COMPLEMENT COERCION AND THE SEMANTICS OF ASPECTUAL VERBS
by
Christopher Harris O’Brien

A THESIS
Submitted
to Michigan State University
in partial fulfillment of the requirements
for the degree of
MASTER OF ARTS
LINGUISTICS
2012

ABSTRACT
COMPLEMENT COERCION AND THE SEMANTICS OF ASPECTUAL VERBS
by
Christopher Harris O’Brien
This thesis investigates the semantics of complement coercion (CC). The most prominent account of this phenomenon is due to Pustejovsky (1991, 1993), who essentially
argues that it can be explained via a type-shift in nominal complements. However,
this analysis suffers from two major empirical drawbacks. For one, it incorrectly
predicts the ungrammaticality of certain sorts of coordinate constructions. It also
overgeneralizes, predicting that the full range of aspectual verbs (AVs) in English
should allow CC. It turns out, however, that only a subset of these verbs do. This
thesis develops a theory of complement coercion which can account for the coordinate
structure data and for the pattern of CC availability in AVs.
This thesis advances a type-based analysis of CC in AVs. Specifically, it is proposed that AVs which allow CC differ in their type from those that do not. This
analysis also captures the insight that the coercing verb, rather than the complement, is responsible for CC via an operator which shifts intransitive verbs that modify properties of times (such as AVs) into transitive verbs. Due to the proposed type
difference between the two AV subclasses, it turns out that only certain AVs will
allow CC. What emerges is a theory of CC which can account for data that eluded
previous analyses and also makes novel predictions about the relationship between
CC and other facets of tense and aspect in English.

Copyright by
Christopher Harris O’Brien
2012

ACKNOWLEDGEMENTS

Most of this thesis emerged out of conversations with Marcin Morzycki during the
2011-2012 school year. Without all of Marcin’s encouragement, not only would I
never have finished this thesis, but I probably never wouldn’t have even started it.
Not only that, but his penetrating criticisms and suggestions were instrumental in
molding this thesis into something worthwhile.
I’d also like to thank the other members of my committee, Alan Munn and
Cristina Schmitt, whose insightful comments and criticisms were ultimately highly
beneficial. And I’d like to thank Curt Anderson, Greg Johnson, Gabriel Rogrigues,
and everyone else who was at the sessions of semantics awkward time where I presented the material contained in this thesis. The input from these meetings was also
very helpful. And Julia Kartush, who besides being just about the best co-TA one
could ask for, also provided many of the judgements that comprise the data for this
thesis.
Finally, I’d like to thank my parents, Cyndy and Tim, and sister, Elizabeth,
whose love and support helped me finish this thesis, along with everything else. And
my roommates- Carrie, Emily, and Evan- whose company in my off-hours this past
year has kept me happy and sane.

iv

TABLE OF CONTENTS

LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

vi

CHAPTER 1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . .

1

CHAPTER 2 SOME PREVIOUS THEORIES OF COMPLEMENT
COERCION . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1 A Simple Syntactic Analysis . . . . . . . . . . . . . . . . . . . . . . .
3
2.2 The NP Type-Shifting Analysis . . . . . . . . . . . . . . . . . . . . .
5
2.3 Problems With The NP Type-Shifting Analysis . . . . . . . . . . . .
8
2.4 Other Possibilities: Underspecification Of The AV . . . . . . . . . . . 12
CHAPTER 3 PRELIMINARIES FOR A
3.1 What Is To Be Discussed . . . . . . . .
3.2 The Syntax And Semantics Of AVs . .
3.3 Some Evidence For This Framework . .

NEW
. . . .
. . . .
. . . .

ANALYSIS
. . . . . . . .
. . . . . . . .
. . . . . . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

19
19
27
30

CHAPTER 4 A TYPE-BASED ANALYSIS OF ASPECTUAL VERBS
AND COMPLEMENT COERCION EXPLAINED . . 41
4.1 What Is To Be Explained And How . . . . . . . . . . . . . . . . . . . 41
4.2 The Semantics Of Aspectual Verbs . . . . . . . . . . . . . . . . . . . 43
4.3 Complement Coercion Explained . . . . . . . . . . . . . . . . . . . . 49
4.4 Some Consequences Of The Type-Based Analysis . . . . . . . . . . . 52
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
CHAPTER 5 WHAT’S NEXT? . . . . . . . . . . . . . .
5.1 Pred And The Interpretation Of Coerced Predicates . .
5.2 Complement Coercion And Enjoy-Type Verbs . . . . .
5.3 The Syntax Of AVs Revisited . . . . . . . . . . . . . .
5.4 Towards A Sortal Analysis Of AVs . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

58
58
59
61
68

BIBLIOGRAPHY . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

v

LIST OF FIGURES

4.1

Derivation for John stopped eating the cookie. . . . . . . . . . . .

45

4.2

Derivation for John finished eating the cookie. . . . . . . . . . . .

48

4.3

Derivation for Finish the cookies. . . . . . . . . . . . . . . . . . . .

51

vi

Chapter 1
INTRODUCTION

English aspectual verbs (AVs), like begin and finish, have been discussed intermittently in both the semantic and syntactic literature (Perlmutter, 1970; Newmeyer,
1975; Portner, 1992; Fukuda, 2008). Even less frequently discussed is a peculiar
alternation which some of these verbs allow:
(1)

a. John began reading the book.
b. John began the book.

(2)

a. The author started writing the book.
b. John started the book.

(3)

a. Phil finished mowing the lawn.
b. Phil finished the lawn.

While aspectual verbs typically take some sort of participle or infinitival1 verb as
their complement (1-3 a), they may also take a nominal complement. These latter cases are remarkable for two reasons. First, the AV can take complements of
different semantic types, even though the AV seems to be performing an identical
semantic function in both cases. Secondly, the grammar can apparently reconstruct
a fully specified interpretation from the combination of the aspectual verb and the
1 These infinitival complements will not be discussed here, but see Portner (1992, chap-

ters 4 and 5) for an extensive discussion.

1

amputated complement. This process is typically called complement coercion.2 It
will henceforth be abbreviated CC.3
The goal of the present work is to develop a new analysis of complement coercion
in AVs. First, in 2, we will review previous analyses of CC, focusing on the NP
type-shifting analysis (2.2-2.3) and an underspecification analysis (2.4) of CC. We
will discover that these analyses run into various problems with certain coordinate
structures, and also fail to account for why CC is possible with certain AVs but not
with others. In 3 we will clarify the data that the new analysis will account for.
In particular, it will also be observed that the availability of CC in AVs correlates
with the availability of imperfective morphology in these AVs. We will also lay out
the syntactic and semantic assumptions underlying this analysis. In 4, the analysis
will be developed. To prefigure this section a little bit, this analysis will construct
a semantics for AVs such that certain AVs have the appropriate type to take an
operator which executes CC. This analysis will also make novel predictions about
the imperfective morphology data introduced in 3. Finally, 5 will lay out some
directions that future research may take.

2 Many researchers, e.g. Pustejovsky (1993), also claim that such verbs as enjoy allow

complement coercion. This class of verbs will not be discussed in the main body of this
thesis, but see 5.2 for some discussion on how the analysis developed here may be extended
to these verbs.
3 Complement Coercion has also recently been the subject of a great deal of experimental
work, which this thesis will not have room to explore. But See Pylkk¨nen and McElree
a
(2006, 2007) and Husband et al. (2010) for some discussion.

2

Chapter 2
SOME PREVIOUS THEORIES OF COMPLEMENT COERCION

2.1

A Simple Syntactic Analysis

The simplest explanation for CC would be to posit a null syntactic head, some sort
of abstract verb, as the complement to the coercing aspectual verbs in (1-3):
(4) John began [VERB the book]
This would allow a semantically uniform treatment of the coercing and non-coercing
aspectual verbs (aside from the detail of filling in the interpretation for the former). This analysis was suggested by Newmeyer (1975), and briefly considered by
Pylkk¨nen and McElree (2006), who reject it on the following grounds. They point
a
out that, were this analysis correct, the null verb should be able to take adverbial
and prepositional modifiers, as the overt verbal complement does in cases like (5)
and (6):
(5)

a. Mary [finished [eating the meal slowly]]
b. Mary [finished [eating the meal] slowly]

(6)

a. Mary [finished [baking the bread in the oven]]
b. Mary [finished [baking the bread] in the oven]

In (5a), slowly modifies the lower VP eating the meal, indicating that Mary finished
an event of eating a meal slowly (and perhaps continued on, now eating the meal

3

quickly). In the second reading (5b), it modifies finish, where it means that the
event of finishing eating the entire meal was slow. Similarly in (6a) the preposition
in the oven modifies the lower VP, where it means that Mary completed the event
of baking the bread in the oven (and then perhaps continued baking the bread, now
in some other contraption). In (6b), on the other hand, in the oven modifies finish,
where it means that the event of finishing baking the bread occurred in the oven.
Now consider (7) and (8):
(7)

a. Mary [finished the meal slowly]
b. *Mary [finished [VERB the meal slowly]]

(8)

a. Mary [finished the bread in the oven]
b. Mary [finished [VERB the bread in the oven]]

In these coerced constructions, only the readings (7a and 8a) where the adjunct
modifies finish is available. If these sentences actually had a structure like (4),
then readings based on structures like (7b) and (8b), where slowly and in the oven
respectively modify the abstract predicate, should be possible. Their impossibility
suggests that a null verb along the lines of the structure in (4) is not present in
coerced sentences.
Pylkk¨nen and McElree (2006) also note that, while passivization of the nominal
a
is impossible with a full verbal complement (9a), it is licit in cases of CC (9b):
(9)

a. *The book was finished reading by John
b. The book was finished by John.

4

Were these two constructions syntactically identical (as (4) predicts), then such an
asymmetry would likely not show up.
Next, we shall consider an NP type-shifting analysis proposed by Pustejovsky
(1991, 1993, 1995) which offers a fully compositional semantic solution to the problem
of CC.

2.2

The NP Type-Shifting Analysis

Pustejovsky (1993)1 argues that complement coercion can be executed through a
type-shifting operation within the nominal complement. Specifically, he argues that
there exists a set of type-shifting functions which enable functional application (FA)
in certain instances where FA would otherwise be predicted to fail. Suppose we have
an expression α, which is of type b,a and another expression β. α may apply to β
if either if β is of type b ; or if among the set Σβ of type shifting operations which
apply to β, there is an operation σ that, when applied to β, yields an expression of
type b . If such an operation σ exists α will correctly yield an expression of type
a when applied to the output of σ(β). This can be formalized as in (10) (from
Pustejovsky (1993, 83-84)):

1 I shall focus on Pustejovsky (1993) since it contains the most explicit presentation of

his theory of CC.

5

(10) Functional Application With Coercion (FAC ):
If α is of type b,a and β is of type c , then:
i if c = b, then α(β) is of type a.
ii if there is a σ ∈ Σβ such that σ(β) yields an expression of type b , then
α(σ(β)) is of type a
iii otherwise, type clash.
In other words, certain expressions β have a set of type-shifting operations Σβ associated with them. Coercion can be executed only if the appropriate type-shifting
operation exists for a given expression.
Lets assume that AVs, such as start, have a type of VP, NP, S 2 . In other
words, it takes a VP complement, and return an intransitive verbal predicate. Now,
for coercion to be possible with a nominal in the complement position of the AV,
there must exist an operation ρ in ΣNP (the set of type-shifting operations which
apply to nominals) which allows the nominal to shift to a type VP .
These type-shifting functions have to be limited; otherwise type-shifts could be
posited that would allow anything to take anything as an argument. The set of
type-shifting operations available to the NP complement is restricted by the nominals qualia structure (Pustejovsky, 1991, 1993, 1995), which consists of complex
structured representations inside a lexical item. These representations then link that
item to various other lexical items. In short, this is a way to partially decompose a
nominal’s meaning into various sub-constituents.
2 In using these non-standard type, I am simply following Pustejovsky (1993). In the

analysis presented in 3 and 4, I shall revert to a more standard type-system.

6

This type-shifting operation ρ may only apply to an expression P if there exists
some appropriate qualia role Q such that Q is a two place-predicate with P as its
internal argument. This can be formalized as in (11):
(11) ∀P[ρ(P) → ∃Q[Q(P)(x)]]
Among the qualia roles Pustejovsky proposes are the agentive role QA (how an object
came into existence) and telic role QT (the function or use of an object). The typeshifting operator ρ forces the nominal to be interpreted as either its telic or agentive
roles:
(12) ρ = [QT (NP), QA (NP)]
These two operations QT and QA essentially shift the interpretation of the nominal
from its basic interpretation (e.g. x is a book) to a more complex function associated
with how the nominal comes into existence (QA ), or how it is used (QT ); following
the meaning postulate in (11), the nominal then becomes the internal argument for
this more complex function (Pustejovsky, 1993, pg. 87):
(13)

a. QT (book) = λxλy[read(x)(y)](book)
b. QT (book) = λxλy[read(x)(y)](book)

It should be easy, at this point, to see how this all adds up to allow complement
coercion. Consider a verb, such as begin, that takes a type VP complement. When
it actually takes a verbal complement as its argument, the function applies as usual.
But when it finds a nominal there instead, it must employ the type-shifting operation
ρ, as (10) dictates. So long as the appropriate qualia role is associated with the

7

nominal, then the type-shifting operation will apply, and the AV will be able to
apply to its complement. A derivation along these lines for Mary began the book may
be seen in (14), (Pustejovsky, 1993, pg. 88):
(14)

a. Mary began a book.
b. begin’(QT (a book)(x)){Mary} →
c. begin’(QT (a book)(Mary))(Mary)

What is more, the type-shifting operation, when combined with Pustejovsky’s theory
of qualia structure, correctly predicts the preferred interpretation for constructions
like (14 a); viz. Mary began either reading or writing a novel. It should be noted,
however, that the two are in principle dissociable. One who finds Pustejovsky’s
arguments for the notion of qualia structure unpersuasive may still adhere to a
theory of complement coercion that relies on some sort of nominal type-shifting,
or vice versa. However, as we shall see, there are reasons to be skeptical of the
type-shifting argument.

2.3

Problems With The NP Type-Shifting Analysis

Pustejovsky’s theory of qualia structure has been criticized on both conceptual3
and empirical grounds. On the empirical side, we find that contextual factors can
override default interpretations. Consider a scenario like the one in (15a) below.
Then consider a sentence like (15b):
3 We won’t go into these here, but see Fodor and Lepore (1998) for a very strong critique

of Pustejovsky’s general theory of qualia structure.

8

(15)

a. The goat was happily eating a tattered old paperback book.
b. Because it was so hungry, the goat finished the book quite quickly.

(15b) most likely means that the goat finished eating the book, not reading or writing
it, although its very unlikely that such a non-standard activity as eating would be
stored in the lexical entry for book. Were coercion operations strictly linked to the
nominal’s qualia structure, then it’s hard to see how such contextual sensitivity would
be possible. Lascarides and Copestake (1998) propose an analysis (to be discussed in
5.1) that allows one to combine the qualia theory of nominal structure with contextual
reinterpretation. However, the type-shifting component of Pustejovskys analysis runs
into other, more severe problems.
For one thing, various writers (Copestake and Briscoe, 1992; Godard and Jayez,
1993; Lascarides and Copestake, 1998) have noted a certain type of construction
which the NP type-shifting analysis predicts to be ungrammatical:
(16) John picked up and started the book.
The first verb in this construction, pick up, takes a nominal argument, and the
second, start, a verbal argument. If we assume that the nominal complement shifts
its type to that of a verb in order to combine with an aspectual verb like start, as
the NP type-shifting analysis predicts, then it’s difficult to see how it would also
be able to combine with a verb like pick up, which must take a NP complement.
Therefore the construction should be ungrammatical. Its well-formedness strongly
suggests that the nominal complement does not undergo a type-shift.
What’s more, this analysis seems to make the wrong prediction about other kinds
of coordinate structure. The NP type-shifting theory would predict that, because
9

the nominal complement has shifted to the same type as a VP complement, a coerced nominal complement and a VP complement should both be able to appear as
arguments of an aspectual verb in a coordinated construction. However, this is not
the case:
(17)

a. #John began the DVD and eating popcorn.
b. #John finished the book and taking notes.

Again, this suggests that the nominal complement is not responsible for complement
coercion. Rather, it seems that the type contrast is not in the nominal, but in the
aspectual verb itself.
It should be noted that a syntactic explanation for the data in (17) is conceivable,
depending upon the syntax we assume for AV complements. Suppose, for the sake
of argument, that these gerundive complements are control clauses.4 This would
mean that coercing AVs can either c-select for a nominal complement, in which case
they assign [ACC], or for a control complement. In this latter case lets follow Munn
(1993, 76) and assume that the verb assigns some sort of +control feature. Now, a
sentence like (17a) would have a structure something like (18)5 :
(18) begin [NP [N the DVD] [BP [B and] [TP PRO eating popcorn]]]
Assume, again following Munn (1993), that [ACC] and +control are privative features
in complementary distribution, so that the AV may assign either [ACC] or [control],
4 Perlmutter (1970) actually argued that AV syntax is ambiguous between being a con-

trol and a raising structure (the syntax of AVs will be discussed at somewhat greater length
in 3.2 and 5.2).
5 The phrase structure in (18) assumes the existence of a B(oolean)P (Munn, 1993).

10

but not both. Now, in order to license the first conjunct of the coordinate structures
in (18), the AV will have to assign [ACC]. This will mean, however, that no [control]
will be available to license the second conjunct. So the derivation will crash due to
the presence of an unlicensed control clause in the second conjunct.
While such a syntactic explanation for the unacceptability of these sentences
is actually quite plausible, the syntax of AVs is rather unclear, as we shall see in
5.2 below. In any case, sentences of the (16) variety remain a problem for the NP
type-shifting analysis. Moreover, this analysis also runs into a problem of overgeneralization, as we shall presently see.
The NP type-shifting analysis also predicts that this type-shifting operation will
be quite general. The type requirements of the verb motivate a type-shift in the
nominal, whose type-shifting possibilities are lexically constrained. Consider, though,
the following data:
(19)

a. John kept reading the book.
b. *John kept the book.

(20)

a. John stopped reading the book.
b. *John stopped the book.

(21)

a. John ceased reading the book.
b. *John ceased the book.

Each of the verbs in (19-21a) requires a type VP argument, and thus should license
coercion in the nominal. Whats more, its clear from sentences like (1b) that the NP
complement book in (19-21b) is capable of undergoing coercion. So, there are no

11

important differences between (19-21) and (1-3) which should render CC impossible.
The unavailability of CC in these sentences (19-21) is a puzzle, for which an NP
type-shifting analysis offers no solution.6
In addition to those in (1-3), but unlike (19-21), other aspectual verbs seem to
allow coercion:
(22)

a. John resumed reading the book.
b. John resumed the book.

(23)

a. John continued reading the book.
b. John continued the book.

We can thus divide our set of AVs into those that allow CC ((1-3), (22-23)), which
we shall call finish-type verbs, and those that do not (19-21), which we shall call
stop-type verbs. This contrast will be discussed at greater length in 3.1.

2.4

Other Possibilities: Underspecification Of The AV

The first of our objections to the NP Type-shifting analysis, that it incorrectly predicts the ungrammaticality of sentences like (16) (repeated as (24) below), has already been raised by authors such as (Copestake and Briscoe, 1992, 1995; Godard
and Jayez, 1993; Lascarides and Copestake, 1998):

6 One might be able to attribute (18a) to a blocking effect, but this solution has difficulty

explaining (19b), where the context doesn’t encourage blocking, and doesn’t explain (20)
at all.

12

(24) John picked up and started the book.
To address this problem, Copestake and Briscoe (1995) and
Lascarides and Copestake (1998) claim that coercing verbs are underspecified for the
type of their complement. Working in an HPSG-derived framework, these authors
argue that verbs which allow complement coercion, such as the AVs7 , may subcategorize for both nominal complements, which they assume to have a type NP , and
gerundive complements of some sort, which they assume to take a type of VP .8
Thanks to a highly structured lexical semantics, the denotations of these verbs may
alter depending upon which complement the verb takes.
Consider a verb like start. These authors are not explicit about the semantics
of the verb itself, but assume that it relates an event e involving some individual
x to a event e which is the starting event of e. Let’s consider the denotation when
start subcategorizes for a VP complement. Here, the verb start takes in some
relation from individuals to properties of events and relates the event to the starting
sub-event. Its denotation may look something like (25). Note that, for the sake of
consistency with the denotations to be employed in chapters 3 and 4, we will construct
the denotation using the lambda-calculus, even though these authors employ a rather
different framework:

7 These authors assume that AVs pattern with attitude verbs of the enjoy-type in al-

lowing for complement coercion. As was mentioned in fn. 2 of chapter 1, this thesis will
ignore verbs of that class. However, the possibility of extending the analysis in 4 to these
verbs will be discussed in 5.2.
8 Again, we are employing these non-standard types simply because the authors under
discussion do. A more standard type system will be used in chapters 3 and 4.

13

(25)

startVP

= λpVP λxλe ∃e’. start(e, x, e’) & p(e’, x)

When it selects for a VP complement, start and other AVs then take in the VP
complement and return a relation from individuals to the starting event. A predicate
like start reading the book might be constructed in the following way. Since the verb
takes a VP complement, it selects the denotation seen in (25). Let’s assume that
reading the book simply has a denotation like (26):
(26)

reading the book

= λxλe’∃y. reading(e’, x, y) & book(y)

Then start reading the book would have a denotation as in (27):
(27)

startVP reading the book

= λxλe∃e’. start(e, x, e’) & ∃y[ reading(e’, x,

y)] & book(y)]
In words, start reading a book takes an individual argument x where is the agent,
and an event argument e, where e is the starting event associated with the event e’
of x reading some y and where y is a book.
Coercion, in this system, works in a fashion similar to Pustejovsky’s, in that the
interpretation is provided by a qualia role associated with the nominal complement.
However, instead of nominal type-shift being responsible for coercion, the operation
is carried out by the verb itself, when it subcategorizes for a NP complement. In
these cases, start will apply to a NP complement. The event e that is being started
will be specified by deriving a predicate from one of the NP’s qualia roles. A constant
act-on-pred is include in the denotation to delimit the set of predicates which may
occupy the relevant slot.9 So a denotation for start NP may look something like (28):
9 Lascarides and Copestake (1998) note that the range of possible predicates in coercion

cases is subject to numerous restrictions. For instance, telic predicates are acceptable, but

14

(28)

startNP

= λpNP λxλe∃y∃e’. start(e, x, e) & QT/A act-on-pred (e’, x, y) &

p(y)
A predicate like start the book might be constructed in the following way. Assume
that book has a denotation something like the following:
(29)

book

= λx. book(x) & [QT (book) = λyλxλe. read(e)(y)(x)(book)]

So start the book will have a denotation like10 :
(30)

startNP the book

= λxλe∃y∃e’. start(e, x, e) &

QT/A act-on-pred (e’, x, y) & book(y) & →
= start(e, x, e) & read(e, x, y) & book(y)
Thanks on the one hand to the underspecification of the coercing verb, and on the
other to the qualia structure of the nominal, both coercing and non-coercing AV
structures will yield the same meanings.
So, sentences like (23) are not a problem for such an underspecification theory,
since start can select for complements of the same type as pick up, NP . In these
cases, the coercion in the AV is predicted to come from the internal lexical semantics
of the verb, as we see in (28).
However, it is not so clear that this system can account for the data in (17),
repeated here as (31):
not inherently atelic ones. Also illicit are predicates referring to covering some distance
(e.g. John began *(crossing) the desert). I take it that act-on-pred is supposed to filter
out such unacceptable predicates. This will be discussed further in sections 4.3 and 5.1.
10 The derivation for these coercing sentences is more easily explained when the HPSGstyle syntax the authors under discussion assume is used. Thus, the denotation in (30)
may appear somewhat unclear.

15

(31)

a. #John began the DVD and eating popcorn.
b. #John finished the book and taking notes.

Typically, theories that allow for underspecification of a verb predict that coordination of unlike categories should be grammatical, as long as the selecting verb accepts
complements of both categories (see Sag et al. (1985) and Munn (1993, 2.5.2) for
discussion). A verb like become, for instance, takes both nominal and adjectival
complements, and so both are acceptable in a coordinate structure11 :
(32) John became a lawyer and good at golf.
Consequently, it does not appear that the sort of syntactic explanation which may
account for the unacceptability of the data in (31) is possible in the sort of framework
assumed by the underspecification analysis, barring some ad hoc stipulation. In other
words, an underspecification theory should predict, incorrectly it turns out, that the
sentences in (31) are grammatical. It is possible that such structures lead to some
problem with semantic interpretation, but it is not clear what that problem would
be.
Another potential problem with the underspecification theory is that it must
attribute the availability of coercion to lexical-level idiosyncrasy. In some sense, this
marks another advantage over the NP-type shift analysis, since the underspecification
11 To account for exceptions to this, underspecification theories must stipulate certain

restrictions on coordinate structures. Consider a verb like remember, which can accept
nominal and sentential complements. To explain the ungrammaticality of (1), Sag et al.
must stipulate that nominal complements must precede sentential ones (example from
Munn (1993, 80)):
i *John remembered that Mary was arriving and the meeting that they had scheduled.

16

theory provides some level of explanation for why only certain AVs allow CC. Its
explanation is, essentially, that the availability of CC comes down to variation in
the sort of lexical entry that an AV has. Verbs which happen to categorize for both
nominal and VP complements, like begin or finish, allow coercion. Verbs which only
take VP complements, like stop or keep, do not. If no other explanation could be
found for such a difference, then such idiosyncrasy would be acceptable. If, however,
other generalizations could be found, which pointed to a deeper explanation for this
difference, then attributing the difference to mere lexical variation would appear
to be inadequate. Although such a theory might capture the right data, it would
remain unsatisfactory from an explanatory viewpoint. As we shall see in 3.1, finishtype verbs also seem to allow imperfective morphology which stop-type verbs do not.
These facts do not follow from the underspecification theory, meaning that it has
less explanatory power.
We have, then, a set of desiderata for a theory of complement coercion. First,
it should predict what sort of interpretation a coerced predicate will receive. The
analysis to be developed in this thesis will not fulfill this requirement (but see 5.1 for
some thoughts on how this might be done). It should also correctly predict the coordination data in (16) and (17/31). We have seen that the NP type-shifting analysis
incorrectly predicts the ungrammaticality of (16), and its semantics also incorrectly
predict the grammaticality of (17/31), although its possible that the ungrammaticality of these sentences stems from independent syntactic considerations. Conversely,
while the underspecification theory makes the correct predictions about the data in
(16), its syntax and semantics do not seem to rule out those in (17/31).

17

Finally, a theory of complement coercion must explain why CC is available in some
aspectual verbs but not in others. The NP type-shifting analysis over-generalizes, and
predicts that all AVs should allow CC. While the underspecification theory allows for
coercion to be available in some AVs but not others, it must explain this at the level
of idiosyncratic lexical variation. It is to this last problem of finding some principled
account for the selective availability of CC among the set of aspectual verbs to which
we shall now turn.

18

Chapter 3
PRELIMINARIES FOR A NEW ANALYSIS

3.1

What Is To Be Discussed

In 2.3, we divided the AVs into two classes: the finish-type verbs, listed in (33)
below, and the stop-type verbs, listed in (34):
(33) Finish-Type Verbs: Begin, Start, Finish, Resume, Continue.
(34) Stop-Type Verbs: Stop, Keep, Cease.
The finish-type verbs allow complement coercion, the stop-type verbs do not. How
else might the two verb classes vary? It turns out that the verbs in (33) and (34)
also differ in their ability to take imperfective morphology. Those in (33) can, as we
see in (35a-d), those in (34) cannot, as seen in (35e-g):
(35)

a. John was beginning to read the book (when the phone rang).
b. John was starting to read the book (when the phone rang).
c. John was finishing reading the book (when the phone rang).
d. John was resuming reading the book (when the phone rang).
e. *John was stopping reading the book (when the phone rang).
f. *John was keeping reading the book (when the phone rang).
g. *John was ceasing reading the book (when the phone rang).

19

There are a few potentially troubling observations to be made about the data
presented in (35). Most obviously, one might suspect that the pattern of data in (35)
falls out from the doubl-ing filter (Ross, 1972; Milsark, 1988; Pullum and Zwicky,
1999). While the exact formulation of this filter varies, it is supposed to block
the appearance of adjacent –ing affixed verbs at PF or s-structure. For instance,
Milsark (1988), whose theory of doubl-ing we shall focus on, formulates the filter in
the following way:
(36) Doubl-ing Filter: At PF, mark as ill-formed any sentence containing
contiguous –ing affixed verbs.
Why might we suspect that the pattern in (35) stems from doubl-ing? Aspectual
verbs like stop and continue, it turns out, are precisely the sorts of verbs discussed
by the authors who have looked at doubl-ing. Consequently, one might argue that
the unacceptability of –ing in stop-type verbs (37) simply stems from a filter like
(36):
(37)

a. *John was stopping reading the book (when the phone rang).
b. *John was keeping reading the book (when the phone rang).
c. *John was ceasing reading the book (when the phone rang).

One might also observe that even some finish-type verbs seem to carry apparent
doubl-ing restrictions:

20

(38)

a. *John is continuing reading the book.1
b. *John was keeping reading the book (when the phone rang).
c. *John was ceasing reading the book (when the phone rang).

This leaves only a handful of AVs as genuine exceptions to doubl-ing, finish and
resume:
(39)

a. John was finishing reading the book.
b. John was resuming reading the book.

Moreover, exceptions to doubl-ing have already been noticed in the literature. Milsark (1988), following earlier observations by Milsark (1972) and Emonds (1973),
discusses cases of apparently adjacent gerundive constructions where the doubl-ing
filter does not apply. These include sentences like those in (40) below:
(40)

a. John was enjoying reading the salacious novel.
b. John was regretting reading the salacious novel.

Their conclusion was that verbs which take nominalizing gerunds may violate the
doubl-ing filter. Tests for the presence of a nominalizing gerund include the ability to take possessive nominals (41a), known as POSS-ing, or nominals licensed by
accusative case (41b), known as ACC-ing:
(41)

a. John enjoyed Fred’s reading the salacious novel aloud.
b. Mary regretted Fred drinking the entire bottle of wine.

1 Curiously, as Ross (1972) originally noted, some AVs yield strong doubl-ing effects only

in the present tense. Thus, John was continuing reading the book sounds only marginally
unacceptable, while (38a) sounds ungrammatical.

21

Nominalizing gerunds also allow passivization (42a, c) and the formation of that-cleft
constructions (42b, d):
(42)

a. Reading the book was enjoyed by John.
b. It was reading the book that John enjoyed the most.
c. Drinking a bottle of wine was regretted by John.
d. It was drinking a bottle of wine that John really regretted.

Conversely, the verbs to which the doubl-ing filter applies, such as start and stop,
do not behave like nominalizing gerunds. For instance, they do not allow ACC-ing
(43b, d) or POSS-ing (43a, c) constructions:
(43)

a. *John started Fred’s reading the book.
b. *John started Mary reading the book.
c. *John stopped Bill’s reading the book.
d. *John stopped Mary reading the book.

Moreover, they do not allow passivization (44) or that-cleft constructions (45):
(44)

a. *Reading the book was started by John.
b. *Reading the book was stopped by John.

(45)

a. *It was reading the book that John started.
b. *It was reading the book that John stopped.

So, the these verbs do not seem to take nominalizing gerundive complements. Rather,
Milsark (1988) proposes that they are control verbs, and so take non-tensed TP
complements.
22

How does Milsark (1988) account for the ability of nominalizing gerunds to block
his formulation of the doubl-ing filter (36)? He proposes that such nominalizations
always allow a cased element to intrude between the two –ing affixed verbs. This
is most obvious in the case of ACC-ing or POSS-ing sentences (41), where an overt
nominal appears. In cases like (40), Milsark (1988) proposes that a case licensed
PRO appears in these positions, taking either genitive or accusative case depending
on the particular construction.
Jaeggli (1980), in a study of restrictions on English to-contraction, had independently proposed that cased PRO is visible to PF filters. If we carry over this
assumption that cased PRO is visible to PF filters, but not PRO of the uncased
variety, then we have an explanation for the exceptions to the doubl-ing filter. With
verbs that take nominalizing gerunds as complements, a cased PRO will appear in
the Spec-NP position, and so the two gerundive verbs are rendered non-contiguous
at PF. Such a phrase structure might look something like (46):
(46) enjoying [NP PRO+GEN [N reading the book]]
Since verbs like stop or start, to which the doubl-ing filter may apply, are assumed
to take tenseless TPs as complements, PRO will not be case licensed, and so will
be invisible at PF. In these cases then, the two ing affixed verbs will appear to be
contiguous at PF, and the doubl-ing filter (36) will cause them to be marked as
ill-formed. A phrase structure for these sorts of verbs might look like (47):
(47) *stopping [TP PRO [VP reading the book]]
If resume and finish were found to pattern with enjoy and regret, then a ready
explanation would be found for the data in (35), repeated below as (48):
23

(48)

a. John was beginning to read the book (when the phone rang).
b. John was starting to read the book (when the phone rang).
c. John was finishing reading the book (when the phone rang).
d. John was resuming reading the book (when the phone rang).
e. *John was stopping reading the book (when the phone rang).
f. *John was keeping reading the book (when the phone rang).
g. *John was ceasing reading the book (when the phone rang).

(48a-b) would be grammatical because the verbs take infinitival complements, so
the doubl-ing filter doesnt apply. (48c-d) would bypass the filter due to an unusual
syntactic structure, much as enjoy does. And (48e-g) would be straightforward
doubl-ing violations.
However, there are at least two reasons to think that a Milsark-style doubl-ing
filter cannot account for the data in (48). First off, it turns out that resume and
finish have a different syntactic structure from enjoy and regret. They do not allow
ACC-ing or POSS-ing constructions (49):
(49)

a. *John finished Fred’s reading the book.
b. *John finished Mary reading the book.
c. *John resumed Bill’s reading the book.
d. *John resumed Mary reading the book.

Moreover, they do not allow passivization (50) or that-cleft constructions (51):

24

(50)

a. *Reading the book was finished by John.
b. *Reading the book was resumed by John.

(51)

a. *It was reading the book that John finished.
b. *It was reading the book that John resumed.

In short, they fail to pattern with the class of verbs which take nominalizing gerunds
as complements. Thus, an explanation of doubl-ing filter exceptions of the sort that
Milsark (1988) employed is unavailable for resume or finish. Indeed, since these two
verbs seem to pattern with the other AVs on syntactic tests (see 3.2 and 5.3 for a
fuller discussion of AV syntax), it seems unlikely that a syntactic explanation along
could account for their unusual behavior relative to stop and keep. In short, verbs
like resume and finish remain a genuine puzzle for Milsark (1988) and other theories
of doubl-ing.
Secondly, while stop-type verbs like stop and keep only take gerundive complements, cease can also take an infinitival complement. So, if a doubl-ing filter like (36)
could account for the data in (35), then a sentence like (52) should be grammatical:
(52) *John was ceasing to read the book (when Mary came in).2
Since this sentence is clearly unacceptable, an independent reason for its ungrammaticality will have to be found.
2 In certain contexts cease seems to be acceptable in a progressive construction:

i

John was ceasing to read many novels (as his graduate studies progressed).

It is interesting that (i) refers to a habitual activity of John’s (reading novels), and not
to a single event, such as (52) refers to. The unusual behavior of cease in such habitual
readings will not be explained by the analysis in 4, and must remain a topic for future
research.

25

How, then, to explain the data we’ve been discussing? I’d like to propose that a
semantic explanation can be found for the data in (35/48). That is, an imperfectivity
operator, morphologically realized by –ing, may apply to finish-type verbs but not
to stop-type verbs.3 So, imperfective morphology will be available on finish-type
verbs, but not on stop-type verbs. This will account both for the pattern of data in
(35/48) and for the apparent unacceptability of imperfective morphology even when
doubl-ing isn’t a factor (52). Moreover, as we shall see in 4, this same semantic
explanation can account for the appearance of CC in the finish-type but not the
stop-type AVs.
This leaves a few questions, neither of which the present work can address in detail. Assuming that the ungrammaticality of (35e-g) arises from a semantic source,
then why do certain finish-type AVs (as seen in (38)) yield apparent doubl-ing effects,
even though they are not subject to the same semantic restrictions as the stop-type
verbs? And, more generally, just what is the doubl-ing filter, if it is anything at all?
Pullum & Zwicky may be on to something with their suggestion that doubl-ing could
be nothing more than a partially grammaticalized “phonesthetic dispreference for jingling sequences of similar sounding endings” (Pullum and Zwicky, 1999, 255-256). It
is notable that only those finish-type verbs which also take infinitival complements
3 One puzzling bit of data comes from Ross (1972, 70):

i *You can go on watching this if you want, but I’m stopping watching it.
ii You can go on watching this if you want, but I’m stopping watching it.
Assuming that elided structure is invisible at PF, then a doubl-ing filter would be able
to account for these data, but not a semantic explanation along the lines to be proposed.
This is a problem which will have to be addressed in later work.

26

seem to display doubl-ing effects. This is consistent with the view that there is
merely a sort of stylistic dispreference for doubl-ing, such that when an infinitival
complement is available, as with continue or start, the speaker strongly disprefers
the –ing complement. When no infinitive is available, as with finish or resume, then
the speaker will be forced to make do with the doubl-ing construction.4 Unfortunately, this thesis doesn’t have room to discuss these possibilities more carefully or
extensively.
The question now is, can we find a similar explanation for both the ability to allow
complement coercion and the ability to take imperfective aspect? These questions
shall form the basis of the following discussion. First, though, the syntactic and
semantic framework the analysis in 4 will employ should be addressed at greater
length.

3.2

The Syntax And Semantics Of AVs

We have already seen that AVs do not seem to have gerundive complements of the
nominalizing sort. What sort of syntax do they have then? The syntax of English AVs
turns out to be rather complicated in ways that dont seem to be immediately relevant
to our concerns (but see 5.3 for some discussion). For this thesis, we shall assume
that AVs are verbs which take as their complements imperfective predicates. We
shall follow Kratzer (1998) and assume that imperfectivity is encoded by an aspect
4 This observation seems to hold more generally. Thus, for verbs like try or attempt,

which take both infinitival and gerundive complements, the doubl-ing filter applies (*John
is trying reading the book, *John is attempting reading the book ). Verbs which only take
gerundive complements, and not infinitival complements, like enjoy or consider, can violate
the doubl-ing filter (John is enjoying reading the book, John is considering reading the book ).

27

operator on the head of Asp(ectual)P. We shall also assume that AspP embeds a v P,
where the external argument of the lower verbal predicate is introduced (Kratzer,
1996). For the analysis in 4, we will simply assume that the complement of the
AV is a kind of modified control clause. Where most control clauses are assumed
to be TPs, we shall assume that AV clauses are genuinely tenseless, and so lack a
TP. Instead we will assume that the complement consists of an AspP embedding a
v P. At the moment, we shall assume that the external argument for the embedded
predicate is introduced by PRO, which is co-indexed with the external argument of
the AV. The syntax we are assuming will thus look like (53) below:
(53) [vP Johni [VP begin [AspP [Asp -ing] [vP PROi eat the apple]]]]
From the perspective of the analysis to be developed in 4, the crucial assumption
encoded in (53) is that AVs apply to the output of an imperfectivity operator (–ing).
This is a semantic, rather than a syntactic assumption. So, while the particular syntactic structure sketched out in (53) may be subject to modifications,5 it is important
for the analyses in 4 to assume that the AV applies to the output of an imperfectivity
operator. Let us now see how a structure like (53) might be interpreted.
We shall assume the theory of tense and aspect developed in Reichenbach (1947),
Klein (1994), and Kratzer (1998, 2002) inter alia. In this framework, the job of an
aspect operator, such as the imperfectivity operator, is to relate the running time
associated with an event- or the event time (ET)- to the reference time (RT), which
might roughly be described as the time from which one views the event in question.
As we saw in (53), we are assuming that aspect operators are introduced by the
5 Indeed, some modifications will be explored in 5.3 below.

28

functional category AspP, which appears above v P. A plausible denotation for our
imperfectivity operator ( realized by the morpheme –ing) might be borrowed from
Kratzer (1998):
(54)

–ing = λP s,t λt∃e. P(e) & t ⊂ τ (e)

According to (54), –ing takes a property of events and returns a property of times. It
asserts that the property of events is true and that t, the RT, is a proper subinterval
of the running time associated with the relevant event (i.e. the ET). The running
time of the event is introduced as the output of the trace function τ , where the event
itself is the input (Link, 1987). Thus –ing relates the ET to the RT, where the latter
is a part of the former.6
To see how the imperfectivity operator in (54) works, lets see how we might go
about constructing a derivation for a predicate like John is eating the cookie. First,
the denotation for the v P, John eat the cookie, might appear as follows:
(55)

John eat the cookie = λe. eat(the cookie, e) & Agent(e) = John

(55) simply states that e is an event of eating the cookie, of which John is the agent.
After application of the imperfectivity operator, we will have a structure like the
following:
(56)

John eating the cookie = λt∃e. John.eat.the.cookie(e) & t ⊂ τ (e)

John eating the cookie returns a property of the RT, and asserts that the RT comprises a proper subinterval of the running time of some event e, and that e is the
6 The reader might note that (54) fails to solve the imperfective paradox, discussed

below in section 3.3. However it shall be adequate for our purposes.

29

event where John eats a cookie. It is such an imperfective property of times that the
AVs will apply to. In 3.3, we shall see some evidence that the framework we have
proposed is on the right track.

3.3

Some Evidence For This Framework

Now, do we have any evidence that such an imperfectivity operator resides in the
gerundive complement of the AVs? Data supporting this hypothesis comes from two
sources: what Dowty (1979) labeled the imperfective paradox, and the distribution
of stative predicates in the complement position of AVs. An imperfectivity operator
with the same –ing morphology and (we will assume) identical semantics makes an
appearance in English progressive constructions. Such constructions are formed by
the application of a copular verb carrying tense information to the output of the
imperfectivity operator. An example of this is seen in (57) below:
(57) John was eating the cookie.
Assuming that was simply denotes a past tense operator, and assuming the imperfectivity operator in (54), then (57) might somewhat informally be translated as
(58):
(58) there is some moment t such that t precedes tnow and t forms the proper
part of an interval where John eats the cookie.
Different entailments arise from a progressive construction depending upon whether
the event encoded within the sentence is telic or atelic. For instance, an atelic

30

progressive sentence (59a) entails that the perfective form of the sentence is also true
(59b):
(59)

a. John was eating.
b. John ate.

However, this entailment does not arise for telic predicates. In other words, a telic
progressive (60a) does not entail that its corresponding perfective form is true (60b):
(60)

a. John was eating the cookie.
b. John ate the cookie.

Various solutions to this problem have been proposed.7 Most of these solutions have
been inspired by Dowty’s original discussion of the paradox, which attributed the
failure of the entailment in telic predicates to the semantics of the imperfectivity
operator itself. According to Dowty, an imperfectivity operator may take in a telic
predicate P, but does not assert that P is true at t in w0 . Rather, it asserts that P
is true at some t’ in some w’ such that t is a proper subinterval of t’ and w’ is an
inertial world of w0 at t. An inertial world is one which is identical to the present
world assuming, to state things somewhat vaguely, that things continue going in the
same direction. So, a statement like (60a) says something like:
(61) There exists some moment t such that t precedes tnow , and t is the proper
subpart of some interval t’ in w0 where for all w’ such that w’ is an inertial
world of t at w0 , John eats the cookie in the interval t’ in w’.
7 See Portner (1992, chapter 5), Zucchi (1999) for discussions of some of the most promi-

nent of these.

31

The details of what constitutes an inertial world turn out to be quite vexing, but
what’s important for our purposes is that the imperfective paradox arises from the
semantics of the imperfectivity operator realized as –ing. Now, if this same imperfectivity operator appeared in the gerundive complements of AVs, then we would
predict the imperfective paradox to arise in AV constructions. This indeed seems
to be the case. As with the progressive construction in (59), both stop-type and
finish-type aspectual verbs entail the perfective form of the complement when the
predicate is atelic:
(62)

a. John stopped eating.
b. John ate.
c. John began pushing the wheelbarrow.
d. John pushed the wheel barrow.

However, we see in (63) that AVs do not entail the perfective form of their complement when the predicate is telic. So (63a) does not entail (63b), and (63c) does not
entail (63d):
(63)

a. John stopped eating the cookie.
b. John ate the cookie.
c. John began building the wheelbarrow.
d. John built the wheel barrow.

These facts would follow if we assumed that an imperfectivity operator, identical to
that appearing in progressive constructions, showed up in the complement position of
AVs. Note that the imperfective paradox is most easily explained if we assume that
32

the imperfectivity operator is intensional. Since such intensionality will not prove
important for the rest of our discussion, we shall ignore it henceforth, and revert to
the non-intensional imperfectivity operator offered (54).
Next, consider the behavior of stative predicates in the complement position of
AVs. Returning to our comparison of progressive constructions and AV constructions, we might note that, in general, stative predicates sound infelicitous when
embedded within progressive constructions:
(64)

a. #John was liking fish.
b. #Bill was loving Patricia.
c. #Fran¸ois was being afraid.
c

Such stative predicates do not have troubling combining with AVs, however:
(65)

a. John started liking fish.
b. Bill kept loving Patricia.
c. Fran¸ois stopped being afraid.
c

Dowty (1979), following work by Taylor (1977), suggests the following explanation
for the infelicity of the sentences in (64). A stative asserts that some property (e.g.
loving Mary) holds for all (contextually relevant) moments of time. A progressive
picks out a moment of time inside of the interval during which some property holds.
Now, since stative properties already hold of all contextually relevant moments of
time, it is simply redundant to pick out a moment of time associated with a stative
and state that the stative holds at that moment.

33

However, Comrie (1976) and Dowty (1979) note that certain stative predicates
allow a progressive reading under the right circumstances. The most common perhaps involve the verb live. This can appear, for instance, in both the simple present
(66a) and progressive (66b) variety, with subtle differences in meaning:
(66)

a. John lives in Chicago.
b. John is living in Chicago.

(66a) seems to indicate that Johns living in Chicago is a more or less permanent
state of affairs. (66b), on the other hand, indicates either a certain degree of uncertainty about John’s place of residence or that John is only temporarily a resident of
Chicago. In other words, the progressive is acceptable with certain stative predicates
as long as the truth of the proposition in question is highly uncertain or the state a
temporary one. Something similar happens with a stative verb like stand. In certain
situations, the progressive form seems infelicitous:
(67)

a. An old oak tree stands in my backyard.
b. #An old oak tree is standing in my backyard.

While (67a) simply indicates that for all contextually relevant times there is an old
oak tree standing in my back yard, the progressive form in (67b) again lends the
state an air of temporariness or uncertainty. (67b) would only be felicitous if the old
oak tree had the capacity to move, which makes it sound odd out of the blue. In
a slightly different context, however, predicates involving stand sound alright in the
progressive form:

34

(68)

a. An old bird bath stands in my backyard.
b. An old bird bath is standing in my backyard.

Again, while (68a) indicates a situation where the old bird bath is a permanent
fixture of my backyard, (68b) indicates a situation where the bird bath’s presence in
my backyard is quite possibly temporary.
To explain this sort of behavior, Dowty invokes Carlson (1977) and the distinction between stage-level and individual-level predicates. Stage-level predicates, to
put things informally, are properties that hold of an individual during a particular interval. Individual-level predicates hold of an individual generally. Carlson’s
own system is rather complicated, and requires ontological assumptions which differ markedly from our own analysis. However, Kratzer (1995) provides a semantics
for individual-level and stage-level predicates more in line with the analysis in 4.
Kratzer claims that the two predicate classes differ in their argument structure: viz.
stage-level predicates have an extra argument for eventualities that individual-level
predicates lack.
Let’s see how this might work for a stage-level/individual-level ambiguous stative
predicate like live in Chicago. For its individual-level interpretation, it would simply
be a property of individuals, which we might represent in a somewhat simplified
denotation like (69):
(69)

live.in.Chicago1 = λx. live.in.Chicago(x)

(69) simply says that an individual x possesses the property of living in Chicago. For
the stage-level reading, live in Chicago would be a relation between individuals and
eventualities, and so might have a denotation like (70):
35

(70)

live.in.Chicago2 = λxλe. live.in.Chicago(x)(e)

(70) says that e is the eventuality of x living in Chicago. Crucially, once the external
argument of (70) has been saturated, it will be a property of eventualities, and so
will be able to interact with our imperfectivity operator. So, a statement like John
is living in Chicago, would have a denotation like (71), ignoring tense:
(71)

John.living.in.Chicago = λt∃e. live.in.Chicago(e) & t ⊂ τ (e)

(71) now states that t, the RT, is a proper subinterval of the running time of some
event e, where is the event of John living in Chicago. Notably, the individual-level
version of live in Chicago, lacking an event argument, will never have the proper type
to interact with an imperfectivity operator, so a progressive construction like (66b)
necessarily has a stage-level reading.
Interestingly, it appears that a stage-level reading is possible for many other
stative predicates, so long as the right context is provided. Take, for instance, a
stative predicate such as like fish. Out of the blue, this predicate accepts simple
present tense (72a), but the corresponding progressive (72b) appears infelicitous:
(72)

a. John likes fish.
b. #John is liking fish.

A statement like (72b) probably feels infelicitous because, under normal circumstances, one’s taste in fish remain constant throughout life, and so an individual-level
reading is more appropriate than a stage-level reading. But imagine speaking to the
parent of a young child. This parent might complain that his daughter is a very
picky eater, which makes planning dinner very difficult. The daughter is particularly
36

difficult when it comes to fish. The parent serves fish for dinner every Friday. One
week, the daughter will claim she loves fish and demand she be served it more often.
The next, she will claim it is disgusting and that she never wants to eat it again.
Then the parent utters (73):
(73) Last week, luckily, my daughter was liking fish (But who knows about this
week).
In such a context, a stage-level reading for like fish sounds perfectly reasonable.
Again, however, this is a context which requires the state to be a possibly temporary
one, necessitating a stage-level reading.
A similar example, noted by Deo (2011), concerns Mary was weighing 110 lbs.
Out of the blue, this sounds infelicitous, since predicates like weigh X lbs are most
naturally employed as individual-level properties. But in circumstances where Mary’s
weight is fluctuating a great deal, such as after a pregnancy, and where her present
weight might be temporary, then a stage-level reading might be appropriate, and so a
statement like Mary was weighing 110 lbs at that point sounds perfectly reasonable.
The imperfectivity operator, then, may accept stative predicates, but only in
contexts where a stage-level reading is appropriate. This accords with our intuitions
about that stative complements of AVs that we saw in (65) above, repeated as (74)
below:
(74)

a. John started liking fish.
b. Bill resumed loving Patricia
c. Fran¸ois stopped being afraid of snakes.
c

37

Intuitively, an AV describes a change in the status of the property encoded in the
complement over time (this will be spelled out more explicitly in 4.2 below). Consequently, only the stage-level variety of stative, which may hold over particular
intervals, should be possible in the complement position. This is consistent with
the hypothesis that an imperfectivity operator resides in the complement position of
AVs.
Interestingly, a subset of stative predicates do not seem to allow stage-level readings. For instance, know, as was observed by Comrie (1976) and Dowty (1979), can
not appear in a progressive construction, even when it is describing a potentially
temporary state. Imagine that Fran¸ois, a native speaker of French, suffers from
c
frequent and temporary lapses into a kind of general aphasia, wherein he displays
absolutely no sign of possessing any knowledge of his native language. In these cases,
a statement like (75) still sounds infelicitous:
(75) #Today, Fran¸ois is knowing French.
c
(75) is infelicitous despite the context, where Franois knowledge of French is a possibly temporary state. Dowty observes that, while the truth of statives that allow
the progressive can be ascertained by observing a large of enough number of instances where the state seems to hold, ascertaining that Fran¸ois knows French does
c
not work in the same way. Instead, something like know French is just a property
that holds of Fran¸ois. It thus seems to always be an individual-level property, and
c
never a stage-level property. This holds with certain other statives like, be American.
Imagine that John managed to renounce and then regain US citizenship repeatedly.
At the moment, he happens to have US citizenship, but it appears that he might
38

renounce it at any moment. Even in such a context, (76b) still sounds odd, just as
(76a) does.
(76)

a. #John is knowing French.
b. #John is being American.

We have reasons to believe that the unacceptability of the sentences in (76) derives
from their incompatibility with imperfective aspect. Briefly, the imperfectivity operator picks out a sub-part of an interval over which some property of events holds.
But individual-level predicates dont hold over intervals, or over events with an associated interval. They hold only over individuals. Moreover, unlike the potentially
stage-level predicates discussed above, one cannot derive from the predicates in (76)
a property which does hold over eventualities. Thus an imperfectivity operator will
simply be unable to interact with such predicates.
If it is the case, as we have hypothesized in 3.2, that the –ing morpheme in
the AV complements encodes a similar imperfectivity operator, then these inherently individual-level statives also ought to appear infelicitous when placed in the
complements of AVs:
(77)

a. #John began knowing French.
b. #John stopped being American.

As we can see in (77), such stative predicates are indeed infelicitous.
It seems to be the case then that AVs may interact with stative predicates, but
only those stative predicates that allow stage-level interpretations and hence may
interact with an imperfectivity operator. The data we have seen for statives, then, is
39

consistent with the hypothesis that the complements of AVs contain an imperfectivity
operator.8

8 The reader might wonder about a stative like be afraid. It can easily combine with

AVs (e.g. John stopped being afraid ). But, as with know French, it seems incapable of
taking a progressive reading. I’d like to suggest this is because it is an inherently stagelevel predicate. Thus, with many statives the progressive allows us to distinguish between
the individual-level and stage-level readings. On the other hand, some statives are only
individual-level predicates, and so they cannot combine with an imperfectivity operator. Be
afraid could combine with such an operator in principle, since it is a stage-level predicate,
but such an operation is redundant because this predicate only holds of particular intervals
anyway. One needn’t say John was being afraid at the haunted house because John was
afraid at the haunted house says the same thing more directly. However, be afraid and other
potentially stage-level predicates can combine with imperfectivity operators in principle,
since they hold over events, and so they can appear in AV constructions.

40

Chapter 4
A TYPE-BASED ANALYSIS OF ASPECTUAL VERBS AND
COMPLEMENT COERCION EXPLAINED

4.1

What Is To Be Explained And How

Our current task, as outlined in 3.1, is to explain why a subset of aspectual verbs,
those of the finish-type, both allow complement coercion and may take imperfective
morphology. Let’s tackle the second of these observations first.
We have already determined that aspectual verbs take as their complement imperfective predicates. Such imperfectivity we have assumed arises from an imperfective
operator like (54), repeated as (78) below:
(78)

–ing = λP s,t λt∃e. P(e) & t ⊂ τ (e)

This operator takes properties of events and returns properties of times. We
also observed in (35), repeated as (79) below, that finish-type AVs may appear in
progressive constructions, while stop-type verbs may not:

41

(79)

a. John was beginning to read the book (when the phone rang).
b. John was starting to read the book (when the phone rang).
c. John was finishing reading the book (when the phone rang).
d. John was resuming reading the book (when the phone rang).
e. *John was stopping reading the book (when the phone rang).
f. *John was keeping reading the book (when the phone rang).
g. *John was ceasing reading the book (when the phone rang).

Following Kratzer (1998), we will assume that such progressive constructions also
require an imperfectivity operator that is morphologically realized by –ing.
It would appear then that, for a sentence like (80) below, the AV is sandwiched
in between two imperfectivity operators:
(80) John is finishing reading the book.
So, if the imperfectivity operator maps properties of events to properties of times
and has the type

s,t , i,t , finish-type verbs simply do the reverse. Taking proper-

ties of times, they return properties of events. Consequently, they should have type
i,t , s,t . This formal somersault, among other things, predicts the possibility of
stacking imperfective operators and finish-type AVs on top of each other indefinitely:
(81) John was resuming finishing reading the book.
(81) is rather hard to parse, but does indeed sound acceptable.
We have also observed that, while the complements of stop-type verbs appear to
be imperfective predicates, the stop-type AV itself cannot appear in a progressive
construction, as we see in (82) below:
42

(82) *John was stopping reading the book.
While one might attribute the ungrammaticality of (82) to a PF-level doubl-ing
filter (Ross, 1972; Milsark, 1988), we have seen that the doubl-ing filter has trouble
accounting for the full spectrum of data seen in (79). Instead, let’s suppose that
stop-type AVs simply modify properties of times, and so have the type

i,t , i,t .

If this were the case, then we would have an explanation for the data in (79). Both
finish-type verbs and stop-type verbs take in predicates of type i,t , but while the
former return predicates of type s,t and so can again combine with an imperfectivity
operator of type

s,t , i,t , stop-type verbs return predicates of type i,t , which

cannot combine with imperfectivity operators again. Let’s see if we can construct
denotations for the AVs which might plausibly capture this distinction.

4.2

The Semantics Of Aspectual Verbs

The following treatment of aspectual verbs is largely inspired by Portner (1992,
chapter 5), which was itself heavily influenced by ter Meulen (1985). Portner’s system
resides in the framework of situation semantics. As we have seen in 4.1, though, our
observations about finish– and stop-type verbs, and their relations to imperfectivity
operators, can be most easily stated when we assume events and times as primitives.
Accordingly, the analysis here will differ somewhat from Portner’s own.
Intuitively, AVs pick out a subinterval corresponding to the interval associated
with the running time of an event and situate that subinterval in time. So, if I
say that John stopped eating the cookies, I am saying in effect that there was some
moment in the past when John was eating the cookies; at the next moment, he was
43

no longer eating the cookies. For John began eating the cookies, I’m saying that there
was no time before the reference time when the event of eating the cookies was being
carried out, but from that time onward the event is being carried out. Similarly John
finished eating the cookies means that the whole interval of times during which the
cookie eating occurs must precede some moment. After that moment, the event of
eating the cookies will be entirely used up, so to speak.
With these intuitive definitions in mind, let’s propose a denotation for the representative stop-type verb:
(83)

stop = λf i,t λt. f(t) & ¬∃t’[t immediately precedes t’ & f(t’)]

Notice that (83) satisfies our desire from 4.1 for a stop-type function to have the
type

i,t , i,t . It says that a property of times is true at an interval t but is not

true at an interval t’ which it immediately precedes. Immediate precedence might be
defined as follows:

Immediate Precedence (i.p.): t immediately precedes t’ iff t precedes t’ and there
does not exist any t” that comes between t and t’.

Now let’s see how (83) might combine with an imperfective predicate, like John
eating the cookies. Recall that this had the following semantic denotation:
(84)

John eating the cookie = λt∃e. John.eat.the.cookie(e) & t ⊂ τ (e)

When stop applies to (84), it will return a structure like the following:

44

AVP
i,t
λt. John.eating.the.cookie(t)
& ¬∃t [t i.p. t & John.eating.the.cookies(t )]
John stop eating the cookie

AV
i,t , i,t
λfλt. f(t) & ¬∃t [t i.p. t & f(t )]
stop

AspP
i,t
λt∃e. John.eat.the.cookie(t)
& t ⊂ τ (e)
John eating the cookie

Asp
vP
s,t , i,t
s,t
λpλt∃e. p(e) λe. eat(the cookie, John, e)
& t ⊂ τ (e)
–ing
John eat the cookie
Figure 4.1: Derivation for John stopped eating the cookie.

(85)

John stop eating the cookies = λt. John.eating.the.cookies(t) & ¬∃t’[t
immediately precedes t’ & John.eating.the.cookies(t’)]

(85) says that at t, the reference time, John is eating the cookie, but at the very
next moment, he is no longer eating the cookie. This can be seen even more clearly
in a tree structure, such as figure 4.1 above.
Lets now consider a denotation for finish:
(86)

finish = λf i,t λe. f(τ (e)) & ¬∃t[ τ (e) < t & f(t) = 1]

(86) says that the property of times f is true for the interval corresponding to the

45

running time of an event e and that there are no times following the running time of
e for which the property of times f is true. So f does not hold for any time intervals
following the running time of e. Notably, verbs like finish are now predicted to have
a type

i,t , s,t .

Note that (83) and (86) are identical except in two crucial respects. First, instead
of taking a time t as an argument as (83) does, (86) takes an event argument e,
whose corresponding interval τ (e) f can apply to. This allows us to explain the
type differences between stop and finish discussed in 4.1. Furthermore, the semantic
differences between these two verbs are explained as the difference between immediate
precedence and precedence (<), where precedence merely indicates that t precedes t’.
This captures our intuition that, when one stops eating the cookies, one may resume
at some point in the future, since it is only the moment immediately following t
in (83) where f(t) cannot be true. Whereas for finish, there cannot be any future
intervals t in which one eats the cookies.
Of course, it is actually too strong to say, as (86) does, that there are no
possible future intervals with the relevant property. For instance, one might resume
eating some cookies at some point in the future, even if one has finished eating the
current batch. Therefore, some sort of contextual restriction is necessary, which is
represented by C in (87):
(87)

finish = λf i,t λe. f(τ (e)) & ¬∃t in C[ τ (e) < t & f(t) = 1]

(87) says essentially the same thing as (86), except that the set of times is now
restricted to those that are contextually relevant. Unlike (86), (87) does not say that
one has finished performing the relevant activity for all time, but merely that one
46

has finished the activity for all contextually relevant times.
Interestingly, we can also develop a denotation for begin that is almost identical
to finish, except for one tweak:
(88)

begin = λf i,t λe. f(τ (e)) & ¬∃t[ t < τ (e) & f(t) = 1]

Again, we have a function of type

i,t , s,t , in accordance with the discussion in

3.1. Unlike (86), τ (e) in (88) follows, rather than precedes, the time t. This captures
the difference between finish and begin. The former indicates the final moment of
doing something, the latter the initial one. Also, as with (86), (88) states that
there are no time intervals with the relevant property which precede τ (e). This is
clearly too strong, so again some context restriction is necessary. Thus a reasonable
denotation for begin might look like (89):
(89)

begin = λf i,t λe. f(τ (e)) & ¬∃t in C[ t < τ (e) & f(t) = 1]

For the sake of completeness, a denotation for finish eating the cookies appears
in figure 4.2. A denotation for begin eating the cookies would work in much the same
way.
These denotations (83-89) illustrate how we might go about constructing semantic
representations for the entire class of aspectual verbs. It seems likely that denotations
for all of the stop-type and finish-type AVs could be constructed along these lines.
For instance, keep might simply be regarded as the inverse of stop. Rather than
saying that a property holds at t and fails to hold at the following time t’, keep
asserts that relevant property still holds at the following time t’. Its denotation
might then look like (90):

47

AVP
s,t
λt. John.eating.the.cookie(t)
& ¬∃t [t < t & John.eating.the.cookies(t )]
John finish eating the cookie

AV
i,t , s,t
λfλe. f(τ (e)) & ¬∃t[τ (e) < t & f(t)]
finish

AspP
i,t
λt∃e. John.eat.the.cookie(t)
& t ⊂ τ (e)
John eating the cookie

Asp
vP
s,t , i,t
s,t
λpλt∃e. p(e) λe. eat(the cookie, John, e)
& t ⊂ τ (e)
–ing
John eat the cookie
Figure 4.2: Derivation for John finished eating the cookie.

(90)

keep = λf i,t λ. f(t) & ∃t’[t immediately precedes t’ & f(t’)]

A denotation for resume and the other AVs might also be constructed along these
lines. But such a project wont be pursued here. What’s important is that we have
developed a system which allows us to capture the apparent type difference between
the two verb classes discussed in 4.1, and thereby to account for their behavior with
imperfective morphology (this will be discussed at greater length in 4.4). Moreover, it
also lays the groundwork for a mechanism that can account for complement coercion.

48

4.3

Complement Coercion Explained

We are now, I think, in a position to explain the transitive alternation seen in (1)
above (reproduced as 91 below):

(91)

a. John began reading the book.
b. John began the book.

While the same AV appears in (91a) and (91b), the altered form (91b) has several curious properties. For one, the aspectual verb now appears to be a transitive
verb. Whats more, as Pustejovsky (1993) inter alia have observed, (91b) seems to
carry some sort of hidden predicate. The sentence is saying that John began doing
something with the book in question.
This doing something is itself restricted in several unusual ways. In particular,
the possible predicates are inherently telic. So, while atelic predicates are acceptable
in (92a) and (92c), their alternations in (92b) and (92d) are not:
(92)

a. John resumed pushing the cart.
b. *John resumed the cart.
c. John began reading books.
d. *John began books.

Other odd restrictions seem to exist on the altered form of the aspectual verb.
For instance, Lascarides and Copestake (1998), following Godard and Jayez (1993),
have observed that complement coercion may not apply to verb phrases referring to
distances, for instance:
49

(93)

a. John began crossing the desert.
b. *John began the desert.

The extent and nature of these thematic restrictions on complement coercion
shall not be pursued in this thesis. Recall from 2.3 that Lascarides and Copestake
(1998) had assumed a constant act-on-pred, which constrained the range of possible interpretations for coercing sentences. We will assume the existence of a similar
abstract predicate, called Pred, which likewise restricts the range of possible interpretations. Being a sort of abstract transitive predicate, we will assume that it is a
relation between individuals and events.
Lascarides and Copestake (1998) assumed the existence of a nominal qualia structure which allowed a fully specified interpretation to emerge. However, they assumed
that such interpretations were defeasible, and subject to reinterpretation depending
upon contextual and world knowledge factors. We will depart from Lascarides and
Copestake with regard to the existence of qualia structure, and instead assume that
the interpretation for Pred is filled in entirely by world-knowledge and contextual
factors. While a fully worked out theory of CC interpretation will not be on offer,
in 5.1 we will discuss how Lascarides and Copestakes work provides an excellent
starting off point for building such a theory.
How might Pred be inserted into the derivation in a compositional manner? Recall that the AVs require imperfective predicates. It makes sense then to assume
that Pred must be embedded inside of some sort of imperfectivity operator. I’d like
now to propose that the insertion of this imperfectived Pred, and hence complement
coercion, is carried out by an operator which applies to finish-type verbs and re-

50

s,t
λe[finish(−ing(Pred(the cookies)))(e)]
FinishT rans the cookies

e, s,t
λxλe[finish(−ing(Pred(x)))(e)]
FinishT rans

i,t , s,t , e, s,t
λf i,t s,t λxλe[f(−ing(Pred(x)))(e)]
Trans

e
the cookies

i,t , s,t
Finish

Figure 4.3: Derivation for Finish the cookies.

turns transitive verbs. We will call this operator Trans. Syntactically, I assume that
this operator appears as a phonologically null affix to the aspectual verb. Here is a
possible denotation for it:
(94)

Trans = λf i,t s,t λxλe∃. f(-ing(Pred (x)))(e)

Essentially, Trans takes an argument of the type of a finish-type aspectual verb
and returns a transitive verb, type e, s, t . It does so by positing an imperfectived
verbal predicate Pred in the argument position of the aspectual verb, and allowing
the individual argument of the aspectual verb to be interpreted as the argument of
this predicate. The construction of finish the cookies is illustrated by the tree in
figure 4.3.
Now, let’s see how this proposal allows us to explain the data discussed in chapters
2 and 3.

51

4.4

Some Consequences Of The Type-Based Analysis

Of the CC analyses discussed in 2, the most cogent were the NP type-shifting analysis
presented in Pustejovsky (1991, 1993) and the underspecification analysis of Copestake and Briscoe (1992, 1995), and Lascarides and Copestake (1998). While the
NP type-shifting analysis succeeded in explaining how these constructions acquire
their default interpretation, it had several drawbacks. For one thing, it made wrong
predictions about the coordinate structure facts in (16) and (17). For another, it
failed to predict the variable availability of complement coercion in aspectual verbs.
While the underspecification theory discussed in 2.4 can account for (16), it still has
problems with (17). Lets consider first (16), repeated as (95) below:
(95) John picked up and started the book.
Pustejovsky’s NP type-shifting analysis predicts that this sentence should be ungrammatical, as the nominal will simultaneously have to have the type of a nominal,
in order for pick up to apply, and a verbal predicate, in order for start to apply.
A theory which holds that complement coercion stems from underspecification in
the coercing verb, such as Copestake and Briscoe (1995) and Lascarides and Copestake (1998), will correctly predict (95) to be grammatical. Likewise, the type-based
analysis developed in this work makes the correct prediction about (95). Under this
theory, an abstract operator, Trans, will be affixed to the AV (as we see in figure
4.3). Since Trans returns an AV that takes a nominal complement, just like pick up
does, no type clash will arise.
(17a), repeated as (96) below, shows that an aspectual verb cannot take NP and
VP complements:
52

(96) # John began the DVD and eating popcorn.
Recall first that, depending upon one’s assumptions, a syntactic explanation for the
unacceptability of (96) is conceivable. Leaving that possibility aside, semantically
the NP type-shifting analysis incorrectly predicts that (96) should be acceptable,
since the DVD will be shifted to have the same type as eating popcorn, so no typeclash should ensue. For the underspecification theory, I do not think that a syntactic
explanation is available. Since AVs like begin accept both nominal and verbal predicates as complements, then coordination of both complements should be possible
for AVs. Thus, this theory also makes incorrect predictions about (96), absent some
stipulations about the possible coordinate structure of AV complements.
However, our type-based analysis of complement coercion provides a semantic
explanation of (96), independent of ones syntactic assumptions. In order for begin
to apply to the DVD, Trans must apply to the verb, generating an AV that takes a
nominal complement. Its semantics would resemble those of figure 4.3. However, for
begin to apply to a verbal predicate like eating popcorn, Trans must not have applied
to it. Thus, for (96) to be grammatical under our theory, Trans must be at once
present and absent. Since this is clearly impossible, the type-based analysis correctly
predicts that (96) is ungrammatical.
The Trans analysis also makes correct predictions about the facts that hindered
the simple syntactic analysis discussed in 2.1. Recall that, in (5) and (6), repeated
below as (97) and (98), modifiers could modify both the verbal complement and the
aspectual verb itself:

53

(97)

a. Mary [finished [eating the meal slowly]]
b. Mary [finished [eating the meal] slowly]

(98)

a. Mary [finished [baking the bread in the oven]]
b. Mary [finished [baking the bread] in the oven]

This follows from the structure in figure 4.2, where both an overt verbal complement
and aspectual verb are available for modification. In the coercing sentences, like
(7) and (8), repeated as (99) and (100) below modification can only occur to the
aspectual verb:
(99)

a. Mary [finished the meal slowly]
b. *Mary [finished [VERB the meal slowly]]

(100)

a. Mary [finished the bread in the oven]
b. Mary [finished [VERB the meal slowly]]

This is because these constructions lack any overt verbal complement, as figure 3
shows. The abstract predicate Pred is instead embedded inside of the AV itself. So,
if an adverb modifies Pred, it must also modify the AV, and so interpretations like
(99b) and (100b) are impossible. The Trans analysis also predicts this outcome.
Now lets turn to the observations made in 3.1: that the finish-type verbs allow
both CC and take imperfective morphology, while the stop-type verbs allow neither. Recall that the NP type-shifting analysis incorrectly predicted all AVs to allow
CC. The underspecification theory left CC availability up to idiosyncratic lexical
variation, so a correlation such as this one would be highly surprising. However,
the analysis developed in 4 accounts for both CC availability and the availability
54

of imperfective morphology with the same mechanism: the type difference between
the two verb classes. Finish-type verbs, we hypothesized, take properties of times
into properties of events, and so have a type of

i,t , s,t . Since imperfectivity

operators, we assume, take properties of events into properties of times, our theory
correctly predicts that they should be able to apply to finish-type predicates. Notice
that a second application of the imperfectivity operator is impossible with stop-type
verbs, which have a type of

i,t , i,t , and therefore block the application of the

imperfectivity operator.
In 4.3, it was argued that complement coercion could be explained by a complex
operator, Trans, which applies to the AV. Trans must apply to an argument of a
specific type. Since the finish-type and stop-type verbs differ in type, it cannot
apply to both. We assumed that it applies only to verbs of type

i,t , s,t , that is

to finish-type verbs. So complement coercion is possible with finish-type verbs, but
not with verbs of the stop-type. Thus a single type difference accounts for both of
these phenomena.

4.5

Summary

When considering the problem of complement coercion, we first observed that previous analyses of the phenomenon fail for one reason or another. In particular,
Pustejovsky’s NP type-shifting analysis of CC seems to make incorrect predictions
about the behavior of AVs in certain coordinate structures. Moreover, it overgeneralizes, and predicts that CC should be available across the set of AVs. An
underspecification-based analysis of CC makes correct predictions about the coor-

55

dinate structures in (16), but not those in (17). Also, while it can account for the
selective availability CC in AVs, it must do so at the level of idiosyncratic lexical
variation.
To develop a new analysis of CC, we first observed that AVs could be divided
into two classes: the finish-type verbs, which allow complement coercion; and the
stop-type verbs, which do not. It was also observed that, while the former also allow
imperfective morphology, the latter do not. We next sketched out some syntactic and
semantic assumptions about AV structures. We assumed that AVs take as their complement imperfective predicates. We assumed that the imperfectivity of the verbal
complement stemmed from the application of an imperfectivity operator, which itself
was the head of a AspP. This imperfectivity operator takes in properties of events
and returns properties of a time which is a part of the running time associated with
that event. We also explored some data consistent with these hypotheses, related to
the imperfective paradox and the behavior of stative predicates in the complement
position of AVs.
Since this imperfectivity operator can apparently apply to finish-type verbs, and
since finish-type verbs can apply to the imperfectivity operator, it makes sense to
assume that finish-type verbs are the inverse of the imperfectivity operator. They
take properties of times and return properties of events. The stop-type verbs, on
the other hand, simply modify properties of times. Consequently, imperfectivity
operators cannot apply to them. We then proceeded to develop a semantics for AVs
consistent with these type distinctions.
This accounts for the imperfective morphology data, but what about complement

56

coercion? For this, we argued for the existence of an abstract predicate Pred, which
is introduced into the derivation by an operator Trans. Trans, which can only apply
to predicates with the type of the finish-type verb, essentially turns that verb into
a transitive verb. The argument of the transitivized finish-type verb then becomes
the logical argument of Pred. Now, since Trans can apply to a finish-type verb, it
cannot apply to a stop-type verb. Thus, a single factor, the type difference between
the two verb classes, explains these two apparently unrelated phenomena.
There are a few puzzling components to this analysis, which will have to be
explored more carefully in future work. In 5, I would like to make some suggestions
about how such work might proceed. First, we have assumed the existence of an
abstract predicate, Pred, but offered no hint about how its interpretation might be
filled in. Some ideas about this will be sketched out in 5.1 below. Also, many authors
(Pustejovsky, 1993; Lascarides and Copestake, 1998) have observed that certain nonAVs, such as enjoy, seem to allow an alternation very similar to AV complement
coercion. In 5.2, we shall see how these enjoy-type verbs resemble finish-type AVs,
and how our analysis of AVs may be extended to these verbs.
Finally, we assumed that AVs are a sort of modified control verb. The AV, we
assumed, embeds an AspP, rather than a full TP. The eternal argument of the embedded predicate is provided by a PRO which is controlled by the external argument
of the matrix clause. While this seems to account for several semantic facts about
AV complements, it is a somewhat non-standard syntactic analysis. In 5.3 we will
explore some other syntactic possibilities for AVs.

57

Chapter 5
WHAT’S NEXT?

5.1

Pred And The Interpretation Of Coerced Predicates

In 4.3, we posited the existence of an abstract predicate, Pred. Its most obvious
limitation is that it is not at all clear how it works. How, for example, does one infer
that in a sentence like John began the book John is most likely beginning reading the
book, rather than doing something else with it? Perhaps the most interesting theory
for such interpretation was proposed by Lascarides and Copestake (1998). Recall
from 2.4 that these authors assume a qualia structure similar to Pustejovsky’s, along
with coercing verbs that are underspecified as to the type of their complement.
Lascarides and Copestake also posit an abstract constant within the coercing verb
called act-on-pred. In order to acquire a default interpretation, a co-indexation mechanism links act-on-pred to either the telic or the agentive qualia role within the verbs
NP complement. What allows for contextual modification along the lines of (15)
above? Lascarides and Copestake argue that qualia information is marked as default
within the lexicon. This default information can then be overridden by contextual
factors. This latter operation is spelled out with a pragmatic theory called Discourse
In Commonsense Entailment, or DICE (Asher and Morreau, 1991; Lascarides and
Asher, 1993; Lascarides et al., 1996).
Lascarides and Copestake provide a very elegant solution to the problems of
CC interpretation. It would be an interesting task to combine their system with

58

the one developed in the present work. In particular, there are suggestive parallels
between the act-on-pred of their analysis and Pred posited above. While integrating
our analysis with their model may yield very interesting results, Lascarides and
Copestake assume a markedly different syntactic-semantic framework from ours, and
the task will be a formidable one.

5.2

Complement Coercion And Enjoy -Type Verbs

As many authors have observed (Pustejovsky, 1991, 1993; Lascarides and Copestake,
1998), complement coercion appears to extend beyond the set of finish-type AVs.
Consider, for instance, verbs like enjoy and attempt:
(101)

a. John enjoyed (reading) the novel.
b. John attempted (reading) the novel.

For convenience, we will call non-AVs that appear to allow complement coercion
enjoy-type verbs. Enjoy-type verbs seem to have many properties in common with
finish-type AVs. In particular, they seem to display the imperfective paradox. Thus,
neither (101a) nor (101b) entail that John actually read the novel. In contrast,
consider verbs like regret and celebrate:
(102)

a. John regretted reading the novel.
b. John celebrated reading the novel.

(102a) and (102b) do entail that John actually completed reading the entire
novel. Their complements thus appear to perfective. Intriguingly, these perfective
regret-type verbs do not seem to allow complement coercion:
59

(103)

a. #John regretted the novel.
b. #John celebrated the novel.

(103a) is clearly unacceptable on a reading where John regrets having read the novel.
(103b) is only acceptable on a different reading, where John is celebrating a novels
existence (for instance, at an event celebrating the centennial anniversary of the
publication of some famous novel). It cannot mean that John is celebrating having
finished the novel.
In another parallel to finish-type verbs, enjoy-type verbs allow imperfective morphology:
(104)

a. John is enjoying reading the novel.
b. John is attempting reading the novel.

Our analysis predicted that two ingredients were necessary for complement coercion
to be possible for a given verb. First, the verb required an imperfective complement.
Second, the output of the verb had be able to combine with another imperfectivity
operator. It is highly suggestive that the enjoy-type verbs, which also appear to
allow complement coercion, have both of these properties. Verbs which lack the
fist of these properties, like regret, on the other hand, seem to resist complement
coercion. While a full treatment of enjoy-type and regret-type verbs is beyond the
scope of this thesis, it seems that their behavior with respect to complement coercion
is amenable to an analysis similar to that developed for finish-type and stop-type
AVs.

60

5.3

The Syntax Of AVs Revisited

Starting in 3.2 above, we assumed the syntactic structure in (53), repeated as (105)
below:
(105) [vP Johni [VP begin [AspP [Asp -ing] [vP PROi eat the apple]]]]
(105) is a sort of modified control structure. Unlike canonical control sentences, the
matrix AV embeds an AspP rather than a nonfinite TP. It turns out that, though,
in some of the earliest discussions of AV syntax, Perlmutter (1968, 1970) actually
proposed that AVs may be ambiguous between raising and control verbs.
For a raising verb analysis, he offered the following arguments. It is well known
that raising verbs like seem allow idioms to maintain their idiomatic meaning (106a),
exhibit synonymy between active and passive constructions (107), and allow expletive subjects (109a). Likewise, a finish-type verb like start also allows idiomatic
constructions (106b), exhibits the active and passive synonymy (108), and allows for
expletive subjects (109b):
(106)

a. The shit seemed to hit the fan.
b. The shit started hitting the fan.

(107)

a. The noise seemed to annoy Joe.
b. Joes seemed to be annoyed by the noise.

(108)

a. The noise started annoying Joe.
b. Joe started to be annoyed by the noise.

61

(109)

a. It seemed to rain.
b. It started to rain.

Stop-type verbs also seem to behave like raising verbs in these respects:
(110) The shit stopped hitting the fan.
(111)

a. The noise stopped annoying Joe.
b. Joe stopped being annoyed by the noise.

(112) It stopped raining.
In other respects, AVs fail to behave like raising verbs. For instance, raising verbs
may not appear in imperatives (113), and they may not be embedded under control
verbs (114):
(113) *Seem to eat the pie.
(114) *John forced Bill to seem to eat the pie.
AVs, on the other hand, allow both imperatives (115) and embedding under control
verbs (116):
(115) Start/Stop eating the pie.
(116) John forced Bill to start/stop eating the pie.
Moreover, some AVs, like finish, fail the tests for raising verbhood (106-109) that
other AVs pass:

62

(117) *The shit finished hitting the fan.
(118)

a. John finished eating the apple.
b. #The apple was finished being eaten by John.

(119) #It finished raining.
These considerations led Perlmutter (1968, 1970) to conclude that most AVs are
ambiguous between raising and control verbs, while some- such as finish- are just
control verbs.
It seems that Perlmutter’s observations could be integrated into our analysis without much difficulty. The crucial difference between the two, of course, is that raising
verbs lack an external argument, while control verbs do not. While these will cause
the denotations between raising and control AVs to differ slightly, it doesn’t look like
these sorts of distinctions would affect the core proposals of this thesis, particularly
with regard to the interactions between AVs and imperfectivity operators.
More recently, Fukuda (2008) has proposed that AVs are actually heads of functional projections within a single clause, rather than biclausal raising or control verbs.
Fukuda offers two arguments against a biclausal analysis of AVs. First, he notes that
biclausal constructions typically allow for a temporal discontinuity between the matrix and embedded clause, as one might see in (120):
(120) Today, John hoped to win the race tomorrow.
In (120), today modifies the matrix clause, and tomorrow the embedded control
clause. Following observations from Akmajian et al. (1979), Fukuda also notes that
biclausal construction often allow the grammatical aspect encoded in the embedded

63

clause to differ from that encoded in the matrix clause (examples from Akmajian
et al. (1979, 43,(125)):
(121)

a. We’ll try to make him [P rogressive be singing Coming through the Rye]
when. . .
b. I will try [P erf ect to have finished the work] by the time. . .

In (121), while the matrix clause has past perfective aspect, the embedded clause
has either imperfective (121a) or perfect (121b) aspect. Again, this provides evidence
that such constructions are biclausal.
Next, Fukuda (2008) observes that AVs do not allow these sorts of discontinuities.
For instance, the event of beginning cannot take place on a different day then the
event described in the complement:
(122) #Today, John began to bake the cake tomorrow.
Moreover, again following Akmajian et al. (1979), Fukuda observes that AVs do not
seem to allow complex aspectual structure in the complement (from Akmajian et al.
(1979, 40, (112))):
(123)

a. *He began [Progressive being running down the road].
b. *He began [Perfective to have finished his homework].

This considerations lead Fukuda to conclude that AV constructions are not biclausal.
Instead, he proposes that they actually head a functional projection. Specifically, he
proposes that they head AspP ( Travis 1991; Borer 1994, 1998)1 .
1 Fukuda (2008) argues that this AspP may appear either below or above vP. We will

ignore this aspect of his proposal since (i) it would seriously complicate the semantics and

64

There are some problems with Fukuda’s arguments for rejecting a biclausal analysis of AVs. For one, although (122) and (123) are indeed unacceptable, its not clear
whether the origin of this unacceptability is actually syntactic, as Fukuda claims, or
semantic. Indeed, attempt, which is commonly analyzed as taking control clauses as
its complements, also seems to forbid these sorts of temporal and aspectual discontinuities:
(124) #Today, John attempted to bake the cake tomorrow.
(125)

a. *He attempted [Progressive being running down the road].
b. *He attempted [Perfective to have finished his homework].

It seems likely that the unacceptability of (124) and (125) stem from semantic,
rather than syntactic, peculiarities of attempt. It seems quite possible that a semantic
explanation could also be found for the unacceptability of the sentences in (122) and
(123), particularly given our assumptions that the event encoded by the AV be a
sub-event of that encoded by the complement.
However, the argument of Fukuda (2012) for Japanese AVs being functional heads
is somewhat more persuasive, and the notion that at least some AVs are non-verbal
functional heads is an interesting one. One possibility which might be worth considering is that perhaps finish-type AVs are control verbs and stop-type verbs are
morphemes associated with a functional category, perhaps an A(spectual)V(erb)P.
(ii) his arguments for it involve a rather detailed discussion of infinitival complements,
which dont fall under the purview of this thesis. Fukuda (2012) extends this analysis to
AVs to Japanese.

65

The proposal would then be that finish has a syntactic structure like (126)2 and stop
a structure like (127):
(126) [vP Johni [VP finish [AspP –ing [vP PROi eat the cookie]]]]
(127) [AVP stop [AspP –ing [vP John eat the cookie]]]
Structures like (126) and (127) seem to accord with the differences in behavior we
have observed. Recall that, under the analysis in 4, finish-type AVs take in a property
of the RT and return a property of events. Such a type is quite strange within a
monoclausal syntax, but would be highly plausible in a structure like (126). The idea
would be that finish-type AVs take a property of the reference time associated with
the embedded clause and convert it into the ET of the event encoded by the matrix
clause. Stop-type verbs, on the other hand, are simply operators which modify the
RT in some fashion.
One consequence of the syntax proposed in (126) and (127) is that finish-type
verbs are predicted to be genuinely biclausal, while stop-type verbs are predicted
to be monoclausal. We might then predict different behavior with manner adverbs,
which modify a verbal predicate. Consider a sentences like (128):
(128) John cheerfully ate the cookie.
(128)’s natural interpretation is that Johns manner of eating the cookie was cheerful.
It therefore seems to be modifying the VP. A subject-oriented reading where it is
cheerful of John to eat the cookie seems somewhat infelicitous. Now consider (129)
and (130):
2 Or perhaps finish-type AVs are ordinary control verbs, selecting for TPs.

66

(129) John cheerfully finished eating the cookie.
(130) #John cheerfully stopped eating the cookie.
While the judgments are rather subtle, it does seem that the manner adverb
behaves appropriately in (129), describing a scenario where Johns finishing eating
the cookie was a cheerful event. In (130), though, the only available reading seems to
be a subject-oriented one, making the sentence sound rather odd. It is difficult to get
a reading where Johns stopping eating the cookie was a cheerful event. This seems
consistent with the hypothesis that finish-type verbs contain a VP which manner
adverbs can modify, but stop-type AVs lack such a level.
The judgments in (129) and (130) are suggestive that the syntax proposed in
(126) and (127) is on the right track, but such data are hardly conclusive. Perhaps
a more thorough exploration of the control literature will turn up more evidence
either for our against our proposal. In any case, this proposal at least has the virtues
of providing markedly different syntactic structures which seem to accord with the
semantic analysis presented in 4.
If this proposal holds up for English, then it would be interesting to see how it
could be extended cross-linguistically. In English, it seems to be the case that control
and raising verbs which select for imperfective predicates- verbs like finish and begin,
or enjoy and attempt - are also the verbs which allow for complement coercion. Does
this same generalization hold for other languages? If so, this would provide exciting
evidence that the semantic mechanisms that this thesis has argued to be responsible
for complement coercion in English actually hold across natural language.

67

5.4

Towards A Sortal Analysis Of AVs

Certain predicates have an intuitive quality of punctuality. One might think of the
achievement verbs, or of certain accomplishments like slam the door or reach the
mountain top. A diagnostic for this might be found by using the adverbial at 3:00
which denotes a single time-point, as we see in (131):
(131)

a. John slammed the door at 3:00 pm.
b. John reached the mountain top at 3:00 pm.
c. #John built the house at 3:00 pm.
d. #John climbed Mount Everest at 3:00 pm.

The aspectual verbs have a similar instantaneous feel to them, as we can see in (132):
(132) John began/finished/stopped building the house at 3:00 pm.
However, there is a difference, I think, between the way in which the two verb classes
come about their instantaneousness. To try and make it explicit, we might first
note that events have natural start-points and end-points. There is a period when
an event is in progress and periods before and after in which it is not in progress.
Even atelic events have these boundaries if we assume, following Kratzer (2002) that
some possibly covert measure phrase closes off the events of atelic predicates. Begin
and finish, in essence, exploit these natural start-points and end-points of events.
Verbs like resume do something similar, essentially by imposing a new starting-point
(resume means something like re-begin).
Stop-type don’t have this luxury. To say that John stopped reading the book or
John kept reading the book is not to refer to either the end or beginning of the read
68

the book event. Instead, these sorts of verbs must look inside of an event and identify
specific instantaneous sub-events within. Punctuality does not come for free, as it
does with verbs of the finish-type, but must be imposed.
It also seems to be the case that AVs carry certain presuppositions. This is
highlighted by Groucho Marx’s famous- and rather misogynistic- joke:
(133) Have you stopped beating your wife yet?
The joke, of course, is that no matter how you answer (133), it will always presuppose
that you have been beating your wife. A less offensive sentence, using a different
test, yields a similar result:
(134) It is not the case that Dave stopped eating the cookie.
In (134), it is still the case that Dave was eating the cookie in the past, indicating
presupposition-hood. The reader will also note that these tests yield comparable
results when applied to the other AVs.
Examine these tests a little more closely, however, and well notice a subtle but
crucial difference holds between the presuppositions supplied by the two verb classes.
Consider again the negation test:
(135)

a. It is not the case that Dave stopped eating the cookie.
b. It is not the case that Dave kept eating the cookie.
c. It is not the case that Dave began eating the cookie.
d. It is not the case that Dave finished eating the cookie.

In (135a) and (135b), it is true that, even if Dave did not stop or keep eating
the cookie, he had been eating the cookie at the time immediately before the time
69

in question. Conversely, (135c) presupposes that Dave had not been eating the
cookie at any time before the time in question, and (135d) merely presupposes that
converse: that he had been eating the cookie at some point before the time in
question. So, while all four constructions presuppose something about the property
of Dave’s eating the cookie, the stop-type verbs appear to restrict the presupposition
much more severely than the finish-type verbs.
These observations suggest that the difference between the two verb classes may
not be between different types, but instead between different sorts of intervals. A
verb like begin, for instance, asserts that a particular property holds of the reference
time, and presupposes that no previous times hold that property. It thus marks
the boundary between the reference time, when some property of times starts to
be true, and all previous times when it was not true. Similarly, finish presupposes
that some property of times has been true before the reference time, and asserts
that following the reference time it will now longer be true. The stop-type verbs, on
the other hand, merely relate a property of times to particular time-points. Stop, for
instance, presupposes that a property of times was true at the time-point immediately
preceding the reference time-point, and that it is false immediately after. Keep
also presupposes that a property of times was true at the time-point immediately
preceding the reference time-point, and asserts that it was true immediately after.
In short, the stop-type verbs return properties of time-points, while the finish-type
verbs do not.
This, if correct, suggests that the source of the stop-type verbs incompatibility with imperfectivity operators may arise because the former return properties of

70

time-points, which by definition cannot be divided into subintervals. The job of an
imperfectivity operator, on the other hand, is to pick out a subinterval from some
interval. Thus the two are inherently incompatible. Making this observation more
formally explicit, and integrating it into the analysis developed above, will require
looking more thoroughly into the literature on tense and aspect and the linguistic
encoding of time. However, it seems that such a project would result in a theory of
AVs with even greater explanatory power.

71

BIBLIOGRAPHY

72

BIBLIOGRAPHY

Akmajian, A., Steele, S., and Wasow, T. (1979). The category aux in universal
grammar. Linguistic Inquiry, 10:1–64.
Asher, N. and Morreau, M. (1991). Commonsense entailment: A modal theory of
non-monotonic reasoning. The 12th International Joint Conference On Artificial
Intelligence (IJCAL-91).
Carlson, G. (1977). Reference To Kinds In English. PhD thesis, University Of
Massachusetts, Amherst.
Comrie, B. (1976). Aspect: An Introduction To The Study Of Verbal Aspect And
Related Problems. Cambridge University Press, Cambridge, UK.
Copestake, A. and Briscoe, T. (1992). Lexical operations in a unification-based
framework. In Lexical Semantics And Knowledge Representation: Proceedings of
the First SIGLEX Workshop, pages 101–119, Berlin. Springer-Verlag.
Copestake, A. and Briscoe, T. (1995). Semi-productive polysemy snd sense extension.
Journal Of Semantics, 12(1):15–67.
Deo, A. (2011). Imperfective readings: Partitions as quantificational domains. In
Proceedings of SALT 19, volume 19. SALT.
Dowty, D. (1979). Word Meaning And Montague Grammar. Reidel, Dordrecht.
Emonds, J. (1973). Alternatives to global constraints. Glossa, 7:39–62.
Fodor, J. and Lepore, E. (1998). The emptiness of the lexicon: Reflections on james
pustejovsky’s The Generative Lexicon. Linguistic Inquiry, 29(2):269–288.
Fukuda, S. (2008). Two syntactic positions for english aspectual verbs. WFCCFL
26.
Fukuda, S. (2012). Aspectual verbs as functional heads: Evidence from japanese
aspectual verbs. Natural Language And Linguistic Theory, In Press.
Godard, D. and Jayez, J. (1993). Towards a proper treatment of coercion phenomena.
Sixth Conference Of the European Chapter of the Association of Computational
Linguistics.
73

Jaeggli, O. (1980). Remarks on To contraction. Linguistic Inquiry, 11(1):239–245.
Klein, W. (1994). Time In Language. Routledge, London.
Kratzer, A. (1995). Stage-level and individual-level predicates. In Carlson, G. and
Pelletier, F. J., editors, The Generic Book, pages 125–175. University of Chicago
Press, Chicago.
Kratzer, A. (1996). Severing the external argument from its verb. In Rooryck, J.
and Zaring, L., editors, Phrase Structure And The Lexicon, pages 109–137. Kluwer,
Dordrecht.
Kratzer, A. (1998). More structural analogies between pronouns and tenses. Proceedings of SALT VIII.
Kratzer, A. (2002). Telicity and the meaning of objective case. In Gu´ron, J. and
e
Lecarme, J., editors, The Syntax Of Time. MIT Press, Cambridge, MA.
Lascarides, A. and Asher, N. (1993). Temoral interpretation, discourse relations, and
commonsense entailment. Linguistics And Philosophy, 16(5):437–493.
Lascarides, A. and Copestake, A. (1998). Pragmatics and word meaning. Journal
Of Linguistics, 34.
Lascarides, A., Copestake, A., and Briscoe, E. (1996). Ambiguity and coherence.
Journal Of Semantics, 13(2):41–65.
Link, G. (1987). Algebraic semantics of event structures. In Groenendijk, J., stokhof,
M., and Veltman, F., editors, Proceedings Of The Sixth Amsterdam Colloquium,
pages 243–272. ITLI, University Of Amsterdam, Amsterdam.
Milsark, G. (1972). Re:doubl-ing. Linguistic Inquiry, 3(4):542–549.
Milsark, G. (1988). Singl-ing. Linguistic Inquiry, 19(4):611–634.
Munn, A. (1993). Topics In The Syntax And Semantics Of Coordinate Structures.
PhD thesis, University Of Maryland, College Park.
Newmeyer, F. (1975). English Aspectual Verbs. Mouton, The Hauge.
Perlmutter, D. (1968). Deep And Surface Structure Constraints In Syntax. PhD
thesis, MIT, Cambridge, MA.

74

Perlmutter, D. (1970). The two verbs begin. In Jacobs, R. A. and Rosenbaum, P. S.,
editors, Readings In English Transformational Grammar, pages 107–119. Blaisdell,
Waltham, MA.
Portner, P. (1992). Situation Theory And The Semantics Of Propositional Expressions. PhD thesis, University Of Massachusetts, Amherst.
Pullum, G. and Zwicky, A. (1999). Gerund participles and head-complement inflection conditions. In Collins, P. and Lee, D., editors, The Clause In English: In
Honour Of Rodney Huddleston. John Benjamins Publishing Company, Amsterdam/Phildadelphia.
Pustejovsky, J. (1991).
17(4):409–441.

The generative lexicon.

Computational Linguistics,

Pustejovsky, J. (1993). Type coercion and lexical selection. In Pustejovsky, J., editor,
Semantics And The Lexicon. Kluwer Academic Publishers.
Pustejovsky, J. (1995). The Generative Lexicon. MIT Press, Cambridge, MA.
Pylkk¨nen, L. and McElree, B. (2006). The syntax-semantics interface: On-line
a
composition of sentence meaning. In Traxler, M. and Gernsbacher, M., editors,
Handbook Of Psycholinguistics. Elsevier, New York.
Pylkk¨nen, L. and McElree, B. (2007). An meg study of silent meaning. Journal Of
a
Cognitive Neuroscience, 19:1905–1921.
Reichenbach, H. (1947). Elements Of Symbolic Logic. Macmillan, New York.
Ross, J. (1972). Doubl-ing. Linguistic Inquiry, 3(1):61–86.
Sag, I., Gazdar, G., Wasow, T., and Weisler, S. (1985). Coordination and how to
distinguish categories. Natural Language And Linguistic Theory, 3:117–171.
Taylor, B. (1977). Tense and continuity. Linguistics And Philosophy, 1(2):199–220.
ter Meulen, A. (1985). Progressives without possible worlds. Proceedings Of The
Twenty First Annual Meeting Of The Chicago Linguistic Society.
Zucchi, S. (1999). Incomplete events, intensionality, and imperfective aspect. Natural
Language Semantics, 7:179–215.

75