants Taylor buys 𝑥.
72 E S C A P I N G O PAC I T Y

There are two important differences between the truth-conditions that

our framework assigns to these two LFs.
Quantifier scope: Since want is a universal quantifier over worlds and
a book about soccer is an existential quantifier over individuals, there’s a
question about the relative scope of the two quantifiers. In the first truth-
conditions we sketched, the existential quantifier scopes under the uni-
versal quantifier: for every world there is an individual such that bla-bla.
In the second LF, the existential quantifier scopes over the universal one:
there is an individual such that in every world yadda-yadda. The most
common terminology for this difference involves the pair specific/non-
specific: the sentence (or the object DP) is interpreted specifically if the
DP takes scope over the intensional operator, and it is interpreted non-
specifically if the DP takes scope under the intensional operator.
Predicate evaluation: When the existential quantifier scopes over the in-
tensional operator, this also has the effect that the predicate contained
in it, book about soccer, is evaluated in the matrix evaluation world. And
when the quantifier takes lower scope, its predicate is evaluated in the
worlds that the intensional operator shifts to. One evocative terminology
for whether the predicate is evaluated with respect to the matrix evalu-
ation world or the worlds shifted to by the intensional operator is trans-
parent/opaque. A predicate evaluated relative to the matrix world is called
transparent. A predicate evaluated in the worlds shifted to by an intensional
operator is receiving an opaque interpretation.
There’s another terminological pair that is very common: de re/de dicto.
One way to conceive of that distinction in our framework is that it stands
for a particular combination of the two distinctions we just introduced:
de re means specific and transparent, and de dicto means non-specific and
opaque.
Caution about terminology: terminological confusion and exuberance
is rampant in this area (and many others). In a way, terminology is just
a short-hand way to pick out salient properties of LFs (or their denota-
tion). It’s the latter that truly matters. One particular problematic aspect
of the terminology is its binary nature, while the relevant distinctions are
actually more complex, especially as soon as we are dealing with nested
intensionality.

Exercise 4.2 Consider the sentence Emma must want Taylor to buy a book
about soccer. One can imagine using this to describe a scenario where we are
seeing Taylor enter a bookstore known to cater to soccer aficionados. For some
reason we won’t go into, we come to the conclusion that there is a specific book
about soccer that Emma must have asked Taylor to buy. But at the same time,
we have no idea what that book might be, so there’s not a specific book about
which we made our deduction. This suggest that we may want to give the object
DP intermediate scope. So, draw an LF that corresponds to this idea and calculate
its truth-conditions. □
 S P E C I F I C I T Y A N D T R A N S PA R E N C Y 73

The notion of specificity is particularly laden with complexity. We use

it here to indicate the scope of a DP with respect to an intensional oper-
ator. There are quite a few other conceptions, usefully surveyed by von
Heusinger 2011. One that is very common is the notion that specific DPs
indicate that the speaker has a referent or referents “in mind”, something
that may even be relevant in the absence of intensional operators. Con-
sider:
(4) A man is knocking on the door.

The speaker might base such a claim on circumstantial evidence (hearing

a particularly masculine knock on the door) or on clear visual evidence
(seeing the man through the window). In the second case, we’re tempted
to say that the speaker has a specific referent “in mind” while in the first
case there is non-specific existential quantification.Whether this is merely For discussion, see for example Ludlow &
a pragmatic distinction about what is grounding the speaker’s claim or Neale 1991.

whether the LF reflects this difference in “specificity” is a debatable issue.

If one adopted the idea that all matrix sentences contain an epistemic op-
erator, such as Meyer’s “Matrix K” operator (Meyer 2013), then we would
have a genuine LF ambiguity if DPs such as a man could take scope over
or under that operator.
Our recommendation is to use the terms specific/non-specific, trans-
parent/opaque, de re/de dicto only with extreme caution. They are some-
times useful shorthands, but unless it is crystal-clear what properties of
LFs/denotations you are using them to pick out, they are more likely to
be a source of obfuscation and confusion.
Let’s look at some more examples of the ambiguity predicted by our
framework as soon as we allow for the possibility of an embedded DP to
take scope either under or over a relevant intensional operator. A classic
kind of example is (5), which is of the same form as our initial example
and contains the DP a plumber inside the infinitival complement of want.
(5) Mesut wants to marry a plumber.

According to the non-specific reading, every possible world in which

Mesut gets what he wants is a world in which there is a plumber whom
he marries. According to the specific reading, there is a plumber in the
actual world whom Mesut marries in every world in which he gets what
he wants. We can imagine situations in which one of the readings is true
and the other one false.
For example, suppose Mesut thinks that plumbers make ideal spouses,
because they can fix things around the house. He has never met one so
far, but he definitely wants to marry one. In this scenario, the non-specific
reading is true, but the specific reading is false. What all of Mesut’s desire-
worlds have in common is that they have a plumber getting married to
Mesut in them. But it’s not the same plumber in all those worlds. In fact,
74 E S C A P I N G O PAC I T Y

there is no particular individual (actual plumber or other) whom he mar-

ries in every one of those worlds.
For a different scenario, suppose that Mesut has fallen in love with
Robin and wants to marry Robin. Robin happens to be a plumber, but
Mesut doesn’t know this; in fact, he wouldn’t like it and might even call
off the engagement if he found out. Here the specific reading is true, be-
cause there is an actual plumber, viz. Robin, who gets married to Mesut
in every world in which he gets what he wants. The non-specific reading
is false, however, because the worlds which conform to Mesut’s wishes
actually do not have him marrying a plumber in them. In his favorite
worlds, he marries Robin, who is not a plumber in those worlds.
When confronted with this second scenario, you might, with equal
justification, say ‘Mesut wants to marry a plumber’, or ‘Mesut doesn’t want
to marry a plumber’. Each can be taken in a way that makes it a true de-
scription of the facts — although, of course, you cannot assert both in the
same breath.This intuition fits well with the idea that we are dealing with Actually, why wouldn’t one be able to
a genuine ambiguity. assert both sentences in the same breath, if
both have a true reading?
Let’s look at another example:
(6) Sari believes that your abstract will be accepted.

Here the relevant DP in the complement clause of the verb believe is your
abstract. Again, we detect an ambiguity, which is brought to light by con-
structing different scenarios.

(i) Sari’s belief may be about an abstract that she reviewed, but since
the abstract is anonymous, she doesn’t know who wrote it. She told
me that there was a wonderful abstract about subjacency in Hindi
that is sure to be accepted. I know that it was your abstract and in-
form you of Sari’s opinion by saying (6). This is the specific reading.
In the same situation, the non-specific reading is false: Among Sari’s
belief worlds, there are many worlds in which your abstract will
be accepted is not true or even false. For all she knows, you might
have written, for instance, that terrible abstract about Antecedent-
Contained Deletion, which she also reviewed and is positive will be
rejected.
(ii) For the other scenario, imagine that you are a famous linguist, and
Sari doesn’t have a very high opinion about the fairness of the ab-
stract selection process. She thinks that famous people never get
rejected, however the anonymous reviewers judge their submis-
sions. She believes (correctly or incorrectly — this doesn’t matter
here) that you submitted a (unique) abstract. She has no specific in-
formation or opinion about the abstract’s content and quality, but
given her general beliefs and her knowledge that you are famous,
she nevertheless believes that your abstract will be accepted. This is
the non-specific reading. Here it is true in all of Sari’s belief worlds
 S P E C I F I C I T Y A N D T R A N S PA R E N C Y 75

that you submitted a (unique) abstract and it will be accepted. The

specific reading of (6), though, may well be false in this scenario.
Suppose — to flesh it out further — the abstract you actually sub-
mitted is that terrible one about ACD. That one surely doesn’t get
accepted in every one of Sari’s belief worlds. There may be some
where it gets in (unless Sari is certain it can’t be by anyone famous,
she has to allow at least the possibility that it will get in despite its
low quality). But there are definitely also belief-worlds of hers in
which it doesn’t get accepted.
We have taken care here to construct scenarios that make one of
the readings true and the other false. This establishes the existence
of two distinct readings. We should note, however, that there are
also many possible and natural scenarios that simultaneously sup-
port the truth of both readings. Consider, for instance, the following
third scenario for sentence (6).
(iii) Sari is your adviser and is fully convinced that your abstract will
be accepted, since she knows it and in fact helped you when you
were writing it. This is the sort of situation in which both the non-
specific and the specific reading are true. It is true, on the one hand,
that the sentence your abstract will be accepted is true in every one of
Sari’s belief worlds (non-specific reading). And on the other hand,
if we ask whether the abstract which you actually wrote will get
accepted in each of Sari’s belief worlds, that is likewise true (specific
reading).
In fact, this kind of “doubly verifying” scenario is very common
when we look at actual uses of attitude sentences in ordinary con-
versation. There may even be many cases where communication
proceeds smoothly without either the speaker or the hearer mak-
ing up their minds as to which of the two readings they intend or
understand. It doesn’t matter, since the possible circumstances in
which their truth-values would differ are unlikely and ignorable
anyway. Still, we can conjure up scenarios in which the two read-
ings come apart, and our intuitions about those scenarios do support
the existence of a semantic ambiguity.

Exercise 4.3 For the two examples just discussed, we can explain their non-
specific (and opaque) interpretation via LFs where the relevant DP remains inside
the scope of the intensional operator at LF:
(7) Mesut wants [ [ a plumber]1 [ ᴘRᴏ2 to marry t1 ]]

(8) Sari believes [ the abstract-by-you will-be-accepted]

To obtain the specific (and transparent) readings, we apparently have to QR

the DP to a position above the intensional predicate, minimally the VP headed
by want or believe.
76 E S C A P I N G O PAC I T Y

(9) [ a plumber]1 [ Mesut wants [ ᴘRᴏ2 to marry t1 ]]

(10) [ the abstract-by-you]1 [ Sari believes t1 will-be-accepted]

Calculate the interpretations of the four structures in (7) — (10), and determine
their predicted truth-values in each of the (types of) possible worlds that we de-
scribed above in our introduction to the ambiguity.
Some assumptions to make the job easier: (i) Assume that (7) and (9) are
evaluated with respect to a variable assignment that assigns Mesut to the num-
ber 2. This assumption takes the place of a worked out theory of how controlled
PRO is interpreted. (ii) Assume that abstract-by-you is an unanalyzed one-
place predicate. This takes the place of a worked out theory of how genitives with
a non-possessive meaning are to be analyzed. □

4.2 Raised subjects

In the examples of ambiguities that we have looked at so far, the surface

position of the DP in question was inside the modal predicate’s clausal or
VP-complement. We saw that if it stays there at LF, a non-specific opaque
reading results, and if it covertly moves up above the modal operator, we
get a specific transparent reading. In the present section, we will look at
cases in which a DP that is superficially higher than a modal operator can
still be read non-specifically. In these cases, it is the specific reading which
we obtain if the LF looks essentially like the surface structure, and it is the
non-specific reading for which we apparently have to posit a non-trivial
covert derivation.

4.2.1 Non-specific readings for raised subjects

Suppose I come to my office one morning and find the papers and books
on my desk in different locations than I remember leaving them the night
before. I say:
(11) Somebody must have been here (since last night).

On the assumptions we have been making, somebody is base-generated as

the subject of the VP be here and then moved to its surface position above
the modal. So (11) has the following S-structure, which is also an inter-
pretable LF.
(12) somebody [ 2 [ [ must 𝑓 ] [ t2 have-been-here]]]

What does (12) mean? The appropriate reading for must here is epistemic,
[
so suppose the variable 𝑓 is mapped to the relation 𝜆𝑤 .𝜆𝑤 ′ . 𝑤 ′ is com-
]
patible with the evidence in 𝑤 . Let 𝑤 0 be the utterance world. Then the
truth-condition calculated by our rules is as follows.
 S P E C I F I C I T Y A N D T R A N S PA R E N C Y 77

(13) (12) is true iff ∃𝑥 [𝑥 is a person in 𝑤 0 &

∀𝑤 ′ [𝑤 ′ is compatible with the evidence in 𝑤 0 → 𝑥 was here in
𝑤 ′ ]]

But this is not the intended meaning. For (13) to be true, there has to be
a person who in every world compatible with the evidence was in my of-
fice. In other words, all the relevant worlds have to have one and the same
person coming to my office. But this is not what you intuitively under-
stood me to be saying about the evidence when I said (11). The context
we described suggests that I do not know (nor have any opinion about)
which person it was that was in my office. For all I know, it might have
been John, or it might have been Mary, or it have been this stranger here,
or that stranger there. In each of the relevant worlds , somebody or other
was in my office, but no one person was there in all of them. I do not
believe of anyone in particular that he or she was there, and you did not
understand me to be saying so when I uttered (11). What you did under-
stand me to be claiming, apparently, was not (13) but (14).
(14) ∀𝑤 ′ [𝑤 ′ is compatible with the evidence in 𝑤 0
→ ∃𝑥 [𝑥 is a person in 𝑤 ′ & 𝑥 was here in 𝑤 ′ ]]

In other words — to use the terminology we introduced in the last sec-

tion — the DP somebody in (11) appears to have a non-specific reading.
How can sentence (11) have the meaning in (14)? The LF in (12), as
we saw, means something else; it expresses a specific reading, which typ-
ically is false when (11) is uttered sincerely. So there must be another LF.
What does it look like and how is it derived? One way to capture the in-
tended reading, it seems, would be to generate an LF that’s essentially the
same as the underlying structure we posited for (11), i.e., the structure
before the subject has raised:
(15) [IP e [I′ [ must 𝑓 ] [ somebody have-been-here]]]

(15) means precisely (14) (assuming that the unfilled Spec-of-IP position
is semantically vacuous), as you can verify by calculating its interpreta-
tion by our rules. So is (15) (one of ) the LF(s) for (11), and what assump-
tion about syntax allow it to be generated? Or are there other — perhaps
less obvious, but easier to generate — candidates for the non-specific LF-
structure of (11)?
Before we get into these question, let’s look at a few more examples.
Each of the following sentences, we claim, has a non-specific reading for
the subject, as given in the accompanying formula. The modal operators
in the examples are of a variety of syntactic types, including modal auxil-
iaries, main verbs, adjectives, and adverbs.
78 E S C A P I N G O PAC I T Y

(16) Everyone in the class may have received an A.

∃𝑤 ′ [𝑤 ′ conforms to what I believe in 𝑤 &
∀𝑥 [𝑥 is in this class in 𝑤 ′ → 𝑥 received an A in 𝑤 ′ ]].

(17) At least two semanticists have to be invited.

∀𝑤 ′ [𝑤 ′ conforms to what is desirable in 𝑤
→ ∃2𝑥 [𝑥 is a semanticist in 𝑤 ′ & 𝑥 is invited in 𝑤 ′ ]].

(18) Somebody from New York is expected to win the lottery.

∀𝑤 ′ [𝑤 ′ conforms to what is expected in 𝑤
→ ∃𝑥 [𝑥 is a person from NY in 𝑤 ′ & 𝑥 wins the lottery in 𝑤 ′ ]]

(19) Somebody from New York is likely to win the lottery.

∀𝑤 ′ [𝑤 ′ is as likely as any other world, given what I know in 𝑤
→ ∃𝑥 [𝑥 is a person from NY in 𝑤 ′ & 𝑥 wins the lottery in 𝑤 ′ ]]

(20) One of these two people is probably infected.

∀𝑤 ′ [𝑤 ′ is as likely as any other world, given what I know in 𝑤
→ ∃𝑥 [𝑥 is one of these two people & 𝑥 is in infected in 𝑤 ′ ]]

To bring out the intended non-specific reading of the last example (to
pick just one) imagine this scenario: We are tracking a dangerous virus
infection and have sampled blood from two particular patients. Unfortu-
nately, we were sloppy and the blood samples ended up all mixed up in
one container. The virus count is high enough to make it quite probable
that one of the patients is infected but because of the mix-up we have no
evidence about which one of them it may be. In this scenario, (20) ap-
pears to be true. It would not be true under a specific reading, because
neither one of the two people is infected in every one of the likely worlds.
Excursus. Hopefully the exact analysis of the modal operators likely
and probably is not too crucial for the present discussion, but you may
still be wondering about it. As you see in our formula, we are thinking
of likely (probably) as a kind of epistemic necessity operator, i.e., a uni-
versal quantifier over a set of worlds that is somehow determined by the
speaker’s knowledge. (We are focussing on the “subjective probability”
sense of these words. Perhaps there is also an “objective probability” read-
ing that is circumstantial rather than epistemic.) What is the difference
then between likely and e.g. epistemic must (or necessary or I believe that)?
Intuitively, ‘it is likely that p’ makes a weaker claim than ‘it must be the
case that p’. If both are universal quantifiers, then, it appears that likely is
quantifying over a smaller set than must, i.e., over only a proper subset of
the worlds that are compatible with what I believe. The difference con-
cerns those worlds that I cannot strictly rule out but regard as remote pos-
sibilities. These worlds are included in the domain for must, but not in the
one for likely. For example, if there was a race between John and Mary,
 S P E C I F I C I T Y A N D T R A N S PA R E N C Y 79

and I am willing to bet that Mary won but am not completely sure she
did, then those worlds where John won are remote possibilities for me.
They are included in the domain of must, and so I will not say that Mary
must have won, but they are not in the domain quantified over by likely,
so I do say that Mary is likely to have won.
This is only a very crude approximation, of course. For one thing,
probability is a gradable notion. Some things are more probable than oth-
ers, and where we draw the line between what’s probable and what isn’t is
a vague or context-dependent matter. Some people even believe that must,
necessary etc. arguably don’t really express complete certainty (because in
practice there is hardly anything we are completely certain of ), but rather
just a very high degree of probability. For more discussion of likely, neces-
sary, and other graded modal concepts in a possible worlds semantics, see
e.g. Kratzer 1981, Yalcin 2010, Lassiter 2017.
A different approach may be that likely quantifies over the same set
of worlds as must, but with a weaker, less than universal, quantificational
force. I.e., ‘it is likely that p’ means something like p is true in most of the
worlds conforming to what I know. A prima facie problem with this idea
is that presumably every proposition is true in infinitely many possible
worlds, so how can we make sense of cardinal notions like ‘more’ and
‘most’ here? But perhaps this can be worked out somehow. End of ex-
cursus.
A word of clarification about our empirical claim: We have been con-
centrating on the observation that non-specific readings are available, but
have not addressed the question whether they are the only available read-
ings or coexist with equally possible specific readings. Indeed, some of
the sentences in our list appear to be ambiguous: For example, it seems
that (18) could also be understood to claim that there is a particular New
Yorker who is likely to win (e.g., because he has bribed everybody). Oth-
ers arguably are not ambiguous and can only be read non-specific. This is
what von Fintel & Iatridou (2003) claim about sentences like (16). They
note that if (16) also allowed a specific reading, it should be possible to
make coherent sense of (21).
(21) Everyone in the class may have received an A. But not everybody
did.
In fact, (21) sounds contradictory, which they show is explained if only Some follow up on von Fintel & Iatri-
the non-specific reading is permitted by the grammar. They conjecture dou 2003 can be found in Tancredi 2008,
Swanson 2010
that this is a systematic property of epistemic modal operators (as opposed
to deontic and other types of modalities). Epistemic operators always have
widest scope in their sentence.
So there are really two challenges here for our current theory. We
need to account for the existence of non-specific readings, and also for
the absence, in at least some of our examples, of specific readings. We will
be concerned here exclusively with the first challenge and will set the sec-
80 E S C A P I N G O PAC I T Y

ond aside. We will aim, in effect, to set up the system so that all sentences
of this type are in principle ambiguous, hoping that additional constraints
that we are not investigating here will kick in to exclude the specific read-
ings where they are missing.
To complicate the empirical picture further, there are also examples
where raised subjects are unambiguously specific. Such cases have been
around in the syntactic literature for a while, and they have received re-
newed attention in the work of Lasnik and others. To illustrate just one
of the systematic restrictions, negative quantifiers like nobody seem to per-
mit only surface scope (i.e., wide scope) with respect to a modal verb or
adjective they have raised over.
(22) Nobody from New York is likely to win the lottery.

(22) does not have a non-specific reading parallel to the one for (19) above, For a thorough investigation of low scope
i.e., it cannot mean that it is likely that nobody from NY will win. It can readings of negative DPs, see Iatridou &
Sichel 2011.
only mean that there is nobody from NY who is likely to win. This too is
an issue that we set aside.
In the next couple of sections, all that we are trying to do is find and
justify a mechanism by which the grammar is able to generate both spe-
cific and non-specific readings for subjects that have raised over modal
operators. It is quite conceivable, of course, that the nature of the addi-
tional constraints which often exclude one reading or the other is ulti-
mately relevant to this discussion and that a better understanding of them
may undermine our conclusions. But this is something we must leave for
further research.

4.2.2 Syntactic “Reconstruction”

Given that the non-specific reading of (11) we are aiming to generate is
equivalent to the formula in (14), an obvious idea is that there is an LF
which is essentially the pre-movement structure of this sentence, i.e., the
structure prior to the raising of the subject above the operator. There are
a number of ways to make such an LF available.
One option, most defended in Sauerland & Elbourne (2002), is to as-
sume that the raising of the subject can happen in a part of the derivation
which only feeds PF, not LF. In that case, the subject simply stays in its
underlying VP-internal position throughout the derivation from DS to
LF. (Recall that quantifiers are interpretable there, as they generally are in
subject positions.)
Another option is a version of the so-called Copy Theory of move-
ment introduced in Chomsky (1993). This assumes that movement gen-
erally proceeds in two separate steps, rather than as a single complex op-
eration as we have assumed so far. Recall that in H&K, it was stipulated
that every movement effects the following four changes:
(i) a phrase 𝛼 is deleted,
 S P E C I F I C I T Y A N D T R A N S PA R E N C Y 81

(ii) an index i is attached to the resulting empty node (making it a so-

called trace, which the semantic rule for “Pronouns and Traces” rec-
ognizes as a variable),
(iii) a new copy of 𝛼 is created somewhere else in the tree (at the “land-
ing site”), and
(iv) the sister-constituent of this new copy gets another instance of the
index i adjoined to it (which the semantic rule of Predicate Abstrac-
tion recognizes as a binder index).
If we adopt the Copy Theory, we assume instead that there are three
distinct operations:
“Copy”: Create a new copy of 𝛼 somewhere in the tree, attach an index i
to the original 𝛼, and adjoin another instance of i to the sister of the
new copy of 𝛼. (= steps (ii), (iii), and (iv) above)
“Delete Lower Copy”: Delete the original 𝛼. (= step (i) above)
“Delete Upper Copy”: Delete the new copy of 𝛼 and both instances of i.
The Copy operation is part of every movement operation, and can hap-
pen anywhere in the syntactic derivation. The Delete operations happen
at the end of the LF derivation and at the end of the PF deletion. We have
a choice of applying either Delete Lower Copy or Delete Upper Copy
to each pair of copies, and we can make this choice independently at LF
and at PF. (E.g., we can do Copy in the common part of the derivation
and than Delete Lower Copy at LF and Delete Upper Copy at PF.) If
we always choose Delete Lower Copy at LF, this system generates ex-
actly the same structures and interpretations as the one from H&K. But if
we exercise the Delete Upper Copy option at LF, we are effectively un-
doing previous movements, and this gives us LFs with potentially new
interpretations. In the application we are interested in here, we would ap-
ply the Copy step of subject raising before the derivation branches, and
then choose Delete Lower Copy at PF but Delete Upper Copy at LF. The
LF will thus look as if the raising never happened, and it will straightfor-
wardly get the desired non-specific reading.
If the choice between the two Delete operations is generally optional,
we in principle predict ambiguity wherever there has been movement.
Notice, however, first, that the two structures will often be truth-conditionally
equivalent (e.g. when the moved phrase is a name), and second, that they
will not always be both interpretable. (E.g., if we chose Delete Upper
Copy after QRing a quantifier from object position, we’d get an uninter-
pretable structure, and so this option is automatically ruled out.) Even so,
we predict lots of ambiguity. Specifically, since raised subjects are always
interpretable in both their underlying and raised locations, we predict all
raising structures where a quantificational DP has raised over a modal op-
erator (or over negation or a temporal operator) to be ambiguous. As we
have already mentioned, this is not factually correct, and so there must
be various further constraints that somehow restrict the choices. (Similar
82 E S C A P I N G O PAC I T Y

comments apply, of course, to the option we mentioned first, of applying

raising only on the PF-branch.)
Yet another solution was first proposed by May (1977): May assumed
that QR could in principle apply in a “downward” fashion, i.e., it could
adjoin the moved phrase to a node that doesn’t contain its trace. Exercis-
ing this option with a raised subject would let us produce the following
structure, where the subject has first raised over the modal and then QRed
below it.
(23) t 𝑗 𝜆𝑖 [ must 𝑓 [ someone 𝜆 𝑗 [ t𝑖 have been here]]]

As it stands, this structure contains at least one free variable (the trace t 𝑗 )
and can therefore not possibly represent any actual reading of this sen-
tence. May further assumes that traces can in principle be deleted, when
their presence is not required for interpretability. This is not yet quite
enough, though to make (23) interpretable, at least not within our frame-
work of assumptions, for (24) is still not a candidate for an actual reading
of (11).
(24) 𝜆𝑖 [ must 𝑓 [ someone 𝜆 𝑗 [ t𝑖 have been here]]]

We would need to assume further that the topmost binder index could
be deleted along with the unbound trace, and also that the indices i and
j can be the same, so that the raising trace t 𝑗 is bound by the binding-
index created by QR. If these things can be properly worked out some-
how, then this is another way to generate the non-specific reading. No-
tice that the LF is not exactly the same as on the previous two approaches,
since the subject ends up in an adjoined position rather than in its origi-
nal argument position, but this difference is presumably without semantic
import.
What all of these approaches have in common is that they place the
burden of generating the non-specific reading for raised subjects on the
syntactic derivation. Somehow or other, they all wind up with structures
in which the subject is lower than it is on the surface and thereby falls
within the scope of the modal operator. They also have in common that
they take the modal operator (here the auxiliary, in other cases a main
predicate or an adverb) to be staying put. I.e., they assume that the non-
specific readings are not due to the modal operator being covertly higher
than it seems to be, but to the subject being lower. Approaches with these
features will be said to appeal to “syntactic reconstruction” of the subject.
This is a very broad notion of “reconstruction”, where basically any
mechanism which puts a phrase at LF in a location nearer to its under-
lying site than its surface site is called “reconstruction”. In some of the
literature, the term is used more narrowly. For example, May’s downward
QR is sometimes explicitly contrasted with genuine reconstruction, since
it places the quantifier somewhere else than exactly where it has moved
from.
 S P E C I F I C I T Y A N D T R A N S PA R E N C Y 83

4.2.3 Some Alternatives to Syntactic Reconstruction

Besides (some version of ) syntactic reconstruction, there are many other
ways in which one can try to generate non-specific readings for raised
subjects. Here are some other possibilities that have been suggested and/or
readily come to mind. We will see that some of them yield exactly the
non-specific reading as we have been describing it so far, whereas others
yield a reading that is very similar but not quite the same. We will con-
fine ourselves to analyses which involve no or only minor changes to our
system of syntactic and semantic assumptions. Obviously, if one departed
from these further, there would be even more different options, but even
so, there seem to be quite a few.
Let us look again at (11) and its apparent surface structure (12), both
repeated here:
(11) Somebody must have been here (since last night).

(12) somebody [ 2 [ [ must 𝑓 ] [ t2 have-been-here]]]

1. Raising the modal operator, variant 1: no trace

Conceivably, an LF for the non-specific reading of (11) might be de-

rived from the S-structure (=(12)) by covertly moving must (and its covert
R-argument) up above the subject. This would have to be a movement
which leaves no (semantically non-vacuous) trace. Given our inventory
of composition rules, the only type that the trace could have to make the
structure containing it interpretable would be the type of the moved op-
erator itself (i.e. ⟨𝑠𝑡, 𝑡⟩). If it had that type, however, the movement would
be semantically inconsequential, i.e., the structure would mean exactly the
same as (12). So this would not be a way to provide an LF for the non-
specific reading. If there was no trace left however (and also no binder
index introduced), we indeed would obtain the non-specific reading.

Exercise 4.4 Prove the claims we just made in the previous paragraph. Why is
no type for the trace other than ⟨𝑠𝑡, 𝑡⟩ possible? Why is the movement semanti-
cally inert when this type is chosen? How does the correct intended meaning arise
if there is no trace and binder index? □

2. Raising the modal operator, variant 2: trace of type s

[Requires slightly modified inventory of composition rules. Derives an in-

terpretation that is not quite the same as the non-specific opaque reading
we have assumed so far. Rather, it is the non-specific transparent “third”
reading discussed in the next chapter.]

3. Higher type for trace of raising, variant 1: type ⟨𝑒𝑡, 𝑡⟩

[Before reading this section, read and do the exercise on p.212/3 in H&K]
84 E S C A P I N G O PAC I T Y

So far in our discussion, we have taken for granted that the LF which
corresponds to the surface structure, viz. (12), gives us the specific read-
ing. This, however, is correct only on the tacit assumption that the trace
of raising is a variable of type e. If it is part of our general theory that all
variables, or at least all interpretable binder indices (hence all bound vari-
ables), in our LFs are of type e, then there is nothing more here to say.
But it is not prima facie obvious that we must or should make this general
assumption, and if we don’t, then the tree in (12) is not really one single
LF, but the common structure for many different ones, which differ in
the type chosen for the trace. Most of the infinitely many semantic types
we might assign to this trace will lead to uninterpretable structures, but
there turns out to be one other choice besides e that works, namely ⟨𝑒𝑡, 𝑡⟩:
(25) somebody 𝜆2,⟨𝑒𝑡,𝑡 ⟩ [ [ must 𝑓 ] [ 𝑡 2,⟨𝑒𝑡,𝑡 ⟩ have-been-here]]

(25) is interpretable in our system, but again, as in the previous approach, That a trace of type ⟨𝑒𝑡, 𝑡 ⟩ does not in
the predicted interpretation is not exactly the non-specific reading as we fact yield the targeted non-specific opaque
reading had not been noticed until we
have been describing it so far, but the non-specific transparent third read-
bothered to calculate the meaning of (25).
ing. For example, Fox 2000, which derives
from a dissertation supervised by us, is un-
Exercise 4.5 Using higher-type traces to “reverse” syntactic scope-relations is a aware of the fact that a high-type but ex-
trick which can be used quite generally. It is useful to look at a non-intensional tensional trace gives a scope-reconstructed
but transparent reading.
example as a first illustration. (26) contains a universal quantifier and a negation,
and it is scopally ambiguous between the readings in (a) and (b).
(26) Everything that glitters is not gold.
a. ∀𝑥 [𝑥 glitters → ¬𝑥 is gold] “surface scope”
b. ¬∀𝑥 [𝑥 glitters → 𝑥 is gold] “inverse scope”

We could derive the inverse scope reading for (26) by generating an LF (e.g. by
some version of syntactic reconstruction”) in which the every-DP is below not.
Interestingly, however, we can also derive this reading if the every-DP is in its
raised position above not but its trace has the type ⟨⟨𝑒, 𝑡⟩, 𝑡⟩.
Spell out this analysis. (I.e., draw the LF and show how the inverse-scope
interpretation is calculated by our semantic rules.) □

Exercise 4.6 Convince yourself that there are no other types for the raising trace
besides e and ⟨𝑒𝑡, 𝑡⟩ that would make the structure in (12) interpretable. (At least
not if we stick exactly to our current composition rules.) □

4. Higher type for trace of raising, variant 2: type ⟨𝑠, ⟨𝑒𝑡, 𝑡⟩⟩

If we want to get exactly the non-specific reading that results from syn-
tactic reconstruction out of a surface-like LF of the form (12), we must
use an even higher type for the raising trace, namely ⟨𝑠, ⟨⟨𝑒, 𝑡⟩, 𝑡⟩⟩, the
type of the intension of a quantifier. As you just proved in the exercise,
this is not possible if we stick to exactly the composition rules that we
 S P E C I F I C I T Y A N D T R A N S PA R E N C Y 85

have currently available. The problem is in the VP: the trace in subject
position is of type ⟨𝑠, ⟨⟨𝑒, 𝑡⟩, 𝑡⟩⟩ and its sister is of type ⟨𝑒, 𝑡⟩. These two
cannot combine by either FA or IFA, but it works if we employ another
variant of functional application.
(27) Extensionalizing Functional Application (EFA) Notice that the problem here is kind of
If 𝛼 is a branching node and {𝛽, 𝛾 } the set of its daughters, then, for the mirror image of the problem that
led to the introduction of “Intensional
any world 𝑤 and assignment 𝑔:
Functional Application” in H&K, ch.
if ⟦𝛽⟧𝑤,𝑔 (𝑤) is a function whose domain contains ⟦𝛾⟧𝑤,𝑔 , 12. There, we had a function looking
then ⟦𝛼⟧𝑤,𝑔 = ⟦𝛽⟧𝑤,𝑔 (𝑤)(⟦𝛾⟧𝑤,𝑔 ). for an argument of type ⟨𝑠, 𝑡 ⟩, but the
sister node had an extension of type t.
IFA allowed us to, in effect, construct an
Exercise 4.7 Calculate the truth-conditions of (12) under the assumption that argument with an added “s” in its type.
the trace of the subject quantifier is of type ⟨𝑠, ⟨⟨𝑒, 𝑡⟩, 𝑡⟩⟩. □ This time around, we have to get rid of
an “s” rather than adding one; and this is
what EFA accomplishes.
Can we choose between all these options? So we now have three different “func-
tional application”-type rules altogether
Two of the methods we tried derived readings in which the raised sub-
in our system: ordinary FA simply ap-
ject’s quantificational determiner took scope below the world-quantifier in plies ⟦𝛽⟧𝑤 to ⟦𝛾 ⟧𝑤 ; IFA applies ⟦𝛽⟧𝑤 to
′
the modal operator, but the raised subject’s restricting NP still was evalu- 𝜆𝑤 ′ .⟦𝛾 ⟧𝑤 ; and EFA applies ⟦𝛽⟧𝑤 (𝑤 )
𝑤
to ⟦𝛾 ⟧ . At most one of them will be ap-
ated in the utterance world (or the evaluation world for the larger sen-
plicable to each given branching node,
tence, whichever that may be), in other words: a non-specific but trans- depending on the type of ⟦𝛾 ⟧𝑤 .
parent interpretation. It is difficult to assess whether such readings are Think about the situation. Might there
actually available for the particular sentences under consideration, and we be other variant functional application
will postpone this question to the next chapter. We would like to argue rules?

here, however, that even if these readings are available, they cannot be
the only readings that are available for raised subjects besides their wide-
scope readings. In other words, even if we allowed one of the mecha-
nisms that generated these sort of hybrid readings, we would still need
another mechanism that gives us, for at least some examples, the “real”
non-specific opaque readings that we obtain e.g. by syntactic reconstruc-
tion. The relevant examples that show this most clearly involve DPs with
more descriptive content than somebody and whose NPs express clearly
contingent properties.
(28) A neat-freak must have been here.

If I say this instead of our original (11) when I come to my office in the
morning and interpret the clues on my desk, I am saying that every world
compatible with the evidence is such that someone who is a neat-freak
in that world was here in that world. Suppose there is a guy, Bill, whom I
know slightly but not well enough to have an opinion on whether or not
he is neat. He may or not be, for all I know. So there are worlds among
the relevant worlds where he is a neat-freak and worlds where he is not.
I also don’t have an opinion on whether he was or wasn’t the one who
came into my office last night. He did in some of the relevant worlds
and he didn’t in others. I am implying with (28), however, that if Bill
isn’t a neat-freak, then it wasn’t him in my office. I.e., (28) is telling you
86 E S C A P I N G O PAC I T Y

that, even if there are relevant worlds in which Bill is a slob and worlds
in which (only) he was in my office, there aren’t any relevant worlds in
which Bill is a slob and the only person who was in my office. This is cor-
rectly predicted if (28) expresses the “genuine” non-specific reading in
(29), but not if it expresses the “hybrid” reading in (30).
(29) ∀𝑤 ′ [𝑤 ′ is compatible with the evidence in 𝑤 0 →
∃𝑥 [𝑥 is a neatfreak in 𝑤 ′ and 𝑥 was here in 𝑤 ′ ]]

(30) ∀𝑤 ′ [𝑤 ′ is compatible with the evidence in 𝑤 0 →

∃𝑥 [𝑥 is a neatfreak in 𝑤 0 and 𝑥 was here in 𝑤 ′ ]]

We therefore conclude the mechanisms 2 and 3 considered above (what-

ever their merits otherwise) cannot supplant syntactic reconstruction or
some other mechanism that yields readings like (29).
This leaves only the first and fourth options that we looked at as po-
tential competitors to syntactic reconstruction, and we will focus the rest
of the discussion on how we might be able to tease apart the predictions
that these mechanisms imply from the ones of a syntactic reconstruction
approach.
As for moving the modal operator, there are no direct bad predictions
that we are aware of with this. But it leads us to expect that we might find
not only scope ambiguities involving a modal operator and a DP, but also
scope ambiguities between two modal operators, since one of them might
covertly move over the other. It seems that this never happens. Sentences
with stacked modal verbs seem to be unambiguous and show only those
readings where the scopes of the operators reflect their surface hierarchy.
(31) a. I have to be allowed to graduate.
b. #I am allowed to have to graduate.

Of course, this might be explained by appropriate constraints on the move- See Lechner 2007 for an early discus-
ment of modal operators, and such constraints may even come for free in sion of semantic effects of head move-
ment. See McCloskey 2016 for a recent
the right syntactic theory. Also, we should have a much more compre-
re-assessment.
hensive investigation of the empirical facts before we reach any verdict.
If it is true, however, that modal operators only engage in scope interac-
tion with DPs and never with each other, then a theory which does not
allow any movement of modals at all could claim the advantage of having
a simple and principled explanation for this fact.
What about the “semantic reconstruction” option, where raised sub-
jects can leave traces of type ⟨𝑠, ⟨𝑒𝑡, 𝑡⟩⟩ and thus get narrow scope seman-
tically without ending up low syntactically? This type of approach has
been explored quite thoroughly and defended with great sophistication.
The main consideration against semantic reconstruction and in favor of
syntactic reconstruction comes from binding theoretic concerns. We give
some crucial examples from Fox 2000 here.
 S P E C I F I C I T Y A N D T R A N S PA R E N C Y 87

Scope reconstruction and Condition C.

Consider:
(32) a. [A first year student] seems to David t to be at the party.
b. [Someone from NY] is very likely t to win the lottery.

These examples are a reminder that we can get non-specific readings of

raised subjects. Fox argues that Binding Theory Condition C can be used
to distinguish syntactic and semantic reconstruction. Assuming that Con-
dition C applies at LF, syntactic reconstruction (where the full DP is in-
terpreted in a lower position) predicts that an R-expression in that DP
should not be allowed if there’s a higher coreferential pronoun. The se-
mantic reconstruction account has no simple way of making the same
prediction.
Here are the crucial tests:
(33) a. [A student of his1 ] seems to David1 to be at the party.
𝑂𝐾 specific, 𝑂𝐾 non-specific

b. [A student of David’s1 ] seems to him1 to be at the party.

𝑂𝐾 specific, *non-specific

(34) a. [Someone from his1 city] seems to David1 t to be likely to win

the lottery. 𝑂𝐾 specific, 𝑂𝐾 non-specific

b. [Someone from David1 ’s city] seems to him1 t to be likely to

win the lottery. 𝑂𝐾 specific, *non-specific

Fox claims that the (b) cases do not have a non-specific reading. If syntac-
tic reconstruction is the mechanism that gives us non-specific readings of
A-moved subjects, the explanation is straightforward.

Exercise 4.8 Use the following observations to bolster Fox’s argument.

(35) a. The cat seems to be out of the bag.
b. Advantage might have been taken of them.

(36) a. For these issues to be clarified, many new papers about his1 philoso-
phy seem to Quine t to be needed.
b. #For these issues to be clarified, many new papers about Quine’s1
philosophy seem to him t to be needed. □

If Fox’s argument is correct, then (some form of ) syntactic reconstruction

is the mechanism by which we get non-specific readings of A-movement
subjects. This means that syntactic reconstruction is possible, but it also
means that semantic reconstruction must not be available. Otherwise,
there would be a way of getting non-specific readings that would not be
sensitive to Condition C. So, the question arises why semantic reconstruc-
88 E S C A P I N G O PAC I T Y

tion is unavailable. Fox (2000: p. 171, fn. 41) discusses two ways of ruling
out the high type traces that would give rise to semantic reconstruction:

(i) “traces, like pronouns, are always interpreted as variables that range
over individuals (type 𝑒)”,
(ii) “the semantic type of a trace is determined to be the lowest type
compatible with the syntactic environment (as suggested in Beck
1996)”.

We will return to this issue in a later chapter when we can raise it again
in a slightly different framework.

4.3 Further readings

In addition to the references about reconstruction in the text, you might
want to look at Boeckx 2001 and Baltin 2010.
5 The Third Reading

5.1 A Problem . . . . . . . . . . . . . . . . . . . . . . . . 89
5.2 The Standard Solution: Overt World Variables . . . . 91
5.3 The third reading with conditionals and modals . . . 96
5.4 Binding Theory for World Variables . . . . . . . . . 97
5.5 Excursus: Semantic reconstruction revisited . . . . . . 98
5.6 Further reading . . . . . . . . . . . . . . . . . . . . . 99

5.1 A Problem

Janet Dean Fodor discussed examples like (1) in her dissertation (1970).
(1) Mary wanted to buy a hat just like mine.

Fodor observes that (1) has three readings, which she labels “specific trans-
parent”, “non-specific transparent”, and “non-specific opaque.”

(i) On the “specific transparent” reading, the sentence says that there
is a particular hat which is just like mine such that Mary has a de-
sire to buy it. Say, I am walking along Newbury Street with Mary.
Mary sees a hat in a display window and wants to buy it. She tells
me so. I don’t reveal that I have one just like it. But later I tell you
by uttering (1).
(ii) On the “non-specific opaque” reading, the sentence says that Mary’s
desire was to buy some hat or other which fulfills the description
that it is just like mine. She is a copycat.
(iii) On the “non-specific transparent” reading, finally, the sentence will
be true, e.g., in the following situation: Mary’s desire is to buy some
hat or other, and the only important thing is that it be a Red Sox
cap. Unbeknownst to her, my hat is one of those as well.

The existence of three different readings appears to be problematic for

the scopal account of specific/non-specific ambiguities that we have been
assuming. It seems that our analysis allows just two semantically distinct
types of LFs: Either the DP a hat just like mine takes scope below want, as
in (2), or it takes scope above want, as in (3).
90 E S C A P I N G O PAC I T Y

(2) Mary wanted [ [a hat-just-like-mine]1 [ PRO to buy 𝑡 1 ]]

(3) [a hat-just-like-mine]1 [ Mary wanted [ PRO to buy 𝑡 1 ]]

In the system we have developed so far, (2) says that in every world 𝑤 ′
in which Mary gets what she wants, there is something that she buys in
𝑤 ′ that’s a hat in 𝑤 ′ and like my hat in 𝑤 ′ . This is Fodor’s “non-specific
opaque” reading. (3), on the other hand, says that there is some thing 𝑥
which is a hat in the actual world and like my hat in the actual world, and
Mary buys 𝑥 in every one of her desire worlds. That is Fodor’s “specific
transparent.” But what about the “non-specific transparent” reading? To
obtain this reading, it seems that we would have to evaluate the predicate
hat just like mine in the actual world, so as to obtain its actual extension (in
the scenario we have sketched, the set of all Red Sox caps). But the exis-
tential quantifier expressed by the indefinite article in the hat-DP should
not take scope over the modal operator want, but below it, so that we can
account for the fact that in different desire-worlds of Mary’s, she buys
possibly different hats.
There is a tension here: one aspect of the truth-conditions of this read-
ing suggests that the DP a hat just like mine should be outside of the scope
of want, but another aspect of these truth-conditions compels us to place it
inside the scope of want. We can’t have it both ways, it would seem, which
is why this has been called a “scope paradox”
Another example of this sort, due to Bäuerle 1983, is (4):
(4) Georg believes that a woman from Stuttgart loves every member of
the Vf B team.
Bäuerle describes the following scenario: Georg has seen a group of men
on the bus. This group happens to be the Vf B team (Stuttgart’s soccer
team), but Georg does not know this. Georg also believes (Bäuerle doesn’t
spell out on what grounds) that there is some woman from Stuttgart who
loves every one of these men. There is no particular woman of whom he
believes that, so there are different such women in his different belief-
worlds. Bäuerle notes that (4) can be understood as true in this scenario.
But there is a problem in finding an appropriate LF that will predict its
truth here. First, since there are different women in different belief-worlds
of Georg’s, the existential quantifier a woman from Stuttgart must be inside
the scope of believe. Second, since (in each belief world) there aren’t dif-
ferent women that love each of the men, but one that loves them all, the
a-DP should take scope over the every-DP. If the every-DP is in the scope
of the a-DP, and the a-DP is in the scope of believe, then it follows that
the every-DP is in the scope of believe. But on the other hand, if we want
to capture the fact that the men in question need not be Vf B-members in
Georg’s belief-worlds, the predicate member of the VfB team needs to be
outside of the scope of believe. Again, we have a “scope paradox”.
 THE THI RD READING 91

Before we turn to possible solutions for this problem, let’s have one
more example:
(5) Mary hopes that a friend of mine will win the race.

This again seems to have three readings. In Fodor’s terminology, the DP

a friend of mine can be “non-specific opaque,” in which case (5) is true iff
in every world where Mary’s hopes come true, there is somebody who
is my friend and wins. It can also have a “specific transparent” reading:
Mary wants John to win, she doesn’t know John is my friend, but I can
still report her hope as in (5). But there is a third option, the “non-specific
transparent” reading. To bring out this rather exotic reading, imagine
this: Mary looks at the ten contestants and says I hope one of the three on the
right wins - they are so shaggy - I like shaggy people. She doesn’t know that
those are my friends. But I could still report her hope as in (5).

5.2 The Standard Solution: Overt World Variables

The scope paradoxes we have encountered can be traced back to a basic

design feature of our system of intensional semantics: the relevant “eval-
uation world” for each predicate in a sentence is strictly determined by
its LF-position. All predicates that occur in the (immediate) scope of the
same modal operator must be evaluated in the same possible worlds. E.g.
if the scope of want consists of the clause a friend of mine (to) win, then ev-
ery desire-world 𝑤 ′ will be required to contain an individual that wins
in 𝑤 ′ and is also my friend in 𝑤 ′ . If we want to quantify over individuals
that are my friends in the actual world (and not necessarily in all the sub-
ject’s desire worlds), we have no choice but to place friend of mine outside
of the scope of want. And if we want to accomplish this by means of QR,
we must move the entire DP a friend of mine.
Not every kind of intensional semantics constrains our options in this
way. One way to visualize what we might want is to write down an LF
that looks promising:
(6) Mary wanted𝑤0 [𝜆𝑤 ′ [ a hat-just-like-mine𝑤0 ] 𝜆𝑥 1 [ PRO to buy𝑤 ′ 𝑥 1 ]]

We have annotated each predicate with the world in which we wish to

evaluate it. 𝑤 0 is the evaluation world for the entire sentence and it is the
world in which we evaluate the predicates want and hat-just-like-mine.
The embedded sentence contributes a function from worlds to truth-
values and we insert an explicit 𝜆-operator binding the world where the
predicate buy is evaluated. The crucial aspect of (6) is that the world in
which hat-just-like-mine is evaluated is the matrix evaluation world and
not the same world in which its clause-mate predicate buy is evaluated.
This LF thus looks like it might faithfully capture Fodor’s third reading.
Logical forms with overt world variables such as (6) are in fact the
standard solution to the problem presented by the third reading. Let us
92 E S C A P I N G O PAC I T Y

spell out some of the technicalities. Later, we will consider a couple of al-
ternatives.
We return to the basic system used in Heim & Kratzer up to chapter
11. The interpretation function is relativized only to an assignment func-
tion, not to any other evaluation parameters such as a world, a time, or an
index. The semantic rules are Functional Application, Predicate Abstrac-
tion, and Predicate Modification, in their formulations from the earlier
part of H&K. There is no rule of Intensional Functional Application. The
only ingredient of intensional semantics that we do retain is the expanded
type system and ontology. We have a third basic type besides 𝑒 and 𝑡, the
type 𝑠. 𝐷𝑠 is the set of all indices, for now possible worlds (later: world-
time pairs).
There are a number of innovations in the lexicon and in the syntax.
As for the lexicon, the main change concerns the treatment of predicates
(verbs, nouns, adjectives). They now all get an additional argument, of
type 𝑠. The decision to make the world-argument
the predicate’s first (lowest) argument
(7) a. ⟦smart⟧ = 𝜆𝑤 ∈ 𝐷𝑠 . 𝜆𝑥 ∈ 𝐷𝑒 . 𝑥 is smart in 𝑤 is arbitrary, and nothing hinges on it.
b. ⟦likes⟧ = 𝜆𝑤 ∈ 𝐷𝑠 . 𝜆𝑥 ∈ 𝐷𝑒 . 𝜆𝑦 ∈ 𝐷𝑒 . 𝑦 likes 𝑥 in 𝑤 For all we know, it could be the highest
argument, or somewhere in between.
c. ⟦teacher⟧ = 𝜆𝑤 ∈ 𝐷𝑠 . 𝜆𝑥 ∈ 𝐷𝑒 . 𝑥 is a teacher in 𝑤
d. ⟦friend⟧ = 𝜆𝑤 ∈ 𝐷𝑠 . 𝜆𝑥 ∈ 𝐷𝑒 . 𝜆𝑦 ∈ 𝐷𝑒 . 𝑦 is 𝑥’s friend in 𝑤

This also applies to attitude predicates, modals, and tenses. We illustrate

with believe and must:
(8) a. ⟦believe⟧ = 𝜆𝑤 ∈ 𝐷𝑠 . 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . 𝜆𝑥 ∈ 𝐷.
∀𝑤 ′ [𝑤 ′ conforms to what 𝑥 believes in 𝑤 → 𝑝 (𝑤 ′ ) = 1]
b. ⟦must⟧ = 𝜆𝑤 ∈ 𝐷𝑠 . 𝜆𝑅 ∈ 𝐷 ⟨𝑠,𝑠𝑡 ⟩ . 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ .
∀𝑤 ′ [𝑅(𝑤)(𝑤 ′ ) = 1 → 𝑝 (𝑤 ′ ) = 1]

Note that predicates (ordinary ones and modal ones), like the ones in (7)
and (8) now have as their semantic values what used to be their intensions.
There is no change to the entries of proper names, determiners, or
truth-functional connectives; these keep their purely extensional (“s-free”)
types and meanings:
(9) a. ⟦Ann⟧ =Ann
b. ⟦and⟧ = 𝜆𝑢 ∈ 𝐷𝑡 . [𝜆𝑣 ∈ 𝐷𝑡 . 𝑢 = 𝑣 = 1]
c. ⟦the⟧ = 𝜆𝑓 ∈ 𝐷 ⟨𝑒,𝑡 ⟩ : ∃!𝑥 . 𝑓 (𝑥) = 1. the 𝑦 such that 𝑓 (𝑦) = 1.
d. ⟦every⟧ = 𝜆𝑓 ∈ 𝐷 ⟨𝑒,𝑡 ⟩ . 𝜆𝑔 ∈ 𝐷 ⟨𝑒,𝑡 ⟩ . ∀𝑥 [𝑓 (𝑥) = 1 → 𝑔(𝑥) = 1]

So, let’s start analyzing a simple sentence.

 THE THI RD READING 93

(10) [𝑉 𝑃 John leaves]

The verb’s type is ⟨𝑠, 𝑒𝑡⟩, so it’s looking for a sister node which denotes a
world. John, which denotes an individual, is not a suitable argument.
We get out of this problem by adding a couple of items to our lexicon,
which are abstract (unpronounced) morphemes. One is a series of pro-
nouns of type 𝑠 (“index pronouns” or, for now, “world pronouns”). In this
chapter, we will write them as 𝑤𝑛 , with a numerical subscript 𝑛, or even
as 𝑤, 𝑤 ′, 𝑤 ′′ . (Later, we sometimes might write them as pro𝑛 and rely on
context to make clear we are not referring to an individual.) Their se-
mantics is what you expect: they get values from the assignment function.
94 E S C A P I N G O PAC I T Y

We will stipulate that a complete (matrix) sentence must not contain In the 2016 edition of this class, Suzana
any free variables of type 𝑠 and must receive a denotation of type ⟨𝑠, 𝑡⟩. Fong noted that this stipulation is prima
facie less appealing than the alternative
This means that we need binders of world pronouns. Many proposals
assumption that type-𝑠 pronouns are ex-
in this line of thought help themselves to freely inserted covert binders. actly like type-𝑒 pronoun in every respect,
We will follow H&K in not doing that. Instead we posit one more lexical including the ability to remain free and
item, analogous to the covert vacuous operator PRO of type 𝑒 in H&K get values from a contextually supplied
assignment. Irene tried to sketch some
(pp.227-228): a semantically vacuous operator, OP, which moves and principled reason why it might not be
leaves a trace of type 𝑠. Its syntactic properties are such that it must end possible to refer to a specific world other
up in C or right below a functional head in the “clausal spine” between C than the world one is in. But as Mitya
Privoznov pointed out, a similar idea is not
and V, and it must get there by a very short movement, a kind of “head
plausible for times, given the existence of
movement”. We are leaving this rather vague. temporal deictics like then. So at best there
So, our sentence John leaves contains OP, generated as the first sister of might be a principled reason why the
the verb and then moved to the “top” of the sentence: world-coordinate of a free index-pronoun
would always have to be 𝑤𝑢 . Irene had
(11) OP 1 [ John [ leaves 𝑡 1 ]] to concede therefore that the ban against
free index-pronouns was just a stipulation.
Our system generates the following denotation for (11): “𝜆𝑤𝑠 . John leaves We want to think more about (a) whether
in 𝑤”, a proposition. We rewrite the definition of truth/falsity of an utter- we really need it, and (b) if we do, what
ance as follows: might explain it.

(12) An utterance of a sentence (=LF) 𝜙 in world 𝑤 is true iff ⟦𝜙⟧(w) =

1.
So, if we utter our sentence in this world (call it 𝑤 @ ), then the utterance
was true iff John leaves in 𝑤 @ .
Now, we have to look at more complex sentences. First, a simple case
of embedding. The sentence is John wants to leave, which now has an LF
like this:
(13) [ OP 1 [ John [ wants 𝑡 1 [ OP 2 [ PRO(= John) [ leave 𝑡 2 ]]]]]]

Exercise 5.1 Calculate the semantic value of (13). □

Next, look at an example involving a complex subject, such as the teacher

left. The verb will need a world argument as before. The noun teacher will
likewise need one, so that the can get the required argument of type ⟨𝑒, 𝑡⟩
(not ⟨𝑠, 𝑒𝑡⟩!). Now, our system makes an interesting prediction: one of
the world arguments has to be OP and one of them has to be a pronoun.
(Why?) We have free choice, it appears, as to which predicate gets which
kind of world argument. Let’s assume, for now, that we insert OP as the
sister of the verb and a world pronoun 𝑤 ? as the sister of the noun. Since
we have stipulated that a complete sentence cannot contain any free world
pronouns, the operator and the pronoun have to be co-indexed. So, after
OP-movement, we will have this LF:
(14) OP 1 [ [ The [ teacher 𝑤 1 ]] [ left 𝑡 1 ]]

This will denote the correct proposition (true of a world 𝑤 iff the unique
individual who is a teacher in 𝑤 left in 𝑤).
 THE THI RD READING 95

Now comes the payoff. Consider what happens when the sentence
contains both a modal operator and a complex DP in its complement.
(15) Mary wants a friend of mine to win.

There are now three predicates that need world arguments. Furthermore,
since want needs a proposition as its second argument (after its world ar-
gument), there needs to be an OP on top of the embedded clause. There
also needs to be an OP on top of the matrix clause. As before, a friend of
mine can stay in the embedded clause or QR into the matrix clause. When
it moves into the matrix clause, the only way to not leave its world argu-
ment free is to co-index it with the matrix OP. But when it stays below,
we can choose to co-index it with either OP, which is how we gener-
ate two non-specific readings, one opaque and one transparent. Here are
the three LFs (to make the structures more readable, we leave off most of
the bracketing and start writing the world arguments as subscripts to the
predicates): So, instead of writing 𝑡 1 for a trace of type
𝑠 that serves as the first argument of leaves,
(16) a. non-specific opaque: say, we write “leaves𝑤1 ”
OP 1 Mary wants𝑤1 [ OP 2 a friend-of-mine𝑤2 leave𝑤2 ]
b. specific transparent:
OP 1 a friend-of-mine𝑤1 3 Mary wants𝑤1 [ OP 2 𝑡 3 leave𝑤2 ]
c. non-specific transparent:
OP 1 Mary wants𝑤1 [ OP 2 a friend-of-mine𝑤1 leave𝑤2 ]

Notice that the third reading is minimally different from the first reading:
all that happened is the choice to co-index the world argument of friend of
mine with the matrix OP.
In this new framework, then, we have a way of resolving the apparent
“scope paradoxes” and of acknowledging Fodor’s point that there are two
separate distinctions to be made when DPs interact with modal operators.
First, there is the scopal relation between the DP and the operator; the
DP may take wider scope (Fodor’s “specific” reading) or narrower scope
(“non-specific” reading) than the operator. Second, there is the choice of
binder for the world-argument of the DP’s restricting predicate; this may
be cobound with the world-argument of the embedded predicate (Fodor
“opaque”) or with the modal operator’s own world-argument (“transpar-
ent”). So the transparent/opaque distinction in the sense of Fodor is not
per se a distinction of scope; but it has a principled connection with scope
in one direction: Unless the DP is within the modal operator’s scope, the
opaque option (= co-binding the world-pronoun with the embedded
predicate’s world-argument) is in principle unavailable. (Hence “specific”
implies “transparent”, and “opaque” implies “non-specific”.) But there is
no implication in the other direction: if the DP has narrow scope w.r.t.
to the modal operator, either the local or the long-distance binding op-
96 E S C A P I N G O PAC I T Y

tion for its world-pronoun is in principle available. Hence “non-specific”

readings may be either “transparent” or “opaque”.

Exercise 5.2 For DPs with extensions of type 𝑒 (specifically, DPs headed by
the definite article), there is a truth-conditionally manifest transparent/opaque
distinction, but no truth-conditionally detectable specific/non-specific distinction.
In other words, if we construct LFs analogous to (16)[a-c] above for an example
with a definite DP, we can always prove that the first option (wide scope DP)
and the third option (narrow scope DP with distantly bound world-pronoun)
denote identical propositions. In this exercise, you are asked to show this for the
example in (17).
(17) John believes that your abstract will be accepted. □

5.3 The third reading with conditionals and modals

So, far our examples of the third reading have all been with attitude pred-
icates but the phenomenon can also be observed in conditionals and with
modals. A famous example is due to Abusch 1994:
(18) Things would be different if every senator had grown up to be a
rancher instead.
What makes conditionals different is that the if -clause is a scope island
for quantifiers so that every senator cannot QR scope out of the if -clause
in (18). But the question of whether its predicate, senator, is interpreted in
the matrix evaluation world (“transparent”) or in the worlds that if takes
us to (“opaque”) remains open. Abusch’s example is constructed to heavily
favor the transparent reading.
Percus 2000 provides a clever minimal pair that shows the expected
ambiguity:
(19) a. If every semanticist owned a villa in Tuscany, there would be
no field at all.
b. If I were a syntactician and if every semanticist owned a villa in
Tuscany, I would be quite envious.

We can also see the ambiguities at work in modal sentences:

(20) a. It could have been that everyone inside was outside.
b. Everyone inside is permitted to be outside.

These latter examples are from Yalcin 2015, who proceeds to discuss the
very puzzling fact that transparent readings do not seem to be available
in certain “epistemic” contexts, neither with indicative conditionals nor
modals.
 THE THI RD READING 97

5.4 Binding Theory for World Variables

One could in principle imagine some indexings of our LFs that we have
not considered so far. In a system (unlike ours) where one freely inserts
“𝜆𝑤” operators on top of every clause, one could generate the following
LF, which indexes the predicate of the complement clause to the matrix
𝜆-operator rather than to the one on top of its own clause.
(21) 𝜆𝑤 0 John wants𝑤0 [𝜆𝑤 1 PRO leave𝑤0 ]

Of course, the resulting semantics would be pathological: what John

would be claimed to stand in the wanting relation to is a set of worlds that
is either the entire set 𝑊 of possible worlds (if the evaluation world is one
in which John leaves) or the empty set (if the evaluation world is one in
which John doesn’t leave). Clearly, the sentence has no such meaning.
Would we need to restrict our system to not generate such an LF? Perhaps
not, if the meaning is so absurd that the LF would be filtered out by some
overarching rules distinguishing sense from nonsense. Nevertheless, it is
gratifying to note that this kind of LF is unavailable in our system: the
only place where the lower world-binder OP could originate is as a sis-
ter to leave and moving it to the lower CP (where we need a proposition
to feed to want) would by necessity make it bind the world argument of
leave.
But there are real problems when we look at more complex examples.
Here is one discussed by Percus in important work (Percus 2000):
(22) Mary thinks that my brother is Canadian.

Since the subject of the lower clause is a type 𝑒 expression, we expect at

least two readings: opaque and transparent, cf. Exercise 5.2. The two LFs
are as follows:
(23) a. opaque
OP 0 Mary thinks𝑤0 [ (that) OP 1 my brother𝑤1 (is) Canadian𝑤1 ]
b. transparent
OP 0 Mary thinks𝑤0 [ (that) OP 1 my brother𝑤0 (is) Canadian𝑤1 ]

But as Percus points out, there is another indexing that might be gen-
erated:
(24) OP 0 Mary thinks𝑤0 [ (that) OP 1 my brother𝑤1 (is) Canadian𝑤0 ]

In (24), we have co-indexed the main predicate of the lower clause with
the matrix 𝜆-operator and co-indexed the nominal predicate brother with
the embedded 𝜆-operator. That is, in comparison with the transparent
reading in (23b), we have just switched around the indices on the two
predicates in the lower clause.
Note that this LF will not lead to a pathological reading. So, is the pre-
dicted reading one that the sentence actually has? No. For the transpar-
98 E S C A P I N G O PAC I T Y

ent reading, we can easily convince ourselves that the sentence does have
that reading. Here is Percus’ scenario: “My brother’s name is Allon. Sup-
pose Mary thinks Allon is not my brother but she also thinks that Allon
is Canadian.” In such a scenario, our sentence can be judged as true, as
predicted if it can have the LF in (23b). But when we try to find evidence
that (24) is a possible LF for our sentence, we fail. Here is Percus:

If the sentence permitted a structure with this indexing, we

would take the sentence to be true whenever there is some
actual Canadian who Mary thinks is my brother — even when
this person is not my brother in actuality, and even when Mary
mistakenly thinks that he is not Canadian. For instance, we would
take the sentence to be true when Mary thinks that Pierre
(the Canadian) is my brother and naturally concludes — since
she knows that I am American — that Pierre too is American.
But in fact we judge the sentence to be false on this scenario,
and so there must be something that makes the indexing in
(24) impossible.

Percus then proposes the following descriptive generalization:

(25) GENERALIZATION X: The situation pronoun that a verb selects for
must be coindexed with the nearest 𝜆 above it. Percus works with situation pronouns
rather than world pronouns, an immaterial
We expect that there will need to be a lot of work done to understand the difference for our purposes here.
deeper sources of this generalization. But note that we could implement
the constraint in our system by brute force: the OP operator can only be
generated as the sister of a main predicate, not as the sister of a predicate
inside an argument nominal.

5.5 Excursus: Semantic reconstruction revisited

Let us look back at the account of non-specific readings of raised subjects

that we sketched earlier in Section 4.2.3. We showed that you can derive
such readings by positing a high type trace for the subject raising, a trace
of type ⟨𝑠, ⟨𝑒𝑡, 𝑡⟩⟩. Before the lower predicate can combine with the trace,
the semantic value of the trace has to be extensionalized by being applied
to the lower evaluation world (done via the EFA composition principle).
Upstairs the raised subject has to be combined with the 𝜆-abstract (which
will be of type ⟨⟨𝑠, ⟨𝑒𝑡, 𝑡⟩⟩, 𝑡⟩) via its intension.
We then saw data suggesting that syntactic reconstruction is actually
what is going on. This, of course, raises the question of why semantic
reconstruction is unavailable (otherwise we wouldn’t expect the data that
we observed).
In this excursus, we will briefly consider whether our new framework
has something to say about this issue. Let’s figure out what we would have
 THE THI RD READING 99

to do in the new framework to replicate the account in the section on

semantics reconstruction.
Downstairs, we would have a trace of type ⟨𝑠, ⟨𝑒𝑡, 𝑡⟩⟩. To calculate its
extension, we do not need recourse to a special composition principle,
but can simply give it a world-argument (co-indexed with the abstractor
resulting from the movement of the 𝑤-OP in the argument position of
the lower verb).
Now, what has to happen upstairs? Well, there we need the subject
to be of type ⟨𝑠, ⟨𝑒𝑡, 𝑡⟩⟩, the same type as the trace, to make sure that its
semantics will enter the truth-conditions downstairs. But how can we do
this?
We need the DP somebody from New York to have as its semantic value
an intension, the function from any world to the existential quantifier
over individuals who are people from New York in that world. This is
actually hard to do in our system. It would be possible if (i) the predicate(s)
inside the DP received 𝑤-PRO as their argument, and if (ii) that 𝑤-PRO
were allowed to moved to adjoin to the DP. If we manage to rule out at
least one of the two preconditions on principled grounds, we would have
derived the impossibility of semantic reconstruction as a way of getting
non-specific readings of raised subjects.

(i) may be ruled out by the Binding Theory for world pronominals,
when it gets developed.
(ii) may be ruled out by principled considerations as well. Perhaps,
world-abstractors are only allowed at sentential boundaries.

See Larson 2002 for some discussion of recalcitrant cases, one of which
is the object position of so-called intensional transitive verbs, a topic for
another occasion.

5.6 Further reading

There is an interesting literature spawned by Percus 2000: Schwager
2009, Romoli & Sudo 2009, Keshet 2010, Keshet 2011, Schwarz 2012,
Keshet & Schwarz 2014.
Fodor also discussed a fourth reading, specific opaque, which is hard to
fit into our framework. Whether it really exists is a question discussed in
some recent work: Szabó 2010 and Francez 2017.
P A R T III

Time
6 Beginnings of tense and aspect

6.1 A first proposal for tense . . . . . . . . . . . . . . . . 103

6.2 Time frame adverbials . . . . . . . . . . . . . . . . . 105
6.3 Are tenses referential? . . . . . . . . . . . . . . . . . . 108
6.4 Referential tense and perfective/imperfective aspect . 110
6.4.1 Referential tense after all . . . . . . . . . . . 110
6.4.2 Event semantics and perfective aspect . . . . 112
6.4.3 The English progressive . . . . . . . . . . . . 113
6.4.4 Stativity effects . . . . . . . . . . . . . . . . . 116
6.5 Formalizing the referential analysis . . . . . . . . . . 118

6.1 A first proposal for tense

Tense logic, or temporal logic, is a branch of logic first developed by the This chapter is even more the outcome of
aptly named Arthur Prior in a series of works, in which he proposed collaborative efforts than other chapters.
We are very much indebted to Roger
treating tense in a way that is formally quite parallel to the treatment of
Schwarzschild, who has used our notes
modality discussed in Chapter 3. Since tense logic (and modal logic) typi- several times in his teaching and is the
cally is formulated at a high level of abstraction regarding the structure of source of many edits and additions.
sentences, it doesn’t concern itself with the internal make-up of “atomic”
sentences and thus treats tenses as sentential operators (again, in parallel
to the way modal operators are typically treated in modal logic). We will
begin by integrating a version of Prior’s tense logic into our framework.
The first step is to switch to a version of our intensional semantic sys- We remain vague for now about what we
tem where instead of a world parameter, the evaluation function is sen- mean by “times” (points in time? time in-
tervals?). This will soon need clarification,
sitive to a parameter that is a pair of a world and a time. Such a pair will
and we will decide that we should mean
also be called an “index”. We use metalanguage variables 𝑖, 𝑖 ′ , … for in- “intervals”.
dices, and write 𝑤𝑖 and 𝑡𝑖 to pick out the world in 𝑖 and the time in 𝑖 re-
spectively (i.e., 𝑖 = ⟨𝑤𝑖 , 𝑡𝑖 ⟩). Predicates will now have lexical entries that
incorporate their sensitivity to both worlds and times:
(1) ⟦tired⟧𝑖 = 𝜆𝑥 ∈ 𝐷. 𝑥 is tired in 𝑤𝑖 at 𝑡𝑖

The composition principles from Heim & Kratzer and the preced-
ing chapters stay the same, except that type 𝑠 is now the type of indices,
and intensions are functions from indices to extensions. For example, the
intension of sentence is now a function from world-time pairs to truth-
104 T I M E

values. We might call this a “temporal proposition”, to distinguish it from This necessitates a slight rewriting of our
a function from just worlds to truth-values, but we will often just call it a previous entries for modals and attitude
verbs. We will attend to this when we get
“proposition”.
to relevant examples later on.
In this framework, we can formulate a very simple-minded first analy-
sis of the present and past tenses and the future auxiliary will. As for (LF)
syntax let’s assume that complete sentences are TPs, headed by T (for
“tense”). There are two morphemes of the functional category T, namely
PAST (past tense) and PRES (present tense). The complement of T is an MP
or a VP. MP is headed by M (for “modal”). Morphemes of the category
M include the modal auxiliaries must, can, etc., which we talked about in
previous chapters, the semantically vacuous do (in so-called “do-support”
structures), and the future auxiliary will. Evidently, this is a semantically
heterogeneous category, grouped together solely because of their com-
mon syntax (they are all in complementary distribution with each other).
The complement of M is a VP. When the sentence contains none of the
items in the category M, we assume that MP isn’t projected at all; the
complement of T is just a VP in this case. (TP is always projected in a Many subordinate clauses — those we
root clause, whether there is an MP or not.) We thus have LF-structures call “finite” — also always have a TP. As
for embedded clauses more generally
like the following. (The corresponding surface sentences are given below,
(including infinitives etc.), we don’t need
and we won’t be explicit about the derivational relation between these to take a stand here.
and the LFs. Assume your favorite theories of syntax and morphology
here.)
(2) [TP Svenja [T′ PRES [VP t [V′ be tired ]]]]
= Svenja is tired.

(3) [TP Svenja [T′ PAST [VP t [V′ be tired ]]]]

= Svenja was tired.

(4) [TP Svenja [T′ PRES [MP t [M′ woll [VP t [V′ be tired ]]]]]]
= Svenja will be tired.

woll in (4) stands for the underlying uninflected form of the auxiliary We use “woll” as the name of the root un-
which surfaces as will in the present tense (and as would in the past tense). derlying will and would, following Abusch
1988 and Ogihara 1989: p.32; Abusch
When we have proper name subjects, we will assume for simplicity that
(1997: fn.14, p.22) attributes the coinage
they are reconstructed into their VP-internal base position. of woll to Mats Rooth in class lectures at
What are the meanings of PRES, PAST, and woll? For PRES, the sim- UT Austin.
plest assumption that seems to work is that it is semantically vacuous. This
means that the interpretation of the LF in (2) is identical to the interpreta-
tion of the bare VP Svenja be tired:
(5) For any index 𝑖: ⟦PRES (Svenja be tired)⟧𝑖 = ⟦Svenja be tired⟧𝑖 = 1
iff Svenja is tired in 𝑤𝑖 at 𝑡𝑖 .

Does this adequately capture the intuitive truth-conditions of the sentence

Svenja is tired? It does if we make the following general assumption:
 BEGI NN I NGS OF T ENSE AND ASPECT 105

(6) Utterance Rule:

An utterance of a sentence (= LF) 𝜙 that is made in a world 𝑤 at a
time 𝑡 counts as true iff ⟦𝜙⟧ ⟨𝑤,𝑡 ⟩ = 1 (and as false if ⟦𝜙⟧ ⟨𝑤,𝑡 ⟩ = 0).

This assumption ensures that (unembedded) sentences are, in effect, in-

terpreted as claims about the time at which they are uttered (“utterance
time” or “speech time”). If we make this assumption and we stick to the
lexical entries we have adopted, then we are driven to conclude that the
present tense has no semantic job to do. A tenseless VP Svenja be tired
would in principle be just as good as (2) to express the assertion that Svenja
is tired at the utterance time. Apparently, it is just not well-formed as an
unembedded structure, but this fact then must be attributed to principles
of syntax rather than semantics.
What about PAST? When a sentence like (3) Svenja was tired is uttered
at a time 𝑡, then what are the conditions under which this utterance is
judged to be true? A quick answer is: an utterance of (3) at 𝑡 is true iff
there is some time before 𝑡 at which Svenja is tired. This suggests the fol-
lowing entry:
(7) For any index 𝑖: ⟦PAST⟧𝑖 = 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . ∃𝑡 before 𝑡𝑖 : 𝑝 (⟨𝑤𝑖,, 𝑡⟩) = 1

So, the past tense seems to be an existential quantifier over times, re-
stricted to times before the utterance time.
For will, we can say something completely analogous: Are there also tenses with universal force ?
Two possible candidates that call for closer
(8) For any index 𝑖: ⟦woll⟧𝑖 = 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . ∃𝑡 after 𝑡𝑖 : 𝑝 (⟨𝑤𝑖,, 𝑡⟩) = 1 examination: gnomic tenses (e.g. in An-
cient Greek), and the (universal reading
Apparently, PAST and woll are semantically alike, even mirror images of of the) English perfect (as in I have been
each other, though they are of different syntactic categories. The fact that tired since yesterday morning). Both have
PAST is the topmost head in its sentence, while woll appears below PRES, been written about in the formal seman-
tics literature (the latter extensively — you
is due to the fact that our syntax happens to require a T-node in every
could start with Iatridou, Anagnostopolou
complete sentence. Semantically, this has no effect, since PRES is vacuous. & Izvorski 2001 and von Fintel & Iatridou
Both (7) and (8) presuppose that the set or times comes with an intrin- 2019.
sic order. For concreteness, assume that the relation ‘precedes’ (in sym- Strict linear orders are transitive, irreflex-
bols: <) is a strict linear order on the set of all times. The relation ‘follows’, ive, asymmetric, and connected. See Sec-
of course, can be defined in terms of ‘precedes’ (𝑡 follows 𝑡 ′ iff 𝑡 ′ precedes tion 3.2.2 on the basics of order theory.
Some rethinking is needed once we move
𝑡).
to intervals.

6.2 Time frame adverbials

In this section, we take a brief look at temporal adverbials, specifically so-

called frame adverbials, such as:
(9) Svenja was tired on February 1, 2001.

There are two ideas that come to mind. One is that phrases like on Febru-
ary 1, 2001 are restrictors of temporal operators (kind of like if -clauses
are restrictors of modals). The other idea is that they are modifiers of the
106 T I M E

proposition in the temporal operator’s scope. If we want to go with the

first idea, we have to make some changes. Our current PAST and woll are
unrestricted (1-place) operators, so there is no place for a restrictor. The
second idea is easier to implement, and we try that first.
A propositional modifier is a function from propositions to truth-values,
where “proposition” for us now means “temporal proposition”. Here is an
entry for on February 1, 2001. Intuitively, this modifier takes a proposi- Technically, the modifier returns a truth-
tion and returns a proposition that puts an added condition on the time- value, not a proposition. We get back a
proposition only when we compute the
coordinate of its index-argument.
intension of the phrase that includes the
(10) ⟦on February 1, 2001⟧𝑖 = 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ . modifier.

[𝑝 (𝑖) = 1 & 𝑡𝑖 is part of February 1, 2001]

(11) LF: PAST [VP [VP Svenja be tired ] [PP on February 1, 2001 ]]

Exercise 6.1 Imagine that sentence (9) is not given the LF in (11), but this
one, with the PP attached higher:
(12) LF: [T′ ᴘᴀSᴛ [VP Svenja be tired ]] [PP on February 1, 2001]

What would the truth-conditions of this LF be? Does this result correspond at
all to a possible reading of this sentence (or any other analogous sentence)? If not,
how could we prevent such an LF from being produced? □

The truth conditions that we derive given (10) and (11) look good: the
sentence is predicted true as uttered if there is a time which is both before
the utterance time and within Feb 1, 2001 and at which Svenja is tired,
and it is predicted false if there is no such time. But arguably this is not
exactly right. Suppose that somebody uttered this sentence at an utter-
ance time that preceded the date in the adverbial, say at some time in the
year 2000. Our analysis predicts that this utterance is false. But in fact it
feels more like a presupposition failure; the speaker is heard to be taking
for granted that Feb 1, 2001 is in the past of his speaking. Standard pre-
supposition tests confirm this. For example, the negated sentence (Svenja
wasn’t tired on Feb 1, 2001) and the polar question (Was Svenja tired on Feb
1, 2001?) also convey that the speaker assumes he is speaking after Feb 1,
2001.
 BEGI NN I NGS OF T ENSE AND ASPECT 107

If we want to account for this more fine-grained intuition, the restric- It also has the virtue of avoiding the
tor approach has an advantage after all. Let’s revise the entries for PAST potential overgeneration issue that you
looked at in the exercise above. Q: How
and woll so that they denote 2-place operators, and moreover they encode
so?
a non-emptiness presupposition.
(13) For any index 𝑖:
⟦PAST⟧𝑖 = 𝜆𝑝 : ∃𝑡 [𝑡 < 𝑡𝑖 & 𝑝 (𝑤𝑖 , 𝑡) = 1].
𝜆𝑞.∃𝑡 [𝑡 < 𝑡𝑖 & 𝑝 (𝑤𝑖 , 𝑡) = 1 & 𝑞(𝑤𝑖 , 𝑡) = 1]

(14) For any index 𝑖: How about present tense? Should we

⟦woll⟧𝑖 = 𝜆𝑝 : ∃𝑡 [𝑡 > 𝑡𝑖 & 𝑝 (𝑤𝑖 , 𝑡) = 1]. make this presuppositonal as well — which
would imply it is not, after all, completely
𝜆𝑞.∃𝑡 [𝑡 > 𝑡𝑖 & 𝑝 (𝑤𝑖 , 𝑡) = 1 & 𝑞(𝑤𝑖 , 𝑡) = 1]
vacuous? Frame adverbials in present tense
sentences do occur. Typically they are ad-
Furthermore, let’s change (i.e., simplify) the meaning of the adverbial so
verbials like today, on this beautiful Monday,
that it has a suitable type to serve as the temporal operator’s first argu- which in virtue of their own meaning al-
ment. The LF-structure must be accordingly different as well. Instead of ready are required to contain the speech
(11) above, we now posit (17), where the adverb forms a constituent with time. The following entry would make
room for them and duplicate this require-
the tense. This requires the surface order to be derived by some reorder-
ment as a presupposition:
ing, perhaps extraposition of the adverbial.
(14) For any index 𝑖: ⟦PRES⟧𝑖 =
(16) ⟦on February 1, 2001⟧𝑖 = 1 iff 𝑡𝑖 is part of February 1, 2001 𝜆𝑝 ∈ 𝐷 ⟨𝑠,𝑡 ⟩ : 𝑝 (𝑤𝑖 , 𝑡𝑖 ). 𝜆𝑞 ∈
𝐷 ⟨𝑠,𝑡 ⟩ . 𝑞 (𝑤𝑖 , 𝑡𝑖 ) = 1
(17) LF: [T′ PAST [PP on February 1, 2001 ]] [VP Svenja be tired ]
For the purposes of these lecture notes, we
The meanings we now derive contain the desired presuppositions: The leave the matter open and stick with the
vacuous meaning for PRES in the discus-
past tense sentence (9) presupposes that Feb 1, 2001 is at least in part be-
sions of the upcoming sections.
fore the utterance time, the future sentence Svenja will be tired on Feb 1,
2001 presupposes that this date is at least in part after the utterance time,
and the present tense sentence Svenja is tired on Feb 1, 2001 presupposes
that the utterance time is on this date. Apart from the presuppositions,
the meanings are the same as before. On the down-side, the new analysis
posits both more complex meanings for the tenses and a less direct corre-
spondence between LF constituency and surface structure. Furthermore,
how is it supposed to apply to simple sentences without adverbials? Not
every tensed sentence contains an obligatory frame adverb, after all. We
are forced to say there is a covert restrictor whenever there isn’t an overt
one. But this, upon reflection, turns out to be a virtue, as we will see in
the next section.
Considering the mismatch between the
LF in (17) and the surface order, note that
Exercise 6.2 When a quantifier appears in a tensed sentence, we expect two we saw a similar issue of apparent mis-
match when we decided to treat if -clauses
scope construals. Consider a sentence like this: as restrictors of modals. Both issues might
(18) Every professor (in the department) was a teenager in the Sixties. be addressed by simply letting modal and
temporal operators take their arguments
We can imagine two LFs: in the opposite order (something sug-
gested by Chierchia 1995). Rewriting
(19) [ᴘᴀSᴛ in the sixties] [every professor be a teenager] the lexical entries in this way is a routine
exercise. We leave this matter open. The
syntax of frame adverbials is a non-trivial
object of study.
108 T I M E

(20) [every professor] 7 [ [ᴘᴀSᴛ in the sixties] [𝑡 7 be a teenager] ]

Describe the different truth-conditions which our system assigns to the two LFs.
Is the sentence ambiguous in this way? If not this sentence, are there analogous
sentences that do have the ambiguity? □

Exercise 6.3 Our official entry for every makes it a time-insensitive (and
world-insensitive) item:
(21) For any 𝑖, ⟦every⟧𝑖 = 𝜆𝑓 ⟨𝑒,𝑡 ⟩ .𝜆𝑔 ⟨𝑒,𝑡 ⟩ . ∀𝑥 : 𝑓 (𝑥) = 1 → 𝑔(𝑥) = 1

Consider now two possible variants (we have boxed the portion where they dif-
fer):
(22) For any 𝑖, ⟦every⟧𝑖 = 𝜆𝑓 ⟨𝑒,𝑡 ⟩ .𝜆𝑔 ⟨𝑒,𝑡 ⟩ . ∀𝑥 at 𝑡𝑖 : 𝑓 (𝑥) = 1 → 𝑔(𝑥) = 1

(23) For any 𝑖, ⟦every⟧𝑖 = 𝜆𝑓 ⟨𝑒,𝑡 ⟩ .𝜆𝑔 ⟨𝑒,𝑡 ⟩ . ∀𝑥 : 𝑓 (𝑥) = 1 at 𝑡𝑖 → 𝑔(𝑥) = 1

Does either of these alternative entries make sense? If so, what does it say? Is it
equivalent to our official entry? Could it lead to different predictions about the
truth-conditions of English sentences? □

6.3 Are tenses referential?

Our first semantics for the past tense, in Section 6.1, treated it as an un-
restricted existential quantifier over times. This seems quite adequate for
examples like (24), which seem to display the expected unrestricted exis-
tential meaning:
(24) Georgia went to a private school.

All we learn from (24) is that at some point in the past, whenever it was
that Georgia went to school, she went to a private school.
Partee in her famous paper “Some structural analogies between tenses
and pronouns in English” (Partee 1973) presented an example where tense
appears to act more “referentially”:
(25) I didn’t turn off the stove.

“When uttered, for instance, halfway down the turnpike, such a sentence
clearly does not mean either that there exists some time in the past at
which I did not turn off the stove or that there exists no time in the past
at which I turned off the stove. The sentence clearly refers to a partic-
ular time — not a particular instant, most likely, but a definite interval
whose identity is generally clear from the extralinguistic context, just as
the identity of the he in [He shouldn’t be in here] is clear from the context.”
Partee argues, in effect, that neither of the two plausible LFs that our
system from Section 6.1 derives can correctly capture the meaning of
(25). Given that the sentence contains a past tense and a negation, there
are two possible scopings of the two operators:
 BEGI NN I NGS OF T ENSE AND ASPECT 109

(26) a. PAST NEG I turn off the stove.

b. NEG PAST I turn off the stove.

Exercise 6.4 Using our old semantics from 6.1, show that neither LF in (26)
captures the meaning of (25) correctly.

In a commentary on Partee’s paper (at the same conference it was pre-

sented at), Stalnaker pointed out that a minor amendment of the Priorean
theory can deal with (25). One just needs to allow the existential quanti-
fier to be contextually restricted to times in a salient interval. Since natural
language quantifiers are typically subject to contextual restrictions, this
is not a problematic assumption. Note that Partee formulated her obser-
vation in quite a circumspect way: “The sentence refers to a particular
time”; Stalnaker’s suggestion was that the reference to a particular time
is part of the restriction to the quantifier over times expressed by tense,
rather than tense itself being a referring item.
Ogihara 1995, 1996 argued that the restricted existential quantifica-
tion view is in fact superior to Partee’s analysis, since Partee’s analysis
needs an existential quantifier anyway. It is clear that the time being re-
ferred to in the stove-sentence (25) is a protracted interval (the time dur-
ing which Partee was preparing to leave her house). But the sentence is
not interpreted as merely saying that this interval is not a time at which
she turned off her stove. That would only exclude a fairly absurd kind of
slow-motion turning-off-of-the-stove (turning off the stove only takes
a moment). Instead, the sentence says that in the salient interval there is
no time at which she turned off the stove. Clearly, we need an existential The alternative is to say that the existential
quantifier in there somewhere and the Priorean theory provides one. quantifier is not expressed by tense but
comes from somewhere else — perhaps
Ogihara made the point with the following example:
from aspect, or from the lexical entry of
(27) Patricia: Did you see Solène? the verb itself. We will come back to these
options.
Lea: Yes, I saw her, but I don’t remember exactly when.

The question and answer in this dialogue concern the issue of whether
Lea saw Solène at some time in a contextually salient interval.
Stalnaker’s and Ogihara’s conclusions converge with what we already
ended up with in Section 6.2, after considering the interaction of tenses
with time frame adverbials. In order to capture presuppositions of tensed
sentences with frame adverbials, we already modified Prior’s original pro-
posal and made room for a restrictor in the semantics of the past tense.
Given this revised analysis of the past tense as a 2-place existential quan-
tifier, it is unsurprising, in fact expected, that an implicit, contextually
salient restrictor should be present when there isn’t an overt one. What
then about example (24), Georgia went to a private school, for which the un-
restricted analysis seemed to do well? Let us say that the covert restrictor
in this case picks out a very long interval, perhaps Georgia’s entire life-
time, or even the entire past from the big bang to the utterance time, or
110 T I M E

all eternity. (What exactly the right restrictor is in this case, and what
makes it contextually available, may be a bit unclear, but we leave it at
that.)

Exercise 6.5 Assuming the restricted existential quantifier analysis of past tense
that we adopted in Section 6.2, which of the scope constellations in (26) captures
the meaning of (25) correctly?

6.4 Referential tense and perfective/imperfective aspect

6.4.1 Referential tense after all

Let us return to Partee’s stove and the prospects of a “referential” theory
of tense. Our discussion of Partee’s example (following Stalnaker and
Ogihara) just now came to the conclusion that we did need past tense to
be an existential quantifier over times, albeit a contextually restricted one.
The stove-example is interpreted as a claim about a particular contextu-
ally relevant interval. But the speaker’s claim is not merely that she didn’t
turn off the stove at that interval. That in itself would be compatible with
her turning off the stove at some smaller interval inside the contextually
relevant interval. The speaker’s claim is stronger: she did not turn off the
stove at any time that is contained in this interval. This is a negative exis-
tential claim. So there needs to be an existential quantifier somewhere in
the LF and below the scope of not, and we concluded that past tense must
be supplying it. But this conclusion is not inescapable. Granted that there
has to be an existential quantifier somewhere — but couldn’t it be some-
where else than in the meaning of tense? One alternative that comes to
mind is to locate it in the lexical meaning of the verb (here turn off ). This
means we abandon the lexical entry in (28) and instead adopt the one in
(29).
(28) ⟦turn-off⟧𝑖 = 𝜆𝑦.𝜆𝑥 .𝑥 turns off 𝑦 in 𝑤𝑖 at 𝑡𝑖

(29) ⟦turn-off⟧𝑖 = 𝜆𝑦.𝜆𝑥 .𝑥 turns off 𝑦 in 𝑤𝑖 in 𝑡𝑖

The difference between ‘at’ and ‘in’ looks small at first, but if we reflect
on the meaning of ‘in’, we see the hidden existential quantifier. When
something happens in an interval, it happens at some part of the interval.
We can make this more transparent in the metalanguage and rewrite (29)
as (30).
(30) ⟦turn-off⟧𝑖 = 𝜆𝑦.𝜆𝑥 .∃𝑡 ⊆ 𝑡𝑖 : 𝑥 turns off 𝑦 in 𝑤𝑖 at 𝑡

The subset sign here stands for the containment relation between time
intervals. A time interval can be defined as a certain kind of set of mo-
ments, as in (31), so the subset relation is well defined.
(31) A set of moments 𝑆 is an interval iff for any two moments that are
in 𝑆, every moment between them is also in 𝑆.
 BEGI NN I NGS OF T ENSE AND ASPECT 111

Another way to clarify the distinction between ‘at’ and ‘in’ is to use the
kind of metalanguage that is familiar from the literature on Davidsonian
event semantics.
(32) abbreviations in “event talk”:
a. turn-off(𝑒, 𝑥, 𝑦) = 𝑒 is an event of turning off 𝑦 by agent 𝑥
b. 𝜏 (𝑒) = the (exact) time-interval occupied by event 𝑒
also called the “run-time” or “temporal trace” of e

(33) event-talk formulation of (28), the old entry with ’at’:

⟦turn-off⟧𝑖 = 𝜆𝑦.𝜆𝑥 .∃𝑒 [turn-off(𝑒, 𝑥, 𝑦) & 𝑒 is in 𝑤𝑖 & 𝜏 (𝑒) = 𝑡𝑖 ]

(34) event-talk formulation of (29), the new entry with ’in’:

⟦turn-off⟧𝑖 = 𝜆𝑦.𝜆𝑥 .∃𝑒 [turn-off(𝑒, 𝑥, 𝑦) & 𝑒 is in 𝑤𝑖 & 𝜏 (𝑒) ⊆ 𝑡𝑖 ]

Let us spell out now how Partee’s proposal for the meaning of past tense
can be upheld after all, once we assume the lexical semantics specified in
(29)/(30)/(34). The first task here is to write new lexical entries for the
tense morphemes, which encode Partee’s idea that tenses refer to spe-
cific time intervals and are semantically and pragmatically akin to per-
sonal pronouns. We will defer the full execution of this task until later
and make do for the time being with a couple of syncategorematic ad hoc
rules for the interpretation of TPs.
′
(35) ⟦PAST 𝜙⟧𝑖 = 1 iff ⟦𝜙⟧ ⟨𝑤𝑖 ,𝑡 ⟩ = 1, where 𝑡 ′ is the contextually salient This contextually salient time is also called
time before 𝑡𝑖 (no truth value defined if there is no such time) the “topic time” (Klein 1994) or the “ref-
erence time” (a term which goes back to
′
(36) ⟦woll 𝜙⟧𝑖 = 1 iff ⟦𝜙⟧ ⟨𝑤𝑖 ,𝑡 ⟩ = 1, where 𝑡 ′ is the contextually salient Reichenbach 1947, but which has various
other uses in the literature).
time after 𝑡𝑖 (no truth value defined if there is no such time)

(PRES remains vacuous as before, i.e., ⟦PRES 𝜙⟧𝑖 = ⟦𝜙⟧𝑖 .)

The stove example has the two potential LFs in (37).
(37) a. PAST NEG I turn off the stove.
b. NEG PAST I turn off the stove.
We can compute the truth conditions for both of these under the new
semantics for PAST and turn off, and it turns out that they are the same.
(38) ⟦(37a)⟧𝑖 = 1 iff ⟦(37b)⟧𝑖 = 1 iff
¬∃𝑒 [turn-off(𝑒, 𝑥, 𝑦) & 𝑒 is in 𝑤𝑖 & 𝜏 (𝑒) ⊆ 𝑡 ′ ],
where 𝑡 ′ is the contextually salient time before 𝑡𝑖
(no truth-value defined if there is no such time)

The fact that both scopal orders yield the same truth conditions is ar-
guably a point in favor of this approach. The English sentence is not in
fact perceived as ambiguous. Our earlier approach, on which past tense
was a contextually restricted existential quantifier, did not make this pre-
diction — at least not without the help of additional assumptions (such as
a syntactic constraint on the position of negation with respect to other
112 T I M E

heads on the clausal spine). Now that the existential quantifier comes
bundled with the lexical verb, its scope is automatically “frozen” below
everything that scopes over the verb.

6.4.2 Event semantics and perfective aspect

Up to now, we have presupposed a pre-Davidsonian view of lexical mean- We share the goal of integrating David-
ings, on which verbs take only individuals or propositions as their argu- sonian event semantics and traditional
intensional semantics with von Stechow
ments. Even when we recently inserted some event-talk into the metalan-
& Beck 2015, from whom we borrow a
guage of our lexical entries, we still defined the denotation of a verb like number of ideas.
turn off as a function from two individuals to a truth value. In this sec-
tion, we switch to a Davidsonian treatment of verbs as predicates of events
and integrate this with our conception of sentence-intensions as temporal
propositions.
In an extensional Davidsonian semantics, lexical entries look like (39). Many practitioners of event semantics as-
sume that the event argument is the only
(39) ⟦laugh⟧ = 𝜆𝑥 .𝜆𝑒. 𝑒 is an event of 𝑥 laughing real argument of the verb, whereas the
abbreviated: ⟦laugh⟧ = 𝜆𝑥 .𝜆𝑒. laugh(𝑒, 𝑥) subject, object, etc. are arguments of ab-
stract theta-role heads that combine with
Assuming that events are not in 𝐷𝑒 , but have their own basic type 𝜈, VPs the verb in the manner of modifiers. Here
thus are of type ⟨𝜈, 𝑡⟩. (All the verb’s non-event arguments are merged we remain agnostic on this matter. For
concreteness, we assume that verbs take
inside the VP.)
all the traditional arguments in addition to
In a semantics that is both Davidsonian and intensional, do we have their event-argument, but the other view
to rewrite these entries? For example, should we perhaps rewrite (39) as is equally compatible with everything we
(40)? will say. We just abstract away from the
internal compositional semantics of the
(40) ⟦laugh⟧𝑖 = 𝜆𝑥 .𝜆𝑒. 𝑒 is an event of 𝑥 laughing in 𝑤𝑖 at 𝑡𝑖 VP.

That depends. Here we follow Kratzer and assume that each event occurs
in just one world and at just one time. It is not possible for a given 𝑒 to be This assumption is made here mostly to
an event of 𝑥 laughing in one world and to be some other kind of event keep things simple. It is not innocuous and
not uncontroversial. See e.g. Hacquard
in another world. Nor is it possible for one and the same 𝑒 to be an event
2009b for an analysis of root modals that
of 𝑥 laughing at one time and something else at another time. Reformu- makes crucial use of the idea that an actual
lations such as (40) are uncalled for then, and we can essentially stick with event exists in non-actual worlds and has
(39). different properties there.

But how then does world and time dependence enter the semantic Strictly speaking, we should now write
computation? And how can tenses and modal operators combine with (41), but since 𝑖 in (39) does not occur on
VPs? VPs are now type ⟨𝜈, 𝑡⟩, which leads to a type-mismatch if we try the right side of =, (39) can be shorthand
for (41).
to combine them directly with a modal operator or with a tense (regard-
less of whether the tense is a Priorian temporal operator or a Partee-style (41) For any index 𝑖, ⟦laugh⟧𝑖 =
𝜆𝑥 .𝜆𝑒. laugh(𝑒, 𝑥 )
referential tense). The way out of this problem is to posit a more complex
clause structure, with a further functional head that intervenes between T
(or M) and V. This is called an “aspect” head (category label “Asp”), and
its semantic job is to existentially bind the event argument of the VP and
return a world- and time-sensitive denotation of type 𝑡.
 BEGI NN I NGS OF T ENSE AND ASPECT 113

One instance of Asp is the so-called “perfective”, for which we posit (42) combines the standard formal analysis
the following entry. of perfective aspect (among many oth-
ers: Klein 1994, Kratzer 1998) with the
(42) ⟦PFV⟧𝑖 = 𝜆𝑃 ⟨𝜈,𝑡 ⟩ .∃𝑒 [𝑃 (𝑒) = 1 & 𝜏 (𝑒) ⊆ 𝑡𝑖 & 𝑒 ≤ 𝑤𝑖 ] semantics of von Stechow & Beck 2015’s
≤ := is part of (= occurs in) Modl head. It locates the event both in a
time interval and in a possible world.
𝜏 := the run time of (temporal trace of )

PFV is morphologically zero in English, so we can posit it in the LFs of

sentences with simple tensed verbs. Here is an example.
(43) a. Barbara turned off the stove.
b. LF: [TP PAST [AspP PFV [VP Barbara turn-off the stove ]]]

Using the syncategorematic rule (35) for referential PAST, our entry (42)
for PFV, and a Davidsonian entry for the verb, we compute the following
interpretation. (Do this as an exercise.)
(44) ⟦(43b)⟧𝑖 = 1 iff ∃𝑒 [turn-off(𝑒, B, the stove) & 𝜏 (𝑒) ⊆ 𝑡 ′ & 𝑒 ≤ 𝑤𝑖 ],
where 𝑡 ′ is the contextually salient time before 𝑡𝑖
(no truth-value defined if there is no such time)

This is the same meaning that we obtained in the previous section, when
we had built the existential quantification into the lexical meaning of the
verb. What used to be the meaning of VP is now the meaning of AspP.
We have located the event-quantifier in its own functional head, but oth-
erwise it is the same analysis.

Exercise 6.6 What about the negated sentence that was Partee’s original ex-
ample? Where can we now generate negation in an interpretable LF? Does the
current analysis still predict that the sentence is not in fact ambiguous? □

6.4.3 The English progressive

Besides perfective aspect, there is imperfective aspect — or more accu- See e.g. Arregui, Rivero & Salanova 2014
rately, there is probably a family of imperfective aspects in different lan- for a recent approach to cross-linguistic
semantic variation in imperfective aspects.
guages that have some shared and some non-shared properties. English
has an imperfective aspect known as the “progressive”, with an overt mor- We are not serious about morphology
phology that consists of a copula which governs a present-participial form here. The meaning may well be carried
by an abstract head and the be a vacuous
of the VP (V-ing). We posit a functional head be-PROG as the aspect head
element.
in the English progressive construction.
The basic intuition behind much work on the perfective-imperfective
distinction is that, whereas perfective aspect locates an event within the
evaluation time, imperfective aspect does the reverse, i.e., it places the
evaluation time within the event time. If we formalize this intuition di-
rectly, without introducing any further differences from the perfective,
we come up with (45).
114 T I M E

(45) First attempt: This is similar to the first formal analysis

⟦be-PROG⟧𝑖 = 𝜆𝑃 ⟨𝜈,𝑡 ⟩ .∃𝑒 [𝑃 (𝑒) = 1 & 𝑡𝑖 ⊆ 𝜏 (𝑒) & 𝑒 ≤ 𝑤𝑖 ] of the progressive, due to Bennett & Par-
tee 1978. They did not work in an event
semantics, however. Also, their semantics
It is well-known, however, that there is also a difference in how the event
required 𝑡𝑖 to be a non-final subinterval of
is related to the evaluation world. While perfective places the event within 𝜏 (𝑒 ), rather than merely 𝑡𝑖 ⊆ 𝜏 (𝑒 ). This
the actual world, the progressive permits it to be partly in another world, requirement seemed too strong in light
so to speak. This point, which was at the center of Dowty 1977’s seminal of examples such as Dowty 1977’s John
was watching TV when he fell asleep (which
work on the progressive, is brought home by examples like (46).
does not say that TV-watching continued
(46) John was going to the store when he ran into Svenja. beyond the point of falling asleep). How-
ever, as Dowty showed, it turned out to
We can’t infer from this sentence that John actually made it to the store, be the right requirement in the context of
the modalized analysis that he proposed,
or will ever make it there. The sentence leaves this open. Perhaps John
see below.
does complete his trip to the store after the encounter, and perhaps he
doesn’t. The truth-conditions of the sentence (46) are compatible with Dowty dubbed this the “imperfective
paradox”, although it’s not really a para-
either scenario. The entry in (45), on the other hand, would require that dox, just a counterexample to a certain
there be a John-going- to-the-store event which occupies a super-interval analysis that looked plausible at first.
of the time of the encounter with Svenja and which occurs in the actual
world. So (45) can’t be quite right.
Dowty’s analysis of the progressive says instead that a John-going-to-
the-store event occurs in certain possible worlds. These possible worlds
are related to the actual world in a particular way: they are worlds which
share a history with the actual world up to a certain point and then de-
velop (possibly counterfactually) in such a way that no events that were
already in progress get interrupted (“inertia worlds”). The idea is, very
roughly, that the sentence tells us: either John actually went to the store,
or if he didn’t, then at least he would have gone there if he hadn’t been in-
terrupted. Since the publication of Dowty’s paper, there has been a suc- See among many others, Landman 1992,
cession of sophisticated counterexamples and refinements to his original Portner 1998.

proposal, but this is beyond the scope of this introduction. Here is a ver-
sion based on Dowty.
(47) second attempt (and final version for us): Apart from introducing quantification
over other worlds, (47) also differs from
[
⟦be-PROG⟧𝑖 = 𝜆𝑃 ⟨𝜈,𝑡 ⟩ .∀𝑤 𝑤 ∈ Inert(𝑖) → ∃𝑒 [𝑃 (𝑒) = 1 & (45) in that it strengthens the requirement
] on the temporal relation between 𝑡𝑖 and
𝑡𝑖 ⊂ < 𝜏 (𝑒) & 𝑒 ≤ 𝑤] 𝜏 (𝑒 ): not only must 𝜏 (𝑒 ) contain all of 𝑡𝑖 ,
but it must moreover extend into the time
where ⊂ < abbreviates: “is a non-final subinterval of” after 𝑡𝑖 . This is intended to capture the
(that is: 𝜏 (𝑒) includes every moment in 𝑡𝑖 as well as some moment intuition that e.g. (46) is not appropriate if
John already reaches the store during his
after the end of 𝑡𝑖 ) encounter with Svenja; see Dowty 1977
for discussion.
(48) Definition: 𝑤 ∈ Inert(𝑖) iff
𝑤 is exactly like 𝑤𝑖 up to the end of 𝑡𝑖 and then develops in such a
way that no events are interrupted.

We will see in a minute that there is a class of VPs for which the truth-
conditions predicted by (47) come very close to those predicted by the
simpler (45). But examples like (46) show that this must not always hold.
 BEGI NN I NGS OF T ENSE AND ASPECT 115

Let’s do a simple example.

(49) a. Sari is laughing.
b. LF: [TP PRES [AspP be-PROG [VP Sari laugh ]]]
c. ⟦(49b)⟧𝑖 = 1 iff
[ ]
∀𝑤 𝑤 ∈ Inert(𝑖) → ∃𝑒 [𝑡𝑖 ⊂ < 𝜏 (𝑒) & 𝑒 ≤ 𝑤 & laugh(𝑒, Sari)]

Just as it stands, (49c) does not logically entail that any laughing happens
in the world 𝑤𝑖 (i.e., in the utterance world 𝑤𝑢 if this is an unembedded
assertion). It only talks about the inertia worlds. However, there is a prop-
erty of the lexical meaning of laugh that permits us to draw further infer-
ences. Laughing events are made up of lots of sub-events which them-
selves are laughing events, down to very little ones that don’t last much
more than an instant. Given this, consider a world in Inert(𝑤𝑢 , 𝑡𝑢 ), say
𝑤. If (49b) is true in 𝑤𝑢 at 𝑡𝑢 , it follows that 𝑤 contains an event of Sari
laughing whose run-time includes 𝑡𝑢 . Among the subevents of this event,
which themselves are events of Sari laughing, there will most likely be
one that is early enough and small enough to have transpired by the end
of 𝑡𝑢 . And since up to the end of 𝑡𝑢 , the histories of 𝑤 and 𝑤𝑢 are identical, The only condition under which this
this small Sari-laughing event in 𝑤 must have a perfectly matching coun- would not hold is if the laughing starts
right at the beginning of 𝑡𝑢 and 𝑡𝑢 itself
terpart in 𝑤𝑢 . That’s why we infer from (49a) that there is actual laughing
is too short to fit even a minimal laughing
at the utterance time. event. This would have to be a very short
This is the kind of example for which (47) and the simpler entry (45) utterance time, shorter than it realistically
predict almost identical truth conditions. (47) demands something slightly takes to say Sari laughs, so we disregard
this possibility. But we will later see a
stronger, namely that moreover the laughing continues at least a little bit problem with this.
beyond the utterance time unless it is interrupted (which means it would
have continued). So they are not quite equivalent, but the difference is
very subtle.
Importantly, however, this almost-equivalence depends on the par-
ticular property of the meaning of the VP that we just exploited in our
reasoning. Had the VP been Sari go to the store, it would have been a very
different matter. Events of Sari going to the store are not made up of lots
of smaller events which each are events of Sari going to store. They are
made up of smaller events which are events of Sari going towards the
store, but since most of these don’t end with Sari at the store, they are
not events of Sari going to the store. So if we are told that every 𝑤 ∈
Inert(𝑖) contains an event of Sari going to the store which occupies a
super-interval of 𝑡𝑖 , we cannot infer that Sari goes to the store in 𝑤𝑖 . We
can merely infer that 𝑤𝑖 contains an event that is indistinguishable from
those parts of the inertia-worldly trips-to-the-store which fall before the
end of 𝑡𝑖 . In other words, we infer that 𝑤𝑖 contains the beginning of a Sari-
go-to-the-store event, but not necessarily anything more.
The attentive reader may have wondered why we used a past tense
example to illustrate the perfective in the previous section, but a present
tense example for the progressive in the current section. Indeed, it is in-
116 T I M E

cumbent upon us to examine what the theory predicts for every possible
combination of a tense and an aspect.

6.4.4 Stativity effects

It is well known that non-stative predicates in the simple present tense The examples in (50) do not have promi-
have a limited range of felicitous uses. Sentences such as those in (50) are nent habitual (generic) readings. Ignore
such readings if you can get them any-
spontaneously judged as odd by speakers of English.
way. The # judgments apply to an in-
(50) a. #Sari laughs. tended episodic reading (describing a
single event).
b. #Sari wakes up. Simple present tense on a non-stative
c. #Sari goes to the store. verb is systematically grammatical when
the sentence has a generic or habitual
Let us see what our theory predicts. We see no progressive morphology, interpretation, or when it describes the
content of a plan or schedule. We don’t
but there has to be an aspect head for the sentence to express a proposi-
worry about these cases here, since they
tion, so the aspect must be PFV. With present tense semantically vacuous, very plausibly involve a covert modal op-
we then have LFs and predicted meanings like (51) for (50c). erator of some kind that applies to the
VP before any tense or (higher) aspect.
(51) LF: [TP PRES [AspP PFV [VP Sari go to the store ]]] (See for example, Copley 2008 or Thomas
true at 𝑖 iff ∃𝑒 [𝜏 (𝑒) ⊆ 𝑡𝑖 & 𝑒 ≤ 𝑤𝑖 & go-to(𝑒, Sari, the store)] 2014.) That modal operator may itself be
an aspect head, or it may create a big-
This says that if Sari goes to the store is asserted in 𝑤𝑢 at 𝑡𝑢 , the assertion is ger VP which is a predicate of states. In
true iff there is a Sari-go-to-the store event in 𝑤𝑢 whose run-time is con- the latter case, whatever explains the ac-
ceptability of stative VPs under present
tained within 𝑡𝑢 . This is a somewhat implausible scenario, given that trips
tense will also explain the acceptability of
to the store typically take longer than the production of such a short sen- present tense generics/habituals. See be-
tence. One may be tempted to attribute the strangeness of (50c) to this low. Some kind of modal analysis might
fact. But upon reflection, that doesn’t look like the right explanation. We also work for some of the cases in the list,
like stage directions and plot summaries,
can set up a scenario that eliminates the implausibility. Imagine Sari was
but less plausibly to e.g. the sportscaster
already very close to the store, and/or she is on a very fast vehicle …. The style or the historical present.
judgment about (50c) is not really affected by such manipulations, but
we would expect it to be if pragmatic plausibility were all that mattered.
And the pragmatic explanation looks even less convincing when we con-
sider the other examples in (50). Waking-up events are very short, if not
instantaneous, so such events should have no problem fitting inside the
utterance time and (50b) should be just fine. As regards (50a), we have al-
ready said that longer laughing events are made up of shorter laughing
events. So if Sari laughs for any duration that overlaps with the utterance
time, there is probably a laughing event within the utterance time, and
(50a) should be fine as well.
Friends of pragmatic approaches like to remind us that the examples in There are few formal semantic analyses
(50) are not ungrammatical. Sentences of this sort are acceptable in a va- of the historical present. An exception is
Zucchi 2005.
riety of special contexts or registers, such as play-by-play sportscasting,
the historical or narrative present, newspaper headlines, stage directions,
plot summaries, explicit performatives, …, to name some. It is appealing
to say that the essence of (at least some of ) these special uses is a pretense
that the utterance time is something other than what it is, a pretense that
one is speaking at a time closer to the events being reported, at a pretend-
 BEGI NN I NGS OF T ENSE AND ASPECT 117

utterance-time that is earlier and/or longer than one’s actual utterance.

This may or may not be right. At any rate, it does not directly address the
question why (50a-c) are unacceptable outside of these special registers or
contexts. One seems to need a concomitant assumption that the “ordi- Cf. Bennett & Partee 1978: “We regard
nary” register involves a different pretense, namely that the utterance time a speech act as occurring at a moment of
time and understand the assertion as being
is shorter than it actually is, in fact, that it is a mere instant in the technical
true at that moment. Accordingly, we are
sense (a singleton of one moment), and hence too short to contain even a inclined to only use the reportive simple
getting-up event or a minimal laughing-event. present when the act being described
For the sake of the argument, let’s see how it may help to stipulate that seems to be almost instantaneous and to
be occurring at the moment of utterance.”
𝑡𝑢 is always an instant. To get the desired mileage out of this assumption, See also Dowty 1979.
we must also sharpen some specifics regarding the lexical meanings of
verbs. These assumptions are not uncontroversial, but widely accepted As Milo Philipps-Brown (pc) pointed out,
one worry about this assumption is that it
in the literature: None of the VPs in (50) describe events that can pos-
undermines our earlier reasoning about
sibly have run-times that are instants. Any VP that entails a change of the progressive Sari is laughing. There we
state — whether it is a change that takes time (like getting from some attributed the intuition that this sentence
place else to the store) or a virtually “instantaneous” change (like from entails the existence of laughter in 𝑤𝑢
to the fact that 𝑡𝑢 was long enough to
asleep to awake) — because of that applies only to events whose run-time
contain a minimal laugh. It is not clear
contains at least two moments (one at which the previous state holds and how to resolve this tension. Perhaps we
one at which the result state holds). Likewise, any VP that describes an can get out of it by convincing ourselves
activity or movement or other happening of some sort (like laughing, os- that we judge the utterance true, after
all, if all that actually happens before the
cillating, raining, even sleeping) describes events that may have very short interruption is an instant sized beginning
run times but never just a single moment. These assumptions about lexi- of a minimal laugh.
cal semantics make the predicted truth-conditions for clauses with PFV (as
These ideas are common in the literature
computed in (51)) impossible to satisfy unless 𝑡𝑖 is a proper interval, i.e.,
and go back at least to Taylor 1977. See
not a singleton. Filip 2012 for a recent and comprehensive
From this perspective there is a straightforward account of what makes survey.
stative predicates different. Once we change the VPs in (50) to stative
ones, the simple present tense becomes perfectly fine (in every register).
(52) a. Sari is tired.
b. Sari is at home.
c. Sari owns a factory.

Suppose the distinguishing semantic feature is precisely that predi-

cates like tired, at home, and own a factory describe eventualities (“states”)
whose run-times can be instants. A state of Sari being tired may be long
or short, but it is necessarily made up of shorter and shorter sub-states
which are also states of Sari being tired. And not only that — it is even
made up of such sub-states that occupy a single instant. The latter makes
tired different from laugh or even move, which apply to eventualities whose
run-times may be infinitesimally short but are still always proper inter-
vals. What does this buy us? It lets us say that the sentences in (52) have
the exact same parses as those in (50), with a perfective aspect head, and
yet they have truth-conditions which can be satisfied by an instant.
118 T I M E

(53) LF: PRES [ PFV [ Sari be tired ]]

meaning: true at 𝑖 iff ∃𝑒 [𝜏 (𝑒) ⊆ 𝑡 &𝑒 ≤ 𝑤𝑖 & tired(𝑒, Sari)]
lexical entry: ⟦tired⟧ = 𝜆𝑥 .𝜆𝑒. 𝑒 is a state of 𝑥 being tired

(Type 𝜈 must be understood in such a way that 𝐷 𝜈 includes states in addi-

tion to “events” in a narrow sense. Bach 1986 coined the term “eventual-
ity” for this broader sense of “event”.)
So, together with the stipulation that the utterance time is treated as It is often said in this context that progres-
an instant, this approach to the stative/non-stative distinction provides an sives pattern with statives in the present
tense because progressive VPs are stative.
explanation for why stativity is required in the simple present tense. We
This is not literally true on our analysis,
can also reassure ourselves that present progressives are still expected to because be-PROG is an aspect head and
be uniformly good even if 𝑡𝑢 must be an instant. This is because be-PROG AspPs are not predicates of states (or of
places the event in a super-interval of 𝑡𝑖 . eventualities of any kind). One might,
however, entertain a different analysis on
Whether or not the assumption that 𝑡𝑢 is an instant can ultimately be which (at least some of ) the operators we
defended, it is important to be aware that the stativity effect we witness in are used to calling “aspects” have meanings
present tense matrix clauses is replicated perfectly in certain environments of type ⟨𝜈𝑡, 𝜈𝑡 ⟩. (Another head higher in
the structure would then have to be re-
which are neither matrix nor (morphologically) present. These environ-
sponsible for binding the state argument
ments include the complements of epistemic necessity modals and the and introducing the world and time.)
infinitival complements of verbs like believe and claim.
(54) a. Sari must sleep/go to the store. The judgment in (a) presupposes an in-
deontic reading only tended episodic reading for the VP. To the
extent that the VP can be read habitually,
b. Sari must be at home/ be sleeping. the epistemic reading becomes available.
epistemic reading okay The point here is that the judgments for
the constructions in (54)–(56) are parallel
(55) Sari claimed to *work/*go to work/be at work/be going to work. to the judgments for the same VPs in the
simple present tense.
(56) Sari believed Svenja to *sleep/*go to the store/be at home/be sleep-
ing.

We will return to this observation in a later section, in connection with

the discussion of so-called “Sequence of Tense”.

6.5 Formalizing the referential analysis

Above we stated Partee’s proposal as follows:

′
(35) ⟦PAST 𝜙⟧𝑖 = 1 iff ⟦𝜙⟧ ⟨𝑤𝑖 ,𝑡 ⟩ = 1, where 𝑡 ′ is the contextually salient
time before 𝑡𝑖 (no truth value defined if there is no such time)

Apart from not being fully compositional, this is a bit vague for us to
work with when (in the next chapter) we consider complex sentences
with several occurrences of past tense. Let us therefore make it a little
more precise.
Partee suggested that past tense was analogous to a pronoun like he.
We are used to representing pronouns as variables (see e.g. Heim & Kratzer),
so Partee-style tenses too should then have denotations that are sensitive
to a variable assignment. So let’s make two assumptions: First, each oc-
 BEGI NN I NGS OF T ENSE AND ASPECT 119

currence of PAST and woll at LF must carry a numerical subscript, like a

pronoun. Second, we add a new type 𝑖 (for “intervals”) and we assume
that variable assignments can assign elements of 𝐷𝑖 to object-language
variables (numerical indices). Notice that this new type 𝑖 is not the same
as our existing type 𝑠 (world-interval-pairs), though the second member
of an element of 𝐷𝑠 is always a member of 𝐷𝑖 .
We now write the following entries. (PRES remains vacuous for now
and therefore needs no entry.)
(57) ⟦PAST𝑛 ⟧𝑖,𝑔 = 𝜆𝑝 ∈ 𝐷𝑠𝑡 : 𝑔(𝑛) ∈ 𝐷𝑖 & 𝑔(𝑛) < 𝑡𝑖 . 𝑝 (𝑤𝑖 , 𝑔(𝑛)) = 1

(58) ⟦woll𝑛 ⟧𝑖,𝑔 = 𝜆𝑝 ∈ 𝐷𝑠𝑡 : 𝑔(𝑛) ∈ 𝐷𝑖 & 𝑔(𝑛) > 𝑡𝑖 . 𝑝 (𝑤𝑖 , 𝑔(𝑛)) = 1

Let’s illustrate how this works in a simple example.

(59) Barbara turned off the stove.
new LF: PAST7 [ PFV [ Barbara turn off the stove ]]

(60) ⟦PAST 7 [ PFV [ Barbara turn off the stove ]]⟧𝑖,𝑔 is defined
iff 7 ∈ dom(𝑔) & 𝑔(7) ∈ 𝐷𝑖 & 𝑔(7) < 𝑡𝑖
when defined, ⟦PAST 7 [ PFV [ Barbara turn off the stove ]]⟧𝑖,𝑔 = 1
iff ∃𝑒 [𝜏 (𝑒) ⊆ 𝑔(7) & 𝑒 ≤ 𝑤𝑖 & turn-off(𝑒, Barbara, the stove)]

If this is a matrix sentence, we evaluate it with respect to the utterance

world and time, and we also rely on the utterance context to furnish a
suitable assignment (call it 𝑔𝑢 ), which maps the free variable 7 to a time
interval that precedes 𝑡𝑢 . Intuitively, this is the salient past interval that the
speaker has in mind when making this past-tense claim (also called the
“topic time”). So we have:
(61) An utterance 𝑢 of the LF “PAST 7 [PFV [Barbara turn off the stove]]”
is felicitous only if
𝑔𝑢 is such that 7 ∈ dom(𝑔𝑢 ) & 𝑔𝑢 (7) ∈ 𝐷𝑖 & 𝑔𝑢 (7) < 𝑡𝑢 ,
and it is moreover true iff
∃𝑒 [𝜏 (𝑒) ⊆ 𝑔(7) & 𝑒 ≤ 𝑤𝑖 & turn-off(𝑒, Barbara, the stove)]

It will also be useful to clarify how frame adverbials might be treated

in a Partee-style approach to tense. A natural idea here is that a frame ad-
verb contributes a further presupposition about the intended topic time.
Recall our treatment of frame adverbs as having extensions of type 𝑡.
(16) ⟦on February 1, 2001⟧𝑖 = 1 iff 𝑡𝑖 is part of February 1, 2001

To make room for this in the LF of a sentence with a Partee-style tense,

we want to revise the entries from (57) and (58) and give PAST and woll a
second argument.
(62) ⟦PAST𝑛 ⟧𝑖,𝑔 = 𝜆𝑝 ∈ 𝐷𝑠𝑡 : 𝑔(𝑛) ∈ 𝐷𝑖 & 𝑔(𝑛) < 𝑡𝑖 & 𝑝 (𝑤𝑖 , 𝑔(𝑛)) = 1.
𝜆𝑞𝑠𝑡 .𝑞(𝑤𝑖 , 𝑔(𝑛)) = 1
120 T I M E

For example:
(63) Barbara turned off the stove on February 1, 2001.
LF: [ PAST 7 on February 1, 2001 [ PFV [ Barbara turn off the stove]]]
uttered felicitously only if
7 ∈ dom(𝑔𝑢 ) & 𝑔𝑢 (7) ∈ 𝐷𝑖 & 𝑔𝑢 (7) < 𝑡𝑢 & 𝑔𝑢 (7) ⊆ Feb 1, 2001,
and uttered truly iff ∃𝑒 [𝜏 (𝑒) ⊆ 𝑔(7) & 𝑒 ≤ 𝑤𝑢 & turn-off(𝑒, Barbara, the stove)]

Notice that there is still some room for context-dependency, in that the
speaker may be referring to either the whole of Feb 1st or to a proper
part of it (e.g. the morning of that day). But the role of context is greatly
reduced by the contribution of the adverb.
The revised, 2-place, entry for the future is analogous. When we have
frame adverbs with present tense, we also need a non-vacuous semantics
for PRES, but this is no different in the Partee-approach than it was in the
Priorian approach.
To conclude, let us highlight how the Partee-style, “referential”, anal-
ysis of tenses differs from the Prior-as-modified-by-Stalnaker-style anal-
ysis, and also what they have in common. The essential difference is that
Partee-style PAST𝑛 and woll𝑛 do not express existential quantification over
times, but instead rely on a contextually furnished variable assignment to
supply a particular time. In the Partee-approach, the denotations of past
and future clauses are always context-dependent; in the modified-Prior
approach, they only are if there happens be a silent restrictor together
with the existential quantifier. Both approaches assume that the extensions
of PAST (𝑛) and woll (𝑛) are sensitive to the evaluation time, and both assume
that these items shift the evaluation time for their complements. The two
approaches also agree on the treatment of PRES, which, according to both,
does not shift evaluation time and makes no semantic contribution other
than a presupposition when there is a frame adverb.
7 Embedded tenses

[what follows are the lecture notes from 2018; to be revised and retypeset]
122 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
21
4. Tense in embedded clauses: relative clauses
In this section we consider predictions that our analyses of tense from the previous sections make
about multiclausal sentences in which an embedded tense is in the scope of another tense.
Specifically, we will look at the interpretation of tenses in relative clauses . 39

For most purposes of this discussion, we can and will remain neutral about the choice between a
Partee-style ("referential") analysis of tenses and a more traditional Priorian ("existential
operator") analysis. We will cast the main discussion in the latter framework – specifically, we
will take PAST and woll to be contextually restricted existential quantifiers (as spelled out in
section 1 above), and assume that PAST and woll combine directly with a VP whose intension is
of type <s,t>. Footnotes or excursions will be added to show how the relevant issues play out
under the referential analysis and the division of labor between aspect and lexical verb meaning
that we introduced in section 3. In the text, we generally stick to stative VPs and assume that a
stative VP is true at an evaluation time if the state holds throughout that time.
Throughout this section, we will also reduce clutter by disregarding the world-component of the
evaluation index. I.e., we will pretend here that type s is just times (intervals) rather than world-
time pairs. (But this simplification will have to be undone again when we turn to tensed
complement-clauses in section 5.)

4.1. A predicted correlation between DP-scope and temporal interpretation

Surface sentences of the form [matrix-clause .. tense [VP . ..... [DP .. [relative-clause ... tense ....]]]]
generally should allow two different LFs, depending on where the DP is QRed: either within the
matrix VP or above the matrix tense. Since the relative clause and its tense are contained in the
DP, the scope of the DP determines whether or not the embedded tense is in the scope of the
matrix tense. Let us look at some concrete examples to see how the relative scopes of the tenses
affect the truth-conditions of the sentence.

We begin with an example with matrix future and embedded present.

(1) John will be married to someone who is famous.
To simplify matters, we disregard the covert restrictor that should come with each tense. The 40

two LFs then are (2a) and (3a), and the truth conditions that we compute for them, evaluating
the sentence at the utterance time tu, are (2b) and (3b). 41

(2) (a) LF: PRES woll [ [some one 7[PRES t7 be famous] ] 8[John be married to t8] ]
42

(b) ∃t [t > tu & ∃x [famous(x, t) & married(j, x, t)] ]

We will also discuss tenses embedded in complement clauses, in a later section.

39

But below we consider variants of the example that have added frame adverbs.
40

We use some hopefully self-explanatory abbreviations in the metalanguage, e.g., married(x, y, t) for x is
41

married to y at t.
W simplilify here by disregarding the pro-head-noun one (i.e, treating it as semantically vacuous).
42

Strictly speaking, this contributes a predicate like 'person' or 'human', which is also evaluated at different
 EMBEDDED TENSES 123

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
22
(3) (a) LF: [some one 7[PRES t7 be famous] ] 8[ PRES woll [John be married to t8] ]
(b) ∃x [famous(x, tu) & ∃t [t > tu & married(j, x, t)] ]
The scopal relations between the two existential quantifiers don't make a logical difference, but
what does differ between the two LFs is where the predicate famous is evaluated: In (2), famous
is evaluated at the same future time as the predicate married-to, whereas in (3), famous is
evaluated at the utterance time. So (2) says that at some future time, John will be married to a
person who will be famous then (a so-called "simultaneous reading"), whereas (3) says that he
will be married to a person who is famous now – let's call this a "now reading". (Neither LF, of
course, denies additional fame at other times than those required for its truth.) The two predicted
readings are logically independent, i.e., there are scenarios that make (2) true and (3) false, and
scenarios that make (3) true and (2) false. The English sentence can indeed be judged as true
under either one of these types of scenarios, so our predictions look satisfactory. 43

Our theory also makes correct predictions about what frame adverbs we can coherently add to
this kind of sentence and how their addition narrows down the range of verifying scenarios.
Consider, for example, the sentences in (4) - (6). (# marks a judgment of incoherence.)
(4) In 2020, John will be married to someone who is in jail then (in 2020).
(5) In 2020, John will be married to someone who is in jail now (in 2018).
(6) #In 2020, John will be married to someone who is in jail in 2019.
In order to interpret the adverbials, we need the 2-place meaning for woll and the non-vacuous,
2-place meaning for PRES (see sec 1.2. and fn 12). Spell out the details as an exercise. (4) has to
have an LF where the someone-DP scopes below the matrix woll, and (5) an LF where it scopes
above it. The opposite scope constellations can also be generated, but they receive unsatisfiable
truth conditions (or presuppositions) due to the adverbs. (Assume that tu is in 2018.)

Next we have an example with past tenses in both matrix and relative clause, (7). The two LFs
and their truth conditions are in (8) and (9).

times dependiing on the DP's scope. See below for an example with a more contentful head noun.
43
In a Partee-style theory with a division of labor between tense and aspect, the predicted truth conditions
also depend on the choices of (silent) aspect heads. We assume that, for type-reasons, the someone-DP
cannot be QRed to the edge of VP (below aspect), but it can scope either just above AspP or above woll.
If the aspects in both matrix and relative clause are imperfective, the predicted truth-conditions will be
essentially the same as in the Priorian theory, modulo the difference that the speaker is referring to a
specific future time. Also, if the DP scopes above woll, then no matter what aspects we have we predict a
now-reading, in the sense that fame must hold at tu for the LF to be true. When the DP scopes below
woll, however, this isn't necessarily a "simultaneous" reading. If both clauses are parsed with perfective
aspect, then the truth-conditions can be satisfied if the speaker's intended future topic time contains a
period marriage and a period of fame, but not necessarily overlapping or in any specific order. We may
wonder if such a truth-condition is attested, and if not, may want to constrain the system in such a way
that at least the embedded stative must be parsed with imperfective. Be that as it may – we do still make
the prediction that a now-interpretation for the embedded clause is contingent on scoping the DP above
woll. This prediction is not affected by the switch from Priorian to Partee-style framework. Neither are
the predictions regarding compatible frame adverbs.
124 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
23
(7) John was married to someone who was famous.
(8) (a) LF: PAST [ [some one 7[PAST t7 be famous] ] 8[John be married to t8] ]
(b) ∃t [t < tu & ∃x [ ∃t'[t' < t & famous(x, t')] & married(j, x, t)] ]
(9) (a) LF: [some one 7[PAST t7 be famous] ] 8[ PAST [John be married to t8] ]
(b) ∃x [ ∃t' [t' < tu & famous(x, t')] & ∃t [t < tu & married(j, x, t)] ]

Again there is a difference in how the time of fame relates to the time of marriage. (8b) requires
for its truth that the person be famous (at least) for some time before the marriage. (Though it
does not deny that the person is also famous during or after the marriage.) (9b), by contrast,
leaves the temporal relation between fame and marriage completely open. Both must obtain
somewhere before the utterance time, but the onset of fame could be before, during, or after the
marriage. In this example, one of the predicted readings entails the other; i.e., any scenario that
44

makes (8) true makes (9) true as well – but not the other way round. Again, speaker's judgments
about the sentence (7) are consistent with our predictions. In particular, (7) can be judged true in
a variety of scenarios, including a scenario where the (first) onset of fame is after the end of the
marriage (but before the speech time). Following Kusumoto (2005), we call this type of scenario
a "later-than-matrix" scenario.

Again we can add frame adverbs to sharpen intuitions and further test our predictions.
(10) In 2016, John was married to someone who was rich then (in 2016).
(11) In 2016, John was married to someone who was rich in 2015.
(12) In 2016, John was married to someone who was rich in 2017.
All these sentences are coherent, and we can account for this – provided again that we assume
suitable scopes for the DP. (12), in particular, receives satisfiable truth conditions only if the DP
scopes above the matrix PAST.

44
What about the Partee-style framework? Here we have two past tenses, each with its own variable
subscript, which in principle could be two occurrences of the same free variable or two different
variables. The speaker accordingly must have either just one or else two different salient ("topic") times
in mind. It turns out, however, that coindexing is only an option if neither past is in the scope of the
other; otherwise we generate a self-contradictory presupposition that the embedded topic time precedes
itself. If the embedded past is in the LF-scope of the matrix past, it must refer to a different, and earlier,
interval than the matrix past. The predicted truth-conditions for the LF with narrow DP-scope are then
essentially the same as in the Priorian analysis (and not detectably affected by choice of aspects).
In the LF where the DP scopes outside the matrix tense, the two past tenses can be coindexed or not and
can effectively refer to any two topic times as long as both of them are before tu. Even if they corefer, the
option of construing both predicates as perfective (if available for statives) implies that marriage and fame
need not stand in any specific temporal relation to each other. Overall, the range of possible verifying
scenarios for the wide-scope-DP LF ends up being the same as in the Priorian analysis. The crucial
prediction we are aiming to highlight here – namely that later-than-matrix scenarios can only verify the
wide-scope-DP LF – is made in the Partee-framework just as much as in the Priorian one.
 EMBEDDED TENSES 125

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
24
4.2. An overgeneration problem for present under past

There are a number of other combinations of a matrix tense and an embedded tense that we could
consider at this point, but mostly this will not teach us much more. The one exception concerns
sentences with a matrix past and an embedded present or will.
(13) John was married to someone who is famous.
(14) John was married to someone who will be famous.
Given that will spells out PRES + woll, both (13) and (14) really are instances of PRES embedded
under PAST. We focus our discussion on (13), and you should be able to work out how it extends
to (14).
As with the earlier examples, our theory predicts a scopal ambiguity due to multiple possible
landing sites for QR of the someone-DP.
(15) (a) LF: PAST [ [some one 7[PRES t7 be famous] ] 8[John be married to t8] ]
(b) ∃t [t < tu & ∃x [famous(x, t) & married(j, x, t)] ]
(16) (a) LF: [some one 7[PRES t7 be famous] ] 8[ PAST [John be married to t8] ]
(b) ∃x [famous(x, tu) & ∃t [t < tu & married(j, x, t)] ]
In (15), famous is evaluated at the same (past) time as the predicate married-to, and in (16),
45

famous is evaluated at the utterance time. Unfortunately for our theory, only (16) corresponds to
an attested reading of the English sentence. Present under past can only have a "now reading".
(13) doesn't have an additional "simultaneous" reading with the truth-conditions (15b) that are
expressed by LF (15a). This fact is further highlighted by the deviance that results from adding
simultaneity-forcing adverbs as in (17).
(17) # In 2016, John was married to someone who is famous at that time/then (in 2016).

So here we have a problem of overgeneration. How might we fix it? For the time being, we will
just hint at three general directions for a solution, without trying to make any one of them
precise. We will return to this task after we have gathered a bit more evidence that tells us which
directions are more likely to be correct.
Since the problem stems from the fact that we generate the LF in (15), one idea is to block the
generation of this LF. In (15), PRES is in the scope of PAST. Perhaps this scope constellation is
for some reason not allowed. Stowell (1993, 1995) suggested an analogy with polarity-sensitive
items such as any and some, which mean the same thing but are subject to different distributional
constraints at LF. any is a negative polarity item (NPI) and as such is required to be in the scope

45
Again, the Partee-framework yields essentially the same prediction. In particular, since present tense in
that framework too is vacuous (except for possible presuppositions), a present AspP within the scope of
the matrix past tense will be evaluated at the past topic time that the speaker intends for the matrix
predicate. Whether this amounts to a genuinely simultaneous reading depends again on the choice of
embedded aspect, but either way, it is a reading according to which fame holds in the past and need not
hold at tu. So the overgeneration problem carries over.
126 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
25
of a negation; some is a positive polarity item (PPI) and as such is required to be outside of the
46

scope of negation. Stowell called the present tense an "anti-past-polarity item", meaning an item
that is barred from occurring in the scope of PAST at LF – just as some is barred from occurring in
the LF-scope of not. If we place some in the surface c-command domain of not, as in John didn't
solve some problems, QR must apply to give it wide scope at LF, and only an inverse-scope
reading is therefore attested. Similarly, Stowell suggested, if PRES is in the surface c-command
domain of PAST, as in our sentence (13), the only way to satisfy its anti-past polarity requirement
is to change scope relations by means of covert movement – in this case by QRing the DP that
contains the PRES above the matrix PAST. So the only licit LF for (13) will be (16) – which
indeed is the LF that captures the attested meaning.

A second, completely different, approach to the problem says that (15a) is basically a fine LF –
except it doesn't get pronounced as (13). Rather, it gets pronounced as John was married to
someone who was famous. In other words, when PRES is in the LF-scope of PAST, it gets
pronounced like a past tense. (This is reminiscent of the "Sequence of Tense" rules of traditional
grammar.) If a PRES that is in the LF-scope of PAST can't be pronounced as a morphological
present, then a morphological present can't be parsed as a PRES that is in the scope of PAST at LF.
So again we get rid of the overgeneration problem and predict correctly that the surface sentence
(13) doesn't have the reading in (15b). This second approach places the burden on a theory of
morphological spell-out which is far from trivial to work out.

Yet another direction to take is to question our assumptions about the semantics of present tense.
So far we have assumed that PRES is either vacuous or, if restricted by an adverb, introduces a
presupposition about the evaluation time. Either way, PRES doesn't shift the evaluation time, and
therefore the sister-VP of a PRES is evaluated at the same time as the material right above it,
whatever time that may be. Our semantics recognizes no intrinsic connection between PRES and
the utterance time tu. Only when a PRES happens to be topmost in a matrix clause (or embedded,
but not under any operators that shift evaluation time), will its sister-VP be interpreted at tu.
This is then due to the Utterance Rule (see (6) in sec 1.1.) and not to the meaning of PRES. But
perhaps what the data in (13) and (17) should be teaching us is that this was wrong. Even when
embedded under an evaluation-time shifter such as the matrix PAST in LF (15a), PRES seems to
select tu as the evaluation time for its VP. So perhaps we must refer to tu in the semantics of
PRES itself rather than just in the Utterance Rule. On this type of approach, (15a) would be a
well-formed LF and also a possible parse of the surface sentence (13), but it wouldn't have the
meaning in (15b).

4.3. Counterexamples to the correlation of DP-scope and temporal interpretation

Even though the overgeneration problem we identified in section 4.2 is the most glaring
shortcoming of our current theory, we will set it aside for a while. Instead, we now take a closer

This is grossly oversimplified. (For one thing, NPIs can also be in the scope of downward-entailing
46

operators other than negation.) We are just pointing to Stowell's analogy here, not being serious about
NPIs and PPIs.
 EMBEDDED TENSES 127

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
26
look at the predictions that we drew out in section 4.1. concerning the systematic connection
between DP-scope and temporal interpretation in relative clauses. Let's remind ourselves of the
predictions.
(18) Prediction 1: In a sentence of the (surface) form
[matrix-clause .. past/will [VP ..... [DP .. [relative-clause ... present ....]]]],
the DP must scope higher than the matrix PAST or woll in order to get a now reading for
the VP in the relative clause.
(19) Prediction 2: In a sentence of the (surface) form
[matrix-clause .. past [VP . ..... [DP .. [relative-clause ... past ....]]]],
the DP must scope higher than the matrix PAST for the sentence to be true in a later-than-
matrix scenario.

We will consider two types of examples that turn out to violate these predictions. One is a kind
of example discussed by Keshet (2010) , in which the head noun of the relative clause gets
47

evaluated at a different time than the predicate in the relative clause.

(20) Five years ago, Jill married a 30-year-old who made partner two years later.
It is easy to understand (20) as saying that Jill’s spouse was 30 at the time of marriage and that
two years later, at the age of 32, the spouse made partner. Given the adverb five years ago, we
can conclude that the spouse is 35 at the time of utterance. The later-than-matrix reading of the
relative clause requires the object DP to take scope above the tense, allowing it to be interpreted
relative to tu. But then the noun phrase 30 year old will also be interpreted relative to tu and
we’ll get the set of individuals who are 30 years old at the time of utterance. Jill’s spouse is not
in that set. Summarizing, the head noun interpretation requires the object DP to be in the scope
of matrix past tense. But the relative clause needs to be outside the scope of the matrix tense,
given the theorem in (19). This contradiction reveals a flaw in the system of interpretation.

The second type of example, due to Kusumoto (2005) , contains a negative polarity item in the
48

49
DP whose licenser is below the matrix tense, as in (21).
(21) Next year, he will try not to rent to anybody who lives in this building now.
(22) They failed to write a single law that was signed by the Governor.
The point of using NPIs (here anybody, a single) is to confine scope of the DP. Since the NPI
must be in the scope of its licencer (here not, fail), it is ipso facto in the scope of the matrix tense.
In (21), NPI anybody must be in the scope of negation and hence it must be in the scope of woll.
But lives is interpreted relative to the time of utterance, which as the theorem in (18) says,

47Keshet, E. (2010) “Situation Economy,” Natural Language Semantics 18.4:385-434

Keshet, E. (2011) “Split Intensionality: A New Scope Theory of De re and De dicto,” Linguistics and
Philosophy 33.4:251-283
Kusumoto, K. (2005) “On the Quantification over Times in Natural Language,” Natural Language
48

Semantics 13.4:317–357
(21) and (22) were provided by Roger Schwarzschild. See below for Kusumoto's own examples.
49
128 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
27
requires the DP to have scope above tense.
(22) describes a dysfunctional state government in which the legislature writes laws and then
nothing happens to them because when they reach the governor’s desk, they don’t get signed.
NPI a single must be in the scope of fail and hence it must be in the scope of PAST on fail. The
potential signings are later than the writing, hence the relative clause has a later-than-matrix
interpretation, which as the theorem in (19) says, requires the DP headed by a single to have
scope above the tense.

Homework: Kusumoto uses failed to rather than didn't in order to make sure that the semantic
negation is clearly within the scope of the matrix PAST and not e.g. right above it. (If the
negation were above the tense, the NPI could be licensed by scoping the DP between them, and
this way we could capture the desired reading that's true in the later-than-matrix scenario.)
However, this makes sense only within a Partee-style referential approach to tense, not in the
modified-Prior framework we have employed here in the main text. To see the issue, consider a
50

simple fail-sentence such as (i).

(i) She failed to turn off the stove.
A reasonable semantics for fail is (ii).
(ii) [[fail]]i = λu ∈ D . u = 0
t
This says that fail is means the same as not, it just has a different syntax. 51

(a) Show that, in the context of the Partee approach to tense (as spelled out above in the
Appendix to section 3), this meaning for fail allows us to assign a reasonable syntactic
analysis and intuitively appropriate interpretation to example (i). Explain why the same
is not true for the modified Priorian approach.
(b) Analyze example (22) within the Partee approach and spell out how it makes Kusumoto's
point.
(c) [To be done after you have read to the end of section 4.4.2.] Show how the problem is
solved in the "extensional" framework (Partee-style version) that we introduce in the next
section below (incl. footnote).

Kusumoto's original examples are the past-under-past sentences (23) and (24), to which we can
also add present-under-past variants (25) and (26).
(23) (At the audition last month), I tried not to hire anybody who put on a terrible performance
(tonight).
(24) (At the Open House last week), she failed to talk to any prospective student who (later)
decided to come to UMass.

Kusumoto's paper does employ a version of the Partee approach, so this is not a criticism of her
50

argumentation.
Perhaps there is an additional modal component to the meaning of fail, something like 'didn't, but should
51

have'. Even so, this won't affect the argument regarding the scope of the negation in fail.
 EMBEDDED TENSES 129

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
28
(25) (At the Open House last week), she failed to talk to any prospective student who is still
undecided (today).
(26) (At the audition last month), I tried not to hire anybody who is putting on a terrible
performance (tonight).
The logic of the argument is the same as above. The (only available) licenser for the NPI any is
clearly within the scope of the matrix past tense. So if the any-DP wanted to QR above the
matrix tense, it would also have to QR above its NPI-licenser. So our theory predicts that none
of these examples should be verifiable by later-than-matrix scenarios, and with the parenthesized
adverbs, they should be incoherent. (On Stowell's proposal, the examples with embedded
present tense (i.e., (25), (26)) should even be plainly ungrammatical.) But (at least for
Kusumoto's informants ), these sentences are coherent descriptions of later-than-matrix
52

scenarios. Kusumoto offers the following verifying scenarios. "Suppose we are watching a play
with a casting director. Some of the cast members are very bad and the play is a failure. The
casting director can truthfully say something like [(23)], claiming no responsibility for the failure
of the play." (Kusumoto 2005, p. 327) For (24), "..., suppose that ten prospective students
53

showed up at the UMass open house, all of whom had not decided whether to come to UMass
yet. A faculty member talked to only five of them, and none of them decided to come. Among
those who she failed to talk to, four decided to come to UMass. In this situation, sentence [(24)]
is judged true." (loc.cit.)

Kusumoto's conclusion is that, to avoid these false predictions, we need a theory which allows
embedded tenses to be evaluated with respect to the utterance time, even when these embedded
tenses are in the LF-scope of higher tenses that shift the evaluation time for their sisters. This
situation should remind you of a problem we discussed earlier in the semester: the "third
readings" of DPs in the scope of modal operators. Indeed we will join a widespread consensus in
the field that the two problems are the same and have a single solution.

4.4. Relative clauses in a framework with index variables in the object language

4.4.1. The extensional system (mostly review ) 54

We return to the basic system used in Heim & Kratzer up to chapter 11. The interpretation
function is relativized only to an assignment function, not to any other evaluation parameters
such as a world, a time, or an index. The semantic rules are Functional Application, Predicate
Abstraction, and Predicate Modification, in their formulations from the earlier part of H&K.
There is no rule of Intensional Functional Application. The only ingredient of intensional
semantics that we do retain is the expanded type system and ontology. We have a third basic

Roger Schwarzschild reports disagreement with these judgments.

52

Since Kusumoto's example (23) has past tenses in both clauses, she imagines the casting director to be
53

speaking after the performance. For the pres-under-past version (26), he would have to be speaking
during the play.
54
from ch. 8 of 2011 lecture notes
130 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
29
type besides e and t, the type s. Ds is the set of all indices, i.e., world-time pairs. A new
assumption (different from both our previous extensional and intensional systems) is that
assignment functions are now (partial) functions from object-language variables (natural
numbers) into De È Ds, i.e., their possible values include world-time pairs as well as individuals.

There are a number of innovations in the lexicon and in the syntax. As for the lexicon, the main
change concerns the treatment of predicates (verbs, nouns, adjectives). They now all get an
additional argument, of type s. 55

(10) [[tired]] = λi ∈ Ds. λx ∈ De. x is tired in wi at ti

[[like]] = λi ∈ Ds. λx ∈ De. λy ∈ De. y likes x in wi at ti
[[person]] = λi ∈ Ds. λx ∈ De. x is a person in wi at ti
etc.
This also applies to attitude predicates, modals, and tenses. We illustrate with PAST and the
56

counterfactual modal would . 57

(11) [[would]] = λi ∈ Ds. λp ∈ D<s,t>. λq ∈ D<s,t>.

"w [ p(w,ti) = 1 & w is otherwise similar to wi ® q(w,ti) = 1 ]
(12) [[PAST]] = λi ∈ Ds. λp ∈ D<s,t>. λq ∈ D<s,t> : $t [ t < ti & p(wi,t) = 1 ].
$t [ t < ti & p(wi,t) = 1 & q(wi,t) = 1 ]
There is no change to the entries of proper names, determiners, or truth-functional connectives;
these keep their purely extensional ("s-free") types and meanings.

Finally, there are two new kinds of abstract (i.e., unpronounced) morphemes. One is a series of
pronouns of type s ("index pronouns" or "world-time pronouns"). We write them as pron, with a
numerical subscript n. (This makes them look just like covert pronouns of type e, but the
environment will disambiguate. Besides, we won't be using examples with both types of
pronouns in them.) Their semantics is what you expect: they get values from the assignment
function.

55
The decision to make the index-argument the predicate's first (lowest) argument is arbitrary, and nothing
hinges on it. For all we know, it could be the highest argument, or somewhere in between.
56
(12) is the version for the Priorian framework (as modified to accommodate frame adverbs). A Partee-
style tense now gets an entry like (i):
(i)(a) full version that accommodates frame adverbs:
[[PASTn]]g = λi ∈ Ds.λp ∈ D⟨s,t⟩: g(n) is a time interval & g(n) < ti & p(wi, g(n)) = 1.
λq ∈ D⟨s,t⟩. q(wi, g(n)) = 1.
(b) simplified (no adverb):
[[PASTn]]g = λi ∈ Ds: g(n) is a time interval & g(n) < ti.
λq ∈ D⟨s,t⟩. q(wi, g(n)) = 1.
Rewriting the entries for aspect heads is also routine, e.g.:
(ii) [[PFV]] = λi ∈ Ds. λP<v,t>. $e [P(e) = 1 & t(e) Í ti & e ≤ wi].
57
We take this modal to be accidentally homophonous with the spell-out of PAST+woll. That's probably
not right, but this is a complex research area beyond the scope of this class. (See Iatridou's 2000 LI-
article and current work by Kai and Sabine.)
 EMBEDDED TENSES 131

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
30
The other new item is a semantically vacuous operator, OP, which moves and leaves a trace of
type s. Its syntactic properties are such that it must end up in C or right below a functional head
in the "clausal spine" between C and V, and it must get there by a very short movement, a kind
of "head movement". We are leaving this rather vague. – We also stipulate that a complete
58

(matrix) sentence must not contain any free variables of type s and must receive a denotation of
type <s,t>. This implies that there is a CP layer in matrix clauses and there is always an instance
of OP in the matrix C. 59

We have everything in place now to return to the discussion of tense. Let's just finish our review
session with a quick reminder of the analysis for third readings in modal contexts.
(13) If everyone in here were outside, Building 56 would be empty.
LF (ignoring tense):
OP1 [[would-t1 [if OP2. everyone in-here-pro1 be outside-t2]] [OP3. Bg56 be empty-t3]]

To make interpretable the node immediately above each lexical predicate or modal, some
pronoun or trace must occupy its innermost (type-s) argument position. We can't just use
pronouns everywhere, because then they would all remain free. There is no free insertion of
lambda-binders in this theory. All binding depends on movement, so we need to generate
operators in at least some of the argument positions and we must move them. The semantic type
of the modal would furthermore demands two arguments of type <s,t>, and constituents of this
semantic type can only be created by moving an operator to their edges. The complete sentence
also must be of type <s,t> and therefore have an operator at the very top. These strictures
together determine almost everything in the LF in (13) – except the fact that the if-clause must
have a pronoun in the subject and an operator in the predicate and not the other way round. We
attribute the latter fact to (vaguely stated) syntactic constraints on the operator's movement path
and landing site (pending a deeper explanation).

4.4.2. Relative clauses and their tenses

The rule of Predicate Modification (PM) has not changed from the original H&K system. It
applies to two phrases of type <e,t>. Therefore a relative clause must have this type in order to
be able to combine with its head noun. This means that we don't want an OP at the edge of a
relative clause, neither below nor above the moved relative pronoun. (We would get the wrong
semantic type either way, be it <e,st> or <s,et>.) Instead we have to use a pronoun in the
argument structure of the highest predicate/modal/tense within the relative clause. This pronoun
is free in the constituent to which we apply PM. (Though, of course, it can be – in fact, must be
– bound eventually in the higher structure.) If the relative clause is past or future, this free

As noted in the lectures on the "third reading", there is a substantial recent literature exploring various
58

ways to give principled explanations for Percus's "Generalization X" and similar constraints on the
distribution and binding of world variables. Here we just assume that some principles or other are in
place to prevent overgeneration (e.g., the unattested reading of Percus's Mary thinks that my brother is
Canadian).
It also requires rewriting the definition of truth/falsity of an utterance. Instead of (6) in section 1.1, we
59

now need this: An utterance of a sentence φ that is made in a world w at a time t is true iff [[φ]](w,t) = 1.
132 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
31
pronoun is the argument of PAST or woll; if the relative clause is in the present tense (which we
still treat as vacuous), the free pronoun is the argument of the verb (or the adjective/noun after a
copula) . On the other hand, at the edge of the VP-complement to PAST or woll we always need a
60

moved OP, because these temporal operators select for arguments of type <s,t>. (Furthermore
61

we need to merge an OP in the inner argument position of the matrix tense, so that we can move
this to the matrix C and satisfy our requirement that the matrix clause denote a proposition of
type <s,t>.)
Let us illustrate all this with one of our examples from section 2.1. 62

(14) John will be married to someone who is famous.

(15) LF1: narrow DP scope, local binding:
OP1 woll-t1
OP2 [ [some who6[PRES t6 be famous-pro2]] 7[John be married-t2 to t7] ]
li. ∃t [t > ti & ∃x [famous(x, t) & married(j, x, t)] ]
("simultaneous" reading, i.e., famous when married)
(16) LF2: narrow DP scope, non-local binding:
OP1 woll-t1
OP2 [ [some who6[PRES t6 be famous-pro1]] 7[John be married-t2 to t7] ]
li. ∃t [t > ti & ∃x [famous(x, ti) & married(j, x, t)] ]
("famous now" reading)
(17) LF3: wide DP scope:
OP1 [ [some who6[PRES t6 be famous-pro1]]
7[woll-t1 OP2 [John be married-t2 to t7]]
li. ∃x [famous(x, ti) & ∃t [t > ti & married(j, x, t)] ]
("famous now" reading)
We are still considering both scopal relations between the some-DP and the matrix tense (for all
we know, nothing rules one out). But the reading we obtain by scoping the DP high (LF3) is
now equivalent to one of the LFs in which the DP scopes low, namely LF2, where the world-
time pronoun in the relative clause is non-locally bound from the matrix C.

60
Or the argument of Asp, if we use the Davidsonian framework in which VPs are type <v,t>.
61
Could we avoid this OP and instead move the tense operator itself from the argument position of the
verb"? That kind of syntax has indeed been explored by Junri Shimada
http://research.nii.ac.jp/~kanazawa/semantics/2007/0817/Head_Movement_Binding_Theory_Ph
rase_Structure.pdf, who credits the idea to Kai von Fintel http://web.mit.edu/fintel/choicepoints.pdf.
62
Here we simplify again in various ways. First, we omit the world-component of the index. Second, we
ignore the tenses' restrictors. Third, we disregard the dummy head noun 'one'. Notice that the relative
clause nevertheless needs to be of type <e,t>, since this is the type that the determiner some combines
with.
Please draw yourself some trees, they will be more readable than the bracketed strings. Some notational
conventions I have used to improve readability at least a little: low numbers (1, 2, ...) and boldface for
variables of type s, higher numbers (6, 7, ...) and plain font for variables of type e; hyphenating type-s
arguments with the predicates they saturate; small italics for semantically vacuous items.
 EMBEDDED TENSES 133

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
32
This equivalence is the key to the solution of Kusumoto's problems. What the system achieves is
the freedom to evaluate the embedded tense with respect to the utterance time even when it is in
the scope of the matrix tense. Therefore, we can trap the DP's scope below the matrix tense – for
example, by introducing an NPI determiner and a suitably low licenser for it – and still evaluate
the relative clause at the matrix index.

4.4.3. Present under past again: the overgeneration problem is still with us
Switching to the extensional framework has addressed Kusumoto's argument against the scopal
theory of later-than-matrix interpretations, but it has not done anything yet to fix our
overgeneration problem. We still predict an unattested simultaneous reading for present
embedded under past. Here is its new LF.
(18) John was married to someone who is famous.
(19) LF with narrow scope DP and local binding:
OP1 PAST-t1
OP2 [ [some who6[PRES t6 be famous-pro2]] 7[John be married-t2 to t7] ]
li. ∃t [t < ti & ∃x [ famous(x, t) & married(j, x, t)] ]
(unattested "simultaneous" reading, i.e., famous when married)

But we have made some progress. For one thing, we have learned something about how not to
fix the problem. We don't want to legislate against the scopal relations that we see in (19).
Whatever we do must not rule out the scopally isomorphic LF in (20).
(20) LF with narrow DP scope and non-local binding:
OP1 PAST-t1
OP2 [ [some who6[PRES t6 be famous-pro1]] 7[John be married-t2 to t7] ]
li. ∃t [ t < ti & ∃x [ famous(x, ti) & married(j, x, t)] ]
(attested "famous now" reading)

This means that Stowell's analogy with polarity sensitivity is coming to look unhelpful. What
about the traditional idea that a Sequence of Tense rule governs morphological spell-out and
ensures that the LF in (19) is paired with the PF John was married to someone who was famous?
This still also looks rather unappealing, given the lack of locality between the affected verb form
and the environment that must trigger the rule. (Again, the rule must not apply indiscriminately
to both (19) and (20), yet these structures are indistinguishable in the immediate local vicinity of
the affected verb.) Nevertheless, something like a Sequence of Tense rule, a non-local
morphological agreement mechanism, turns out to be the favored solution in much of the recent
literature. We present this in the next section. And we will see that the switch to our current
63

extensional framework, while not by itself the solution to the problem, was a necessary
prequisite for it. The solution to be presented is one that could not have gotten off the ground
without a syntax that posits world-time pronouns and traces as part of syntactic representations.

For a dissenting view and counterproposal, see two recent papers by Altshuler & Schwarzschild
63

(Amsterdam Colloquium 2013, Sinn und Bedeutung 2013).

134 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
33
4.5. Sequence of Tense as feature-agreement under semantic binding
Kratzer (1998, SALT) proposed that the phenomenon of Sequence of Tense has the same nature
and explanation as a certain phenomenon in the morphosemantics of personal pronouns that she
dubbed "fake indexicality". We take a brief excursion into this topic before we return to tense.
64

4.5.1. Fake indexicals and feature transmission

Fake indexicals are first or second person pronouns that are interpreted as bound variables. (This
makes them "fake" because a genuine indexical refers, by definition, to a specific contextually
determined individual.) An example is (21), on the prominent "sloppy" reading that denies that
other people brushed their teeth.
(21) Only I brushed my teeth.
The challenge such examples pose emerges, for example, if one tries to extend the
presuppositional account of gender features to other phi-features including person. A
65

presuppositional semantics for 1st-person presumably would say that [[myn]]g is undefined
unless g(n) is the speaker. But this would make it impossible for my in (21) to take on a range of
alternative values, as it has to if it is to be bound by the quantifier only I.
We will not engage here in a serious discussion of fake indexicality, just sketch the analysis
developed in Kratzer (1998) and related work. This assumes that grammar does not produce a
perfect match between semantically interpreted and phonologically realized phi-features. In
particular, what we witness in (21) is a 1st-person feature on my that is present at PF but absent
at LF – hence not contributing to the meaning of this sentence (whatever the actual semantics of
1st person may be). Implementations use either a mechanism that deletes certain base-generated
features in the LF-branch of the derivation while retaining them in the PF-branch , or else a
66

mechanism that adds (copies) features in the PF-branch onto nodes that are feature-less
underlyingly and at LF. Either way, the mechanism is crucially sensitive to a syntactic
representation of semantic binding relations (such as coindexing) and it (probably ) operates
67

non-locally. For concreteness we state the following rule (22), which employs the concept of
"binding" defined in (23). 68

64
Philippe Schlenker pursued a related but distinct approach to SOT-phenomena in his 1999 MIT thesis.
For Schlenker, the essential analogy was between SOT and indexical shifting (in languages like Amharic)
– or rather, actually, between indexical shifting and the absence of SOT (in non-SOT languages such as
Japanese and Russian). How different the two approaches ultimately are depends on how one views the
relation between fake indexicals and shifted indexicals, a question on which both Kratzer and Schlenker,
as well as a number of later authors, have taken evolving positions over the years.
65
Cooper, Heim & Kratzer. If one doesn't treat 1st-person pronouns as variables in the first place, but as
indexicals in the sense of Kaplan, one has an even more basic problem. They cannot be bound variables
then.
66
see von Stechow 2003
67
Kratzer (1998, 2005) argues that it obeys certain locality constraints after all, but this is controversial.
Most other authors (Schlenker 1999, von Stechow 2003, Heim 2005, 2008, Wurmbrand 2015) assume or
argue for non-local versions.
68
See also H&K p. 263.
 EMBEDDED TENSES 135

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
34
(22) Feature transmission under semantic binding:
In the derivation of PF, copy the features of a phrase onto any pronouns and traces it
binds.
(23) a binds bn iff a's sister node (not counting semantically vacuous material ) has the
69

daughters n (a binder index) and g, such that g dominates bn (and does not dominate any
other occurrence of n that c-commands bn).
Applying this rule to examples requires a few more ancillary assumptions about feature traffic.
E.g., for (21) we must assume that the 1st-person feature base-generated on I first percolates to
the quantifier only I, from whence it then can be transmitted down to the possessive pronoun by
rule (22). So the derivation of (21) goes like this.
(24) base-generate: [only [pro7 1st]] brush pro8's teeth
subject moves spec-V to spec-I: [only [pro7 1st]] 8[t8 brush pro8's teeth]
derivation to LF: no further changes
derivation to PF: percolate in only-DP: [only [pro7 1st]]-1st 8[t8 brush pro8's teeth]
transmit under binding: [only [pro7 1st]]-1st 8[t8-1st brush pro8-1st's teeth]
The point is that the trace and possessive pronoun bound by only I have 1st-person features at
PF, but not at LF. At LF, the only 1st-person feature is on the subject I, where indeed it is
interpreted and constrains the reference of the free variable 7 to pick out the speaker. The trace
and bound pronoun are feature-less variables and thus have well-defined denotations under any
assignment that assigns something to the variable 8.

4.5.2. Interpreted and uninterpreted tense features

Let us return to tense now and begin to spell out Kratzer's analogy between fake indexicals and
Sequence of Tense. The first step is to clarify the relation between abstract tense morphemes
(such as PAST) and tense morphology (such as an -ed suffix or a suppletive form like was). The
idea is that this is similar to the relation between interpreted and uninterpreted phi features.
There is the item, or – as we will now say – the "(interpreted) feature" PAST, which is part of the
underlying structure and of the LF, and which is semantically contentful. (Its meaning is what
we have been assuming, e.g., (12) in section 4.4.1). And there is an uninterpreted twin of it –
we'll write it PAST in little italics – which shows up in various places at PF and informs the spell-
out of verbs in its vicinity. The actual spell-out rules (e.g. go PAST ® went) can work on very
local configurations (and we won't say much more about them here – that's what we have
morphologists for). But the mechanism by which the uninterpreted tense features get to be
where they are is sensitive to a not necessarily local structural configuration with abstract
ingredients like variables and binders. It is, in fact, the very mechanism that is behind fake
indexicals. Let us see how our rule (22) "feature transmission under semantic binding" can apply
to tense features.

This proviso becomes relevant later: the rule should apply also if there is a vacuous operator together
69

with the binder index. For the moment you can ignore it.
136 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
35
First some mono-clausal examples.
(25) [PAST-OP C] [John like-OP Mary]
move operators: OP1 [[PAST-t1 C] OP2 [John like-t2 Mary]]
percolate feature: OP1 [[PAST-t1 C]-PAST OP2 [John like-t2 Mary]]
transmit feature: OP1 [[PAST-t1 C]-PAST OP2 [John like-[t2-PAST] Mary]]
spell out: John liked Mary.
(26) [PAST-OP C] [John be rich-OP]
move operators: OP1 [[PAST-t1 C] OP2 [John be rich-t2]]
percolate feature: OP1 [[PAST-t1 C]-PAST OP2 [John be rich-t2]]
transmit feature: OP1 [[PAST-t1 C]-PAST OP2 [John be rich-[t2-PAST] ]]
spell out: John was rich.
We included the tense's restrictor (C) here to be fully explicit about the structure, but will go
right back to ignoring it below. In the copular example, the uninterpreted tense feature has ended
up a bit lower perhaps than we would want it. Tense gets spelled out on the copula after all, not
on the adjective. We won't fret over this here. 70

What about present tense? As long as we are treating it as vacuous, it isn't binding any variables
and thus can't be a source of transmitted features. The most natural move at this point is to
abolish the item or feature PRES altogether, and leave it to the morphology to spell out verbs
without a tense feature in the form that we traditionally call their "present tense" form . Given71

that present tense morphology is actually zero (once we factor out subject agreement), this is also
reasonable from a morphologist's perspective. But bear in mind that we are not really wedded to
the vacuous treatment of the present, but have entertained a non-vacuous version too (when we
considered frame adverbs). There might then also be an uninterpreted twin PRES of a non-
vacuous PRES, and the the analysis of John likes Mary might be (28). But here we go with (27)
for simplicity.
(27) John like-OP Mary
move operator: OP2 [John like-t2 Mary]]
spell out: John likes Mary.

The issue disappears, or at least changes, when we integrate aspect into the structure. The PAST feature
70

will then end up no lower than the type-s argument of the Asp head – in both copular sentences and those
with regular main verbs. When Asp is morphologically zero, the tense feature next to Asp gets spelled
out on the verb immediately below. For a recent discussion of verbal morphology and the role of
semantically vacuous auxiliaries like do and be, see Bjorkman's 2011 MIT thesis.
At least it does this in finite environments. We must assume that the morphology somehow knows
71

whether it is dealing with e.g. an infinitive or a participle, where all tense distinctions are systematically
neutralized.
 EMBEDDED TENSES 137

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
36
(28) [PRES-OP C] [John like-OP Mary]
move operators: OP1 [[PRES-t1 C] OP2 [John like-t2 Mary]]
percolate: OP1 [[PRES-t1 C]-PRES OP2 [John like-t2 Mary]]
transmit: OP1 [[PRES-t1 C]-PRES OP2 [John like-[t2-PRES] Mary]]
spell out: John likes Mary.

2.4.3. Sequence of Tense explained

Now we reap the benefits of this new morpho-semantic theory. Look again at our problematic
sentence John was married to someone who is famous. First, here is a derivation for the attested
"famous now" reading.
(29)(a) base-generate:
PAST-OP [ John be married-OP to [some one [who be famous-pro1]] ]

(b) move operators (note non-local binding of pronoun!):

OP1 PAST-t1
OP2 [ John be married-t2 to some one who6[t6 be famous-pro1]] ]

(c) from (b), by percolation and transmission:

OP1 [PAST-t1]-PAST
OP2 [ John be married-t2-PAST to some one who6[t6 be famous-pro1] ] ]

(d) from (b), by QR:

OP1 PAST-t1
OP2 [ [some one who6[t6 be famous-pro1]] 7[John be married-t2 to t7] ]

The only trace or pronoun that receives a PAST feature by transmission is the argument of the
matrix predicate be married, so the embedded verb remains without a tense feature and surfaces
as present tense.
Now let's convince ourselves that we no longer generate the unattested simultaneous reading.
An LF that expresses this reading must have the argument of the embedded predicate
semantically bound by the matrix tense, as in the LF in (30).
(30) OP1 PAST-t1
OP2 [ [some one who6[t6 be famous-pro2]] 7[John be married-t2 to t7] ]
But with semantic binding comes feature transmission, and therefore the embedded predicate
would then have to surface as a past tense form. We can't have semantic binding without a
morphological reflex. This is what is behind the phenomenon called "Sequence of Tense".
The system makes further predictions. One (closely related to the above) pertains to sentences
with embedded future. Consider (31).

(31) John was married to someone who will be famous.

Our old theory predicted, falsely, that this sentence, as spoken in 2016, could be verified by a
scenario in which John and Mary were married from 2004 to 2006 and Mary was famous from
138 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
37
2008 to 2010. (In other words, the temporal order in the scenario is marriage < fame <
utterance.) When the sentence was given an LF where the some-DP scoped below the matrix
past, the old theory made that LF true in the described scenario. As a matter of fact, however,
the sentence unambiguously entails the existence of a person who is famous sometime after the
utterance time.
The new theory generates the attested reading by coindexing the pronoun in the argument
position of woll with the topmost operator in the matrix. (Exercise: Show this.) (It doesn't
matter in this case how high the some-DP scopes.) For the unattested reading (that would be true
in the above scenario), the inner argument of woll would have to be bound by the matrix PAST.
In that configuration, however, the transmission rule applies and drops a PAST feature onto the
argument of woll. For morphology, this means that woll spells out to would, not to will. So this
is the LF of a different sentence, namely (32).

(32) John was married to someone who would be famous.

Indeed (32) sentence has a different meaning from (31) and can describe our scenario. (Perhaps
it's most felicitous with added adverbials, e.g., John was married to someone who would later be
famous, or In 2005, John was married to someone who would be famous in 2009.)

The new theory does not affect predictions for any of the sentences that have a matrix present. It
also does not affect predictions for sentences with matrix future, i.e. will – provided we assume
that woll (unlike PAST) is not a feature in the first place and has no uninterpreted twin-feature.
(So if there is a feature at all that enters into spelling out will, it's just PRES.)
As for configurations with matrix past, we have already looked at embedded present and
embedded future. What about matrix past embedding another past?

(33) John was married to someone who was famous.

We already know that one of the derivations for this surface configuration involves an
underlying structure that has no tense (the equivalent of vacuous present) in the lower clause.
This gave a simultaneous reading. (We call this the "SOT derivation".) But is that the only
derivation which outputs (33) at PF? Suppose we generate an embedded PAST in the base
structure. Then, after the operators have moved, we have a structure of the form in (34).
(34) OP1 PAST-t1
OP2 [ John be married-t2 to some one who6[PAST-pron OP3[t6 be famous-t3]] ]
I left the subscript on the pronoun unspecified, because we have choices there. We can have n=1
or n=2.
First consider n=1. Then the pronoun is bound by the highest OP and no features get transmitted
onto it. Each of the PAST's, however, transmits a PAST to the trace it binds, so we get PAST on t2
and t3, hence was married and was famous. The meaning has two independent quantifiers over
times before g(1), i.e., before the utterance time if this is used as a matrix sentence. So this is a
reading where the verifying marriage and fame can in principle be in any temporal order with
respect to each other. (Adverbs can be added to constrain it further.) We are happy to generate
 EMBEDDED TENSES 139

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
38
this derivation, since the surface sentence does allow this kind of interpretation.

Now consider n=2 in (35). Meaning-wise, this imposes a requirement that the fame held before
the marriage (i.e., either the fame began and ended before the marriage or it continued from
before the marriage into it or beyond). Since that kind of scenario is already allowed as a special
case of the truth conditions of the n=1 LF, it is hard to determine with truth-value judgments
whether the grammar ought to generate this as a separate reading. But it won't hurt, it seems.
Now let's look at the PF side for this case. Given what we have made explicit about percolation
and transmission, we get the following pre-spell-out structure in the lower clause.
(35) lower clause for (34) with n=2, after percolation and transmission:
... who6 [ [PAST-[pro2-PAST]]-PAST OP3 [t6 be famous-[t3-PAST] ] ]
The left-most PAST (the one on pro2) has been transmitted from the higher clause. The second
PAST, the one attached to the complex phrase [PAST-[pro2-PAST]], got there by percolation from
the PAST-head of that phrase itself. And the third PAST (on t3) has been transmitted from that
complex phrase. How does all this get spelled out? Presumably the only place where
morphology does something is in the predicate (be) famous, so we expect was famous. The
structure in the complex phrase headed by PAST presumably must stay silent, because there is no
verbal element there to carry tense inflection. So this is homophonous with the outputs of the
other two derivations – an okay prediction, if not one that we can distinguish empirically at this
point from another possibility that might be entertained as well, namely that the structure in (35)
is not well-formed or not spell-out-able at all (for some reason to be identified).
A question that tends to come up at this point is whether we should work out a morphology
which spells (35) out as a pluperfect, i.e., who had been famous. After all, this is intuitively the
English sentence which best matches the interpretation we computed for the n=2-version of (34):
John was married to someone who had been famous does entail fame before the marriage.
Morphologically, however, it would seem to take a bit of ad hoc machinery to get from (35) to
had been famous. We may have a more elegant way to generate the pluperfect. Assume (as was
briefly mentioned earlier) that our lexicon contains an item that is synonymous with PAST, but
syntactically different in that it is generated lower (below M, if any) and qualifies as "verbal" in
the sense that's relevant to whether morphology can realize tense inflection on it. Then we can
72

generate a minimal variant of (35):

(36) ... who6 [ [have-[pro2-PAST]] OP3 [t6 be famous-t3] ]
(matrix same as in (34))
The meaning of (36) is the same as with (35), but the morphology is straightforward and outputs
had. (The copula then surfaces as a past participle form, but this is among the further
73

morphological details that we are not attending to here.)

In the latter respect it is thus like woll.

72

We assume, as we did for woll, that have is not a feature and ipso facto doesn't have an uninterpreted
73

twin.
140 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
39
5. Embedded tense in complement clauses
We have only looked at relative clauses so far, but there are other multiclausal constructions to
consider in which tenses occur in the scope of other tenses. In this section, we look at
complement clauses to attitude and speech verbs.
A prerequisite for this discussion is the semantic analysis of verbs such as believe, think, know,
hope, tell, say. We rely on earlier in the semester, where we developed lexical entries like (1).
(1) For any world w,
[[think]]w = λp⟨s,t⟩. λx. ∀w′ [w' is compatible with x′s beliefs in w ® p(w′) = 1]
We adapt this to our current framework as follows. 74

(2) [[think]] = λi. λp⟨s,t⟩. λx. ∀w [w is compatible with x′s beliefs in wi at ti ® p(w, ti) = 1]

5.1 Simple cases: non-past matrix

Let's warm up with a sentence that has present tense in both clauses.
(3) John thinks that Mary has the key.
Since we have abolished PRES, the structure of (3) is quite simple. believe calls for an argument
of type <s,t>, so we must generate an OP in the lower clause. The representation after OP-
movement will be (4).
(4) OP1[John think-t1 OP2[Mary have-t2 the key] ]
This structure is interpretable as it stands, so it can be our LF and we compute the meaning in
(5a). It is also ready to be spelled out by the morphology, since there are no relevant features
here to percolate and transmit. Recall that verbs with no tense feature are spelled out (in finite
environments) as present tense.
(5) (a) λi. ∀w [w compatible with J′s beliefs in wi at ti ® M has the key in w at ti]
(b) morphology: [think-t1] ® thinks, [have-t2] ® has
Examples with a non-present tense in the embedded clause but a present in the matrix are also
straightforward, as you can verify yourself. Let's move on to the interesting cases, with non-
present tenses in the matrix.

The first case we examine is a future embedding a present.

(6) John will think that Mary has the key.
Again the structure after operator movement can serve without further derivation as LF and as
input to morphology (there are still no features to copy around).

74
There is another natural way to go, which is what you find in most of the contemporary literature
(Ogihara, Abusch, Kratzer, von Stechow, Kusumoto, etc).
(i) [[think]] = λi. λp⟨s,t⟩. λx. ∀i′ [i′ is compatible with x′s beliefs in wi at ti ® p(i′) = 1]
(i), unlike (2), quantifies over both worlds and times. The difference between (2) and (i) becomes
important in the analysis of so-called de se readings, but it won't matter for the modest purposes of this
introduction.
 EMBEDDED TENSES 141

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
40
(7) OP1[woll-t1 OP2[John think-t2 OP3[Mary have-t3 the key] ] ]
(a) λi. $t [ t > ti &
∀w [w compatible with J′s beliefs in wi at t ® M has the key in w at t] ]
(b) morphology : [woll-t1] ® will, [have-t2] ® has
75

The meaning we computed in (7a) amounts to a "simultaneous" reading. The worlds compatible
with John's beliefs at the future time we are talking about are worlds in which Mary has the keys
then, at the time of his thinking. This is indeed what the English sentence (6) means. In fact, it
is its only possible reading. In distinction to the case of a present relative clause in a future
matrix, there is not an additional reading here on which the present in the complement clause
evaluates at the utterance time. The theory predicts this unambiguity. Because the attitude verb
selects an argument of type <s,t>, we must generate an operator with the lower verb and move it
to the embedded C. We cannot put a pronoun there and bind it non-locally. That would result in
a type-mismatch, because the sister of think would be type t.

If we do want to talk about a belief that someone will hold in the future about our current
utterance time, how do we express this in English? We have to use an embedded past.
(8) scenario: Arriving at the office one morning, Mary realizes that she left her keys at
home. But as luck would have it, someone forgot to lock up the night before, so she can
get in anyway. As she is entering the office, she thinks about the boss (John), who will
arrive later and, not knowing any of the above, will not be surprised to find her sitting at
her desk. She says:
John will think that I had my key.
(The analysis of this sentence in our theory is straightforward; do it as an exercise.) These
examples provide an illustration of Ogihara's (1996) principle of "temporal directionality
isomorphism", i.e., the generalization that tenses in an attitude complement must always reflect
the attitude holder's temporal perspective on the embedded event (rather than the speaker's).
Kratzer (1998) and Kusumoto (2005) proposed to derive temporal directionality isomorphism
from the semantic type of attitude verbs, and we have implemented this approach.

5.2. Past matrix and sequence of tense

Next we ask ourselves what underlying structures we can build with a matrix PAST and what will
happen to them in their LF and PF derivations and their semantic and morphological
interpretation. First let's base-generate PAST operators in both clauses.
(9) OP1[PAST-t1 OP2[John think-t2 OP3[PAST-t3 OP4[Mary have-t4 the key] ] ] ]

Interpretability requirements once again determine the presence and landing sites of the OP's. (9)
is interpretable and expresses what is called a "back-shifted" ("earlier-than-matrix") reading.

75
think does not spell out as thinks, because it is not in a finite VP. woll (like the other auxiliaries of
syntactic category M) governs infinitival morphology on its complement VP, and it is irrelevant in this
case what tense features, if any, might have been transmitted to the type-s-argument of V. Even when
there is a PAST on the inner argument of an M (as in 'John could see the ocean', or 'John thought he would
see the ocean'), the VP still spells out as a bare infinitive.
142 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
41
(10) λi. $t [ t < ti &
∀w [w compatible with J′s beliefs in wi at t ® $t' [ t' < t & M has the key in w at t'] ]
What about morphology? Each PAST operator certainly will transmit a feature to its trace, and
the outcome will be past tense morphology in both clauses.
(11) OP1[[PAST-t1]-PAST OP2[J think-t2-PAST OP3[ [PAST-t3]-PAST OP4[M have-t4-PAST key] ] ] ]
spell out: John thought that Mary had the key.
So we generate a back-shifted reading for this surface sentence of English, and this is good. The
sentence can be understood in this way. Imagine, e.g., that we are talking about the story in (8)
at a later time. We could then say John thought that Mary had her key meaning that when John
arrived in the late morning, he thought that Mary had her key at the earlier time of her arrival.

But the same surface sentence also has a "simultaneous" reading. In fact, this may be the most
prominent reading out of context. The simultaneous reading is standardly attributed to Sequence
of Tense. Let's see how our implementation of SOT as feature-transmission-under-binding
might cover this case. To get the simultaneous truth-conditions, our LF must not have a PAST
operator in the lower clause, so we don't base-generate one there. The structure needs to be as in
(12).
(12) OP1[ PAST-t1 OP2[ John think-t2 OP3[Mary have-t3 the key] ] ]
λi. $t [ t < ti & ∀w [w compatible with J′s beliefs in wi at t ® M has the key in w at t] ]
Does this structure wind up at PF with a transmitted PAST feature on the argument of the
embedded verb have? That actually doesn't quite yet follow from the assumptions we have in
place. Percolating the feature in the phrase headed by PAST and then transmitting it down to
variables bound by that phrase only gets us as far as (13).
(13) OP1[[PAST-t1]-PAST OP2[John think-[t2-PAST] OP3[Mary have-t3 the key] ] ]

We need a further assumption here, namely that there is percolation in the phrase headed by the
verb think, in such a way that the tense feature from the argument of think gets passed up.
(14) Percolation in the verbal complex: 76

[V [arg F]] becomes [V [arg F]]-F

If we apply (14) in (13), this will then feed another application of the transmission rule, and we
arrive at (15).
(15) OP1[[PAST-t1]-PAST OP2[John [think-[t2-PAST]]-PAST OP3[Mary have-t3-PAST the key] ] ]
spell out: John thought that Mary had the key.
So we generate the simultaneous reading for this surface sentence.
Two remarks are in order. First, the system does not generate any reading for the past-under-past
sentence that could describe a "later-than-matrix" scenario. Suppose that someone at some time
in the past had a thought about a time which was still in their future at the time of their thinking,

76
For this to work as intended, it is actually crucial that the index-argument be the innermost argument of
the verb – a decision which up to now was an arbitrary matter of exposition.
 EMBEDDED TENSES 143

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
42
but which is now past in relation to our utterance time. This sort of thought cannot be reported
by a sentence with a simple past in the lower clause, but requires instead an embedded would
(another difference between complement clauses and relative clauses, and another illustration of
Ogihara's "temporal directionality isomorphism"). The following minimal pairs make the point.
(16) (a) * We knew long before the test that you flunked.
(b) We knew long before the test that you would flunk.
(17) I wanted to go shopping before the stores closed last night,
(a) * because I was afraid you were hungry this morning.
(b) because I was afraid you would be hungry this morning.
(18) By the 1970s, people no longer expected
(a) * that Germany was reunified before the end of the century.
(b) that Germany would be reunified before the end of the century.
Examples with would (such as the (b)-cases above or John thought that Mary would have the
keys) are generated with appropriate meanings by our current analysis, as the reader is invited to
work out.

A second remark is that the analysis as it now stands does not generate the sentence in (19) at all.
(19) John thought that Mary has the keys.
If we base-generate no temporal operator in the lower clause, the only possible derivation is the
one we saw above in (12) - (15), which leads to past tense morphology in both clauses.
Interpretability constrains us to generate a locally bound operator in the embedded clause, not a
pronoun that might be bound from higher up, i.e., from the matrix C. There is simply no way to
derive (19), and the prediction is therefore that present tense in the complement of a past tense
matrix attitude verb should be ungrammatical. Unfortunately the empirical facts are not so
simple – a complication that we discuss in the following section.

5.3. Present under past and double-access readings

Present-under-past attitude and speech reports are not simply another way to express a
simultaneous reading (in which case we could simply have dealt with them by making the new
percolation rule in (14) optional). Nor do they simply report someone's past thought about the
present time which was in the future of their thinking (which would make them counterexamples
to temporal directionality isomorphism). Rather these sentences have a distinctive and peculiar
meaning of their own, known in the literature as a "double access" reading. We introduce the
phenomenon by quoting from a paper by Altshuler & Schwarzschild : 77

"Suppose that ... at the mall, I ask Sylvia where her friend, Mary, is. She replies: “Mary is at
home today”. Later that day, when I’m at the beach and asked for Mary’s whereabouts, I can
truthfully say:
(9) Sylvia said that Mary is at home.

77
(2013, Amsterdam Colloquium)
144 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
43
Taking our cue from Abusch’s (1997) discussion of present complements of believed, we
observe that (9) is true on the so-called double access reading because two conditions are met.
To describe those conditions, we will need to refer to the time and world at which Sylvia replied
to me at the mall. We symbolize those as w@ and tmall respectively. Below, in (10) - (11), we
first give each condition in descriptive terms ...
(10) Relative Present Condition
a. If Sylvia’s utterance was true, then Mary is at home in w@ at tmall.
[...]
(11) Deictic Present Condition
a. If Sylvia’s utterance was true, then Mary is at home throughout an interval that
includes the time at which (9) is uttered.
[...]
We call (10) ‘Relative Present Condition’ because it makes the present tense look like a relative
present – a sort of Priorian present tense (Prior 1967) [...] We call (11) ‘Deictic Present
Condition’ because it makes the present tense look like a deictic present – a tense that in any
context picks out the utterance time of that context. [...] This would explain why (9) is true if
uttered on the same day as tmall but not if uttered on the following day.
In sum we propose that the present tense in English is an amalgam of both a relative and a deictic
present. More concretely, we propose that the English present demands truth at the local
evaluation time (relative tense component) and at or after the speech time (deictic tense
component). In a simple present tense clause [...], the local evaluation time is the speech time so
the two components cannot be told apart."

A&S propose, in effect, that "double access" is hard-wired into the meaning of the English
present tense: this tense shifts to a new evaluation time that is constrained by its relations to two
other times. We adopt a related idea , but without actually departing from our assumption that
78

the present tense is vacuous and therefore does not shift evaluation time at all. We instead draw
a parallel to a phenomenon with personal plural pronouns known as 'split antecedents' or 'split
binding'. An example is (20a), which has an LF like (20b), in which the +-sign denotes a
79

function (of type <e,<e,e>>) which maps two individuals to the smallest (plural) individual that
contains them both as parts. The truth-conditions of (20a) on this parse are fulfilled by a
scenario in which everyone told someone "You and I should get together".
(20) (a) Everyone told someone that they should get together.
(b) everyone 1[someone 2[ t1 told t2 that [pro1 + pro2] should get together ] ]

The idea that we want to spell ou here is that the topmost type-s argument in a double-access
complement is a world-time pair in which the time is kind of a plural time. This requires a
special sum-operator for world-time pairs, whose definition is admittedly a little funny, since it
uses both input times but effectively "throws away" one of the worlds.

78
Among other differences from A&S's proposal, we are not requiring either of the two arguments to be
the utterance index. Empirical consequences of this difference will show up only in sentences with
multiple levels of embedding, whose examination we leave to another occasion. There are several other
differences too.
79
refs
 EMBEDDED TENSES 145

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
44
(21) [[+]] = λi ∈ Ds. λj ∈ Ds. <wi, [ti,tj]>,
where [ti,tj] is the interval from ti to tj, which is the smallest interval that contains both ti
and tj
What we have in mind for the example sentence Sylvia said that Mary is at home is a
representation of the form in (22).
(22) OP1[ [PAST t1 ] OP2[Sylvia say-t2 OP3[ Mary be at-home-[_+_] ] ] ]
We have filled in the OP's and traces that are required for the usual reasons of interpretability.
Note that these include an OP (OP3) at the edge of the complement of say, to meet the latter's
requirement for a proposition. One of the two blanks in the "plural" argument of the embedded
verb thus has be occupied by the trace of OP3. The other one is up for grabs and could in
80

principle be bound by any one of the three higher OP's (OP1, OP2, or OP3). There is, however, one
more principled constraint that we can identify before we examine the remaining options one by
one. The trace of the operator OP3 will have to go to the left of +, not to the right. This has to do
with the asymmetry regarding the worlds in definition (21). Let's compute what would happen if
OP3 were to bind (only) a variable on the right. (23) shows the result of interpreting the say-VP,
assuming n ≠ 3.
(23) [[Sylvia say-t2 OP3. Mary be at-home-(pron + t3)]]g = 1 iff
∀w [w is compatible with what Sylvia says in wg(2) at tg(2)
® Mary is at home in wg(n) at [tg(n), tg(2)] ]
The universal quantifier over worlds binds vacuously here, so this is a pathological meaning.
We are down to three ways of filling the blanks in (22).
(24) OP1[ [PAST t1] OP2[Sylvia say-t2 ...
(a) ... OP3[ Mary be at-home-[t3 + pro3] ] ] ]
(b) ... OP3[ Mary be at-home-[t3 + pro2] ] ] ]
(c) ... OP3[ Mary be at-home-[t3 + pro1] ] ] ]
(24a), with the two arguments of + coindexed, amounts to the same meaning as if we had simply
put t3 instead of t3 + pro3. (Definition (20) implies that [[+]](i)(i) = i.) This expresses a
simultaneous reading. (24b) turns out to be equivalent with this as well (prove as exercise).
(24c) is the only interesting choice. We compute the proposition in (25).
(25) λi. $t [ t < ti & ∀w [w is compatible with what Sylvia says in wi at t
® Mary is at home in w at [t, ti] ] ]
This represents our desired double-access reading. It implies that, according to what Sylvia said
at tmall, Mary was at home at the interval from tmall to the utterance time.

80
It is worth noting that this is another illustration of Ogihara's "temporal directionality isomorphism".
Ogihara (1996) actually introduced that principle in the context of discussing the double access
phenomenon. He thereby drew attention to the fact that even in this case – which may superficially look
as if an embedded tense were chosen to reflect solely the utterer's perspective – we have a tense that upon
careful examination turns out to relate the embedded event also to the subjective "now" of the subject in
the reported past thought/speech act.
146 T I M E

24.973, MIT Spring 2018

Kai von Fintel & Irene Heim
45
What about morphology? It is one thing to have a semantics that generates the double-access
reading, and another to predict correctly that this reading – and only this reading – surfaces with
a present tense in the embedded clause. Here the story that we need in order to get the facts right
is not yet determined by our existing assumptions, unfortunately. Given those assumptions,
including Percolation in the Verbal Complex from (14) above, the lower clause under the three
disambiguations in (24a - c) will receive the transmitted features shown in (25).
(25) OP1[ [PAST-t1]-PAST OP2[Sylvia [say-[t2-PAST]]-PAST ...
(a) ... OP3[ Mary be at-home-[t3-PAST + pro3-PAST] ] ] ]
(b) .... OP3[ Mary be at-home-[t3-PAST + pro2-PAST] ] ] ]
(c) ... OP3[ Mary be at-home-[t3-PAST + pro1] ] ] ]
In the two structures that express the simultaneous reading, (25a) and (25b), both arguments of +
have a transmitted PAST. We would want to make sure that in this case, PAST percolates to the
whole "plural" argument and is seen by the morphology and expressed on the verb, giving us
was at home. (If that didn't happen and we got is at home, we would wrongly pair this
morphology with an unattested simultaneous reading!) In the structure (25c), on the other hand,
which expresses the double access reading, only one of the arguments of + is marked PAST and
the other one is unmarked. In this case, we want to say that PAST does not percolate to the plural
argument and is not spelled out, so that the embedded verb surfaces as is at home. Our structures
are rich enough for us to be able to make the desired distinction, though the assumptions required
don't look as principled as we might have hoped. More research is needed here. If this approach
to double access is on the right track, the morphological mechanisms should ultimately fall
together with independently attested mechanisms for assigning phi-features to plural personal
pronouns and coordinated DPs. 81

81
refs include Podobryaev 2014 MIT PhD
P A R T IV

Questions
8 Interrogative clauses

8.1 A little pragmatics . . . . . . . . . . . . . . . . . . . . 150

8.2 Compositional computation of semantic values for in-
terrogative clauses . . . . . . . . . . . . . . . . . . . . 155
8.2.1 An updated version of Karttunen 1977 . . . . 155
8.2.2 Back to pragmatics: Mapping interrogative
denotations to partitions . . . . . . . . . . . . 158
8.2.3 Syntax: the role of the [WH] feature . . . . . 161
8.3 Embedded questions and question-embedding verbs . 163
8.3.1 Responsive verbs, a type-mismatch, and Dayal’s
strategy . . . . . . . . . . . . . . . . . . . . . 163
8.3.2 An ANS operator inspired by Karttunen . . . 165
8.3.3 Problems, and an ANS operator inspired by Groe-
nendijk & Stokhof . . . . . . . . . . . . . . . 166
8.3.4 ANS in matrix questions? . . . . . . . . . . . . 168
8.3.5 Embedding under rogative verbs . . . . . . . 170
8.4 Wh-movement with pied-piping and reconstruction 170
8.4.1 Background on plurals and many . . . . . . . 171
8.4.2 How many-questions . . . . . . . . . . . . . . 173
8.4.3 More on reconstruction . . . . . . . . . . . . 175

So far, we have developed a semantic system that compositionally de-

rives semantic values for “declarative” clauses, whether standing on their
own as main or matrix sentences or embedded as constituent clauses (rel-
ative clauses, complement clauses, adjunct clauses). Extensionally speak-
ing, declaratives are of type 𝑡 and their denotation is determined relative
to an evaluation index (a world, or a world-time pair) and a variable as-
signment: ⟦𝜙⟧𝑖,𝑔 . The system also supplies (possibly relative to a variable
assignment) an intension of type ⟨𝑠, 𝑡⟩: a proposition, or a function from In set talk: a set of indices. In the simplest
indices to truth-values, ⟦𝜙⟧𝑔¢ = 𝜆𝑖.⟦𝜙⟧𝑖,𝑔 . system: a set of worlds.

In this chapter, we will expand the system to deal with a central kind of Another important kind of non-
non-declarative clause: interrogatives. These also can stand on their own: declarative are imperatives, if you’re in-
terested in which, you should consult
(1) a. Did Sakina see Emily? Portner 2007, M. Kaufmann 2012, von
b. Did Sakina see Emily or did Sakina see Julie? Fintel & Iatridou 2017.

c. Who did Sakina see?

150 Q U E S T I O N S

And they can appear as embedded clauses:

(2) a. Melanie wonders whether Sakina saw Emily.
b. Inessa is surprised who Sakina saw.
c. Whether Sakina saw Emily or not, she definitely saw Julie.

The semanticist’s task in the analysis of interrogative clauses is to pro-

pose suitable types of semantic values for them and to show how these
semantic values are built up compositionally. But before we apply our-
selves to this task, it is useful to say a little bit about speech acts, and about
the relation between semantic values and speech acts.

8.1 A little pragmatics

Leaving semantics out of the picture for the moment, interrogative clauses
differ from declarative clauses in their syntax and in their pragmatics. In
English, interrogative clauses can be told apart syntactically from declar-
ative clauses by the presence of a clause-initial wh-phrase and/or the
inverted order of subject and auxiliary. The terminology “interroga-
tive”/“declarative”, however, alludes to a distinction not in grammatical
form but in communicative function. When uttered as main clauses (i.e.,
not embedded in a larger structure), declarative sentences typically serve
to make assertions, whereas interrogative sentences serve to ask questions.
These two clause-types serve to perform different kinds of speech acts.
Stalnaker 1978 outlined an influential formal model of what happens
in a conversation and what it means to make an assertion. The central
concept is that of a body of publicly shared information, or “common
ground”, which evolves as the conversation proceeds. A proposition is
in the common ground if each interlocutor is “disposed to act as if he as-
sumes or believes the proposition is true, and as if he assumes or believes
that his audience assumes or believes that it is true as well”. The common
ground can be characterized by the set of worlds in which every propo-
sition in the common ground is true; Stalnaker calls this set of worlds the
“context set”. The act of making an assertion is a proposal to update — in
fact, to shrink — the context set. The particular way in which the context Not “shrink” in the sense of reducing car-
set is to be shrunk depends on the semantic value of the asserted sentence, dinality. The cardinality of the context set
may never become less than uncountably
more specifically, on its intension. To assert a sentence 𝜙 is to propose that
infinite.
the current context set 𝑐 be replaced by a new context set which is the
intersection of 𝑐 with the intension of 𝜙, i.e., 𝑐 ∩ ⟦𝜙⟧¢ . We’re leaving off the variable assignment
Against this backdrop, how might we think about the speech act of parameter, which is legitimate if there are
no free variables in 𝜙.
asking a question? What is the point of this speech act? Questions, un-
like assertions, don’t provide information about the world and therefore At least this is true if we stick to ques-
do not in themselves lead to updates of the common ground. Their pur- tions that don’t have presuppositions (or
to questions whose presuppositions are al-
pose rather is to constrain the future course of the conversation in a cer-
ready common ground at the point when
tain way. In the absence of any particular question under consideration, they are asked). When a question that
has a presupposition is asked in a context
where the presupposition is not yet in
the common ground, a listener can get
new information by accommodating this
presupposition.
 I N T E R R O G AT I V E C L A U S E S 151

a speaker might assert whatever they find interesting or useful. But if

someone asks a question, they thereby express an expectation that the next
conversational move will be one out of a very limited set of possible as-
sertions — intuitively, the next assertion will “address the question” rather
than provide some random other information.
Here is a way to formally model this intuition. A question partitions
the context set in a certain way, and it amounts to a request for assertions
which, so to speak, “carve up” the context set along the lines determined
by this partition. The requested next assertion should not make any dis-
tinctions between worlds which the partition lumps together.
Here’s the formal definition of a partition:

Definition 1 (Partition) A set P is a partition of a set 𝐴 iff

(i) every element of P is a non-empty subset of 𝐴,
(ii) any two non-identical elements of P are disjoint from each other,
and
(iii) the union of all the elements of P is 𝐴.
The elements of a partition are also called its cells.

With this definition in hand, we refine Stalnaker’s model of conversa- The idea of refining the Stalnakerian
tion. Each stage in a conversation is now characterized not by its context model of the context by moving from
the context set to a partitioned context
set, but by its “partitioned context set”, i.e., by a partition of some set of
set goes back to Groenendijk 1999. A
possible worlds. The union of this partition corresponds to Stalnaker’s old formally equivalent way is to model the
context set, i.e., it contains the worlds compatible with every proposi- context as an equivalence relation (see, for
tion in the common ground. The partitioning represents the interlocu- example, Jäger 1996 and Hulstijn 1997).
We will later see approaches that give a
tors’ shared commitment not to make distinctions between worlds that partition semantics to interrogatives. It’s
are “cell-mates” — at least not for the time being. This is a renegotiable important to distinguish the two moves.
commitment and, as we will see, it only stays in force until someone raises As we’ll see in this section, one can have
a partition pragmatics without a partition
a new question. The definition of “relevant assertion” in Definition 3 be-
semantics. This theme is explored in much
low spells out precisely what the commitment amounts to. In Definition more sophisticated detail in Fox 2018.
2, we adapt Stalnaker’s characterization of the context-changing role of
assertions to our new, more elaborate set-up. Starr 2020 adds to the partitioned context
set model a way of tracking the speech act
Definition 2 (Update by assertion) To assert a sentence 𝜙 is to propose effect of imperatives.
that the current partitioned context set 𝐶 be replaced by a new partitioned
Our proposed context model is simple
context set which is constructed by intersecting each cell of C with the compared to some other well-known
intension of 𝜙. More precisely, C is to be replaced by systems, such as Roberts 2012 and Farkas
& Bruce 2010.

{𝑝 : 𝑝 ≠ ∅ & ∃𝑝 ′ ∈ C. 𝑝 = 𝑝 ′ ∩ ⟦𝜙⟧¢ }

Definition 3 (Relevance) An assertion is relevant w.r.t. to a partitioned

context set C iff the update it proposes results in a subset of C.

The notion of “relevance” that is formalized here calls for a bit of clari-
fication. First, here are a couple of more commonly used (and more trans-
parent) definitions.
152 Q U E S T I O N S

Definition 4 (Relevance′ ) A proposition 𝑝 (i.e. set of worlds) is relevant

w.r.t. to a partition Q of a set of worlds iff 𝑝 is identical to a cell in Q or to
a union of cells in Q.

Definition 5 (Relevance′′ ) A proposition 𝑝 is relevant w.r.t. to a parti-

tion Q of a set of worlds iff there is no pair of worlds 𝑤 and 𝑤 ′ such that
both
(i) there is a cell 𝑐 ∈ Q such 𝑤 ∈ 𝑐 & 𝑤 ′ ∈ 𝑐
(i.e., 𝑤 and 𝑤 ′ are cell-mates),
(ii) and 𝑤 ∈ 𝑝 & 𝑤 ′ ∈ 𝑝
(i.e. 𝑝 has different truth values in 𝑤 and 𝑤 ′ ).

These two definitions differ slightly in that Definition 5 (unlike Def-

inition 4) allows 𝑝 to include arbitrary worlds that are not in any cell of
Q at all. But both express the essential requirement that a relevant propo-
sition must not discriminate among cell-mates. It must not “cut up” the
space of worlds (in the union of the partition) in any way that does not
coincide with the boundaries between cells.
Definitions 4 and 5 apply to a proposition, whereas our Definition 3
applies to an assertion and presupposes the rule for update-by-assertion.
But apart from this difference, Definition 3 captures the same notion of
relevance. (To make the connections fully precise: If C is a partitioned
context set and 𝜙 is an asserted sentence, then the assertion of 𝜙 is “rel-
∪
evant” w.r.t. C in the sense of Definition 3 iff C ∩ ⟦𝜙⟧¢ is “relevant”
w.r.t. C in the sense of Definition 4, which holds, in turn, iff ⟦𝜙⟧¢ is “rel-
evant” w.r.t. C in the sense of Definition 5. Prove these equivalences as an
exercise.)
To get an intuitive appreciation of this concept, it is important to see
that “relevance” in this technical sense is a purely negative requirement, so
to speak: it is just the absence of any irrelevant (i.e., not specifically asked
for) information. By this criterion, a “relevant” answer may be totally un-
informative; it could be a tautology, or any other proposition that is al-
ready entailed by the common ground. So evidently, relevance is not a
sufficient condition for appropriate answers; there are additional condi-
tions, notably that an answer — or any assertion, for that matter — should
be informative. It is also worth pointing out that the formal requirement
of relevance is prima facie violated by many answers which in actual fact
are judged felicitous, e.g., when somebody responds to the question Is
it raining? by saying: Well, Becky just came in dripping wet, or Maybe, or I
don’t know. To make sense of this wider range of intuitively felicitous an-
swers, a proponent of the present formal model needs to say that in these
cases, the answerer is implicitly changing or amending the question that
was explicitly asked, and instead is addressing a different (though related)
question that they don’t formulate explicitly but expect their audience to
infer (“accommodate”).
 I N T E R R O G AT I V E C L A U S E S 153

Now the only piece of our story that is missing is a recipe for “update
by question”. The idea, as we have said, is that asking a question amounts
to proposing a specific replacement of the current partitioned context set
by a new one. As in the case of assertions, the construction of the new
partitioned context set should be determined by the semantic value of the
sentence that was uttered. We know what we have in mind for particular
examples. E.g., when someone asks Is it raining?, the intended outcome is
a two-membered partition, with one cell consisting of worlds in which
it is raining and the other cell of worlds in which it is not raining. If the
question is a so-called “alternative question” like Is Julie in Barcelona or did
Emily call them on the phone?, the result should be a four-celled partition
corresponding to the four possible configurations of truth-values for the
two component sentences. If the question asked is Who among these two
(Carli and Jodie) came to the party?, the result should be a partition with
four cells: one with worlds in which Carli and Jodie both came, one with
worlds in which Carli but not Jodie came, one with worlds in which Jodie
but not Carli came, and one with worlds in which neither Carli nor Jodie
came.
What is the general recipe by which these outcomes are obtained as a
function of the semantic value of each interrogative sentence? To answer
this question, we have to carry out the semanticist’s task, i.e., assign se-
mantic values to interrogative clauses. We will do so in the next section,
but for now we will articulate a basic intuition about the semantic value of
interrogatives and explore how they can be used to construct a partitioned
context set.
Perhaps the most widespread idea about the semantic value of inter-
rogatives is that they denote sets of propositions. The particular structural
features of interrogatives are taken to serve to both “lift” the type of the
denotation from the type of propositions (the denotation of declaratives)
to the type of sets of propositions, and to generate the individual proposi-
tions in the set. We will look at the mechanics of this in the next section.
For now, let’s assume the following semantic value for the sample wh-
questions Who among these two (Carli and Jodie) came to the party?:
(3) ⟦Who among these two (Carli and Jodie) came to the party?⟧ =
{𝜆𝑤 . Carli came to the party in 𝑤, 𝜆𝑤 . Jodie came to the party in 𝑤 }

Now, notice that the propositions in this set are not disjoint (they can
both be true of the same world) and that their union leaves out the worlds
where neither Carli nor Jodie came to the party. So, this set of proposi-
tions is not a partition of the set of all possible worlds and also not a par-
tition of any context set with minimal information about who came to
the party. So, there’s apparently not a seamless link between our partition
pragmatics for question acts and our semantic values for interrogatives.
154 Q U E S T I O N S

Should we therefore redo our semantics so that it does, after all, map
interrogative clauses to semantic values that are partitions? This can be
done and has been argued for, and we will return to a discussion of this
possibility. For now, however, we observe that it is not necessary. An
equally reasonable take on the discrepancy between our semantics and
our pragmatics is that it simply reflects the division of labor between se-
mantics and pragmatics. We don’t necessarily need a semantics that di-
rectly delivers partitions as semantic values. All we need is to make precise
how the semantic values delivered by the semantics help determine the
partitions that figure in the pragmatics. In other words, we need to for-
mulate our pragmatic rule for “update by question” in such a way that
it spells out how the partition effected by a question-act is determined
by the semantic value of the uttered sentence. We will work up to this
formulation in a few steps. As we will see, there is a straightforward and
general recipe for converting an arbitrary set of propositions into a par-
tition of a given set of worlds, and our update principle will make use of
that recipe.
Informally speaking, each proposition in the question-denotation de-
fines a “dividing line” across the space of worlds in the context set, and
as we use multiple propositions to draw multiple lines across this space,
we get increasingly fine-grained partitions as the number of propositions
goes up. The first proposition cuts the space into two cells (the region This assumes that the propositions are
where it is true and the region where it is false). The second proposition all “logically independent” of each other
and of the propositions in the original
subdivides these two cells further into four; the third proposition leaves us
common ground. That is, none of them
with eight cells; and so on. contradicts or is entailed by any other (or
Another way of describing this process is that each new proposition any conjunction or disjunction of oth-
introduces a new condition that cell-mates must satisfy. At the beginning, ers). If there are logical relations between
the propositions, then the total number
before any line has been drawn, all of the context set is one big cell and of cells is smaller. Note that the defini-
every world counts as a cell-mate of every other one. Then we draw a tion of “partition” requires all cells to be
line for the first proposition (call it 𝑝 1 ), and if two worlds differ with re- non-empty, so the intersection of two
incompatible propositions is not a cell.
spect to the truth-value of 𝑝 1 , they now are no longer cell-mates. 𝑤 and
𝑤 ′ are cell-mates now only if 𝑝 1 (𝑤) = 𝑝 1 (𝑤 ′ ). With the second propo-
sition, 𝑝 2 , we draw a second line and “break up” yet more of the original
cell- mate relations. At this point, only those pairs of worlds remain cell-
mates which are treated alike by both 𝑝 1 and 𝑝 2 . I.e., 𝑤 and 𝑤 ′ are cell-
mates now iff 𝑝 1 (𝑤) = 𝑝 1 (𝑤 ′ ) & 𝑝 2 (𝑤) = 𝑝 2 (𝑤 ′ ). And so on for the third
and all other propositions in the given set of propositions.
So a set of propositions can be used to define a “cell-mate relation”, and
thus a partition.
 I N T E R R O G AT I V E C L A U S E S 155

(4) Let P be a set of propositions and 𝑐 a set of worlds.

a. ∼ P,𝑐 , the cellmate relation in 𝑐 based on P, is defined as follows:
∀𝑤, 𝑤 ′ : 𝑤 ∼ P,𝑐 𝑤 ′ iff 𝑤 ∈ 𝑐 & 𝑤 ′ ∈ 𝑐 & ∀𝑝 ∈ P : 𝑝 (𝑤) = 𝑝 (𝑤 ′ ).
b. PARTS(P, 𝑐), the partition of 𝑐 based on P, is defined as follows:
{ }
PARTS(P, 𝑐) = 𝑝 : ∃𝑤 ∈ 𝑐.𝑝 = {𝑤 ′ : 𝑤 ∼ P,𝑐 𝑤 ′ }

The cellmate-relation is an equivalence relation (reflexive, symmetric,

and transitive). It holds between any two worlds that give the same truth-
value to each proposition in the “seed set”. The corresponding partition
consists of the subsets of 𝑐 that are made up of worlds that are cell-mates.
We are ready to propose a preliminary rule of “Update by question”: Note that this update rule implies that
each new question “wipes out” the current
Definition 6 (Update by question (draft)) To ask a question by utter- partitioning of the context set and replaces
ing a sentence 𝜙 that denotes a set of propositions P is to propose that the it by a new one. This may not be right
and isn’t the only way we can go. We may
current partitioned context set C be replaced by the new partitioned con-
∪ want a system where successive questions
text set PART(P, C). can serve to refine the partition. Even more
complex possibilities exists, as for example
In the next section, we turn to the task of compositionally deriving the in Roberts 2012.
semantic value of an interrogative. The implementation here differs from
Karttunen 1977 in not using the Mon-
tague Grammar framework; instead we
8.2 Compositional computation of semantic values for interroga- presuppose a division of labor between
tive clauses syntax and semantics in line with Heim
& Kratzer 1998. Aside from implemen-
tation, there are some small substantive
8.2.1 An updated version of Karttunen 1977 differences as well. Most saliently, the ex-
tensions we will assign to interrogative
Interrogative clauses are built largely with the same lexicon and syntactic
clauses will include both true and false
rules as declarative clauses; but they also contain certain interrogative- propositions. (In Hagstrom 2003’s ter-
specific morphemes, syntactic features, or functional heads that make cru- minology, they will correspond to ANS-
cial contributions to their non-declarative meanings. Following Kart- POSS rather than ANSTRUE.) Moreover,
we will depart from Karttunen in the
tunen 1977, our syntax for English posits an abstract (i.e., silent) comple-
treatment of polar questions.
mentizer that has a non-vacuous semantics, and a feature on wh-phrases Karttunen’s motivation for including a
that is semantically vacuous but subject to a certain distributional con- restriction to true propositions had to do
straint. with the semantics of question-embedding
constructions. So the time to discuss it
(5) Karttunen’s “proto-question” operator, syntactically a C-head: is when we get to embedded questions.
⟦?⟧𝑤 = 𝜆𝑝𝑠𝑡 .𝜆𝑞𝑠𝑡 . 𝑝 = 𝑞. Right now we focus on matrix questions.

The official denotation in (5) maps propo-

(6) lexical entries for interrogative words, e.g.: sitions to functions from propositions
[WH] 𝑤
⟦who ⟧ = 𝜆𝑓𝑒𝑡 . ∃𝑥 [𝑥 is human in 𝑤 & 𝑓 (𝑥) = 1] to truth-values, or equivalently, it maps
propositions to characteristic functions of
The meaning of who is exactly the same as the meaning of somebody. We sets of propositions. If we replaced these
will address the role of the feature [WH] in Section 8.2.3. characteristic functions by the corre-
sponding sets, (5) would read as follows:
In the syntactic derivation of a constituent question, wh-movement
applies and puts the wh-word above the operator in C. For the example (5’) ⟦?⟧𝑤 = 𝜆𝑝.{𝑝 }

Who did Sakina see? or who Sakina saw (we will ignore tense and auxil- In other words, it would be a function that
iary do, and thereby ignore the difference between matrix and embedded maps each proposition to the singleton set
versions), this gives rise to a structure like (7): that has it as its only member.
156 Q U E S T I O N S

(7) who 2 [? Sakina see 𝑡 2 ]

Unfortunately, when we examine the semantic types at each node in (7),

we discover a type-mismatch that prevents us from interpreting the top-
most node. We fix this problem by positing a slightly more complex C-
head: We base-generate ? together with another, covert operator as its
sister. This covert operator is semantically vacuous, but it can move and Other vacuous covert operators of this sort
leave a trace of type ⟨𝑠, 𝑡⟩. The full syntactic derivation for a constituent are H&K’s relative pronouns (and their
PRO in ch. 8).
question then proceeds as in (8).
(8) Who did Sakina see? The movement of who is evidently overt
(pre-SS) movement (in this example). For
the movement of the operator, which is
[𝐶 ? OP] Sakina see who not pronounced, we can’t tell whether it
who 2 [ [? OP] Sakina see 𝑡 2𝑒 ] (wh-movement) happens before SS or between SS and LF.
OP 1 [ who 2 [ [? 𝑡 1𝑠𝑡 ] Sakina see 𝑡 2𝑒 ] ] (operator movement)

The resulting LF looks thus like this in tree format:

(9)

OP1

who2

? 𝑡1 Sakina

see 𝑡2

We will usually drop the type-labels on the traces, but keep in mind
that the operator leaves a trace of type ⟨𝑠, 𝑡⟩ (i.e., a variable over proposi-
tions, not individuals). Accordingly the topmost application of Predicate
Abstraction will yield a function from propositions to truth-values (type
⟨𝑠𝑡, 𝑡⟩, the characteristic function of a set of propositions). Let’s compute We leave out the vacuous operator OP
the meaning of this LF: and just interpret the predicate abstract it
creates. ∅ in the superscript stands for the
(10) Computation for LF (9): empty variable assignment, see H&K ch.5
⟦1. who 2. [? 𝑡 1 ] Sakina see 𝑡 2 ⟧𝑤,∅
= (by Predicate Abstraction)
𝜆𝑝. ⟦who 2. [? 𝑡 1 ] Sakina see 𝑡 2 ⟧𝑤,[ 1→𝑝 ]
= (by entry for who and lambda reduction)
𝜆𝑝. ∃𝑥 [ 𝑥 is human in 𝑤 & ⟦2. [? 𝑡 1 ] Sakina see 𝑡 2 ⟧𝑤,[ 1→𝑝 ] (𝑥) = 1 ]
= (by Predicate Abstraction and lambda reduction)
𝜆𝑝. ∃𝑥 [ 𝑥 is human in 𝑤 & ⟦[? 𝑡 1 ] Sakina see 𝑡 2 ⟧𝑤,[ 1→𝑝,2→𝑥 ] = 1 ]
= (by IFA)
′
𝜆𝑝. ∃𝑥 [ 𝑥 is human in 𝑤 & ⟦? 𝑡 1 ⟧𝑤,[ 1→𝑝,2→𝑥 ] (𝜆𝑤 ′ . ⟦Sakina see 𝑡 2 ⟧𝑤 ,[ 1→𝑝,2→𝑥 ] ) = 1 ]
 I N T E R R O G AT I V E C L A U S E S 157

= (by FA and dropping irrelevant superscripts)

′
𝜆𝑝. ∃𝑥 [ 𝑥 is human in 𝑤 & ⟦?⟧(⟦𝑡 1 ⟧ [ 1→𝑝,2→𝑥 ] )(𝜆𝑤 ′ . ⟦Sakina see 𝑡 2 ⟧𝑤 ,[ 1→𝑝,2→𝑥 ] ) = 1 ]
= (by Traces rule)
′
𝜆𝑝. ∃𝑥 [ 𝑥 is human in 𝑤 & ⟦?⟧(𝑝)(𝜆𝑤 ′ . ⟦Sakina see 𝑡 2 ⟧𝑤 ,[ 1→𝑝,2→𝑥 ] ) = 1 ]
= (by entry for ?)
′
𝜆𝑝. ∃𝑥 [ 𝑥 is human in 𝑤 & 𝑝 = 𝜆𝑤 ′ . ⟦Sakina see 𝑡 2 ⟧𝑤 ,[ 1→𝑝,2→𝑥 ] ]
= (by FA, entries for Sakina, see, Traces Rule)
𝜆𝑝. ∃𝑥 [ 𝑥 is human in 𝑤 & 𝑝 = 𝜆𝑤 ′ . Sakina sees 𝑥 in 𝑤 ′ ]

This characterizes a set of propositions that contains one proposition per This is the ANSPOSS in Hagstrom 2003’s
human-in-𝑤: the proposition that that human was seen by Sakina. terminology.

How about polar and alternative questions? Can we just posit the same
operators in the C-head? Let’s try.
Here we show an LF and computation
that includes the vacuous operator. As
(11) Did Sakina see Emily? an exercise, convince yourself that in the
case of polar questions, a simpler structure
[𝐶 ? OP] Sakina see Emily without OP is likewise interpretable, with
equivalent results.
OP 1 [ ? 𝑡 1 ] Sakina see Emily

(12) Computation for LF of the polar question (11):

⟦1. [? 𝑡 1 ] Sakina see Emily⟧𝑤,∅
= (by Predicate Abstraction)
𝜆𝑝𝑠𝑡 .⟦[? 𝑡 1 ] Sakina see Emily⟧𝑤,[ 1→𝑝 ]
= (by IFA)
′
𝜆𝑝𝑠𝑡 .⟦? 𝑡 1 ⟧𝑤,[ 1→𝑝 ] (𝜆𝑤 ′ .⟦Sakina see Emily⟧𝑤 ,[ 1→𝑝 ] )
= (by FA)
′
𝜆𝑝𝑠𝑡 .⟦?⟧𝑤,[ 1→𝑝 ] (⟦𝑡 1 ⟧𝑤,[ 1→𝑝 ] ) (𝜆𝑤 ′ .⟦Sakina see Emily⟧𝑤 ,[ 1→𝑝 ] )
= (by Traces Rule and dropping irrelevant assignment superscripts)
′
𝜆𝑝𝑠𝑡 .⟦?⟧(𝑝) (𝜆𝑤 ′ .⟦Sakina see Emily⟧𝑤 )
= (by entry for ? and lambda reduction)
′
𝜆𝑝𝑠𝑡 . [𝑝 = 𝜆𝑤 ′ .⟦Sakina see Emily⟧𝑤 ]
= (by FA and entries for Sakina, Emily, see)
𝜆𝑝𝑠𝑡 . [𝑝 = 𝜆𝑤 ′ . Sakina see Emily in 𝑤 ′ ]

This is the characteristic function of a singleton set containing one propo-

sition. Is this a good result? If we have in mind that our denotations for
interrogative clauses should directly correspond to an intuitive notion of
“possible answer”, then this is problematic. There is certainly more than
one possible answer to a polar question! But we think of the relation be-
tween the semantics and the pragmatics of interrogative clauses in a less
simple-minded way. A singleton set of a proposition is used to induce a
two-way partition of the context set: one cell containing worlds where
the proposition is true and another where it isn’t. Any relevant response
needs to eliminate one of the cells.
158 Q U E S T I O N S

To complete the current section, we consider an alternative question: The slashes are intended as a crude rep-
resentation of the distinctive intonational
(13) Did Sakina see Emily /or did he see Julie\? contour that characterizes the alternative-
DS: [[𝐶 ? OP] Sakina see Emily] or [[𝐶 ? OP] Sakina see Julie] question reading.
LF: OP 1 [ [ [? 𝑡 1 ] Sakina see Emily] or [ [? 𝑡 1 ] Sakina see Julie] ]

In the alternative question in (13), operator movement must be “across the

board” (ATB), with the result that a single binder binds two coindexed
traces. This is the only way to obtain an interpretable structure, given the
semantic type of or, which is ⟨𝑡, ⟨𝑡, 𝑡⟩⟩.
(14) Computation for LF of the alternative question in (13):
⟦1. [? 𝑡 1 ] Sakina see Emily or [? 𝑡 1 ] Sakina see Julie⟧𝑤,∅
= (by Predicate Abstraction)
𝜆𝑝𝑠𝑡 . ⟦[? 𝑡 1 ] Sakina see Emily or [? 𝑡 1 ] Sakina see Julie⟧𝑤,[ 1→𝑝 ]
= (by FA twice and entry for or)
𝜆𝑝𝑠𝑡 . ⟦[? 𝑡 1 ] Sakina see Emily⟧𝑤,[ 1→𝑝 ] ∨ ⟦[? 𝑡 1 ] Sakina see Julie⟧𝑤,[ 1→𝑝 ]
= (by IFA, FA, entry for ?, etc — see previous computations)
′ ′
𝜆𝑝𝑠𝑡 . [𝑝 = 𝜆𝑤 ′ .⟦Sakina see Emily⟧𝑤 ,[ 1→𝑝 ] ∨ 𝑝 = 𝜆𝑤 ′ .⟦Sakina see Julie⟧𝑤 ,[ 1→𝑝 ] ]
= (by FA and entries for see etc.)
𝜆𝑝𝑠𝑡 . [𝑝 = 𝜆𝑤 ′ . Sakina see Emily in 𝑤 ′ ∨ 𝑝 = 𝜆𝑤 ′ . Sakina see Julie in 𝑤 ′ ]

This is the characteristic function of a set containing two propositions —

the same set, in fact, that is denoted by the constituent question who did
Sakina see if the set of humans happens to be just {Emily, Julie}.

8.2.2 Back to pragmatics: Mapping interrogative denotations to parti-

tions
We can now make good on the promise to provide a bridge between the
compositional semantics of interrogatives and their use in updating a con-
text. We gave a draft version earlier in Definition 6, repeated here:

Definition 6 (Update by question (draft)) To ask a question by utter-

ing a sentence 𝜙 that denotes a set of propositions P is to propose that the
current partitioned context set C be replaced by the new partitioned con-
∪
text set PART(P, C).

To finalize the definition, we need to use our semantics to supply a set

of propositions to serve as the seed for the partition. This is straightfor-
ward in the case of polar and alternative questions, each of which has as
its semantic value a set of propositions that is evaluation-world indepen-
dent. But in the case of wh-questions, an issue arises: their semantic value
depends non-trivially on the evaluation world.
(15) ⟦Who did Sakina see?⟧𝑤
= 𝜆𝑝. ∃𝑥 : 𝑥 is human in 𝑤 . 𝑝 = 𝜆𝑤 ′ . Sakina saw 𝑥 in 𝑤 ′
 I N T E R R O G AT I V E C L A U S E S 159

What set of propositions the interrogative denotes in a world 𝑤 depends

on who the humans in that world are. In case this seems esoteric, consider
a more substantively restricted wh-phrase:
(16) ⟦Which students called?⟧𝑤
= 𝜆𝑝. ∃𝑥 : 𝑥 is a student in 𝑤 . 𝑝 = 𝜆𝑤 ′ . Sakina saw 𝑥 in 𝑤 ′

The semantic value of (16) depends on who the students in the evaluation
world are. It is easy to imagine circumstances where it is not common
ground who the students are. So, how should we update the context with
a question like (16)?
Option 1: The world of utterance. We could say that the context gets par- Option 1 would say something like: To
titioned based on the set of propositions denoted by the interrogative rel- ask a question by uttering a sentence 𝜙
in world 𝑤 is to propose that the cur-
ative to the world in which the question is asked. However, this would
rent partitioned context set C be re-
miss the fact that the participants in the conversation do not know which placed by the new partitioned context
∪
world they are in. The context set is precisely meant to model that they set PART(⟦𝜙⟧𝑤 , C).
have some common ground on what world they’re in but no more: there
are always multiple candidates for what the actual world is. Furthermore,
it is presumably inescapable that participants make some false presuppo-
sitions, in which case the actual world in which the utterance occurs isn’t
even part of the context set (even though everyone is acting as if the com-
mon ground contains the actual world).
So, it seems clear that we need to interpret the interrogative with re-
spect to the worlds in the context set and use the resulting set of propo-
sitions to partition the context set. But now we need to face the possi-
bility that the participants in the conversation where (16) is uttered are
uncertain about who the (relevant) students are. To set up a minimal test
scenario, assume that everyone knows that 𝑎 is a student but there’s un-
certainty about whether 𝑏 is. Now, assume that (16) is uttered against that
background. There are worlds in the context set where 𝑎 and 𝑏 are the
students, and other worlds where only 𝑎 is a student. Our semantics will
deliver two different sets of propositions as the semantic value of the in-
terrogative for the two different kinds of worlds in the context set: for the
𝑎, 𝑏-worlds, we get the set containing the proposition that 𝑎 called and
the proposition that 𝑏 called; while for the 𝑎-worlds, we get the set that
contains only the proposition that 𝑎 called. How should the interrogative
partition the context set?
160 Q U E S T I O N S

Option 2: Collecting all the propositions. To obtain the set of propositions

that is used to partition the context set, we could collect all the proposi-
tions that for some world in the context set are in the semantic value of the
interrogative:
Equivalently, form the big union of all the
sets of propositions that are the semantic
(17) The set of propositions that is used to partition a context set 𝑐 upon value of the interrogative in worlds of the
the utterance of an interrogative 𝜙 is context sets:
∪
⟦𝜙⟧𝑤
{𝑝 : ∃𝑤 ∈ 𝑐. 𝑝 ∈ ⟦𝜙⟧𝑤 } 𝑤 ∈𝑐

For our minimal test scenario, that means that the context set will be
partitioned by the the set containing the proposition that 𝑎 called and the
proposition that 𝑏 called. By our definition of relevance, this means that
the answer “𝑏 called” will be relevant. But notice that when the context
set is updated by the assertion that 𝑏 called, there will still be worlds left
in the context set where 𝑏 is not a student (while it is now accepted that 𝑏
called). This is incorrect: answering that 𝑏 called when what was asked is
Which students called? surely commits one to 𝑏 being a student.
Option 3: Presupposition of consensus. Stalnaker 1978 in his analysis of
the speech act of assertion formulated three principles that he views as
“essential conditions of rational communication”:

1. A proposition asserted is always true in some but not all of the possi-
ble worlds in the context set.
2. Any assertive utterance should express a proposition, relative to each
possible world in the context set, and that proposition should have a
truth-value in each possible world in the context set.
3. The same proposition is expressed relative to each possible world in
the context set.

What’s relevant for us here is the third principle, which ensures that there
is common ground on which proposition the speaker is proposing to add
to the common ground. We suggest that essentially the same principle
applies to the speech act of asking a question: the same set of propositions
needs to be expressed by the interrogative relative to each possible world
in the context set. This then amounts to the assumption that for each wh-
phrase that is used, the interlocutors agree on a specific set of individuals
as its intended range. This means that in our test scenario, the question
Which students called? is a pragmatic error: asking the question presupposes
that it is common ground who the students are.
Hence, the final version of our definition is as follows:

Definition 7 (Update by question (final)) Asking a question by utter-

ing a sentence 𝜙 is only felicitous in a partitioned context set C if ∀𝑤, 𝑤 ′ ∈
∪ ′
𝐶 : ⟦𝜙⟧𝑤 = ⟦𝜙⟧𝑤 .
 I N T E R R O G AT I V E C L A U S E S 161

If felicitous, asking a question by uttering a sentence 𝜙 is to propose

that the current partitioned context set C be replaced by the new par-
∪
titioned context set PART(⟦𝜙⟧𝑤 , C), where 𝑤 is an arbitrary world in
∪
C.

8.2.3 Syntax: the role of the [WH] feature

In the theory as it stands so far, the semantic identification of wh-words

with indefinites (e.g. ⟦who⟧𝑤 = ⟦somebody⟧𝑤 ) leads to overgeneration of
readings. Consider the minimal pair in (18).
(18) a. Who did Sakina see?
b. Did Sakina see somebody?

We have shown in Section 8.2.1 how the Karttunen theory derives

the intended denotation for (18a) (= (10)). But from the Deep Structure
we posited in (10), we could have derived more than one interpretable
structure, namely not just (19a) (which we considered in (10) above) but
also (19b).
(19) a. LF-high: OP 1 [ who 2 [ [? 𝑡 1 ] Sakina see 𝑡 2 ] ]
b. LF-low: OP 1 [[? 𝑡 1 ] who 2 [ Sakina see 𝑡 7 ] ]

Similarly we can derive two interpretable LFs for (18b).

(20) DS: [𝐶 ? OP] Sakina see somebody
a. LF-high: OP 1 [ somebody 2 [ [? 𝑡 1 ] Sakina see 𝑡 2 ] ]
b. LF-low: OP 1 [[? 𝑡 1 ] somebody 2 [ Sakina see 𝑡 2 ] ]

In all of these potential LFs, the IP-sister of C, as well as the mother-

node of C, are semantically of type 𝑡. Therefore, both should be suitable
adjunction sites for the generalized quantifiers who and somebody as far as
type-compatibility goes. (The mnemonic labels ‘high’ and ‘low’ indicate
the difference in the quantifier’s scope.)
But what do these LFs mean? We have already computed (19a) and
seen that it expresses an appropriate meaning for (18a). Specifically, if we
combine our semantics with our pragmatics, and if the domain of peo-
ple that who ranges over consists of just Emily and Julie, (19a) determines
a 4-cell partition. This matches the intuition that the possible fully ex-
haustive answers to (18a) are that Sakina saw only Emily, that he saw only
Julie, that he saw both Emily and Julie, and that he saw neither. But such
a 4-cell partition is clearly not what an utterance of (18b) sets up. (18b)
is a polar question and elicits answers such as Yes or No, expressing the
propositions that Sakina saw somebody and that he saw nobody. So here
we want our semantics and pragmatics to determine a 2-cell partition.
The LF-low in (20b) does deliver precisely this. (Compute this as an exer-
cise.)
162 Q U E S T I O N S

So LF-high in (19a) captures the attested meaning of the wh-question

in (18a), and LF-low in (20b) captures the attested meaning of the polar
question in (18b). In each case, this is the only attested reading for the En-
glish sentence: the wh-question cannot be read as a polar question, and
the polar question cannot be read as a wh-question. Our theory so far fails
to derive this un-ambiguity. The LF-low in (19b) is semantically equiva-
lent to the LF-low in (20b) and thus represents an unattested polar read-
ing for the wh-question. Similarly, the LF-high in (20a) is equivalent to
the LF-high in (19a) and thus represents an unattested wh-question read-
ing for the polar question with somebody. We need to amend our theory
so that only one of the LFs is generated for each of the sentences.
A standard solution — effectively Karttunen’s — is to invoke a syntactic
constraint that regulates scopal relations between existential DPs (who,
somebody) and the interrogative complementizer. As our example teaches
us, who apparently can only scope above ? (more specifically, between
? and its associated OP), whereas somebody is barred from scoping there
and must instead scope below ?. The following stipulation enforces this
generalization.
(21) Wh-Licensing Principle:
At LF, a phrase 𝛼 occupies a specifier position of ? if and only if 𝛼
has the feature [WH].

(21) relies on appropriate assumptions about which phrases have the fea-
ture [WH]. For the time being, assume that certain words such as who,
what, how are marked as [WH] in the lexicon. These then will be the
only phrases that can be located right above ? in well-formed LFs, and
moreover, they cannot be located anywhere else. LF-high for the wh-
question in (19a) complies with (21), as does LF-low for the polar ques-
tion in (20b). LF-low in (19b) is ruled out because it has a phrase marked
[WH] in a location other than spec-of-?, and LF-high in (20a) is prohib-
ited because it has a phrase in spec-of-? which lacks [WH].

Exercise 8.1 Consider the multiple wh-question (22).

(22) Who likes who?

Propose an LF, say how it is derived in the syntax, compute its semantic inter-
pretation, and then compute the partition it imposes. For simplicity, assume in
this last part that the domain of people that who ranges over is just {s, e}, and
that the common ground before the question is totally uninformative, i.e, it is a
partitioned context set whose union is W (the set of all worlds whatsoever). How
many cells does the partition have and what are its cells? Regarding syntax, at-
tend in particular to the satisfaction of the Wh-Licensing Principle.

Exercise 8.2 In written language, a question containing or, such as (23), can be
ambiguous.
 I N T E R R O G AT I V E C L A U S E S 163

(23) Did you talk to David or Norvin?

The questioner may want to know which of the two professors you talked to
(“alternative-question reading”) or just whether you talked to at least one of
them (“polar-question reading”). Perhaps the alternative reading is more salient
out of the blue, but the polar reading can be facilitated by an appropriate context.
(E.g. imagine that your squib is on multiple-wh-constructions, and you have pre-
viously been told that you ought to consult a faculty member who has published
on this topic, namely David or Norvin.)
Your task in the exercise is to analyze the two readings of (23), by proposing
an appropriate LF for each reading and discussing its syntactic derivation as well
as its semantic and pragmatic interpretation. You should assume that English has
only one unambiguous word or, and its semantic type is ⟨𝑡, ⟨𝑡, 𝑡⟩⟩. This means
that, for both readings, you must posit some amount of elided material in the
right disjunct, since or can only coordinate constituents of type 𝑡.

8.3 Embedded questions and question-embedding verbs

A first superficial survey of English verbs that take complement clauses

turns up three groups. One group, exemplified by the verb believe, con- White 2021 argues that things are not that
sists of verbs that take that-clauses but are ungrammatical with an embed- simple.

ded interrogative clause.

(24) a. Sakina believes that Becky called.
b. *Sakina believes who called.
Another rather large second group consists of verbs that can take ei-
ther. This includes know, remember, and tell.
(25) a. Sakina knows/remembers/told Emily that Becky called.
b. Sakina knows/remembers/told Emily who called.

Third, there are verbs which take interrogative complements but are
ungrammatical with that-clauses.
(26) a. *Sakina asked/is wondering that Becky called.
b. Sakina asked/is wondering who called.

We will focus at first on the middle group, which Lahiri 2002 dubbed
the class of “responsive” verbs.

8.3.1 Responsive verbs, a type-mismatch, and Dayal’s strategy

Let’s look at the responsive verb know. We have at least a preliminary
analysis for verbs like know in sentences like John knows that Ann called.
Such verbs take a proposition and a person as their arguments, and their
meaning encodes universal quantifcation over a certain set of possible
worlds. Since know is a factive verb, it also triggers the presupposition that
164 Q U E S T I O N S

its complement is true. For simplicity, we assume here that apart from the
factive presupposition, the meaning of know is the same as the meaning of
believe, and so we can write the lexical entry in (27).
(27) ⟦know⟧𝑤 = 𝜆𝑝𝑠𝑡 : 𝑝 (𝑤) = 1. 𝜆𝑥𝑒 . ∀𝑤 ′ [𝑤 ′ ∈ DOX(𝑥, 𝑤) → 𝑝 (𝑤 ′ ) = 1] Here, DOX maps an individual 𝑥 and a
world 𝑤 into the set of worlds compatible
This entry was designed to work for know with a that-clause. What would with what 𝑥 believes in 𝑤.
happen if we tried to interpret an LF with an interrogative clause as the
sister of know? Evidently, we would run into a type-mismatch. On our
current analysis of interrogative clauses, their extensions are of type ⟨𝑠𝑡, 𝑡⟩
and their intensions of type ⟨𝑠, ⟨𝑠𝑡, 𝑡⟩⟩. Neither type can compose with the
type of know in (27) by means of any of our semantic rules.
What shall we do? Older literature on question-embedding made verbs
like know lexically ambiguous, with distinct (though not unrelated) lexical
entries for declarative-taking know and interrogative-taking know. More
recent work has pursued the strategy of positing a single unambiguous
verb and readjusting the semantic type of the interrogative complement.
Groenendijk & Stokhof 1982 did this first, and another influential ver-
sion of this approach originates with Dayal 1996. Dayal proposed that
the combination of the verb with the Karttunen-denotation of its com-
plement is mediated by an “answer operator”, which maps sets of propo-
sitions to propositions. We will follow Dayal’s general strategy in these
notes. We will entertain a couple of possible meanings for the answer op-
erator and talk about the empirical considerations that bear on the ques-
tion which meaning is correct.
The rough intuition to be implemented is that “to know who called”
means something like “to know the answer to the question ‘who called?”’.
(Paraphrases of this form work for all the verbs in this group, hence Lahiri’s
term “responsive verbs”.) The answer to a question is a proposition, hence
an object of a suitable semantic type to feed to the meaning of know in
(27). If the LFs of sentences like (25b) contain an operator that maps a
question-denotation to the proposition that’s the answer to that question,
we have a solution to the type-mismatch problem.
Since our syntax for interrogative clauses already happens to posit a
silent operator at the top edge of the clause (albeit one that we have so far
treated as semantically vacuous) we need not actually make the structure
more complex. Instead we can assume (following a suggestion by Danny
Fox) that our new answer operator appears instead of the previous vac-
uous one. This means that we base-generate it inside C as the sister of ?
and move it up for interpretability, leaving a type-⟨𝑠, 𝑡⟩ trace as before.
Our LF-structure for a sentence with know and an interrogative comple-
ment then looks as in (28b).
 I N T E R R O G AT I V E C L A U S E S 165

(28) a. Sakina knows who called.

b. Sakina [𝑉 𝑃 knows [𝐶𝑃 ANS 1 [ who 2 [[𝐶 ? 𝑡 1 ] 𝑡 2 called]]]]

We will now focus on the task of proposing a meaning for ANS which
not only fixes the type-mismatch, but also yields reasonable truth condi-
tions for the know-sentence.

8.3.2 An ᴀNS operator inspired by Karttunen

Our initial proposal for the semantics of ANS effectively follows Karttunen
1977 in terms of the truth conditions it predicts for the know-sentence as
a whole (although the compositional implementation is different). The
ANS operator defined in (29) maps the set of propositions denoted by the
interrogative to the single proposition which is the conjunction of all its
true elements.
(29) ⟦ANS⟧𝑤 = 𝜆Q ⟨𝑠𝑡,𝑡 ⟩ . 𝜆𝑤 ′ . ∀𝑝 [Q (𝑝) = 1 & 𝑝 (𝑤) = 1 → 𝑝 (𝑤 ′ ) = 1]
∩
(29’) ⟦ANS⟧𝑤 = 𝜆Q ⟨𝑠𝑡,𝑡 ⟩ . {𝑝 ∈ Q : 𝑤 ∈ 𝑝} equivalent formulation (modulo sets vs.
(the intersection of all members of Q that are true in 𝑤) characteristic functions)

Let’s put this entry to work in a computation for the example sentence.
(30) a. computation of presupposition:
Let 𝑤 be a world. Then
⟦Sakina knows who called⟧𝑤 is defined
iff (by FA twice) Observe that the mother node of know
⟦know⟧𝑤 (⟦ANS 1 who 2 ?-𝑡 1 𝑡 2 called⟧𝑤 )(𝑠) is defined is interpreted by plain Functional Ap-
plication (not by Intensional Functional
iff (by entry for know)
Application, which would have applied if
⟦ANS 1 who 2 ?-𝑡 1 𝑡 2 called⟧𝑤 (𝑤) = 1 know were taking a declarative comple-
iff (by FA) ment). This is because the extension of the
( ) constituent headed by ANS is of type ⟨𝑠, 𝑡 ⟩.
⟦ANS⟧𝑤 ⟦1 who 2 ?-𝑡 1 𝑡 2 called⟧𝑤 (𝑤) = 1
iff by entry for ANS
∀𝑝 [⟦1 who 2 ?-𝑡 1 𝑡 2 called⟧𝑤 (𝑝) = 1 &𝑝 (𝑤) = 1 → 𝑝 (𝑤) = 1]
This is a tautology, so we know that ⟦Sakina knows who called⟧𝑤
is defined for all 𝑤.
b. computation of truth condition:
Let 𝑤 be a world. Then
⟦Sakina knows who called⟧𝑤 = 1
iff (by FA twice)
⟦know⟧𝑤 (⟦ANS 1 who 2 ?-𝑡 1 𝑡 2 called⟧𝑤 )(𝑠) = 1
iff (by entry for know and truth of presupposition)
∀𝑤 ′ [𝑤 ′ ∈ DOX(𝑠, 𝑤) → ⟦ANS 1 who 2 ?-𝑡 1 𝑡 2 called⟧𝑤 (𝑤 ′ ) = 1]
c. We interrupt for an embedded computation:
⟦ANS 1 who 2 ?-𝑡 1 𝑡 2 called⟧𝑤 (𝑤 ′ ) = 1
small iff (by entry for ANS)
∀𝑝 [⟦1 who 2 ?-𝑡 1 𝑡 2 called⟧𝑤 (𝑝) = 1 & 𝑝 (𝑤) = 1 → 𝑝 (𝑤 ′ ) = 1]
166 Q U E S T I O N S

iff (by earlier computations)

∀𝑝 [∃𝑥 [𝑥 is a human in 𝑤 & 𝑝 = 𝜆𝑤 ′′ . 𝑥 called in 𝑤 ′′ ] & 𝑝 (𝑤) =
1 → 𝑝 (𝑤 ′ ) = 1]
iff (by logic of quantifiers)
∀𝑝∀𝑥 [𝑥 is a human in 𝑤 & 𝑝 = 𝜆𝑤 ′′ . 𝑥 called in 𝑤 ′′ & 𝑝 (𝑤) =
1 → 𝑝 (𝑤 ′ ) = 1]
iff (by logic of identity)
∀𝑥 [𝑥 is a human in 𝑤 & [𝜆𝑤 ′′ . 𝑥 called in 𝑤 ′′ ] (𝑤) = 1 →
[𝜆𝑤 ′′ . 𝑥 called in 𝑤 ′′ ] (𝑤 ′ ) = 1]
iff (by 𝜆-reduction)
∀𝑥 [𝑥 is a human in 𝑤 & 𝑥 called in 𝑤 → 𝑥 called in 𝑤 ′ ]
d. resuming main computation: …
iff (by plugging in result of embedded computation)
∀𝑤 ′ [𝑤 ′ ∈ DOX(𝑠, 𝑤) → ∀𝑥 [𝑥 is a human in 𝑤 & 𝑥 called in 𝑤 →
𝑥 called in 𝑤 ′ ]]
iff (by logic of quantifiers)
∀𝑥 [𝑥 is a human in 𝑤 & 𝑥 called in 𝑤 → ∀𝑤 ′ [𝑤 ′ ∈ DOX𝑠, 𝑤 →
𝑥 called in 𝑤 ′ ]]

In other words, the sentence Sakina knows who called is true in 𝑤 if and
only if, for every person who in fact called in 𝑤, Sakina believes (in 𝑤)
that this person called.

8.3.3 Problems, and an ᴀNS operator inspired by Groenendijk & Stokhof

The essence of ANS in our present analysis is that for any evaluation world
𝑤, when applied to the semantic value Q of an interrogative (which is a
set of proposition), it gives the proposition that is the conjunction of all
the propositions in Q that are true in 𝑤. This is well known to make some
troublesome predictions. Let’s start with the most glaring case, then move
to subtler ones, before we diagnose the general problem and proceed to
solve it.
The example we analyzed above involved an embedded constituent
question. What if we embedded a polar question instead?
(31) a. Sakina knows whether Emily called.
b. LF: Sakina knows [𝐶𝑃 ANS 1[ [𝐶 ? 𝑡 1 ] Emily called ] ]

Recall that our denotation for a polar question is a singleton set. The sister
of ANS in (31b) denotes the set whose only member is the proposition that
Emily called. If we complete the calculation, we get the following truth
condition.
(32) presupposition of (31b): tautological
⟦(31b)⟧𝑤 = 1 iff
Emily called in 𝑤 → ∀𝑤 ′ [𝑤 ′ ∈ DOX(𝑠, 𝑤) → Emily called in 𝑤 ′ ]
 I N T E R R O G AT I V E C L A U S E S 167

This says that the sentence (31a) is true in 𝑤 if either one of the following
two conditions is met: either (i) Emily did not call in 𝑤, or else (ii) she did
call in w and Sakina believes in w that she did. This is not satisfactory.
What it gets right is that, if Emily called but Sakina is unaware of this,
then the sentence is false. But it also predicts that, if Emily didn’t call, then
the sentence is true no matter what Sakina believes — even if she wrongly
believes that she did call.
A gut reaction to this problem is that the culprit is our semantics for
polar questions, not our ANS operator. This is what Karttunen would have
said. Indeed, he gave a different semantics for polar questions and did not
have this problem. In our variant of his theory, if we minimally changed How might we do that? Perhaps by giving
the semantics of polar questions so that the sister of ANS were to denote a meaning to whether, letting it denote a
function that maps a singleton set {𝑝 } to
the 2-membered set {that Emily called, that Emily didn’t call}, the truth
the set {𝑝, ¬𝑝 }.
conditions would come out correct without any revision to our entry
for ANS. (Exercise: Convince yourself of this.) This looks like a good way
out — at least at first. But when we look at further problem cases, we will
come to see it is a move that is neither sufficient nor necessary.
Let’s return to the case of the embedded constituent question in (28) The problem we’re discussing here was
and scrutinize the truth conditions we derived in (30) a bit more care- noticed by Karttunen 1977 in a footnote,
and he fixed it by complicating his lexi-
fully. Suppose that w is a world in which nobody called. Then the univer-
cal entry for interrogative-taking know.
sal quantification we computed in (30b) is trivially true: Whatever Sak- He ended up stating the truth condi-
ina’s beliefs in 𝑤 may be, the material conditional ‘[𝑥 called in 𝑤 → Sakina tion in the form of a disjunction, with
believes in 𝑤 that 𝑥 called]’ is true for every 𝑥 (since the antecedent is al- a special clause for the case where the
question-denotation only contains false
ways false). So the sentence (28) is predicted to be true, for example, in propositions. Heim 1994 showed how to
a world where nobody called but Sakina falsely believes that Emily and generalize Karttunen’s special clause to a
Julie called. This does not conform to our intuitions. general solution for all the problem cases
we consider in this section. The solution
Finally, as Groenendijk & Stokhof 1982 forcefully pointed out, even if
we will present in these lecture notes is
we only consider worlds in which some people did in fact call, the truth not quite the same as Heim’s. See papers
conditions imposed by our current (and Karttunen’s) semantics are too by Rullmann & Beck 1998, Beck & Rull-
lax. Suppose that only Emily called, but Sakina thinks that Emily, Julie, mann 1999, Sharvit 2002, and Sharvit &
Guerzoni 2003 for discussion and compar-
and Delphine all called. Would we say that Sakina knows who called?
ison.
We’d be reluctant to. But our semantics deems the sentence Sakina knows
who called to be true in this scenario. After all, Sakina does believe of ev-
ery person who in fact called (namely, of Emily), that that person called.
This is all that our predicted truth conditions require. If our analysis
were right, it simply shouldn’t matter how many false beliefs Sakina has
about people calling who did not in fact call.
Groenendijk & Stokhof argued that the correct semantics for Sakina
knows who called is what they dubbed “strongly exhaustive” — i.e., the
know-sentence is true only when Sakina is fully informed about who
called and who didn’t. She believes that they called of all the people who
did in fact call, and she believes that they didn’t call of all the ones who
didn’t. Can we revise our entry for the answer operator so that it delivers
this more stringent truth condition? Yes, here is how.
168 Q U E S T I O N S

(33) strongly exhaustive answer operator:

⟦ANS⟧𝑤 = 𝜆Q ⟨𝑠𝑡,𝑡 ⟩ .𝜆𝑤 ′ . ∀𝑝 [𝑝 ∈ Q → 𝑝 (𝑤) = 𝑝 (𝑤 ′ )]

(33’) ⟦ANS⟧𝑤 = 𝜆Q ⟨𝑠𝑡,𝑡 ⟩ . {𝑤 ′ : 𝑤 ′ ∼ Q,𝑊 𝑤 } equivalent formulation, using the defini-

tion of “cell-mate” from (4a). Here, capital
What does this semantics for ANS do? It takes the set of propositions Q 𝑊 in the subscript to the ∼-relation is the
denoted by the underlying interrogative and maps the evaluation world set of all possible worlds.

to the worlds that agree with it on the truth-value of all propositions in Q.

In other words, it maps the evaluation world to its cell-mates relative to
Q.
This new semantics solves all the problems that we saw in this section.
For embedded polar questions, it delivers the prediction that if Emily
didn’t call, then Sakina knows whether Emily called is only true if Sakina
knows that Emily didn’t call. For embedded constituent questions, it pre-
dicts that if 𝑥 didn’t call, then Sakina knows who called is not true unless
Sakina knows that 𝑥 didn’t call. As a special case of this latter prediction,
we derive that if nobody called, then Sakina knows who called is only true
if Sakina knows about each person that they didn’t call.

Exercise 8.3 Verify these claims, by doing the computations.

8.3.4 ᴀNS in matrix questions?

Our latest, strongly exhaustive answer operator bears an obvious logical
relation to the pragmatic rule “update by question” (Definition 7) that we
posited earlier to link the semantic values of matrix interrogative clauses
to the speech acts that they serve to perform. In both cases, we use the
set of propositions Q to lump together any evaluation world with those
worlds that agree with it on the truth-values of all propositions in Q.
In the case of interrogatives embedded under responsive predicates like
know, this set of worlds that are cell-mates with the evaluation world is fed
to the attitude predicate as its prejacent proposition. In the case of matrix
questions, we use the set of disjoint but exhaustive cells to partition the
context set.
Given this similarity, we may want to consider positing ANS in matrix
questions as well, instead of the vacuous covert operator 𝑂𝑃. This would
mean rewriting the “update by question” rule. The idea would be that the
pragmatic rule can use the intension of the uttered sentence to construct
the cells of the new partition. The recipe, informally, is to apply this in-
tension to each world in the current context set and intersect each result
with the current context set.

Definition 8 (Update by question, draft revision) To ask a question Note that the relevant sentences 𝜙 are
by uttering a sentence 𝜙 is to propose that the current partitioned context of the form “ANS 𝑄” and relative to any
world 𝑤 denote the proposition that is
set C be replaced by the new partitioned context set
true of any world that agrees with 𝑤 on
∪ ∪ all the propositions in the set denoted by
{𝑝 : ∃𝑤 ∈ C. 𝑝 = ⟦𝜙⟧𝑤 ∩ C}. 𝑄 relative to 𝑤.
 I N T E R R O G AT I V E C L A U S E S 169

We take each world 𝑤 in the prior context set in turn. We evaluate the
matrix question 𝜙 (= ANS 𝑄) in 𝑤 and intersect the resulting proposi-
tion with the context set, thus finding those worlds in the context set that
agree with 𝑤 on all the propositions in ⟦𝑄⟧𝑤 . We collect the resulting set
of sets of worlds to serve as the new partitioned context set.
There is a tricky issue here, which is a reprise of what we discussed on
page 159ff. Take again a test scenario where everyone knows that 𝑎 is a
student but there’s uncertainty about whether 𝑏 is. And there’s maximal
uncertainty about who called. Now, assume that (34) is uttered against
that background:
(34) Which students called?

There are worlds in the context set where 𝑎 and 𝑏 are the students, and
other worlds where only 𝑎 is a student. We saw earlier that in such a con-
text, it is best to rule (34) out as pragmatically infelicitous because the set
of propositions denoted by the interrogative varies among the worlds in
the context set. This was easy enough to do in the system we were work-
ing with at that point. Now, however, we are considering the possibil-
ity that what is being uttered is really “ANS which students called”. That
structure has different semantic values across the worlds in the context set
simply because there are different true answers in those worlds (otherwise
why ask the question?). So, we can’t enforce a presupposition that the in- In an earlier version of these lecture notes,
terrogative have the same value (the same strong true answer) across the it was erroneously claimed that we could
evade the problem this way.
context set.
The set of sets of worlds we would get from Definition 8 in our test
scenario would not in fact be a partition of the context set:

(i) a world 𝑤 where 𝑎 called and 𝑏 is not a student will be lumped with
any world where 𝑎 called and 𝑏 is a student, no matter whether 𝑏
called (since the proposition that 𝑏 called is not in the denotation in
𝑤 of the underlying interrogative);
(ii) a world 𝑤 ′ where 𝑎 called and 𝑏 is a student who called will be
lumped only with other worlds where 𝑎 called and 𝑏 is a student
who called;
(iii) the two sets generated by 𝑤 and 𝑤 ′ are not disjoint, since they both
contain worlds where 𝑎 called and 𝑏 is a student who called.

The only way we can see to prevent this situation is to state directly in
the “update by question” rule that the update is only felicitous if it results
in a partition:

Definition 9 (Update by question, final revision) To ask a question by

uttering a sentence 𝜙 is to propose that the current partitioned context set
C be replaced by the new context set
∪ ∪
{𝑝 : ∃𝑤 ∈ C. 𝑝 = ⟦𝜙⟧𝑤 ∩ C}
170 Q U E S T I O N S

∪
If the resulting set of sets of worlds is not a partition of C, the utter-
ance is infelicitous.

We conclude this subsection by noting that we have now encountered In particular, note that we have derived a
a kind of presupposition that is not grounded in the denotational seman- condition on any wide-scope restriction
on wh-phrases that requires the set char-
tics of the sentences uttered but emerges from the rules and principles of
acterized by that restriction to be “settled”
pragmatics. in the prior context set. It would be inter-
esting to compare this result to ideas about
“d-linking” that are found in the literature
8.3.5 Embedding under rogative verbs on questions.

Access to the intension of the LF headed by ANS also figures plausibly in

the semantics of sentences with non-responsive (“rogative”) question-
embedders, such as the verbs ask and wonder. These verbs have lexical
meanings that suggest rough paraphrases in which know or tell is in the
scope of another intensional operator, e.g., ask = ‘request to be told’, won-
der = ‘want to know’. Following Groenendijk & Stokhof, let’s hypothesize
that these verbs differ in semantic type from responsive verbs. Their (in-
ternal) argument is of type ⟨𝑠, ⟨𝑠, 𝑡⟩⟩ rather than ⟨𝑠, 𝑡⟩. A concrete entry
for wonder along these lines is (35).
DES maps an individual and a world to
those worlds that satisfy the individual’s
(35) ⟦wonder⟧𝑤 = 𝜆𝑞𝑠,𝑠𝑡 .𝜆𝑥𝑒 . desires as they are in the given world.
∀𝑤 ′ [𝑤 ′ ∈ DES(𝑥, 𝑤) → ∀𝑤 ′′ [𝑤 ′′ ∈ DOX(𝑥, 𝑤 ′ ) → 𝑞(𝑤 ′ ) (𝑤 ′′ ) = 1]]

The higher semantic type straightforwardly gives us the prediction

that wonder cannot combine with a that-clause. If we attempted to inter-
pret such an LF, we would encounter a type-mismatch. A that-clause has
an extension of type 𝑡 and an intension of type ⟨𝑠, 𝑡⟩. Whether we used
plain FA or IFA, we wouldn’t obtain the type-⟨𝑠, 𝑠𝑡⟩-function that won-
der is looking for. But if we embed an interrogative clause (headed by
ANS) under wonder, we will succeed. The extension of the mother-node
of ANS is type ⟨𝑠, 𝑡⟩, and its intension is type ⟨𝑠, 𝑠𝑡⟩. So IFA allows us to
interpret the structure.

Exercise 8.4 Convince yourself that, given entry (35), the predicted truth con-
dition for Sakina wonders who called matches the informal paraphrase that we
gave in the text.

8.4 Wh-movement with pied-piping and reconstruction

As an example of the phenomenon of so-called pied-piping, we will an-

alyze the sentence How many cats did you adopt? Before we get to the
point that makes the example interesting for our purposes, we must fill
in a rudimentary account of plurals and gradability.
 I N T E R R O G AT I V E C L A U S E S 171

8.4.1 Background on plurals and many

The basic idea in most current semantic treatments of plural DPs is that
plural definites and pronouns denote entities in 𝐷𝑒 , just like singular defi-
nites, pronouns, and proper names. The only difference is that the entities
denoted by plurals are more complex (and typically spatially discontin-
uous). A distinction is made within the domain 𝐷𝑒 , between so-called
“atoms” or “atomic individuals” (the referents of singular DPs) and “plu-
ralities” or “non-atomic individuals” (the referents of plural DPs). Non-
atomic individuals contain atomic individuals as (proper) parts; e.g., if
Lucy is one of the players, then ⟦Lucy⟧ (i.e., Lucy) is an atomic part of
⟦the players⟧. An atomic individual, on the other hand, has no atomic
parts other than itself. (An atom counts as an atomic part of itself.)
Given that 𝐷𝑒 contains pluralities along with atoms, predicate exten-
sions of type ⟨𝑒, 𝑡⟩ are functions that apply to both atoms and pluralities.
In the case of common nouns, English has a morphological number dis-
tinction which seems to have semantic import:
(36) a. ⟦cat⟧𝑤 = 𝜆𝑥 . is a cat in 𝑤
b. ⟦cats⟧𝑤 = 𝜆𝑥 . every atomic part of 𝑥 is a cat in 𝑤

Being a cat entails being an atomic individual (this is how we agree to Notice that an entity that ⟦cats⟧𝑤 maps
understand our metalanguage). Therefore, the denotation of the singular to 1 is not necessarily a plurality. As in-
terpreted in (36b), the plural noun cats is
noun as defined in (36a) maps every plurality to 0. The pluralized noun
also true of a single cat, because of the fact
in (36b), on the other hand, maps to 1 those pluralities whose atomic parts that an atom is an atomic part of itself. It
are all cats. would be possible to revise (36b) so that
Verbs can show morphological number too, but we assume that this it requires 𝑥 to be non-atomic. But as we
will see below, the current formulation
is always due to morphological agreement with a number-marked sub- actually works better. The reason is, in a
ject, and that the number morphology on the verb is not interpreted it- nutshell, that “One.” is a perfectly good
self. As far as semantics is concerned, verbs are “number neutral” and typ- answer to a how-many question.
ically can be true indiscriminately of both atoms and pluralities. This is
reflected, for example, in a lexical entry like (37).
(37) ⟦meow⟧𝑤 = 𝜆𝑥 . every atomic part of 𝑥 meows in 𝑤

The condition in (37) can be met by both pluralities and atoms. A plu-
rality is mapped to 1 iff all its atomic parts meow, and an atom is mapped
to 1 if it itself meows. (Recall that every atom counts as an atomic part of
itself.)
We can count the atomic parts of a plural individual. For example, the
plural individual composed of Lucy, Kathellen, and Shanice has 3 atomic
parts. Let’s have a concise notation for this.
172 Q U E S T I O N S

(38) Let 𝑥 be an element of 𝐷𝑒 . Then

#(𝑥) := the cardinality of the set {𝑦 : 𝑦 is an atomic part of 𝑥 }.

With this little bit of plural semantics in place, we can now introduce Actually, Hackl assumes (with most of the
Hackl 2001’s semantics for many and an appropriate semantics for inter- literature on adjectives and gradability)
that there is an additional basic type 𝑑
rogative how that will go with it. Hackl proposes that many is not by itself
(for “degrees”) separate from type 𝑒. The
a quantificational determiner of type ⟨𝑒𝑡, ⟨𝑒𝑡, 𝑡⟩⟩. Rather it is looking for number argument of many is a special case
an argument which denotes a number, and only after it has been saturated of a degree argument, and the type for
with such an argument, the resulting phrase is a quantificational deter- many is then ⟨𝑑, ⟨𝑒𝑡, ⟨𝑒𝑡, 𝑡 ⟩⟩ ⟩.

miner. So the type of many is type ⟨𝑒, ⟨𝑒𝑡, ⟨𝑒𝑡, 𝑡⟩⟩⟩ — assuming that num- Hackl’s analysis implies that the superfi-
bers are abstract individuals of some kind, hence members of 𝐷𝑒 — and its cially simplest uses of many, as in Many
entry is as in (39). cats meowed, are actually more complex
at LF: the argument position of many is
(39) ⟦many⟧ = 𝜆𝑛 : 𝑛 is a number. 𝜆𝑓𝑒𝑡 .𝜆𝑔𝑒𝑡 . ∃𝑥 [#(𝑥) = 𝑛 & 𝑓 (𝑥) = bound by a covert POS (“positive opera-
1 & 𝑔(𝑥) = 1] tor”), which is a quantifier over numbers
(degrees) and means something like ‘a
When building a sentence with many, in the simplest case we would fill number (degree) above the contextually
the first argument slot of many with a word that refers to a number. This specified threshold’.

might be an anaphoric demonstrative pronoun that, which in appropriate

discourse contexts can refer to a previously mentioned number, say the
number 3. Or it could be a numeral word, like three, which we take here
to have a meaning of type 𝑒 and be a proper name of the number 3. These
options give us interpretable syntactic representations like (40a,b).
(40) a. [ [ that7 many ] cats ] meowed
b. [ [ three many ] dogs ] barked

(40a) is straightforwardly pronounced as it stands, whereas for (40b),

Hackl assumes that many is unpronounced after numeral words, so this
structure surfaces as Three dogs barked. Let us compute truth-conditions,
using (39). Suppose we have a contextually given assignment for (40a)
which maps the variable 7 to the number 3, and we evaluate the sentence
in the actual world @. Then, by using FA three times to apply ⟦many⟧ to
its three arguments, we obtain the truth-condition in (41).
(41) ∃𝑥 [#(𝑥) = 3 & ⟦cats⟧@ (𝑥) = 1 & ⟦meow⟧@ (𝑥) = 1]

Now we use our entries for cats and meow, and this becomes (42).
(42) ∃𝑥 [#(𝑥) = 3
& every atomic part of 𝑥 is a cat in @
& every atomic part of 𝑥 meows in @ ]

In other words, there is a plurality composed of three meowing cats.

Which is loud but correct.
 I N T E R R O G AT I V E C L A U S E S 173

8.4.2 How many-questions

In a how-many question, the argument slot that was saturated by that or The main source for the argument in this
three in (40) is instead occupied by the wh-word how. In Karttunen’s the- section is Arnim von Stechow’s paper
“Against LF pied-piping” (von Stechow
ory, this will be an existential quantifier, equivalent to some number.
1996).
(43) ⟦how⟧ = 𝜆𝑓𝑒𝑡 . ∃𝑛[𝑛 is a number & 𝑓 (𝑛) = 1]
(type ⟨𝑒𝑡, 𝑡⟩, a generalized quantifier) For Hackl, it would be type ⟨𝑑𝑡, 𝑡 ⟩, a
generalized quantifier over degrees.
This semantic type is not interpretable in situ as the sister of many, and
must undergo (covert) movement for interpretability. With this in mind,
let’s attempt a syntactic derivation for the question How many cats did you
adopt?
(44) a. [𝐶 ? OP] [you adopted how many cats]
b. OP 5 [ [? 𝑡 5 ] [you adopted how many cats] ]
c. OP 5[how many cats 1 [[? 𝑡 5 ] you adopted 𝑡 1 ]
d. OP 5[how 2[𝑡 2 many cats 1[[? 𝑡 5 ] you adopted 𝑡 1 ] ] ]

We have three movements that derive the LF from the base generated
structure in (44a): the movement of the empty OP in (44b), the wh-
movement of the wh-phrase how many cats in (44c), and the QR of the
quantifier how in (44d).
We can check the semantic types to confirm that we have derived an
interpretable structure (do this as an exercise). Let’s compute what (44d)
means. (The details are left as an exercise. Here we conflate sets of propo-
sitions with their characteristic functions.)
(45) ⟦(44d)⟧@ =
…
= {𝑝 : ∃𝑛 [𝑛 is a number & ∃𝑥 [#(𝑥) = 𝑛 & ⟦cats⟧@ (𝑥) = 1] &
𝑝 = 𝜆𝑤 .⟦adopt⟧𝑤 (𝑥)(you)]}

With a little bit of Predicate Logic reasoning, we can rewrite this equiva- We are exploiting the equivalence of vari-
lently as follows: ous scopal arrangements in a formula with
existential quantifiers and conjunctions.
(46) {𝑝 : ∃𝑥 [ ∃𝑛[𝑛 is a number & #(𝑥) = 𝑛] & ⟦cats⟧@ (𝑥) = 1 & The following four are all equivalent.
𝑝 = 𝜆𝑤 .⟦adopt⟧𝑤 (𝑥) (you)]} ∃𝑥 [𝐹𝑥 & ∃𝑦 [𝑅𝑥 𝑦 & 𝐺 𝑦 ] ]
∃𝑥 ∃𝑦 [𝐹𝑥 & 𝑅𝑥 𝑦 & 𝐺 𝑦 ]
We can now contemplate the underlined part and convince ourselves that ∃𝑦∃𝑥 [𝐹𝑥 & 𝑅𝑥 𝑦 & 𝐺 𝑦 ]
∃𝑦 [∃𝑥 [𝐹𝑥 & 𝑅𝑥 𝑦 ] & 𝐺 𝑦 ]
this part is a tautology. It just says that 𝑥 has some number or other of
atomic parts, which cannot fail to be true. So we might as well drop this
conjunct and rewrite (46) as (47).
(47) {𝑝 : ∃𝑥 [⟦cats⟧@ (𝑥) = 1 & 𝑝 = 𝜆𝑤 .⟦adopt⟧𝑤 (𝑥) (you)]}

Interestingly now, this is precisely the meaning we would have derived

for the question Which cats did you adopt? In other words, if (44d) really
were the LF (or one of the LFs) for the how- many-question How many
cats did you adopt?, then we would be making the bad prediction that this
174 Q U E S T I O N S

question is synonymous with (or at least shares a reading with) the which-
question Which cats did you adopt? This would be unfortunate for our the-
ory.
Upon closer inspection of (44d), however, it turns out that our theory
does not actually generate this LF. We have neglected to check whether
(44d) conforms to the Wh-Licensing Principle, repeated here.
(48) At LF, a phrase 𝛼 occupies a specifier position of ? if and only if 𝛼
has the feature [WH].

In (44d), we have two phrases that are scoped above ?, namely how and 𝑡 2
many cats. Let’s say that the positions they occupy both count as specifier
positions of ? (similar to what one has to say for multiple questions like
who ate what? — see Exercise 8.1). Then (48) would require that both of
these phrases have the feature [WH]. But only how actually does, at least
on our current assumption that [WH] is a lexical property possessed by
only a small set of words. The other phrase that is scoped above ? in (44d)
is 𝑡 2 many cats, which does not carry the feature [WH]. So the structure
(44d) is filtered out by the Wh-Licensing principle as syntactically ill-
formed. And this is a good thing, because it means we don’t generate the
unwelcome reading in (47).
We still have to worry, however, about how we generate the read-
ing that our example actually does have. Is there a second, well-formed,
LF for our example? The answer is yes if our syntax allows for “recon-
struction”, i.e. some mechanism by which overtly moved phrases can be
restored to (one of ) their pre-moved positions at LF. In order to satisfy
the Wh-Licensing Principle, reconstruction must apply to the phrase 𝑡 2
many cats (though not, of course, to how). Where can this phrase be re-
constructed to? Well, if we restored it all the way down to its original base
position as the object of adopt, it wouldn’t be interpretable there, because
quantifiers are not interpretable in object positions. But if we can assume
that wh-movement proceeds (or at least is allowed to proceed) successive
cyclically, and that an object wh-phrase can stop over e.g. at the edge of
VP before it moves on to Spec-CP, we can target this intermediate land-
ing site for reconstruction. (Assuming the VP-internal subject hypothe-
sis, VPs are semantically of type 𝑡 and hence quantifiers are interpretable
at their edges.) By means of reconstruction, we can thus obtain another
LF that is both interpretable and in compliance with the WH-Licensing
principle.
(49) OP 5[how 2[[? 𝑡 5 ] 𝑡 2 many cats 1[you adopted 𝑡 1 ] ] ]

The denotation of (49) (which you should compute as an exercise!) is (50).

 I N T E R R O G AT I V E C L A U S E S 175

(50) {𝑝 : ∃𝑛[𝑛 is a number & 𝑝 = 𝜆𝑤 . ∃𝑥 [#(𝑥) = 𝑛 & ⟦cats⟧𝑤 (𝑥) =

1 & ⟦adopt⟧𝑤 (𝑥) (you)]}

This set contains one proposition per number. It contains the proposi-
tion that you adopted 1 cat, the proposition that you adopted 2 cats, the
proposition that you adopted 3 cats, etc. This prediction accords well with
what we feel are expected answers to the question how many cats did you
adopt?

8.4.3 More on reconstruction

Earlier in these notes in Section 4.2.2, we briefly introduced a version of
the “copy theory of movement”, in which reconstruction can be imple-
mented as deletion of a higher copy. We reproduce the relevant passage
here:
This assumes that movement generally proceeds in two separate
steps, rather than as a single complex operation as we have assumed
so far. Recall that in H&K, it was stipulated that every movement
effects the following four changes:

(i) a phrase 𝛼 is deleted,

(ii) an index i is attached to the resulting empty node (making it
a so-called trace, which the semantic rule for “Pronouns and
Traces” recognizes as a variable),
(iii) a new copy of 𝛼 is created somewhere else in the tree (at the
“landing site”), and
(iv) the sister-constituent of this new copy gets another instance of
the index i adjoined to it (which the semantic rule of Predicate
Abstraction recognizes as a binder index).

If we adopt the Copy Theory, we assume instead that there are

three distinct operations:

“Copy”: Create a new copy of 𝛼 somewhere in the tree, attach an

index i to the original 𝛼, and adjoin another instance of i to the
sister of the new copy of 𝛼. (= steps (ii), (iii), and (iv) above)
“Delete Lower Copy”: Delete the original 𝛼. (= step (i) above)
“Delete Upper Copy”: Delete the new copy of 𝛼 and both instances of
i.

The Copy operation is part of every movement operation, and can

happen anywhere in the syntactic derivation. The Delete opera-
tions happen at the end of the LF derivation and at the end of the PF
deletion. We have a choice of applying either Delete Lower Copy
or Delete Upper Copy to each pair of copies, and we can make this
choice independently at LF and at PF. (E.g., we can do Copy in the
common part of the derivation and than Delete Lower Copy at LF
and Delete Upper Copy at PF.)
176 Q U E S T I O N S

When we try to apply this machinery to reconstruction of pied-piped

material in wh-movement, there is a detail that needs attention: After
having created two copies (or more) of how many cats, what happens
when we “move” how? Presumably we create another, higher, copy of
how and coindex it with the lower copy of how. With which lower copy
of how? With the one in the uppermost copy of how many cats? If so, then
after deletion of that uppermost copy of how many cats, the binder index
next to the top copy of how will no longer bind any trace and the struc-
ture will not be interpretable. What we want is to end up with is an LF
in which the top copy of how binds a variable in a lower (retained) copy
of how many cats. There may be various ways to achieve this. Perhaps we
can simply choose a derivation in which how “moves” out of a lower copy
of how many cats, not out the highest copy. (Richards 1998’s Principle of
Minimal Compliance might explain why this long movement is legiti-
mate and yet relies for its legitimacy on the previous movement of the
larger phrase.) Alternatively, we may posit a general principle “Copies
must remain copies”, which says that, as long as a structure contains more
than one copy of a given phrase, every alteration to one of these copies
must identically affect them all. In our case, this would mean that when
we create the top copy of how, we must coindex it with all its existing
copies, in the highest as well as all lower instances of how many cats. This
principle could hopefully be made to fall out from a suitable formalization
of copy theory, perhaps in a multi-dominance framework.
Bibliography
178 B I B L I O G R A P H Y

Abusch, Dorit. 1988. Sequence of tense, intensionality and scope. West

Coast Conference on Formal Linguistics (WCCFL) 7. 1–14.
Abusch, Dorit. 1994. The scope of indefinites. Natural Language Semantics
2(2). 83–135. https://doi.org/10.1007/BF01250400.
Abusch, Dorit. 1997. Sequence of tense and temporal de re. Linguistics and
Philosophy 20(1). 1–50. https://doi.org/10.1023/A:1005331423820.
Abusch, Dorit & Mats Rooth. 2017. The formal semantics of free per-
ception in pictorial narratives. Amsterdam Colloquium 21. https : / /
semanticsarchive.net/Archive/DVlMGQwY/abusch_rooth_ams_
2017.pdf.
Aikhenvald, Alexandra Y. 2004. Evidentiality. Oxford: Oxford University
Press.
Aloni, Maria. 2007. Free choice, modals, and imperatives. Natural Lan-
guage Semantics 15(1). 65–94. https://doi.org/10.1007/s11050- 007-
9010-2.
Alonso-Ovalle, Luis. 2006. Disjunction in alternative semantics. University
of Massachusetts at Amherst PhD thesis. http : / / alonso - ovalle . net /
index.php?page_id=28.
Anand, Pranav & Valentine Hacquard. 2013. Epistemics and attitudes.
Semantics and Pragmatics 6(8). 1–59. https://doi.org/10.3765/sp.6.8.
Arregui, Ana, María Luisa Rivero & Andrés Salanova. 2014. Cross-linguistic
variation in imperfectivity. Natural Language and Linguistic Theory
32(2). 307–362. https://doi.org/10.1007/s11049-013-9226-4.
Arvan, Marcus. 2015. The Lewis-Stalnaker analysis of counterfactuals:
What about Sprigge? https://philosopherscocoon.typepad.com/the_
campaign_for_better_c/2015/03/the- lewis- stalnaker- analysis- of-
counterfactuals-what-about-sprigge.html.
Asher, Nicholas. 1987. A typology for attitude verbs and their anaphoric
properties. Linguistics and Philosophy 10(2). 125–197. https://doi.org/
10.1007/BF00584317.
Auwera, Johan Van Der & Alfonso Zamorano Aguilar. 2015. The history
of modality and mood. In Jan Nuyts & Johan Van Der Auwera (eds.),
The Oxford handbook of modality and mood. Oxford University Press.
https://doi.org/10.1093/oxfordhb/9780199591435.013.4.
Bach, Emmon. 1986. The algebra of events. Linguistics and Philosophy
9(1). 5–16. https://doi.org/10.1007/BF00627432.
Baltin, Mark. 2010. The copy theory of movement and the binding-
theoretic status of A-traces: You can’t get there from here. http : / /
as.nyu.edu/content/dam/nyu- as/linguistics/documents/nyuwpl/
NYUWPL_vol_3_Baltin.pdf.
Barcan, Ruth C. 1946. A functional calculus of first order based on strict
implication. Journal of Symbolic Logic 11(1). 1–16. https://doi.org/10.
2307/2269159.
 BI BLIOGRAPHY 179

Barwise, John & Robin Cooper. 1981. Generalized quantifiers and natural
language. Linguistics and Philosophy 4(2). 159–219. https://doi.org/10.
1007/BF00350139.
Bassi, Itai & Moshe E. Bar-Lev. 2017. A unified existential semantics for
bare conditionals. Sinn und Bedeutung 21. https://sites.google.com/site/
sinnundbedeutung21/proceedings- preprints/SuB%2021%20Bar-
Lev%20and%20Bassi%20final.pdf.
Bauer, Matthias & Sigrid Beck. 2014. On the meaning of fictional texts.
In Daniel Gutzmann, Jan Köpping & Cécile Meier (eds.), Approaches to
meaning: Composition, values, and interpretation (Current Research in the
Semantics / Pragmatics Interface 32), 250–275. Brill. https://doi.org/
10.1163/9789004279377_012.
Bäuerle, Rainer. 1983. Pragmatisch-Semantische Aspekte der NP-Interpretation.
In M. Faust, R. Harweg, W. Lehfeldt & G. Wienold (eds.), Allgemeine
Sprachwissenschaft, Sprachtypologie und Textlinguistik: Festschrift für Peter
Hartmann, 121–131. Tübingen: Narr.
Beck, Sigrid. 1996. Wh-constructions and transparent Logical Form. Univer-
sität Tübingen PhD Thesis.
Beck, Sigrid & Hotze Rullmann. 1999. A flexible approach to exhaustivity
in questions. Natural Language Semantics 7(3). 249–298. https://doi.org/
10.1023/A:1008373224343.
Bennett, Jonathan. 2003. A philosophical guide to conditionals. Oxford Uni-
versity Press.
Bennett, Michael & Barbara Partee. 1978. Toward the logic of tense and as-
pect in English. Indiana University Linguistics Club.
Bhatt, Rajesh. 1997. Obligation and possession. In Heidi Harley (ed.),
Papers from the upenn/mit roundtable on argument structure and aspect,
vol. 32 (MIT Working Papers in Linguistics), 21–40. http : / / people .
umass.edu/bhatt/papers/bhatt-haveto.pdf.
Bhatt, Rajesh & Roumyana Pancheva. 2006. Conditionals. In The Black-
well companion to syntax, vol. 1, 638–687. Blackwell. http : / / www -
rcf.usc.edu/~pancheva/bhatt-pancheva_syncom.pdf.
Blain, Eleanor M. & Rose-Marie Déchaine. 2007. Evidential types: Evi-
dence from Cree dialects. International Journal of American Linguistics
(IJAL) 73(3). 257–291. https://doi.org/10.1086/521728.
Blümel, Andreas. 2019. Adnominal conditionals in German. Linguistics
Vanguard 5. https://doi.org/10.1515/lingvan-2019-0004.
Blumson, Ben. 2009. Pictures, perspective and possibility. Philosophical
Studies. https://doi.org/10.1007/s11098-009-9337-2.
Boeckx, Cedric. 2001. Scope reconstruction and A-movement. Natural
Language and Linguistic Theory 19(3). 503–548. https://doi.org/10.1023/
a:1010646425448.
180 B I B L I O G R A P H Y

Bogal-Allbritten, Elizabeth. 2013. Modification of DPs by epistemic modal

adverbs. Proceedings of the 19th Amsterdam Colloquium. 51–58. http :
//blogs.umass.edu/eba/files/2009/05/BogalAllbritten-2013-AC.pdf.
Bogal-Allbritten, Elizabeth & Andrew Weir. 2017. Sentential and possibly
subsentential modification: The ambiguity of Collins conjunctions.
North East Linguistics Society (NELS) 47. https : / / elizabethba . files .
wordpress.com/2017/04/eba-aw-nels.pdf.
Bonomi, Andrea & Sandro Zucchi. 2003. A pragmatic framework for
truth in fiction. Dialectica 57(2). Preprint http://filosofia.dipafilo.unimi.
it/~bonomi/Pragmatic.pdf, 103–120. https://doi.org/10.1111/j.1746-
8361.2003.tb00259.x.
Butler, Jonny. 2003. A minimalist treatment of modality. Lingua 113(10).
967–996. https://doi.org/10.1016/S0024-3841(02)00146-8.
Carnap, Rudolf. 1947. Meaning and necessity: A study in semantics and modal
logic. Chicago: University of Chicago Press.
Carr, Jennifer. 2014. The if p, ought p problem. Pacific Philosophical Quar-
terly 95(4). 555–583. https://doi.org/10.1111/papq.12048.
Cheke, Lucy G. & Nicola S. Clayton. 2010. Mental time travel in animals.
Wiley Interdisciplinary Reviews: Cognitive Science 1(6). 915–930. https:
//doi.org/10.1002/wcs.59.
Chierchia, Gennaro. 1995. Dynamics of meaning: Anaphora, presupposition,
and the theory of grammar. University of Chicago Press.
Chierchia, Gennaro & Sally McConnell-Ginet. 2000. Meaning and gram-
mar: An introduction to semantics (2nd edition). MIT Press.
Childe, V. Gordon. 1936. Man makes himself. London: C.A. Watts & Co.
Chomsky, Noam. 1993. A minimalist program for linguistic theory. In
Kenneth Hale & Samuel Jay Keyser (eds.), The view from Building 20:
Essays in linguistics in honor of Sylvain Bromberger, 1–52. Cambridge,
MA: MIT Press.
Condoravdi, Cleo & Sven Lauer. 2016. Anankastic conditionals are just
conditionals. Semantics and Pragmatics 9(8). 1–69. https://doi.org/10.
3765/sp.9.8.
Copeland, B. Jack. 2002. The genesis of possible worlds semantics. Journal
of Philosophical Logic 31(2). 99–137. https : / / doi . org / 10 . 1023 / A :
1015273407895.
Copley, Bridget. 2008. The plan’s the thing: Deconstructing futurate
meanings. Linguistic Inquiry 39(2). 261–274. https://doi.org/10.1162/
ling.2008.39.2.261.
Cormack, Annabel & Neil Smith. 2002. Modals and negation in English.
In Sjef Barbiers, Frits Beukema & Wim van der Wurff (eds.), Modality
and its interaction with the verbal system, 133–163. Benjamins.
Cresswell, Max. 1973. Logics and languages. London: Methuen.
Davey, B. A. & H. A. Priestley. 2002. Introduction to lattices and order. 2nd edn.
Cambridge University Press. https://doi.org/10.1017/CBO9780511809088.
 BI BLIOGRAPHY 181

Dayal, Veneeta. 1996. Locality in WH quantification: Questions and rela-

tive clauses in Hindi (Studies in Linguistics and Philosophy). Springer
Netherlands. https://doi.org/10.1007/978-94-011-4808-5.
Dekker, Paul. 2001. On if and only. Semantics and Linguistic Theory (SALT)
11. 114–133. https://doi.org/10.3765/salt.v11i0.3097.
Del Pinal, Guillermo & Brandon Waldon. 2019. Modals under epistemic
tension. Natural Language Semantics 27(2). 135–188. https://doi.org/10.
1007/s11050-019-09151-w.
DeRose, Keith. 1991. Epistemic possibilities. The Philosophical Review
100(4). 581–605. https://doi.org/10.2307/2185175.
Dowty, David R. 1977. Toward a semantic analysis of verb aspect and the
english ‘imperfective’ progressive. Linguistics and Philosophy 1(1). 45–
77. https://doi.org/10.1007/BF00351936.
Dowty, David R. 1979. Word meaning and Montague grammar: The se-
mantics of verbs and times in generative semantics and in Montague’s PTQ.
Dordrecht: Reidel.
Dowty, David R., Robert Wall & Stanley Peters. 1981. Introduction to
Montague semantics. Kluwer.
Drubig, Hans Bernhard. 2001. On the syntactic form of epistemic modal-
ity. ms, Universität Tübingen. http://www.sfb441.uni-tuebingen.de/
b2/papers/DrubigModality.pdf.
Ebert, Christian, Cornelia Ebert & Stefan Hinterwimmer. 2014. A unified
analysis of conditionals as topics. Linguistics and Philosophy 37(5). 353–
408. https://doi.org/10.1007/s10988-014-9158-4.
Edgington, Dorothy. 1995. On conditionals. Mind 104(414). 235–329.
10.1093/mind/104.414.235.
Egan, Andy. 2007. Epistemic modals, relativism, and assertion. Philosophi-
cal Studies 133(1). 1–22. https://doi.org/10.1007/s11098-006-9003-x.
Egan, Andy, John Hawthorne & Brian Weatherson. 2005. Epistemic
modals in context. In Gerhard Preyer & Georg Peter (eds.), Contex-
tualism in philosophy: Knowledge, meaning, and truth, 131–170. Oxford:
Oxford University Press.
Egré, Paul & Mikaël Cozic. 2011. If -clauses and probability operators.
Topoi 30(1). 17–29. https://doi.org/10.1007/s11245-010-9087-y.
Emonds, Joseph. 2011. From primate to human in two easy steps. In Nicholas
Saul & Simon J. James (eds.), The evolution of literature: Legacies of Dar-
win in European cultures (Internationale Forschungen zur Allgemeinen
und Vergleichenden Literaturwissenschaft 152), 51–72. Brill | Rodopi.
https://doi.org/10.1163/9789401206846_006.
Farkas, Donka F. & Kim B. Bruce. 2010. On reacting to assertions and
polar questions. Journal of Semantics 27(1). 81–118. https://doi.org/10.
1093/jos/ffp010.
182 B I B L I O G R A P H Y

Filip, Hana. 2012. Lexical aspect. In Robert I. Binnick (ed.), The Oxford
handbook of tense and aspect. Oxford University Press. https://doi.org/
10.1093/oxfordhb/9780195381979.013.0025.
von Fintel, Kai. 1994. Restrictions on quantifier domains. Amherst, MA:
University of Massachusetts PhD thesis. http://semanticsarchive.net/
Archive/jA3N2IwN/fintel-1994-thesis.pdf.
von Fintel, Kai. 1997. Bare plurals, bare conditionals, and only. Journal of
Semantics 14(1). 1–56. https://doi.org/10.1093/jos/14.1.1.
von Fintel, Kai. 1998. Quantifiers and if -clauses. The Philosophical Quar-
terly 48(191). 209–214. https://doi.org/10.1111/1467-9213.00095.
von Fintel, Kai. 1999. NPI licensing, Strawson entailment, and context
dependency. Journal of Semantics 16(2). 97–148. https: // doi. org/ 10.
1093/jos/16.2.97.
von Fintel, Kai. 2005. Modality and language. In Donald M. Borchert
(ed.), Encyclopedia of philosophy – second edition. MacMillan. http://mit.
edu/fintel/fintel-2005-modality.pdf.
von Fintel, Kai. 2011. Conditionals. In Claudia Maienborn, Klaus von
Heusinger & Paul Portner (eds.), Semantics: An international handbook
of meaning, vol. 2, 1515–1538. de Gruyter. https://doi.org/10.1515/
9783110255072 . 1515. http : / / mit . edu / fintel / fintel - 2011 - hsk -
conditionals.pdf.
von Fintel, Kai. 2012. Subjunctive conditionals. In Gillian Russell & Delia
Graff Fara (eds.), The Routledge companion to philosophy of language,
466–477. New York: Routledge. https : / / doi . org / 1721 . 1 / 95784.
http://mit.edu/fintel/fintel-2012-subjunctives.pdf.
von Fintel, Kai & Anthony S. Gillies. 2007. An opinionated guide to epis-
temic modality. In Tamar Szabó Gendler & John Hawthorne (eds.),
Oxford studies in epistemology: Volume 2, 32–62. Oxford University
Press. http://mit.edu/fintel/fintel-gillies-2007-ose2.pdf.
von Fintel, Kai & Anthony S. Gillies. 2008a. CIA leaks. The Philosophical
Review 117(1). 77–98. https://doi.org/10.1215/00318108-2007-025.
von Fintel, Kai & Anthony S. Gillies. 2008b. MIGHT made right. To ap-
pear in a volume on epistemic modality, edited by Andy Egan and
Brian Weatherson, Oxford University Press. http : / / mit . edu / fintel /
fintel-gillies-2008-mmr.pdf.
von Fintel, Kai & Anthony S. Gillies. 2010. Must …stay …strong! Natural
Language Semantics 18(4). 351–383. https://doi.org/10.1007/s11050-
010-9058-2.
von Fintel, Kai & Anthony S. Gillies. 2011. MIGHT made right. In Andy
Egan & Brian Weatherson (eds.), Epistemic modality, 108–130. Oxford:
Oxford University Press. http://mit.edu/fintel/fintel- gillies- 2011-
mmr.pdf.
von Fintel, Kai & Anthony S. Gillies. 2015. Hedging your ifs and vice
versa. http://mit.edu/fintel/fintel-gillies-2015-hedging.pdf.
 BI BLIOGRAPHY 183

von Fintel, Kai & Anthony S. Gillies. 2020. Still going strong. Natural
Language Semantics. to appear.
von Fintel, Kai & Sabine Iatridou. 2002. If and when if -clauses can re-
strict quantifiers. ms, MIT. http://mit.edu/fintel/fintel-iatridou-2002-
ifwhen.pdf.
von Fintel, Kai & Sabine Iatridou. 2003. Epistemic containment. Linguistic
Inquiry 34(2). 173–198. https://doi.org/10.1162/002438903321663370.
von Fintel, Kai & Sabine Iatridou. 2008. How to say ought in Foreign:
The composition of weak necessity modals. In Jacqueline Guéron &
Jacqueline Lecarme (eds.), Time and modality (Studies in Natural Lan-
guage and Linguistic Theory 75), 115–141. Springer. https://doi.org/
10.1007/978-1-4020-8354-9.
von Fintel, Kai & Sabine Iatridou. 2017. A modest proposal for the mean-
ing of imperatives. In Ana Arregui, María Luisa Rivero & Andrés
Salanova (eds.), Modality across syntactic categories, 288–319. Oxford
University Press. https://doi.org/10.1093/acprof:oso/9780198718208.
003.0013. http://mit.edu/fintel/fintel-iatridou-2017-modest.pdf.
von Fintel, Kai & Sabine Iatridou. 2019. Since since. In Daniel Altshuler &
Jessica Rett (eds.), The semantics of plurals, focus, degrees, and times: Essays
in honor of Roger Schwarzschild, 305–333. Springer. https://doi.org/10.
1007/978-3-030-04438-1_15. http://mit.edu/fintel/fintel-iatridou-
2019-since.pdf.
Fodor, Janet Dean. 1970. The linguistic description of opaque contexts. Cam-
bridge, MA: Massachusetts Institute of Technology PhD thesis. https:
//doi.org/1721.1/12970.
Fox, Danny. 2000. Economy and semantic interpretation. MIT Press.
Fox, Danny. 2007. Free choice and the theory of scalar implicatures. In
Uli Sauerland & Penka Stateva (eds.), Presupposition and implicature in
compositional semantics (Palgrave Studies in Pragmatics, Language and
Cognition), 71–120. London: Palgrave Macmillan. https://doi.org/10.
1057/9780230210752_4. http://web.mit.edu/linguistics/people/faculty/
fox/free_choice.pdf.
Fox, Danny. 2018. Partition by exhaustification: Comments on Dayal
1996. Sinn und Bedeutung 22(1). 403–434. https://ojs.ub.uni-konstanz.
de/sub/index.php/sub/article/download/98/41/.
van Fraasen, Bas C. 1980. Critical notice of Brian Ellis, Rational Belief
Systems. Canadian Journal of Philosophy 10(3). 497–511.
Frana, Ilaria. 2017. Modality in the nominal domain: The case of ad-
nominal conditionals. In Ana Arregui, María Luisa Rivero & Andrés
Salanova (eds.), Modality across syntactic categories, 49–69. Oxford Uni-
versity Press.
Francez, Itamar. 2017. Summative existentials. To appear in Linguistic In-
quiry. https : / / lucian . uchicago . edu / blogs / ifrancez / files / 2017 / 05 /
Summex-preprint.pdf.
184 B I B L I O G R A P H Y

Frank, Anette. 1996. Context dependence in modal constructions. Universität

Stuttgart PhD thesis. http://www.cl.uni-heidelberg.de/~frank/papers/
header.pdf.
Gajewski, Jon. 2005. Neg-raising: Polarity and presupposition. Massachusetts
Institute of Technology PhD thesis. https://doi.org/1721.1/33696.
Gajewski, Jon. 2007. Neg-raising and polarity. Linguistics and Philosophy.
https://doi.org/10.1007/s10988-007-9020-z.
Gamut, L. T. F. 1991. Logic, language, and meaning. Chicago University
Press.
Garson, James. 2018. Modal logic. In Edward N. Zalta (ed.), The Stanford
encyclopedia of philosophy. https://plato.stanford.edu/archives/fall2018/
entries/logic-modal/.
Geurts, Bart. 1998. Presuppositions and anaphors in attitude contexts.
Linguistics and Philosophy 21(6). 545–601. https://doi.org/10.1023/A:
1005481821597.
Geurts, Bart. 2004. On an ambiguity in quantified conditionals. ms, Uni-
versity of Nijmegen. http://ncs.ruhosting.nl/bart/papers/conditionals.
pdf.
Giannakidou, Anastasia. 1999. Affective dependencies. Linguistics and Phi-
losophy 22(4). 367–421. https://doi.org/10.1023/A:1005492130684.
Gillies, Anthony S. 2010. Iffiness. Semantics and Pragmatics 3(4). 1–42.
https://doi.org/10.3765/sp.3.4.
Goodman, Nelson. 1947. The problem of counterfactual conditionals.
Journal of Philosophy 44(5). 113–128. https://doi.org/10.2307/2019988.
Greenberg, Gabriel. 2018. Content and target in pictorial representation.
Ergo, an Open Access Journal of Philosophy 5(33). https : / / doi . org / 10 .
3998/ergo.12405314.0005.033.
Groenendijk, Jeroen. 1999. The logic of interrogation: Classical version.
Semantics and Linguistic Theory (SALT) 9. 109–126. https://doi.org/10.
3765/salt.v9i0.2835.
Groenendijk, Jeroen & Martin Stokhof. 1982. Semantics analysis of wh-
complements. Linguistics and Philosophy 5(2). 175–233. https://doi.org/
10.1007/BF00351052.
Hacking, Ian. 1967. Possibility. The Philosophical Review 76(2). 143–168.
https : / / doi . org / 10 . 2307 / 2183640. http : / / www . jstor . org / stable /
2183640.
Hackl, Martin. 2001. Comparative quantifiers. Massachusetts Institute of
Technology PhD thesis. https://doi.org/1721.1/8765.
Hacquard, Valentine. 2006. Aspects of modality. Massachusetts Institute of
Technology PhD thesis. http://people.umass.edu/hacquard/hacquard_
thesis.pdf.
Hacquard, Valentine. 2009a. Modality. ms, prepared for Semantics: An in-
ternational handbook of meaning, edited by Klaus von Heusinger, Clau-
 BI BLIOGRAPHY 185

dia Maienbon, and Paul Portner. http : / / ling . umd . edu / ~hacquard /
papers/HoS_Modality_Hacquard.pdf.
Hacquard, Valentine. 2009b. On the interaction of aspect and modal aux-
iliaries. Linguistics and Philosophy 32(3). 279–315. https://doi.org/10.
1007/s10988-009-9061-6.
Hacquard, Valentine. 2010. On the event relativity of modal auxiliaries.
Natural Language Semantics 18(1). 79–114. https://doi.org/10.1007/
s11050-010-9056-4.
Hacquard, Valentine. 2013. The grammatical category of modality. Pro-
ceedings of the 19th Amsterdam Colloquium. http://www.illc.uva.nl/AC/
AC2013/uploaded_files/inlineitem/03_Hacquard.pdf.
Hacquard, Valentine. 2016. Actuality entailments. to appear in L. Matthew-
son, C. Meier, H. Rullmann, T. E. Zimmermann (eds.) Companion to
Semantics. Wiley. http://ling.umd.edu/~hacquard/papers/Hacquard_
Actuality%20Entailments_July%202016.pdf.
Hagstrom, Paul. 2003. What questions mean. Glot International 7(7/8).
188–201.
Hanley, Richard. 2004. As good as it gets: Lewis on truth in fiction. Aus-
tralasian Journal of Philosophy 82(1). 112–128. https://doi.org/10.1080/
713659790.
Hawthorne, John. 2007. Eavesdroppers and epistemic modals. ms, Rutgers
University, to appear in the proceedings of the 2007 Sofia Conference
in Mexico, in a supplement to Noûs.
Heim, Irene. 1992. Presupposition projection and the semantics of atti-
tude verbs. Journal of Semantics 9(3). 183–221. https://doi.org/10.1093/
jos/9.3.183.
Heim, Irene. 1994. Interrogative semantics and Karttunen’s semantics for
know. In Rhonna Buchalla & Anita Mittwoch (eds.), The proceedings of
the conference of the israel association for theoretical linguistics (iatl 1), 128–
144. Hebrew University of Jerusalem. http : / / semanticsarchive . net /
Archive/jUzYjk1O.
Heim, Irene & Angelika Kratzer. 1998. Semantics in generative grammar.
Oxford: Blackwell.
Herburger, Elena. 2015. Only if : If only we understood it. Sinn und Be-
deutung 19. 284–301. https://www.uni-goettingen.de/de/document/
download/139914c001e533bc9ce6f1f01127253e.pdf/herburger_sub19.
pdf.
Herburger, Elena. 2016. Conditional perfection: The truth and the whole
truth. Semantics and Linguistic Theory (SALT) 25. 615–635. https://doi.
org/10.3765/salt.v25i0.3079.
von Heusinger, Klaus. 2011. Specificity. In Claudia Maienborn, Klaus von
Heusinger & Paul Portner (eds.), Semantics: An international handbook
of meaning, vol. 2 (Handbücher zur Sprach- und Kommunikationswis-
186 B I B L I O G R A P H Y

senschaft (HSK) 33.2), 1025–1058. de Gruyter. https: // doi. org/ 10.

1515/9783110255072.1025.
Higginbotham, James. 1986. Linguistic theory and Davidson’s program
in semantics. In Ernest LePore (ed.), Truth and interpretation: Perspec-
tives on the philosophy of Donald Davidson, 29–48. Basil Blackwell.
Higginbotham, James. 2003. Conditionals and compositionality. Philo-
sophical Perspectives 17(1). 181–194. https://doi.org/10.1111/j.1520-
8583.2003.00008.x.
Hintikka, Jaako. 1969. Semantics for propositional attitudes. In J.W. Davis,
D.J. Hockney & W.K. Wilson (eds.), Philosophical logic, 21–45. Dor-
drecht: Reidel.
Hirsch, Aron. 2017. An inflexible semantics for cross-categorial operators.
Massachusetts Institute of Technology PhD. https://hirsch.mit.edu/
papers.
Hockett, Charles F. 1960. Logical considerations in the study of animal
communication. In W.E. Lanyon & W.N. Tavolga (eds.), Animal
sounds and communication, 392–430. Washington, DC: American In-
stitute of Biological Sciences.
Hockett, Charles F. & Stuart A. Altmann. 1968. A note on design fea-
tures. In Thomas A. Sebeok (ed.), Animal communication: Techniques of
study and results of research, 61–72. Indiana University Press.
Horn, Laurence R. 1972. On the semantic properties of the logical operators in
English. UCLA PhD thesis.
Hughes, G.E. & M.J. Cresswell. 1968. An introduction to modal logic. Lon-
don: Methuen.
Hughes, G.E. & M.J. Cresswell. 1996. A new introduction to modal logic.
London: Routledge.
Huitink, Janneke. 2010. Quantified conditionals and compositionality.
Language and Linguistics Compass 4(1). 42–53. https://doi.org/10.1111/j.
1749-818X.2009.00175.x.
Hulstijn, Joris. 1997. Structured information states: Raising and resolv-
ing issues. In Anton Benz & Gerhard Jäger (eds.), MunDial’97: Mu-
nich workshop on formal semantics and pragmatics of dialogue, University of
Munich, March 10–12, 1997, Proceedings (CIS-Berichte), 99–117. Cen-
trum für Informations- und Sprachverarbeitung, University of Mu-
nich. http://www.sfs.uni-tuebingen.de/~gjaeger/conferences/volume.
pdf.
Iatridou, Sabine, Elena Anagnostopolou & Roumyana Izvorski. 2001. Ob-
servations about the form and meaning of the perfect: A life in lan-
guage. In Michael Kenstowicz (ed.), Ken Hale: A life in language. MIT
Press.
Iatridou, Sabine & Ivy Sichel. 2011. Negative DPs, A-movement, and
scope diminishment. Linguistic Inquiry 42(4). 595–629. https : / / doi .
org/10.1162/LING_a_00062.
 BI BLIOGRAPHY 187

Ichikawa, Jonathan Jenkins & Matthias Steup. 2018. The analysis of knowl-
edge. In Edward N. Zalta (ed.), The stanford encyclopedia of philosophy,
Summer 2018. Metaphysics Research Lab, Stanford University. https:
//plato.stanford.edu/archives/sum2018/entries/knowledge-analysis/.
Jäger, Gerhard. 1996. Only updates: On the dynamics of the focus par-
ticle only. Amsterdam Colloquium 10. https : / / semanticsarchive . net /
Archive/DgwMDZjZ/gjAC95.pdf.
Kadmon, Nirit. 1987. On unique and non-unique reference and asymmetric
quantification. University of Massachusetts at Amherst PhD thesis.
Kadmon, Nirit & Fred Landman. 1993. ANY. Linguistics and Philosophy
16(4). 353–422. https://doi.org/10.1007/BF00985272.
Kamp, Hans. 1973. Free choice permission. Proceedings of the Aristotelian
Society, New Series 74. 57–74. http://www.jstor.org/stable/4544849.
Karttunen, Lauri. 1977. Syntax and semantics of questions. Linguistics and
Philosophy 1(1). 3–44. https://doi.org/10.1007/BF00351935.
Kaufmann, Magdalena. 2012. Interpreting imperatives (Studies in Linguistics
and Philosophy (SLAP) 88). Dordrecht: Springer. https://doi.org/10.
1007/978-94-007-2269-9.
Kaufmann, Stefan. 2017. The Limit Assumption. Semantics and Pragmatics
10(18). https://doi.org/10.3765/sp.10.18.
Kaufmann, Stefan & Magdalena Schwager. 2009. A unified analysis of
conditional imperatives. Semantics and Linguistic Theory (SALT) 19.
239–256. https://doi.org/10.3765/salt.v19i0.2545.
Keshet, Ezra. 2010. Situation economy. Natural Language Semantics 18(4).
385–434. https://doi.org/10.1007/s11050-010-9059-1.
Keshet, Ezra. 2011. Split intensionality: A new scope theory of de re and
de dicto. Linguistics and Philosophy 33(4). 251–283. https://doi.org/10.
1007/s10988-011-9081-x.
Keshet, Ezra & Florian Schwarz. 2014. De re / de dicto. http://florianschwarz.
net/wp-content/uploads/papers/De_Re___De_Dicto.pdf.
Khoo, Justin. 2011. Operators or restrictors? A reply to Gillies. Semantics
and Pragmatics 4(4). 1–25. https://doi.org/10.3765/sp.4.4.
Klein, Wolfgang. 1994. Time in language.
Klinedinst, Nathan. 2011. Quantified conditionals and conditional ex-
cluded middle. Journal of Semantics 28(1). 149–170. https://doi.org/10.
1093/jos/ffq015.
Knuuttila, Simo. 2003. Medieval theories of modality. In Edward N. Zalta
(ed.), The stanford encyclopedia of philosophy. http://plato.stanford.edu/
archives/fall2003/entries/modality-medieval/.
Kratzer, Angelika. 1977. What must and can must and can mean. Linguis-
tics and Philosophy 1(3). 337–355. https://doi.org/10.1007/BF00353453.
Kratzer, Angelika. 1978. Semantik der Rede: Kontexttheorie – Modalwörter –
Konditionalsätze. Königstein/Taunus: Scriptor.
188 B I B L I O G R A P H Y

Kratzer, Angelika. 1981. The notional category of modality. In Hans-

Jürgen Eikmeyer & Hannes Rieser (eds.), Words, worlds, and contexts:
New approaches in word semantics (Research in Text Theory 6), 38–74.
Berlin: de Gruyter.
Kratzer, Angelika. 1986. Conditionals. Chicago Linguistic Society (CLS)
22(2). 1–15. https : / / semanticsarchive . net / Archive / ThkMjYxN /
Conditionals.pdf.
Kratzer, Angelika. 1991. Modality. In Arnim von Stechow & Dieter
Wunderlich (eds.), Semantics: An international handbook of contempo-
rary research, 639–650. Berlin: de Gruyter. https://udrive.oit.umass.edu/
kratzer/kratzer-modality.pdf.
Kratzer, Angelika. 1998. More structural analogies between pronouns
and tenses. Semantics and Linguistic Theory (SALT) 8. 92–110. https :
//doi.org/10.3765/salt.v8i0.2808.
Kratzer, Angelika. 2012. Modals and conditionals: New and revised perspec-
tives. Oxford University Press. https : / / doi . org / 10 . 1093 / acprof :
oso/9780199234684.001.0001.
Kratzer, Angelika. 2015. Chasing hook: Quantified indicative condi-
tionals. To appear in a festschrift for Dorothy Edgington. http : / /
semanticsarchive . net / Archive / WU3ODIwZ / kratzer - chasing -
hook.pdf.
Kripke, Saul. 1980. Naming and necessity. Oxford: Blackwell.
Lahiri, Utpal. 2002. Questions and answers in embedded contexts (Oxford
Studies in Theoretical Linguistics 2). Oxford University Press.
Landman, Fred. 1991. Structures for semantics (Studies in Linguistics and
Philosophy 45). Springer. https://doi.org/10.1007/978-94-011-3212-
1.
Landman, Fred. 1992. The progressive. Natural Language Semantics 1(1).
1–32. https://doi.org/10.1007/BF02342615.
Larson, Richard. 2002. The grammar of intensionality. In Gerhard Preyer
(ed.), On logical form. Oxford University Press. https://semanticsarchive.
net/Archive/WVhMGRhN/larson.intens.pdf.
Lasersohn, Peter. 1996. Adnominal conditionals. Semantics and Linguistic
Theory (SALT) 6. 154–166. https://doi.org/10.3765/salt.v6i0.2759.
Lassiter, Daniel. 2016. Must, knowledge, (in)directness. Natural Language
Semantics 24. 117–163. https://doi.org/10.1007/s11050-016-9121-8.
Lassiter, Daniel. 2017. Graded modality: Qualitative and quantitative per-
spectives (Oxford Scholarship Online). Oxford University Press. https:
//doi.org/10.1093/oso/9780198701347.001.0001.
Lauer, Sven & Prerna Nadathur. 2016. Quantified indicative conditionals
and the relative reading of most. Sinn und Bedeutung 21. https://sites.
google . com / site / sinnundbedeutung21 / proceedings - preprints /
LauerNadathur-most.pdf.
 BI BLIOGRAPHY 189

Leahy, Brian P. & Susan E. Carey. 2020. The acquisition of modal con-
cepts. Trends in Cognitive Sciences 24(1). 65–78. https://doi.org/10.1016/
j.tics.2019.11.004.
Lechner, Winfried. 2007. Interpretive effects of head movement. http :
//ling.auf.net/lingBuzz/000178.
Leslie, Sarah-Jane. 2009. If , unless, and quantification. In Robert J. Stain-
ton & Christopher Viger (eds.), Compositionality, context and semantic
values: Essays in honour of Ernie Lepore, 3–30. Springer. https://doi.org/
10.1007/978-1-4020-8310-5_1.
Lewis, Clarence Irving & Cooper Harold Langford. 1932. Symbolic logic.
New York: Century.
Lewis, David. 1970. General semantics. Synthese 22(1-2). 18–67. https :
//doi.org/10.1007/BF00413598.
Lewis, David. 1973. Counterfactuals. Oxford: Blackwell.
Lewis, David. 1974. Semantic analyses for dyadic deontic logic. In Sören
Stenlund (ed.), Logical theory and semantic analysis: Essays dedicated to
Stig Kanger on his fiftieth birthday, 1–14. Dordrecht: Reidel.
Lewis, David. 1975. Adverbs of quantification. In Edward Keenan (ed.),
Formal semantics of natural language, 3–15. Cambridge University Press.
Lewis, David. 1978. Truth in fiction. American Philosophical Quarterly
15(1). 37–46. http://www.jstor.org/stable/20009693.
Lewis, David. 1982. Logic for equivocators. Noûs 16(3). 431–441. https:
//doi.org/10.2307/2216219.
Lewis, David. 1986. On the plurality of worlds. Oxford: Blackwell.
Ludlow, Peter & Stephen Neale. 1991. Indefinite descriptions: In defense
of Russell. Linguistics and Philosophy 14(2). 171–202. https://doi.org/10.
1007/bf00627402.
MacFarlane, John. 2005. Logical constants. In Edward N. Zalta (ed.), The
stanford encyclopedia of philosophy. http://plato.stanford.edu/archives/
win2005/entries/logical-constants/.
MacFarlane, John. 2006. Epistemic modals are assessment-sensitive. ms,
University of California, Berkeley, forthcoming in an OUP volume
on epistemic modals, edited by Brian Weatherson and Andy Egan.
http://sophos.berkeley.edu/macfarlane/epistmod.pdf.
Maier, Emar & Sofia Bimpikou. 2019. Shifting perspectives in pictorial
narratives. Proceedings of Sinn und Bedeutung 23(2). 91–106. https://doi.
org/10.18148/sub/2019.v23i2.600.
May, Robert. 1977. The grammar of quantification. Massachusetts Institute
of Technology PhD thesis. https://doi.org/1721.1/16287.
McCloskey, Jim. 2016. Interpretation and the typology of head move-
ment: A re-assessment. Handout for a presentation at the Workshop
on the Status of Head Movement in Linguistic Theory, Stanford Uni-
versity, September 16th and 17th 2016. https://drive.google.com/file/
d/0BzSn1AcNdRX2VmpWZHJwRXRaZEk/view.
190 B I B L I O G R A P H Y

McCready, Eric & Norry Ogata. 2007. Evidentiality, modality and prob-
ability. Linguistics and Philosophy 30(2). 147–206. https://doi.org/10.
1007/s10988-007-9017-7.
Menzel, Christopher. 2016. Possible worlds. In Edward N. Zalta (ed.),
The Stanford encyclopedia of philosophy, Spring 2016. http : / / plato .
stanford.edu/entries/possible-worlds/.
Meyer, Marie-Christine. 2013. Ignorance and grammar. Massachusetts Insti-
tute of Technology PhD thesis. https://doi.org/1721.1/84420.
Montague, Richard. 1973. The proper treatment of quantification in or-
dinary English. In Jaako Hintikka, Julius Moravcsik & Patrick Suppes
(eds.), Approaches to natural language, 221–242. Dordrecht: Reidel. http:
//www.blackwellpublishing.com/content/BPL_Images/Content_
store/Sample_chapter/9780631215417/Portner.pdf.
Nauze, Fabrice. 2008. Modality in typological perspective. Universiteit van
Amsterdam PhD thesis. http : / / www . illc . uva . nl / Publications /
Dissertations/DS-2008-08.text.pdf.
Nute, Donald. 1984. Conditional logic. In Dov Gabbay & Franz Guen-
thner (eds.), Handbook of philosophical logic. volume ii, 387–439. Dor-
drecht: Reidel.
Ogihara, Toshiyuki. 1989. Temporal reference in English and Japanese. Austin,
TX: University of Texas PhD Thesis.
Ogihara, Toshiyuki. 1995. The semantics of tense in embedded clauses.
Linguistic Inquiry 26(4). 663–679. http://www.jstor.org/stable/4178918.
Ogihara, Toshiyuki. 1996. Tense, attitudes, and scope. Springer. https://doi.
org/10.1007/978-94-015-8609-2.
Papafragou, Anna. 2006. Epistemic modality and truth conditions. Lingua
116(10). 1688–1702. https://doi.org/10.1016/j.lingua.2005.05.009.
Partee, Barbara H. 1973. Some structural analogies between tenses and
pronouns in English. The Journal of Philosophy 70(18). 601–609. https:
//doi.org/10.2307/2025024.
Partee, Barbara H. & Herman L.W. Hendriks. 1997. Montague grammar.
In Johan van Benthem & Alice ter Meulen (eds.), Handbook of logic and
language, 5–91. Elsevier.
Percus, Orin. 2000. Constraints on some other variables in syntax. Nat-
ural Language Semantics 8(3). 173–229. https : / / doi . org / 10 . 1023 / A :
1011298526791.
Portner, Paul. 1997. The semantics of mood, complementation, and con-
versational force. Natural Language Semantics 5(2). 167–212. https://doi.
org/10.1023/A:1008280630142.
Portner, Paul. 1998. The progressive in modal semantics. Language 74(4).
760–787. https://doi.org/10.2307/417002.
Portner, Paul. 2007. Imperatives and modals. Natural Language Semantics
15(4). 351–383. https://doi.org/10.1007/s11050-007-9022-y.
Portner, Paul. 2009. Modality. Oxford University Press.
 BI BLIOGRAPHY 191

Portner, Paul. 2017. Mood. To appear as a OUP book.

Prakken, Henry & Marek Sergot. 1996. Contrary-to-duty obligations.
Studia Logica 57(1). 91–115. https://doi.org/10.1007/BF00370671.
Prior, Arthur N. 1958. Escapism: The logical basis of ethics. In A. I. Melden
(ed.), Essays in moral philosophy, 135–146. Seattle: University of Wash-
ington Press.
Proudfoot, Diane. 2006. Possible worlds semantics and fiction. Journal of
Philosophical Logic 35(1). 9–40. https://doi.org/10.1007/s10992-005-
9005-8.
Quine, Willard Van Orman. 1960. Word and object. Cambridge, MA: MIT
Press.
Redshaw, Jonathan. 2014. Does metarepresentation make human mental
time travel unique? Wiley Interdisciplinary Reviews: Cognitive Science
5(5). 519–531. https://doi.org/10.1002/wcs.1308.
Redshaw, Jonathan & Thomas Suddendorf. 2020. Temporal junctures in
the mind. Trends in Cognitive Sciences 24(1). 52–64. https://doi.org/10.
1016/j.tics.2019.10.009.
Reich, Ingo. 2009. What asymmetric coordination in German tells us
about the syntax and semantics of conditionals. Natural Language Se-
mantics 17(3). 219–244. https://doi.org/10.1007/s11050-009-9041-y.
Reichenbach, Hans. 1947. Elements of symbolic logic.
Rett, Jessica. 2019. A typology of semantic entities. To appear in Daniel
Altshuler (ed.) Linguistics meets Philosophy, Cambridge. https://linguistics.
ucla . edu / people / Rett / Rett % 202019 % 20typology % 20semantic %
20entities.pdf.
Richards, Norvin. 1998. The principle of minimal compliance. Linguistic
Inquiry 29(4). 599–629. https://doi.org/10.1162/002438998553897.
Roberts, Craige. 2012. Information structure: Towards an integrated for-
mal theory of pragmatics. Semantics and Pragmatics 5(6). Reissue of a
1998 amendation of the original 1996 manuscript, 1–69. https://doi.
org/10.3765/sp.5.6.
Romoli, Jacopo & Yasutada Sudo. 2009. De re/de dicto ambiguity and pre-
supposition projection. Proceedings of Sinn und Bedeutung 13. 426–438.
http://semanticsarchive.net/Archive/DhhOTI2Z/sub13proc.pdf#
page=447.
van Rooij, Robert. 2006. Free choice counterfactual donkeys. Journal of
Semantics 23(4). 383–402. https://doi.org/10.1093/jos/ffl004.
Ross, Jeff. 1997. The semantics of media (Studies in Linguistics and Philoso-
phy (SLAP) 64). Dordrecht: Kluwer.
Rothschild, Daniel. 2011. A note on conditionals and restrictors. Paper
written for a conference in honor of Dorothy Edgington at the Insti-
tute of Philosophy, London in May 2011. http://semanticsarchive.net/
Archive/WY1M2MxM/conditionals.pdf.
192 B I B L I O G R A P H Y

Rubinstein, Aynat. 2012. Roots of modality. University of Massachusetts at

Amherst PhD thesis. http://semanticsarchive.net/Archive/zg2ZWFhN/
Rubinstein2012Diss.pdf.
Rubinstein, Aynat. 2014. On necessity and comparison. Pacific Philosophi-
cal Quarterly 95(4). 512–554. https://doi.org/10.1111/papq.12047.
Rubinstein, Aynat. 2017. Straddling the line between attitude verbs and
necessity modals. In Ana Arregui, María Luisa Rivero & Andrés Salanova
(eds.), Modality across syntactic categories, 109–131. Oxford University
Press. https://doi.org/10.1093/acprof:oso/9780198718208.003.0007.
Rullmann, Hotze & Sigrid Beck. 1998. Presupposition projection and
the interpretation of which-questions. Semantics and Linguistic Theory
(SALT) 8. 215–232. https://doi.org/10.3765/salt.v8i0.2811.
Russell, Bertrand. 1940. An inquiry into meaning and truth. London: George
Allen & Unwin.
Sauerland, Uli & Paul Elbourne. 2002. Total reconstruction, PF move-
ment, and derivational order. Linguistic Inquiry 33(2). 283–319. https:
//doi.org/10.1162/002438902317406722.
Schlenker, Philippe. 2004. Conditionals as definite descriptions (A refer-
ential analysis). Research on Language and Computation 2(3). 417–462.
https://doi.org/10.1007/s11168-004-0908-2.
Schulz, Katrin. 2005. A pragmatic solution for the paradox of free choice
permission. Synthese 147(2). 343–377. https://doi.org/10.1007/s11229-
005-1353-y.
Schwager, Magdalena. 2009. Speaking of qualities. Semantics and Linguistic
Theory (SALT) 19. 395–412. https://doi.org/10.3765/salt.v19i0.2534.
Schwarz, Florian. 2012. Situation pronouns in determiner phrases. Natural
Language Semantics 20(4). 431–475. https://doi.org/10.1007/s11050-
012-9086-1.
Sen, Amartya K. 1970. Collective choice and social welfare. San Francisco,
CA: Holden-Day.
Sharvit, Yael. 2002. Emedded questions and de dicto readings. Natural
Language Semantics 10(2). 97–123. https : / / doi . org / 10 . 1023 / a :
1016573907314.
Sharvit, Yael & Elena Guerzoni. 2003. Reconstruction and its problems.
Amsterdam Colloquium 14. 205–210.
Speas, Peggy. 2008. On the syntax and semantics of evidentials. Language
and Linguistics Compass 2(5). 940–965. https://doi.org/10.1111/j.1749-
818X.2008.00069.x.
Spencer, Mary. 1971. Why the “s” in “intension”? Mind 80(317). 114–115.
https://doi.org/10.1093/mind/LXXX.317.114.
Stalnaker, Robert. 1968. A theory of conditionals. In Nicholas Rescher
(ed.), Studies in logical theory (American Philosophical Quarterly Mono-
graph Series 2), 98–112. Oxford: Blackwell.
 BI BLIOGRAPHY 193

Stalnaker, Robert. 1978. Assertion. In Peter Cole (ed.), Syntax and seman-
tics, vol. 9, 315–332. New York: Academic Press.
Stalnaker, Robert. 1984. Inquiry. MIT Press.
Stalnaker, Robert. 1999. Context and content. Oxford: Oxford University
Press.
Stanley, Jason & Zoltán Gendler Szabó. 2000. On quantifier domain re-
striction. Mind and Language 15(2/3). 219–261. https : / / doi . org / 10 .
1111/1468-0017.00130.
Starr, William. 2019. Counterfactuals. In Edward N. Zalta (ed.), The
Stanford encyclopedia of philosophy, Fall 2019. Metaphysics Research
Lab, Stanford University. https://plato.stanford.edu/archives/fall2019/
entries/counterfactuals/.
Starr, William. 2020. A preference semantics for imperatives. Semantics
and Pragmatics 13(6). https://doi.org/10.3765/sp.13.6.
von Stechow, Arnim. 1991. Intensionale Semantik: Eingeführt anhand
der Temporalität. http : / / www . sfs . uni - tuebingen . de / ~astechow /
Aufsaetze/Int.SemTemp91.pdf.
von Stechow, Arnim. 1996. Against LF pied-piping. Natural Language
Semantics 4(1). 57–110. https://doi.org/10.1007/BF00263537.
von Stechow, Arnim & Sigrid Beck. 2015. Events, times and worlds: An
LF architecture. In Christian Fortmann, Anja Lübbe & Irene Rapp
(eds.), Situationsargumente im Nominalbereich (Linguistische Arbeiten
562), 13–46. De Gruyter. https://doi.org/10.1515/9783110432893-
002.
Steiner, George. 1975. After Babel: Aspects of language and translation. Ox-
ford University Press.
Stephenson, Tamina. 2007a. Judge dependence, epistemic modals, and
predicates of personal taste. Linguistics and Philosophy 30(4). 487–525.
https://doi.org/10.1007/s10988-008-9023-4.
Stephenson, Tamina. 2007b. Towards a theory of subjective meaning. Mas-
sachusetts Institute of Technology PhD thesis. http://semanticsarchive.
net/Archive/2QxMjk0O/Stephenson-2007-thesis.pdf.
Strawson, P. F. 1950. On referring. Mind 59(235). 320–344. https://doi.
org/10.1093/mind/lix.235.320.
Swanson, Eric. 2008. Modality in language. Philosophy Compass 3(6).
1193–1207. https://doi.org/10.1111/j.1747-9991.2008.00177.x.
Swanson, Eric. 2010. On scope relations between quantifiers and epis-
temic modals. Journal of Semantics 27(4). 529–540. https://doi.org/10.
1093/jos/ffq010.
Swanson, Eric. 2011. Propositional attitudes. In Claudia Maienborn, Klaus
von Heusinger & Paul Portner (eds.), Semantics: An international hand-
book of meaning, vol. 2, 1538–1561. de Gruyter. https://doi.org/10.
1515/9783110255072.1538.
194 B I B L I O G R A P H Y

Szabó, Zoltán Gendler. 2010. Specific, yet opaque. In Maria Aloni, Harald
Bastiaanse, Tikitu de Jager & Katrin Schulz (eds.), Logic, language and
meaning: 17th Amsterdam Colloquium, Amsterdam, The Netherlands, De-
cember 16-18, 2009, revised selected papers (Lecture Notes in Computer
Science 6042), 32–41. Springer. https://doi.org/10.1007/978-3-642-
14287-1_4. https://campuspress.yale.edu/zoltanszabo/files/2015/10/
Specific-yet-Opaque-uhcqrh.pdf.
Tancredi, Chris. 2008. Multiple models. Unpublished report for The
Center for Advanced Research on Logic and Sensibility.
Taylor, Barry. 1977. Tense and continuity. Linguistics and Philosophy 1.
199–220. https://doi.org/10.1007/BF00351103.
Teller, Paul. 1972. Epistemic possibility. Philosophia 2(4). 302–320. https:
//doi.org/10.1007/BF02381591.
Thomas, Guillaume. 2014. Nominal tense and temporal implicatures: Ev-
idence from Mbyá. Natural Language Semantics 22(4). 357–412. https:
//doi.org/10.1007/s11050-014-9108-2.
Varzi, Achille. 1997. Inconsistency without contradiction. Notre Dame
Journal of Formal Logic 38(4). 621–638. https://doi.org/10.1305/ndjfl/
1039540773.
Viebahn, Emanuel & Barbara Vetter. 2016. How many meanings for
may?: The case for modal polysemy. Philosophers’ Imprint 16(10). 1–
26. https://doi.org/2027/spo.3521354.0016.010.
White, Aaron Steven. 2021. On believing and hoping whether. Accepted
with minor revision for publication in Semantics and Pragmatics. https:
//ling.auf.net/lingbuzz/005665/current.pdf.
Willett, Thomas. 1988. A cross-linguistic survey of the grammaticaliza-
tion of evidentiality. Studies in Language 12(1). 51–97.
Williamson, Timothy. 2000. Knowledge and its limits. Oxford: Oxford
University Press.
Wurmbrand, Susi. 1999. Modal verbs must be raising verbs. West Coast
Conference on Formal Linguistics (WCCFL) 18. 599–612. http://citeseerx.
ist.psu.edu/viewdoc/download?doi=10.1.1.35.7442&rep=rep1&type=
pdf.
Yalcin, Seth. 2010. Probability operators. Philosophy Compass 5(11). 916–
937. https://doi.org/10.1111/j.1747-9991.2010.00360.x.
Yalcin, Seth. 2015. Epistemic modality de re. Ergo: An open access journal
of philosophy 2(19). https://doi.org/10.3998/ergo.12405314.0002.019.
Zach, Richard. 2019. Boxes and diamonds: An open introduction to modal
logic. https://bd.openlogicproject.org/bd-screen.pdf.
Zimmermann, Thomas Ede. 2000. Free choice disjunction and epistemic
possibility. Natural Language Semantics 8(4). 255–290. https://doi.org/
10.1023/A:1011255819284.
 BI BLIOGRAPHY 195

Zimmermann, Thomas Ede & Wolfgang Sternefeld. 2013. Introduction to

semantics: An essential guide to the composition of meaning. de Gruyter
Mouton.
Zucchi, Sandro. 2001. Tense in fiction: Reference, anaphora and aspect.
In Carlo Cecchetto, Gennaro Chierchia & Maria Teresa Guasti (eds.),
Semantic interfaces: Reference, anaphora and aspect, 320–355. CSLI Publi-
cations. http://tinyurl.com/5ulwxwg.
Zucchi, Sandro. 2005. The present mode. In Gregory N. Carlson & Fran-
cis Jeffry Pelletier (eds.), Reference and quantification: The Partee effect.
CSLI.
Zvolenszky, Zsófia. 2002. Is a possible-worlds semantics of modality pos-
sible? A problem for Kratzer’s semantics. Semantics and Linguistic The-
ory (SALT) 12. 339–358. https://doi.org/10.3765/salt.v12i0.2866.