c© 2003 Association for Computational Linguistics
A Probabilistic Account of Logical
Metonymy
Maria Lapata
∗
Alex Lascarides
†
University of Sheffield University of Edinburgh
In this article we investigate logical metonymy, that is, constructions in which the argument
of a word in syntax appears to be different from that argument in logical form (e.g., enjoy the
book means enjoy reading the book, and easy problem means a problem that is easy to solve).
The systematic variation in the interpretation of such constructions suggests a rich and complex
theory of composition on the syntax/semantics interface. Linguistic accounts of logical metonymy
typically fail to describe exhaustively all the possible interpretations, or they don’t rank those
interpretations in terms of their likelihood. In view of this, we acquire the meanings of metonymic
verbs and adjectives from a large corpus and propose a probabilistic model that provides a ranking
on the set of possible interpretations. We identify the interpretations automatically by exploiting
the consistent correspondences between surface syntactic cues and meaning. We evaluate our
results against paraphrase judgments elicited experimentally from humans and show that the
model’s ranking of meanings correlates reliably with human intuitions.
1. Introduction
Much work in lexical semantics has been concerned with accounting for regular poly-
semy, that is, the regular and predictable sense alternations to which certain classes of
words are subject (Apresjan 1973). It has been argued that in some cases, the different
interpretations of these words must arise from the interaction between the semantics
of the words during syntactic composition, rather than by exhaustively listing all the
possible senses of a word in distinct lexical entries (Pustejovsky 1991). The class of
phenomena that Pustejovsky (1991, 1995) has called logical metonymy is one such
example. In the case of logical metonymy additional meaning arises for particular
verb-noun and adjective-noun combinations in a systematic way: the verb (or adjec-
tive) semantically selects for an event-type argument, which is a different semantic
type from that denoted by the noun. Nevertheless, the value of this event is pre-
dictable from the semantics of the noun. An example of verbal logical metonymy is
given in (1) and (2): (1a) usually means (1b) and (2a) usually means (2b).
(1) a. Mary finished the cigarette.
b. Mary finished smoking the cigarette.
(2) a. Mary finished her beer.
b. Mary finished drinking her beer.
∗ Department of Computer Science, University of Sheffield, Regent Court, 211 Portobello Street, Sheffield
S1 4DP, UK. E-mail: mlap@dcs.shef.ac.uk.
† School of Informatics, University of Edinburgh, 2 Buccleuch Place, Edinburgh EH8 9LW, UK. E-mail:
alex@inf.ed.ac.uk.
262
Computational Linguistics Volume 29, Number 2
Note how the events in these examples correspond to the purpose of the object denoted
by the noun: the purpose of a cigarette is to smoke it and the purpose of a beer is to
drink it. Similarly, (3a) means a problem that is easy to solve, (3b) means a language
that is difficult to learn, speak, or write, (3c) means a cook that cooks well, (3d) means
a soup that tastes good, (3e) means someone who programmes fast, and (3f) means a
plane that flies quickly.
(3) a. easy problem
b. difficult language
c. good cook
d. good soup
e. fast programmer
f. fast plane
The interpretations of logical metonymies can typically be rendered with a para-
phrase, as we have indicated for the above examples. Verb-nouns are paraphrased
with a progressive or infinitive VP that is the complement of the polysemous verb
(e.g., smoking in (1b)) and whose object is the NP figuring in the verb-noun combina-
tion (e.g., cigarette in (1b)). Adjective-noun combinations are usually paraphrased with
a verb modified by the adjective in question or its corresponding adverb. For example,
an easy problem is a problem that is easy to solve or a problem that one can solve easily
(see (3a)).
Logical metonymy has been extensively studied in the lexical semantics literature.
Previous approaches have focused on descriptive (Vendler 1968) or theoretical (Puste-
jovsky 1991, 1995; Briscoe, Copestake, and Boguraev 1990) accounts, on the linguistic
constraints on the phenomenon (Godard and Jayez 1993; Pustejovsky and Bouillon
1995; Copestake and Briscoe 1995; Copestake 2001), and on the influence of discourse
context on the interpretation of metonymies (Briscoe, Copestake, and Boguraev 1990;
Lascarides and Copestake 1998; Verspoor 1997). McElree et al. (2001) investigated the
on-line processing of metonymic expressions; their results indicate that humans dis-
play longer reading times for sentences like (1a) than for sentences like (1b).
There are at least two challenges in providing an adequate account of logical
metonymy. The first concerns semi-productivity: There is a wealth of evidence that
metonymic constructions are partially conventionalized, and so resolving metonymy
entirely via pragmatic reasoning (e.g., by computing the purpose of the object that
is denoted by the noun according to real-world knowledge) will overgenerate the
possible interpretations (Hobbs et al. 1993). For example, the logical metonymies in
(4) are odd, even though pragmatics suggests an interpretation (because real-world
knowledge assigns a purpose to the object denoted by the NP):
(4) a. ?John enjoyed the dictionary.
b. ?John enjoyed the door.
c. ?John began/enjoyed the highway.
d. ?John began the bridge.
Sentence (4a) is odd because the purpose of dictionaries is to refer to them, or to
consult them. These are (pointlike) achievements and cannot easily combine with en-
joy, which has to be true of an event with significant duration. Domain knowledge
assigns doors, highways, and bridges a particular purpose, and so the fact that the
sentences in (4b)–(4d) are odd indicates that metonymic interpretations are subject to
conventional constraints (Godard and Jayez 1993).
263
Lapata and Lascarides Logical Metonymy
The second challenge concerns the diversity of possible interpretations of metony-
mic constructions. This diversity is attested across and within metonymic construc-
tions. Metonymic verbs and adjectives are able to take on different meanings depend-
ing on their local context, namely, the noun or noun class they select as objects (in the
case of verbs) or modify (in the case of adjectives). Consider the examples in (1), in
which the meaning of the verb finish varies depending on the object it selects. Simi-
larly, the adjective good receives different interpretations when modifying the nouns
cook and soup (see (3c) and (3d)).
Although we’ve observed that some logical metonymies are odd even though
pragmatics suggests an interpretation (e.g., (4c)), Vendler (1968) acknowledges that
other logical metonymies have more than one plausible interpretation. In order to
account for the meaning of adjective-noun combinations, Vendler (1968, page 92) points
out that “in most cases not one verb, but a family of verbs is needed”. For example,
fast scientist can mean a scientist who does experiments quickly, publishes quickly, and
so on.
Vendler (1968) further observes that the noun figuring in an adjective-noun combi-
nation is usually the subject or object of the paraphrasing verb. Although fast usually
triggers a verb-subject interpretation (see (3e) and (3f)), easy and difficult trigger verb-
object interpretations (see (3a) and (3b)). An easy problem is usually a problem that one
solves easily (so problem is the object of solve), and a difficult language is a language that
one learns, speaks, or writes with difficulty (so language is the object of learn, speak,
and write). Adjectives like good allow either verb-subject or verb-object interpretations:
a good cook is a cook who cooks well, whereas good soup is a soup that tastes good.
All of these interpretations of fast scientist, difficult language,orgood soup seem highly
plausible out of context, though one interpretation may be favored over another in a
particular context. In fact, in sufficiently rich contexts, pragmatics can even override
conventional interpretations: Lascarides and Copestake (1998) suggest that (5c) means
(5d) and not (5e):
(5) a. All the office personnel took part in the company sports day last
week.
b. One of the programmers was a good athlete, but the other was
struggling to finish the courses.
c. The fast programmer came first in the 100m.
d. The programmer who runs fast came first in the 100m.
e. The programmer who programs fast came first in the 100m.
The discourse context can also ameliorate highly marked logical metonymies, such
as (4c):
(6) a. John uses two highways to get to work every morning.
b. He first takes H-280 and then H-101.
c. He always enjoys H-280,
d. but the traffic jams on H-101 frustrate him.
Arguably the most influential account of logical metonymy is Pustejovsky’s
(1991, 1995) theory of the generative lexicon. Pustejovsky avoids enumerating the
various senses for adjectives like fast and verbs like finish by exploiting a rich lexical
semantics for nouns. The lexical entry for an artifact-denoting noun includes a qualia
structure: this specifies key features of the word’s meaning that are in some sense
264
Computational Linguistics Volume 29, Number 2
derivable from real-world knowledge but are lexicalized so as to influence conven-
tional processes. The qualia structure includes a telic role (i.e., the purpose of the object
denoted by the noun) and an agentive role (i.e., the event that brought the object into
existence). Thus the lexical entry for book includes a telic role with a value equivalent
to read and an agentive role with a value equivalent to write, whereas for cigarette the
telic role is equivalent to smoke and the agentive role is equivalent to roll or manufacture.
When finish combines with an object-denoting NP, a metonymic interpretation
is constructed in which the missing information is provided by the qualia structure
of the NP. More technically, semantic composition of finish with cigarette causes the
semantic type of the noun to be coerced into its telic event (or its agentive event),
and the semantic relation corresponding to the metonymic verb (finish) predicates
over this event. This results in an interpretation of (1a) equivalent to (1b). Verbs like
begin and enjoy behave in a similar way. Enjoy the book can mean enjoy reading the
book, because of book’s telic role, or enjoy writing the book, because of book’s agentive
role. In fact, the agentive reading is less typical for book than the telic one, but for
other nouns the opposite is true. For instance, begin the tunnel can mean begin building
the tunnel, but the interpretation that is equivalent to begin going through the tunnel
is highly marked. There is also variation in the relative likelihood of interpretations
among different metonymic verbs. (We will return to this issue shortly.) The adjective-
noun combinations are treated along similar lines. Thus the logical polysemy of words
like finish and fast is not accounted for by exhaustive listing.
1
In contrast to the volume of theoretical work on logical metonymy, very little
empirical work has tackled the topic. Briscoe et al. (1990) investigate the presence of
verbal logical metonymies in naturally occurring text by looking into data extracted
from the Lancaster-Oslo/Bergen corpus (LOB, one million words). Verspoor (1997)
undertakes a similar study in the British National Corpus (BNC, 100 million words).
Both studies investigate how widespread the use of logical metonymy is, and how
far the interpretation for metonymic examples can be recovered from the head noun’s
qualia structure, assuming one knows what the qualia structure for any given noun is.
Neither of these studies is concerned with the automatic generation of interpretations
for logical metonymies and the determination of their likelihood.
Although conceptually elegant, Pustejovsky’s (1995) theory of the generative lex-
icon does not aim to provide an exhaustive description of the telic roles that a given
noun may have. However, these roles are crucial for interpreting verb-noun and
adjective-noun metonymies. In contrast to Vendler (1968), who acknowledges that
logical metonymies may trigger more than one interpretation (in other words, that
there may be more than one possible event associated with the noun in question),
Pustejovsky implicitly assumes that nouns or noun classes have one (perhaps default)
telic role without, however, systematically investigating the relative degree of ambigu-
ity of the various cases of logical metonymy (e.g., the out-of-context possible readings
for fast scientist suggest that fast scientist exhibits a higher degree of semantic ambiguity
than fast plane). One could conceivably represent this by the generality of the semantic
type of the telic role in the various nouns (e.g., assign the telic role of scientist a rel-
atively general type of event compared with that for plane). But this simply transfers
the problem: The degree of generality in lexical representation is highly idiosyncratic
and ideally should be acquired from linguistic evidence; furthermore, for nouns with
1 Other lexical accounts, such as Copestake and Briscoe (1995), differ from Pustejovsky’s (1995) in that
the coercion is treated as internal to the semantics of the metonymic verb or adjective rather than the
noun; motivation for this comes from copredication data, such as the acceptability of fast and intelligent
typist and John picked up and finished his beer.
265
Lapata and Lascarides Logical Metonymy
a general telic role, pragmatics would have to do ‘more work’ to augment the gen-
eral interpretation with a more specific one. Even in theories in which more than one
interpretation is provided (see Vendler 1968), no information is given with respect to
the relative likelihood of these interpretations.
Pustejovsky’s account also doesn’t predict the degree of variation of interpretations
for a given noun among the different metonymic verbs: for example, the fact that begin
the house is, intuitively at least, more likely to resolve to an agentive-role interpretation
(i.e., begin building the house) than a telic-role interpretation (i.e., begin living in the
house), whereas the reverse is true of enjoy the house. Ideally, we would like a model
of logical metonymy that reflects this variation in interpretation.
In this article we aim to complement the theoretical work on the interpretation
of logical metonymy by addressing the following questions: (1) Can the meanings
of metonymic adjective-noun and verb-noun combinations be acquired automatically
from corpora? (2) Can we constrain the number of interpretations of these combina-
tions by providing a ranking on the set of possible meanings? (3) Can we determine
whether a particular adjective has a preference for a verb-subject or a verb-object in-
terpretation? We provide a probabilistic model that uses distributional information
extracted from a large corpus to interpret logical metonymies automatically without
recourse to pre-existing taxonomies or manually annotated data.
The differences among the various theoretical accounts—for example, that Copes-
take and Briscoe (1995) treat type coercion as internal to the metonymic word, whereas
Pustejovsky (1995) treats it as part of the noun—do not matter for our purposes, be-
cause we aim to provide information about metonymic interpretations that is compati-
ble with either account. More specifically, we are concerned with using a real-language
corpus to acquire automatically the semantic value of the event that is part of the inter-
pretation. We abstract away from theoretical concepts such as semantic type coercion
and instead utilize co-occurrence frequencies in the corpus to predict metonymic inter-
pretations. Very roughly, we acquire a ranked set of interpretations enjoy V-ing the book
for the construction enjoy the book by estimating the probabilities that V is enjoyed and
that it is something done to books; and we estimate these probabilities on the basis of
the corpus frequencies for V’s appearing as a (verbal) complement to enjoy and for V’s
taking book as its object. Similarly, we acquire a ranked set of verb-subject interpreta-
tions of fast plane by estimating the likelihood of seeing the plane Vs and Vs quickly in
the corpus. (See Sections 2 and 3 for more details and motivation of these models.)
Our results show not only that we can predict meaning differences when the same
adjective or verb is associated with different nouns, but also that we can derive—taking
into account Vendler’s (1968) observation—a cluster of meanings for a single verb- or
adjective-noun combination. We can also predict meaning differences for a given noun
associated with different metonymic verbs and adjectives. We evaluate our results by
comparing the model’s predictions to human judgments and show that the model’s
ranking of meanings correlates reliably with human intuitions.
However, the model is limited in its scope. It is suited for the interpretation of well-
formed metonymic constructions. But it does not distinguish odd metonymies (see (4))
from acceptable ones: in both cases, paraphrases will be generated, at least in principle
(see Section 2.1 for explanation and motivation). In particular, the model does not
learn conventional constraints, such as that enjoy must take an event of duration as its
argument (Godard and Jayez 1993). However, such constraints are potentially captured
indirectly: If the above conventional constraint is right, enjoy referring to should not
be attested in the corpus, and hence according to our model it won’t be part of a
possible paraphrase for enjoy the dictionary (see Sections 2.4.2 and 2.5.3 for further
discussion). Further, since the model abstracts away from semantic type coercion, it
266
Computational Linguistics Volume 29, Number 2
does not distinguish between uses of a verb or adjective that are claimed in the lexical
semantics literature to be metonymic uses (e.g., enjoy the book, fast programmer) and
those that are claimed to be nonmetonymic uses (e.g., enjoy the marriage, fast run-time).
Again, the model of interpretation presented here will generate paraphrases for all
these cases (e.g., it will paraphrase enjoy the marriage as enjoy going to or participating
in the marriage and fast run-time as run-time that goes by or passes quickly). The model
also does not take discourse context into account; for example, it will not predict the
intuitive interpretation of (5e). Rather, it determines the most dominant meanings for
a given metonymic construction overall, across all of the instances of it in the corpus.
The remainder of this article is organized as follows: in the first part (Section 2) we
present our probabilistic model of verbal logical metonymy and describe the model
parameters. In Experiment 1 we use the model to derive the meaning paraphrases for
verb-noun combinations randomly selected from the BNC (see Section 2.3) and for-
mally evaluate our results against human intuitions. Experiment 2 demonstrates that
when compared against human judgments, our model outperforms a naive baseline
in deriving a preference ordering for the meanings of verb-noun combinations, and
Experiment 3 evaluates an extension of the basic model. In the second part (Section 3),
we focus on adjectival logical metonymy. Section 3.1 introduces our probabilistic for-
malization for polysemous metonymic adjectives and Sections 3.3–3.6 present our ex-
periments and evaluate our results. Overall, the automatically acquired model of log-
ical metonymy reliably correlates with human intuitions and also predicts the relative
degree of ambiguity and acceptability of the various metonymic constructions. In Sec-
tion 4 we discuss our results, we review related work in Section 5, and we conclude
in Section 6.
2. Metonymic Verbs
2.1 The Model
Consider the verb-noun combinations in (7) and (8). Our task is to come up with
(7b) and (8b) as appropriate interpretations for (7a) and (8a). Although the interpreta-
tions of (7a) and (8a) are relatively straightforward for English speakers, given their
general knowledge about coffees and films and the activities or events associated
with them, a probabilistic model requires detailed information about words and their
interdependencies in order to generate the right interpretation. Examples of such in-
terdependencies are verbs co-occurring with coffee (e.g., drink, make, prepare) or verbs
that are related to begin (e.g., make, realize, understand).
(7) a. John began the coffee.
b. John began drinking the coffee.
(8) a. Mary enjoyed the film.
b. Mary enjoyed watching the film.
A relatively straightforward approach to the interpretation of (7a) and (8a) would
be to extract from the corpus (via parsing) paraphrases in which the additional in-
formation (e.g., drinking and watching), which is absent from (7a) and (8a), is fleshed
out. In other words, we would like to find in the corpus sentences whose main verb is
begin followed either by the progressive VP complement drinking or by the infinitive to
drink, selecting for the NP coffee as its object. In the general case we would like to find
the activities or events related both to the verb begin and the noun coffee (e.g., drinking,
buying, making, preparing). Similarly, in order to paraphrase (8a) we need information
267
Lapata and Lascarides Logical Metonymy
about the VP complements that are associated with enjoy and can take film as their
object (e.g., watching, making, shooting).
The above paraphrase-based model is attractive given its simplicity: All we need
to do is count the co-occurrences of a verb, its complements, and their objects. The
approach is unsupervised, no manual annotation is required, and no corpus-external
resources are used. Such a model relies on the assumption that the interpretations
of (7a) and (8a) can be approximated by their usage; that is, it assumes that the likeli-
hood of uttering the metonymic construction is equal to that of uttering its interpre-
tation. However, this assumption is not borne out. Only four sentences in the BNC
are relevant for the interpretation of begin coffee (see (9)); likewise, four sentences are
relevant for the interpretation of enjoy film (see (10)).
(9) a. Siegfried bustled in, muttered a greeting and began to pour his
coffee.
b. She began to pour coffee.
c. Jenna began to serve the coffee.
d. Victor began dispensing coffee.
(10) a. I was given a good speaking part and enjoyed making the film.
b. He’s enjoying making the film.
c. Courtenay enjoyed making the film.
d. I enjoy most music and enjoy watching good films.
e. Did you enjoy acting alongside Marlon Brando in the recent
film The Freshman?
The attested sentences in (9) are misleading if they are taken as the only evidence
for the interpretation of begin coffee, for on their own they suggest that the most likely
interpretation for begin coffee is begin to pour coffee, whereas begin to serve coffee and begin
dispensing coffee are less likely, as they are attested in the corpus only once. Note that
the sentences in (9) fail to capture begin to drink coffee as a potential interpretation for
begin coffee. On the basis of the sentences in (10), enjoy making the film is the most likely
interpretation for (8a), whereas enjoy watching the film and enjoy acting in the film are
equally likely.
This finding complies with Briscoe, Copestake, and Boguraev’s (1990) results for
the LOB corpus: begin V NP is very rare when the value of V corresponds to a highly
plausible interpretation of begin NP. Indeed, one can predict that problems with finding
evidence for begin V NP will occur on the basis of Gricean principles of language
production, where the heuristic be brief (which is part of the maxim of manner) will
compel speakers to utter begin coffee as opposed to begin V coffee if V is one of the
plausible interpretations of begin coffee. Thus on the basis of this Gricean reasoning,
one might expect metonymies like (7a) and (8a) to occur with greater frequencies than
their respective paraphrases (see (7b) and (8b)). Tables 1–3 show BNC counts of verb-
noun metonymies (commonly cited in the lexical semantics literature (Pustejovsky
1995; Verspoor 1997)) and their corresponding interpretations when these are attested
in the corpus. The data in Tables 1–3 indicate that metonymic expressions are more
often attested in the BNC with NP rather than with VP complements.
The discrepancy between an interpretation and its usage could be circumvented
by using a corpus labeled explicitly with interpretation paraphrases. Lacking such a
corpus, we will sketch below an approach to the interpretation of metonymies that
retains the simplicity of the paraphrase-based account but no longer assumes a tight
correspondence between a metonymic interpretation and its usage. We present an
268
Computational Linguistics Volume 29, Number 2
Table 1
BNC frequencies for begin.
Examples begin NP begin V-ing NP
begin book 35 17
begin sandwich 4 0
begin beer 2 1
begin speech 21 4
begin solo 1 1
begin song 19 8
begin story 31 15
Table 2
BNC frequencies for enjoy.
Examples enjoy NP enjoy V-ing NP
enjoy symphony 34 30
enjoy movie 5 1
enjoy coffee 8 1
enjoy book 23 9
like movie 18 3
Table 3
BNC frequencies for want.
Examples want NP want V-ing NP
want cigarette 18 3
want beer 15 8
want job 116 60
unsupervised method that generates interpretations for verbal metonymies without
recourse to manually annotated data or taxonomic information; it requires only a
part-of-speech-tagged corpus and a partial parser.
We model the interpretation of a verbal metonymy as the joint distribution P(e, o, v)
of three variables: the metonymic verb v (e.g., enjoy), its object o (e.g., film), and the
sought-after interpretation e (e.g., making, watching, directing). By choosing the ordering
〈e, v, o〉 for the variables e, v, and o, we can factor P(e, o, v) as follows:
P(e, o, v)=P(e)· P(v | e)· P(o | e, v)(11)
The probabilities P(e), P(v | e), and P(o | e, v) can be estimated using maximum likeli-
hood as follows:
ˆ
P(e)=
f(e)
N
(12)
ˆ
P(v | e)=
f(v, e)
f(e)
(13)
ˆ
P(o | e, v)=
f(o, e, v)
f(e, v)
(14)
269
Lapata and Lascarides Logical Metonymy
Table 4
Most frequent complements of enjoy and film.
f(enjoy, e) f(ﬁlm, e)
play 44 make 176
watch 42 be 154
work with 35 see 89
read 34 watch 65
make 27 show 42
see 24 produce 29
meet 23 have 24
go to 22 use 21
use 17 do 20
take 15 get 18
Although P(e) and P(v | e) can be estimated straightforwardly from a corpus (f(e)
amounts to the number of the times a given verb e is attested, N is the number of
verbs found in the corpus (excluding modals and auxiliaries), and P(v | e) can be ob-
tained through parsing, by counting the number of times a verb v takes e as its comple-
ment), the estimation of P(o | e, v) is problematic. It presupposes that co-occurrences of
metonymic expressions and their interpretations are to be found in a given corpus, but
as we’ve seen previously, there is a discrepancy between a metonymic interpretation
and its usage. In fact, metonymies occur more frequently than their overt interpreta-
tions (expressed by the term f(o, e, v) in (14)), and the interpretations in question are not
explicitly marked in our corpus. We will therefore make the following approximation:
P(o | e, v) ≈ P(o | e)(15)
ˆ
P(o | e)=
f(o, e)
f(e)
(16)
The rationale behind this approximation is that the likelihood of seeing a noun o as the
object of an event e is largely independent of whether e is the complement of another
verb. In other words, v is conditionally independent of e, since the likelihood of o is
(largely) determined on the basis of e and not of v. Consider again example (8a): Mary
enjoyed the film. Here, film, the object of enjoy, is more closely related to the underspeci-
fied interpretation e rather than to enjoy. For example, watching movies is more likely
than eating movies, irrespective of whether Mary enjoyed or liked watching them.
We estimate P(o | e) as shown in (16). The simplification in (15) results in a compact
model with a relatively small number of parameters that can be estimated straightfor-
wardly from the corpus in an unsupervised manner. By substituting equations (12),
(13), and (16) into (11) and simplifying the relevant terms, (11) can be rewritten as
follows:
P(e, o, v)=
f(v, e)· f(o, e)
N · f(e)
(17)
Assume we want to generate meaning paraphrases for the verb-noun pair enjoy
film (see (8a)). Table 4 lists the most frequent events related to the verb enjoy and the
most frequent verbs that take film as their object (we describe how the frequencies
f(v, e) and f(o, e) were obtained in the following section). We can observe that seeing,
watching, making, and using are all events associated with enjoy and with film and will
270
Computational Linguistics Volume 29, Number 2
be therefore generated as likely paraphrases for the metonymic expression enjoy film
(see Table 4, in which the underlined verbs indicate common complements between
the metonymic verb and its object).
Note that the model in (17) does not represent the fact that the metonymic verb v
may have a subject. This in practice means that the model cannot distinguish between
the different readings for (18a) and (18b): in (18a) the doctor enjoyed watching the film,
whereas in (18b) the director enjoyed making or directing the film. The model in (17)
will generate the set of events that are associated with enjoying films (e.g., watching,
making, seeing, going to), ignoring the contribution of the sentential subject. We present
in Section 2.5.1 an extension of the basic model that takes sentential subjects into
account.
(18) a. The doctor enjoyed the film.
b. The director enjoyed the film.
It is important to stress that the probabilistic model outlined above is a model of the
interpretation rather than the grammaticality of metonymic expressions. In other words,
we do not assume that it can distinguish between well-formed and odd metonymic
expressions (see the examples in (4)). In fact, it will generally provide a set of interpre-
tation paraphrases, even for odd formulations. The model in (11) has no component
that corresponds to the occurrence of v and o together. Choosing the ordering 〈o, v, e〉
for the variables o, e, and v would result in the following derivation for P(e, o, v):
P(e, o, v)=P(o)· P(v | o)· P(e | o, v)(19)
The term P(v | o) in (19) explicitly takes into account the likelihood of occurrence of
the metonymic expression. This means that no interpretation will be provided for odd
metonymies like enjoy the highway as long as they are not attested in the corpus. Such a
model penalizes, however, well-formed metonymies that are not attested in the corpus.
A striking example is enjoy the ice cream, which is a plausible metonymy not attested
at all in the BNC and thus by (19) would be incorrectly assigned no interpretations.
This is because the maximum-likelihood estimate of P(v | o) relies on the co-occurrence
frequency f(v, o), which is zero for enjoy the ice cream. But the probabilistic model in (11)
will generate meaning paraphrases for metonymic verb-object pairs that have not been
attested in the corpus as long as the co-occurrence frequencies f(v, e) and f(o, e) are
available.
Finally, note that our model is ignorant with respect to the discourse context
within which a given sentence is embedded. This means that it will come up with
the same ranked set of meanings for (20b), irrespective of whether it is preceded by
sentence (20a) or (21a). The model thus does not focus on the meaning of individ-
ual corpus tokens; instead it determines the most dominant meanings for a given
verb-noun combination overall, across all of its instances in the corpus.
(20) a. Who is making the cigarettes for tomorrow’s party?
b. John finished three cigarettes.
c. John finished making three cigarettes.
(21) a. Why is the room filled with smoke?
b. John finished three cigarettes.
c. John finished smoking three cigarettes.
271
Lapata and Lascarides Logical Metonymy
2.2 Parameter Estimation
We estimated the parameters of the model outlined in the previous section from a part-
of-speech-tagged and lemmatized version of the BNC, a 100-million-word collection
of samples of written and spoken language from a wide range of sources designed to
represent current British English (Burnard 1995). The counts f(v, e) and f(o, e) (see (17))
were obtained automatically from a partially parsed version of the BNC created using
Cass (Abney 1996), a robust chunk parser designed for the shallow analysis of noisy
text. The parser’s built-in function was employed to extract tuples of verb-subjects
and verb-objects (see (22)). Although verb-subject relations are not relevant for the
present model, they are important for capturing the influence of the sentential subject
(see Section 2.5) and modeling the interpretations of polysemous adjectives (which we
discuss in Section 3).
(22) a. change situation SUBJ
b. come off heroin OBJ
c. deal with situation OBJ
(23) a. isolated people SUBJ
b. smile good SUBJ
The tuples obtained from the parser’s output are an imperfect source of informa-
tion about argument relations. Bracketing errors, as well as errors in identifying chunk
categories accurately, results in tuples whose lexical items do not stand in a verb-
argument relationship. For example, inspection of the original BNC sentences from
which the tuples in (23) were derived reveals that the verb be is missing from (23a)
and the noun smile is missing from (23b) (see the sentences in (24)).
(24) a. Wenger found that more than half the childless old people in
her study of rural Wales saw a relative, a sibling, niece, nephew
or cousin at least once a week, though in inner city London
there were more isolated old people.
b. I smiled my best smile down the line.
In order to compile a comprehensive count of verb-argument relations, we dis-
carded tuples containing verbs or nouns attested in a verb-argument relationship only
once. Instances of the verb be were also eliminated, since they contribute no semantic
information with respect to the events or activities that are possibly associated with
the noun with which the verb is combined. Particle verbs (see (22b)) were retained
only if the particle was adjacent to the verb. Verbs followed by the preposition by
and a head noun were considered instances of verb-subject relations. The verb-object
tuples also included prepositional objects (see (22c)). It was assumed that PPs adja-
cent to the verb headed by any of the prepositions in, to, for, with, on, at, from, of, into,
through, and upon were prepositional objects.
2
This resulted in 737,390 distinct types of
verb-subject pairs and 1,078,053 distinct types of verb-object pairs (see Table 5, which
presents information about the tuples extracted from the corpus before and after the
filtering).
2 The POS tagging of the BNC (Leech, Garside, and Bryant 1994) distinguishes between verb particle
constructions like down in climb down the mountain and up in put up the painting, on the one hand, and
prepositions, on the other. So this allowed us to distinguish PP complements from NP ones.
272
Computational Linguistics Volume 29, Number 2
Table 5
Number of tuples extracted from the BNC.
Tokens Types
Relation Parser Filtering Tuples Verbs Nouns
SUBJ 4,759,950 4,587,762 737,390 14,178 25,900
OBJ 3,723,998 3,660,897 1,078,053 12,026 35,867
The frequency f(v, e) represents verbs taking progressive or infinitive VP comple-
ments. These were extracted from the parser’s output by looking for verbs followed by
progressive or infinitival complements (a special tag, VDG, is reserved in the BNC for
verbs in the progressive). The latter were detected by looking for verbs followed by in-
finitives (indicated by the marker to (TO0) and a verb in base form (VVI)). The examples
below illustrate the information extracted from the parser’s output for obtaining the
frequency f(v, e), which collapsed counts for progressive and infinitive complements.
(25) a. I had started to write a love-story. start write
b. She started to cook with simplicity. start cook
c. The suspect attempted to run off. attempt run off
(26) a. I am going to start writing a book. start write
b. I’ve really enjoyed working with you. enjoy work with
c. The phones began ringing off the hook. begin ring off
Note that some verbs (e.g., start) allow both an infinitival and a progressive com-
plement (see (25a) and (26a), respectively), whereas other verbs (e.g., attempt) allow
only one type of complement (see (25c)). Even for verbs that allow both types of
complements, there exist syntactic contexts in which the two complement types are in
complementary distribution: to start writing occurs 15 times in the BNC, whereas to start
to write does not occur at all. The situation is reversed for starting writing and starting
to write, for the former does not occur and the latter occurs seven times. Choosing to
focus only on one type of complement would result in a lower count for f(v, e) than
collapsing the counts observed for both types of complements.
Once we have obtained the frequencies f(v, e) and f(o, e), we can determine the
most likely interpretations for metonymic verb-noun combinations. Note that we may
choose to impose thresholds on the frequencies f(v, e) and f(o, e) (e.g., f(v, e) > 1, and
f(o, e) > 1), depending on the quality of the parsing data or the type of meaning
paraphrases we seek to discover (e.g., likely versus unlikely ones).
As an example of the paraphrases generated by our model, consider the sentences
in Table 6, which were cited as examples of logical metonymy in the lexical semantics
literature (Pustejovsky 1995; Verspoor 1997). The five most likely interpretations for
these metonymies (and their respective log-transformed probabilities) are illustrated in
Table 7. Note that the model comes up with plausible meanings, some of which overlap
with those suggested in the lexical semantics literature (underlined interpretations
indicate agreement between the model and the literature). Also, the model derives
several meanings, as opposed to the single interpretations provided in most cases in
the literature. Consider, for example, the pair begin story in Table 7. Here, not only the
interpretation tell is generated, but also write, read, retell, and recount. Another example
273
Lapata and Lascarides Logical Metonymy
Table 6
Paraphrases for verb-noun combinations taken from the literature.
John began the story → telling (Verspoor 1997, page 189)
John began the song → singing (Verspoor 1997, page 189)
John began the sandwich → eating/making (Verspoor 1997, page 167)
Mary wants a job → to have (Pustejovsky 1995, page 45)
John began the book → reading/writing (Verspoor 1997, page 167)
Bill enjoyed Steven King’s last book → reading (Pustejovsky 1995, page 88)
John began the cigarette → smoking (Verspoor 1997, page 167)
Harry wants another cigarette → to smoke (Pustejovsky 1995, page 109)
Table 7
Model-derived paraphrases for verbal metonymies, ranked in order of likelihood.
begin story begin song begin sandwich want job
tell −16.34 sing −15.14 bite into −18.12 get −14.87
write −17.02 rehearse −16.15 eat −18.23 lose −15.72
read −17.28 write −16.86 munch −19.13 take −16.40
retell −17.45 hum −17.45 unpack −19.14 make −16.52
recount −17.80 play −18.01 make −19.42 create −16.62
begin book enjoy book begin cigarette want cigarette
read −15.49 read −16.48 smoke −16.92 smoke −16.67
write −15.52 write −17.58 roll −17.63 take −18.23
appear in −16.98 browse through −18.56 light −17.76 light −18.45
publish −17.10 look through −19.68 take −18.88 put −18.51
leaf through −17.35 publish −19.93 twitch −19.17 buy −18.64
is begin song, for which the model generates the interpretations rehearse, write, hum,
and play, in addition to sing.
The model also exhibits slight variation in the interpretations for a given noun
among the different metonymic verbs (compare begin book and enjoy book and begin
cigarette and want cigarette in Table 7). This is in line with claims made in the lexical
semantics literature (Copestake and Briscoe 1995; Pustejovsky 1995; Verspoor 1997),
and it ultimately contributes to an improved performance against a “naive baseline”
model (see Section 2.4).
In some cases, the model comes up with counterintuitive interpretations: bite into
is generated as the most likely interpretation for begin sandwich (although the latter
interpretation is not so implausible, since eating entails biting into). The model also
fails to rank have as one of the five most likely interpretations for want job (see Ta-
ble 7). The interpretations get and take are, however, relatively likely; note that they
semantically entail the desired interpretation—namely, have—as a poststate. The inter-
pretations make and create imply the act of hiring rather than finding a job. Our model
cannot distinguish between the two types of interpretations. It also cannot discover
related meanings: for example, that get and take mean have or that tell, retell, and recount
(see Table 7) mean tell. (We return to this issue in Section 4.)
In the following section we test our model against verb-noun pairs randomly
selected from the BNC and evaluate the meaning paraphrases it generates against
human judgments. We explore the linear relationship between the subjects’ rankings
and the model-derived probabilities using correlation analysis.
274
Computational Linguistics Volume 29, Number 2
2.3 Experiment 1: Comparison against Human Judgments
Although there is no standard way to evaluate the paraphrases generated by the model
(there is no gold standard for comparison), a reasonable way to judge the model’s per-
formance would seem to be its degree of agreement with human paraphrase ratings.
This can be roughly measured by selecting some metonymic constructions, deriving
their paraphrase interpretations using the model outlined in Section 2.1, eliciting hu-
man judgments on these paraphrases, and then looking at how well the human ratings
correlate with the model probabilities for the same paraphrases.
In the following section we describe our method for assembling the set of exper-
imental materials and eliciting human-subject data for the metonymy paraphrasing
task. We use correlation analysis to compare the model probabilities against human
judgments and explore whether there is a linear relationship between the model-
derived likelihood of a given meaning and its perceived plausibility.
In Section 2.4.1 we introduce a naive model of verbal metonymy that does not
take the contribution of the metonymic verb into account; metonymic interpretations
(i.e., verbs) are simply expressed in terms of their conditional dependence on their
objects. We investigate the naive model’s performance against the human judgments
and the paraphrases generated by our initial model (see Section 2.4).
2.3.1 Method.
2.3.1.1 Materials and Design. From the lexical semantics literature (Pustejovsky 1995;
Verspoor 1997; McElree et al. 2001) we compiled a list of 20 verbs that allow logical
metonymy. From these we randomly selected 12 verbs (attempt, begin, enjoy, finish,
expect, postpone, prefer, resist, start, survive, try, and want). The selected verbs ranged in
BNC frequency from 10.9 per million to 905.3 per million. Next, we paired each one of
them with five nouns randomly selected from the BNC. The nouns had to be attested
in the corpus as the object of the verbs in question. Recall that verb-object pairs were
identified using Abney’s (1996) chunk parser Cass (see Section 2.2 for details). From
the retrieved verb-object pairs, we removed all pairs with BNC frequency of one, as we
did not want to include verb-noun combinations that were potentially unfamiliar to
the subjects. We used the model outlined in Section 2.1 to derive meaning paraphrases
for the 60 verb-noun combinations.
Our materials selection procedure abstracts over semantic distinctions that are
made in linguistic analyses. For instance, current models of lexical semantics typically
assign verbs such as enjoy a nonmetonymic sense when they are combined with NPs
that are purely temporal or eventive in nature, as in enjoy the marriage or enjoy the lecture
(Copestake and Briscoe 1995; Verspoor 1997). This is largely because a logical form
can be constructed in such cases without the use of semantic type coercion; the event-
denoting NP itself is the argument to the predicate enjoy. We did not rule out such
nouns from our materials, however, as our evaluation was conducted on randomly
selected verb-noun pairs.
More generally, we abstract over several criteria that Verspoor (1997) used in dis-
tinguishing metonymic from nonmetonymic uses within the corpus, and we adopt
a linguistically naive approach for two reasons. First, whereas Verspoor (1997) could
deploy more refined criteria because she was hand-selecting the materials from the
corpus and was focusing only on two metonymic verbs (begin and finish), our materials
were randomly sampled and covered a wider range of metonymic constructions. And
second, paraphrases for nonmetonymic cases (e.g., that enjoy the lecture can be para-
phrased as enjoy attending the lecture or enjoy listening to the lecture) may be useful for
some potential NLP applications (see the discussion in Section 4.2), since they provide
275
Lapata and Lascarides Logical Metonymy
Table 8
Number of generated interpretations as frequency cutoff for
f(v, e) and f(o, e) is varied.
Verb-noun
f(v,e) ≥ 1
f(o,e) ≥ 1
f(v,e) ≥ 2
f(o,e) ≥ 2
f(v,e) ≥ 3
f(o,e) ≥ 3
f(v,e) ≥ 4
f(n,e) ≥ 4
finish gig 11 4 3 1
finish novel 31 11 5 3
finish project 65 20 8 6
finish room 79 25 16 10
finish video 44 16 9 6
more detailed information about meaning than would be given by a logical form that
simply features enjoy(e, x, e
prime
), where e
prime
is the (event) variable that denotes the lecture.
Recall from Section 2.2 that thresholding is an option for the counts f(v, e) and
f(o, e). We derived model paraphrases without employing any thresholds for these
counts. Obtaining f(v, e) from the parsed data was relatively straightforward, as there
was no structural ambiguity involved. The parser’s output was postprocessed to re-
move potentially erroneous information, so there was no reason to believe that the
frequencies f(v, e) and f(o, e) were noisy. Furthermore, recent work has shown that
omitting low-frequency tuples degrades performance for language-learning tasks such
as PP attachment (Collins and Brooks 1995; Daelemans, van den Bosch, and Zavrel
1999), grapheme-to-phoneme conversion, POS tagging, and NP chunking (Daelemans,
van den Bosch, and Zavrel 1999). For our task, employing thresholds for f(v, e) and
f(o, e) dramatically decreases the number of derived interpretations. Table 8 shows the
decrease in the number of interpretations as the cutoff for f(v, e) and f(o, e) is varied for
five verb-object pairs that were included in our experimental study. Note that discard-
ing counts occurring in the corpus only once reduces the number of interpretations by
a factor of nearly three. Furthermore, applying frequency cutoffs reduces the range of
the obtained probabilities: only likely (but not necessarily plausible) interpretations are
obtained with f(o, e) ≥ 4 and f(v, e) ≥ 4. However, one of the aims of the experiment
outlined below was to explore the quality of interpretations with varied probabilities.
Table 9 displays the 10 most likely paraphrases (and their log-transformed probabili-
ties) for finish room as the cutoff for the frequencies f(v, e) and f(o, e) is varied. Notice
that applying a cutoff of three or four eliminates plausible interpretations such as dec-
orate, wallpaper, furnish, and tidy. This may be particularly harmful for verb-noun (or
adjective-noun) combinations that allow for a wide range of interpretations (like finish
room).
We estimated the probability P(e, o, v) for each verb-noun pair by varying the term
e. In order to generate stimuli covering a wide range of paraphrases corresponding to
different degrees of likelihood, for each verb-noun combination we divided the set of
generated meanings into three “probability bands” (high, medium, and low) of equal
size and randomly chose one interpretation from each band. This division ensured
that subjects saw a wide range of paraphrases with different degrees of likelihood.
Our experimental design consisted of two factors: verb-noun pair (Pair) and proba-
bility band (Band). The factor Pair included 60 verb-noun combinations, and the factor
Band had three levels, high, medium, and low. This yielded a total of Pair × Band
=60× 3 = 180 stimuli. In order to limit the size of the experiment, the 180 stimuli
were administered to two separate groups of subjects. The first group saw meaning
paraphrases for the verbs attempt, begin, want, enjoy, try, and expect, whereas the sec-
276
Computational Linguistics Volume 29, Number 2
Table 9
Ten most likely interpretations for finish room (with log-transformed
probabilities) as frequency threshold is varied.
f(v,e) ≥ 1 f(v,e) ≥ 2 f(v,e) ≥ 3 f(v,e) ≥ 4
f(o,e) ≥ 1 f(o,e) ≥ 2 f(o,e) ≥ 3 f(o,e) ≥ 4
decorate −18.47 decorate −18.47 fill −18.88 fill −18.88
wallpaper −19.07 fill −18.89 clean −19.08 clean −19.08
clean −19.09 clean −19.08 pack −20.17 pack −20.17
paper −19.09 search −20.13 make −20.36 make −20.36
furnish −19.31 pack −20.17 view −20.78 check −21.24
tidy −19.92 make −20.36 check −21.24 use −21.78
search −20.13 dress −20.55 pay −21.53 build −21.96
pack −20.17 view −20.78 use −21.78 give −22.29
make −20.36 check −21.24 build −21.96 prepare −22.45
view −20.78 paint −21.38 give −22.29 take −23.11
Table 10
Randomly selected example stimuli with log-transformed probabilities derived by the
model.
Probability Band
Verb-noun
High Medium Low
attempt peak climb −20.22 claim −23.53 include −24.85
begin production organize −19.09 influence −21.98 tax −22.79
enjoy city live in −20.77 come to −23.50 cut −24.67
expect reward collect −21.91 claim −23.13 extend −23.52
finish room wallpaper −19.07 construct −22.49 want −24.60
postpone payment make −21.85 arrange −23.21 read −25.92
prefer people talk to −20.52 sit with −22.75 discover −25.26
resist song whistle −22.12 start −24.47 hold −26.50
start letter write −15.59 study −22.70 hear −24.50
survive course give −22.88 make −24.48 write −26.27
try drug take −17.81 grow −22.09 hate −23.88
want hat buy −17.85 examine −21.56 land on −22.38
ond group saw paraphrases for finish, prefer, resist, start, postpone, and survive. Example
stimuli are shown in Table 10.
Each experimental item consisted of two sentences, a sentence containing a meto-
nymic construction (e.g., Peter started his dinner) and a sentence paraphrasing it (e.g.,
Peter started eating his dinner). The metonymic sentences and their paraphrases were
created by the authors as follows. The selected verb-noun pairs were converted into
simple sentences by adding a sentential subject and articles or pronouns where appro-
priate. The sentential subjects were familiar proper names (BNC corpus frequency >
30 per million) balanced for gender. All sentences were in the past tense. In the para-
phrasing sentences, the metonymy was spelled out by converting the model’s output
to a verb taking either a progressive or infinitive VP complement (e.g., started to eat or
started eating). For verbs allowing both a progressive and an infinitive VP complement,
we chose the type of complement with which the verb occurred more frequently in
the corpus. A native speaker of English other than the authors was asked to confirm
that the metonymic sentences and their paraphrases were syntactically well-formed
(items found syntactically odd were modified and retested). Examples of the experi-
277
Lapata and Lascarides Logical Metonymy
mental stimuli the subjects saw are provided in (27) and (28). A complete list of the
experimental items is given in Appendix B.
(27) a. high: Michael attempted the peak
Michael attempted to climb the peak
b. medium: Michael attempted the peak
Michael attempted to claim the peak
c. low: Michael attempted the peak
Michael attempted to include the peak
(28) a. high: Jean enjoyed the city
Jean enjoyed living in the city
b. medium: Jean enjoyed the city
Jean enjoyed coming to the city
c. low: Jean enjoyed the city
Jean enjoyed cutting the city
2.3.1.2 Procedure. The experimental paradigm was magnitude estimation (ME), a
technique standardly used in psychophysics to measure judgments of sensory stim-
uli (Stevens 1975). The ME procedure requires subjects to estimate the magnitude of
physical stimuli by assigning numerical values proportional to the stimulus magni-
tude they perceive. Highly reliable judgments can be achieved in this fashion for a
wide range of sensory modalities, such as brightness, loudness, or tactile stimulation.
The ME paradigm has been extended successfully to the psychosocial domain
(Lodge 1981), and recently Bard, Robertson, and Sorace (1996) and Cowart (1997)
showed that linguistic judgments can be elicited in the same way as judgments of
sensory or social stimuli. ME requires subjects to assign numbers to a series of linguistic
stimuli in a proportional fashion. Subjects are first exposed to a modulus item, to
which they assign an arbitrary number. All other stimuli are rated proportional to the
modulus. In this way, each subject can establish his own rating scale, thus yielding
maximally fine-grained data and avoiding the known problems with the conventional
ordinal scales for linguistic data (Bard, Robertson, and Sorace 1996; Cowart 1997;
Sch ¨utze 1996). In particular, ME does not restrict the range of the responses. No matter
which modulus a subject chooses, he or she can subsequently assign a higher or lower
judgment by using multiples or fractions of the modulus.
In the present experiment, each subject took part in an experimental session that
lasted approximately 20 minutes. The experiment was self-paced, and response times
were recorded to allow the data to be screened for anomalies. The experiment was con-
ducted remotely over the Internet. Subjects accessed the experiment using their Web
browser, which established an Internet connection to the experimental server running
WebExp 2.1 (Keller, Corley, and Scheepers 2001), an interactive software package for
administering Web-based psychological experiments. (For a discussion of WebExp and
the validity of Web-based data, see Appendix A).
2.3.1.3 Instructions. Before participating in the actual experiment, subjects were pre-
sented with a set of instructions. The instructions explained the concept of numeric
magnitude estimation of line length. Subjects were instructed to make estimates of line
length relative to the first line they would see, the reference line. Subjects were told
to give the reference line an arbitrary number, and then assign a number to each fol-
lowing line so that it represented how long the line was in proportion to the reference
278
Computational Linguistics Volume 29, Number 2
line. Several example lines and corresponding numerical estimates were provided to
illustrate the concept of proportionality.
The subjects were instructed to judge how well a particular sentence paraphrased
another sentence, using the same technique that they had applied to judging line
length. Examples of plausible (see (29a)) and implausible (see (29b)) sentence para-
phrases were provided, together with examples of numerical estimates.
(29) a. Peter started his dinner Peter started eating his dinner
b. Peter started his dinner Peter started writing his dinner
Subjects were informed that they would initially have to assign a number to a reference
paraphrase. For each subsequent paraphrase, subjects were asked to assign a number
indicating how good or bad that paraphrase was in proportion to the reference.
Subjects were told that they could use any range of positive numbers for their
judgments, including decimals. It was stressed that there was no upper or lower limit
on the numbers that could be used (exceptions being zero or negative numbers).
Subjects were urged to use a wide range of numbers and to distinguish as many
degrees of paraphrase plausibility as possible. It was also emphasized that there were
no “correct” answers and that subjects should base their judgments on first impressions
and not spend too much time thinking about any one paraphrase.
2.3.1.4 Demographic Questionnaire. After the instructions, a short demographic ques-
tionnaire was administered. The questionnaire asked subjects to provide their name,
e-mail, address, age, sex, handedness, academic subject or occupation, and language
region. Handedness was defined as “the hand you prefer to use for writing”; language
region was defined as “the place (town, federal state, country) where you learned your
first language.”
2.3.1.5 Training Phase. The training phase was meant to familiarize subjects with the
concept of numeric magnitude estimation using line lengths. Items were presented as
horizontal lines, centered in the window of the subject’s Web browser. After viewing
an item, the subject had to provide a numerical judgment via the computer keyboard.
After the subject pressed Return, the current item disappeared and the next item was
displayed. There was no opportunity to revisit previous items or change responses
once Return had been pressed. No time limit was set either for the item presentation
or for the response, although response times were recorded for later inspection.
Subjects first judged the modulus item, and then all the items in the training
set. The modulus was the same for all subjects, and it remained on the screen all
the time to facilitate comparison. Items were presented in random order, with a new
randomization being generated for each subject.
The training set contained six horizontal lines. The range of the shortest to longest
item was one to ten (that is, the longest line was ten times the length of the shortest).
The items were distributed evenly over this range, with the largest item covering the
maximal window width of the Web browser. A modulus item in the middle of the
range was provided.
2.3.1.6 Practice Phase. The practice phase enabled subjects to practice magnitude esti-
mation of verb-noun paraphrases. Presentation and response procedure was the same
as in the training phase, with linguistic stimuli being displayed instead of lines. Each
subject judged the whole set of practice items, again presented to him or her in random
order.
279
Lapata and Lascarides Logical Metonymy
The practice set consisted of eight paraphrase sentences that were representative
of the test materials. The paraphrases were based on the three probability bands and
represented a wide range of probabilities. A modulus item selected from the medium
probability band was provided.
2.3.1.7 Experimental Phase. The presentation and response procedure in the exper-
imental phase were the same as in the practice phase. Subjects were assigned to
groups at random, and a random stimulus order was generated for each subject (for
the complete list of experimental stimuli, see Appendix B). Each group of subjects
saw 90 experimental stimuli (i.e., metonymic sentences and their paraphrases). As in
the practice phase, the paraphrases were representative of the three probability bands
(high, medium, low). Again a modulus item from the medium probability band was
provided (see Appendix B). The modulus was the same for all subjects and remained
on the screen the entire time the subject was completing the task.
2.3.1.8 Subject. Sixty-three native speakers of English participated in the experiment.
The subjects were recruited over the Internet through advertisements posted to news-
groups and mailing lists. Participation was voluntary and unpaid. Subjects had to be
linguistically naive (i.e., neither linguists nor students of linguistics were allowed to
participate).
The data of two subjects were eliminated after inspection of their response times
showed that they had not completed the experiment in a realistic time frame (i.e., they
provided ratings too quickly, with average response time < 1000 ms). The data of one
subject were excluded because she was a non-native speaker of English.
This left 60 subjects for analysis. Of these, 53 subjects were right-handed, 7 left-
handed; 24 subjects were female, 36 male. The age of subjects ranged from 17 to 62;
the mean was 26.4 years.
2.3.2 Results. The data were first normalized by dividing each numerical judgment
by the modulus value that the subject had assigned to the reference sentence. This
operation created a common scale for all subjects. Then the data were transformed by
taking the decadic logarithm. This transformation ensured that the judgments were
normally distributed and is standard practice for magnitude estimation data (Bard,
Robertson, and Sorace 1996; Lodge 1981). All further analyses were conducted on the
resulting normalized, log-transformed judgments.
We performed a correlation analysis to determine whether there was a linear rela-
tion between the paraphrases generated by the model and their perceived likelihood.
This tested the hypothesis that meaning paraphrases assigned high probabilities by the
model are perceived as better paraphrases by the subjects than meaning paraphrases
assigned low probabilities. For each experimental item we computed the average of
the normalized and log-transformed subject ratings. The mean subject ratings were
then compared against the (log-transformed) probabilities assigned by the model for
the same items.
The comparison between the absolute model probabilities and the human judg-
ments yielded a Pearson correlation coefficient of .64 (p <.01, N = 174; six items were
discarded because of a coding error). The mean subject ratings and the model prob-
abilities are given in Appendix B. Appendix C presents descriptive statistics for the
model probabilities and the human judgments. The relationship between judgments
and probabilities is plotted in Figure 1.
An important question is how well humans agree in their paraphrase judgments
for verb-noun combinations. Intersubject agreement gives an upper bound for the
280
Computational Linguistics Volume 29, Number 2
-30 -25 -20 -15 -10
log-transformed model probabilities
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
mean subject ratings
Figure 1
Correlation of elicited judgments and model-derived probabilities for metonymic verb-noun
pairs.
task and allows us to interpret how well the model is doing in relation to humans.
To calculate intersubject agreement, we used leave-one-out resampling. The technique
is a special case of n-fold cross-validation (Weiss and Kulikowski 1991) and has been
previously used for measuring how well humans agree on judging semantic similarity
(Resnik and Diab 2000; Resnik 1999).
For each subject group we divided the set of the subjects’ responses with size m
into a set of size m − 1 (i.e., the response data of all but one subject) and a set of size
one (i.e., the response data of a single subject). We then correlated the mean ratings
of the former set with the ratings of the latter. This was repeated m times. Since each
group had 30 subjects, we performed 30 correlation analyses and report their mean.
For the first group of subjects, the average intersubject agreement was .74 (Min = .19,
Max = .87, StdDev = .12), and for the second group it was .73 (Min = .49, Max
= .87, StdDev = .09). Our model’s agreement with the human data is not far from the
average human performance of .74.
In the following section we introduce a naive model of verbal metonymy. We
compare the naive model’s performance against the human judgments and the para-
phrases generated by our initial model. We discuss extensions of the basic model in
Section 2.5.1.
2.4 Experiment 2: Comparison against Naive Baseline
2.4.1 Naive Baseline Model. In the case of verbal metonymy a naive baseline model
can be constructed by simply taking verb-noun co-occurrence data into account, ig-
noring thus the dependencies between the polysemous verb and its progressive or
infinitival VP complements. Consider the sentence John began the book. In order to gen-
erate appropriate paraphrases for begin book, we will consider solely the verbs that take
book as their object (i.e., read, write, buy, etc.). This can be simply expressed as P(e | o),
the conditional probability of a verb e given its object o (i.e., the noun figuring in the
metonymic expression), which we estimate as follows:
ˆ
P(e | o)=
f(e, o)
f(o)
(30)
281
Lapata and Lascarides Logical Metonymy
The model in (30) treats metonymic verbs as semantically empty and relies on their
object NPs to provide additional semantic information. The counts f(e, o) and f(o) can
be easily obtained from the BNC: f(o) amounts to the number of times a noun is
attested as an object, and f(e, o) are verb-object tuples extracted from the BNC using
Cass (Abney 1996) as described earlier.
2.4.2 Results. We used the naive model to calculate the likelihood of the meaning
paraphrases that were presented to the subjects (see Experiment 1). Through correla-
tion analysis we explored the linear relationship between the elicited judgments and
the naive baseline model. We further directly compared the two models: that is, our
initial, linguistically more informed model and the naive baseline.
Comparison between the probabilities generated by the naive model and the
elicited judgments yielded a Pearson correlation coefficient of .42 (p <.01, N = 174).
(Recall that our initial model yielded a correlation coefficient of .64.) We conducted a
one-tailed t-test to determine if the correlation coefficients were significantly different.
The comparison revealed that the difference between them was statistically significant
(t(171)=1.67, p <.05), indicating that our model performs reliably better than the
naive baseline. Comparison between the two models (our initial model introduced in
Section 2.1 and the naive baseline model) yielded an intercorrelation of .46 (p <.01,
N = 174). These differences between the “full” probabilistic model and the naive base-
line model confirm claims made in the literature: Different metonymic verbs have a
different semantic impact on the resolution of metonymy.
2.5 Experiment 3: Comparison against Norming Data
Our previous experiments focused on evaluating the plausibility of meaning para-
phrases generated by a model that does not take into account the contribution of the
sentential subject. However, properties of the subject NP appear to influence the inter-
pretation of the metonymic expression in otherwise neutral contexts, as is illustrated
in (31), in which the interpretation of enjoy the book is influenced by the sentential
subject: Authors usually write books, whereas critics usually review them.
(31) a. The critic enjoyed the book.
b. The author enjoyed the book.
In this section we present an extension of the basic model outlined in Section 2.1
that takes sentential subjects into account. We evaluate the derived paraphrases and
their likelihood again by comparison with human data. This time we compare our
model against paraphrase data generated independently by subjects that participated
in an experimental study (McElree et al. 2001) that was not designed specifically to
test our model.
2.5.1 The Extended Model. We model the meaning of sentences like (31) again as the
joint distribution of the following variables: the metonymic verb v, its subject s, its
object o, and the implicit interpretation e. By choosing the ordering 〈e, v, s, o〉,wecan
factor P(e, o, s, v) as follows:
P(e, o, s, v)=P(e)· P(v | e)· P(s | e, v)· P(o | e, v, s)(32)
The terms P(e) and P(v | e) are easy to estimate from the BNC. For P(e) all we need
is a POS-tagged corpus and P(v | e) can be estimated from Cass’s output (see equa-
tions (12) and (13)). The estimation of the terms P(s | e, v) and P(o | e, v, s) is, however,
282
Computational Linguistics Volume 29, Number 2
problematic, as they rely on the frequencies f(s, e, v) and f(o, e, v, s), respectively. Re-
call that there is a discrepancy between a metonymic interpretation and its usage. As
we discussed earlier, metonymic interpretation is not overtly expressed in the corpus.
Furthermore, the only type of data available to us for the estimation of P(s | e, v) and
P(o | e, v, s) is the partially parsed BNC, which is not annotated with information re-
garding the interpretation of metonymies. This means that P(s | e, v) and P(o | e, v, s)
need to be approximated somehow. We first assume that the sentential subject s is con-
ditionally independent of the metonymic verb v; second, we assume that the sentential
object o is conditionally independent of v and s:
P(s | e, v) ≈ P(s | e)(33)
P(o | e, v, s) ≈ P(o | e 34)
The rationale behind the approximation in (33) is that the likelihood of a noun s being
a subject of a verb e is largely independent of whether e is the complement of a
metonymic verb v. For example, authors usually write, irrespective of whether they
enjoy, dislike, start, or finish doing it. The motivation for the approximation in (34)
comes from the observation that an object is more closely related to the verb that
selects for it than a subject or a metonymic verb. We are likely to come up with book
or letter for o if we know that o is the object of read or write. Coming up with an object
for o is not so straightforward if all we know is the metonymic verb (e.g., enjoy, finish)
or its sentential subject. It is the verbs, rather than other sentential constituents, that
impose semantic restrictions on their arguments. We estimate P(s | e) and P(o | e) using
maximum likelihood:
ˆ
P(s | e)=
f(s, e)
f(e)
(35)
ˆ
P(o | e)=
f(o, e)
f(e)
(36)
The count f(s, e) amounts to the number of times a noun s is attested as the
subject of a verb e; f(o, e) represents the number of times a noun is attested as an
object of e. Verb-argument tuples can be easily extracted from the BNC using Cass (see
Section 2.2 for details). Table 11 illustrates the model’s performance for the metonymic
constructions in (37). The table shows only the five most likely interpretations the
model came up with for each construction. Interestingly, different interpretations are
derived for different subjects. Even though pianists and composers are semantically
related, pianists are more likely to begin playing a symphony, whereas composers
are more likely to conduct or write a symphony. Similarly, builders tend to renovate
houses and architects tend to design them.
(37) a. The composer/pianist began the symphony.
b. The author/student started the book.
c. The builder/architect started the house.
d. The secretary/boss finished the memo.
In the following section we compare the interpretations generated by the model
against paraphrases provided by humans. More specifically, we explore whether there
is a linear relationship between the frequency of an interpretation as determined in
a norming study and the probability of the same interpretation as calculated by the
model.
283
Lapata and Lascarides Logical Metonymy
Table 11
Subject-related model interpretations, ranked in order of likelihood.
begin symphony start book
composer pianist author student
write −22.2 play −24.20 write −14.87 read −16.12
conduct −23.79 hear −25.38 publish −16.94 write −16.48
hear −25.38 give −28.44 compile −17.84 study −17.59
play −25.81 do −29.23 read −17.98 research −18.86
create −25.96 have −30.13 sign −18.59 translate −17.85
start house finish memo
builder architect secretary boss
renovate −15.43 design −16.87 write −19.79 draft −20.73
build −17.56 build −17.20 type −20.13 send −21.97
demolish −18.37 restore −19.08 send −20.87 sign −22.04
dismantle −19.69 purchase −19.32 sign −22.01 hand −22.12
erect −19.81 site −19.73 make −22.74 write −22.74
2.5.2 Method. For the experiment described in this section, we used the norming data
reported in McElree et al. (2001). In McElree et al.’s study subjects were given sentence
fragments such as (38) and were asked to complete them. Potential completions for
fragment (38a) include writing or reading. The study consisted of 142 different sentences
similar to those shown in (38) and included 15 metonymic verbs. Thirty sentences were
constructed for each of the metonymic verbs start, begin, complete, and finish, and a total
of 22 sentences for attempt, endure, expect, enjoy, fear, master, prefer, resist, savor, survive,
and try.
(38) a. The writer finished the novel.
b. The soldier attempted the mountain.
c. The teenager finished the novel.
The completions can be used to determine interpretation preferences for the metonymic
constructions simply by counting the verbs that human subjects use to complete sen-
tences like those in (38). For example, five completions were provided by the subjects
for fragment (38b): climb, hike, scale, walk, and take. Of these climb was by far the most
likely, with 78 (out of 88) subjects generating this interpretation.
3
The most likely in-
terpretations for (38a) and (38c) were, respectively, write (13 out of 28 subjects) and
read (18 out of 22).
For each of the sentences included in McElree et al.’s (2001) study, we derived in-
terpretation paraphrases using the model presented in Section 2.5.1. We next compared
the interpretations common in the model and the human data.
2.5.3 Results. In Experiment 1 we evaluated the paraphrases generated by the model
by eliciting plausibility judgments from subjects and showed that our model produces
an intuitively plausible ranking of meanings. Here, we evaluate the quality of the
3 McElree et al.’s (2001) linguistic materials were manually constructed and not controlled for frequency.
For example, one would expect (38b) to be relatively rare, even in a large corpus. This is true for the
BNC, in which the combination attempt mountain is not attested at all.
284
Computational Linguistics Volume 29, Number 2
produced paraphrases by directly comparing them to norming data acquired inde-
pendently of our model and the particular corpus we are using.
The comparison between (log-transformed) model probabilities and (log-trans-
formed) completion frequencies yielded a Pearson correlation coefficient of .422 (p<.01,
N = 341). We also compared the completion frequencies against interpretation proba-
bilities derived using the model presented in Section 2.3, which does not take subject-
related information into account. The comparison yielded a correlation coefficient
of .216 (p <.01, N = 341). We carried out a one-tailed t-test to determine if the differ-
ence between the two correlation coefficients is significant. The comparison revealed
that the difference is statistically significant (t(338)=2.18, p <.05). This means that
the fit between the norming data and the model is better when the model explicitly
incorporates information about the sentential subject. The two models are, as expected,
intercorrelated (r = .264, p <.01, N = 341).
2.6 Discussion
We have demonstrated that the meanings acquired by our probabilistic model corre-
late reliably with human intuitions. These meanings go beyond the examples found
in the theoretical linguistics literature. The verb-noun combinations we interpret were
randomly sampled from a large, balanced corpus providing a rich inventory for their
meanings. We have shown that the model has four defining features: (1) It is able
to derive intuitive meanings for verb-noun combinations, (2) it generates clusters of
meanings (following Vendler’s (1968) insight), (3) it predicts variation in interpretation
among the different nouns: The same verb may carry different meanings depending
on its subject or object (compare begin book to begin house and the author began the house
to the architect began the house), and (4) it represents variation in interpretation among
the different metonymic verbs (e.g., begin book vs. enjoy book). This latter property
demonstrates that although the model does not explicitly encode linguistic constraints
for resolving metonymies, it generates interpretations that broadly capture linguistic
differences (e.g., attempt imposes different constraints on interpretation from begin or
enjoy). Furthermore, these interpretations for metonymic verb-noun pairs are discov-
ered automatically, without presupposing the existence of a predefined taxonomy or
a knowledge base.
Note that the evaluation procedure to which we subject our model is rather strict.
The derived verb-noun combinations were evaluated by subjects naive to linguistic
theory. Although verbal logical metonymy is a well-researched phenomenon in the
theoretical linguistics literature, the experimental approach advocated here is, to our
knowledge, new. Comparison between our model and human judgments yielded a re-
liable correlation of .64 when the upper bound for the task (i.e., intersubject agreement)
is on average .74. Furthermore, our model performed reliably better than a naive base-
line model, which achieved a correlation of only .42. When compared against norming
data, an extended version of our model that takes subject information into account
reached a correlation of .42. Comparison against norming data is a strict test on unseen
data that was not constructed explicitly to evaluate our model but is independently
motivated and does not take our corpus (i.e., the BNC) or our particular task into
account.
We next investigate whether such an approach generalizes to other instances of
logical metonymy by looking at adjective-noun combinations. Adjectives pose a greater
challenge for our modeling task, as they can potentially allow for a wider range of
interpretations and can exhibit preferences for a verb-subject or verb-object paraphrase
(see Section 1). Following the approach we adopted for verbal metonymy, we define
the interpretation of polysemous adjective-noun combinations as a paraphrase gener-
285
Lapata and Lascarides Logical Metonymy
ation task. We provide a probabilistic model that not only paraphrases adjective-noun
pairs (e.g., fast plane) with a related verb (e.g., fly) but also predicts whether the noun
modified by the adjective (e.g., plane) is likely to be the verbal object or subject. The
model achieves this by combining distributional information about how likely it is
for any verb to be modified by the adjective in the adjective-noun combination or its
corresponding adverb with information about how likely it is for any verb to take the
modified noun as its object or subject. We obtain quantitative information about verb-
adjective modification and verb-argument relations from the BNC and evaluate our
results by comparing the model’s predictions against human judgments. Consistent
with our results on verbal metonymy, we show that the model’s ranking of meanings
correlates reliably with human intuitions.
3. Metonymic Adjectives
3.1 The Model
Consider again the adjective-noun combinations in (39). In order to come up with the
interpretation of plane that flies quickly for fast plane, we would like to find in the corpus
a sentence whose subject is the noun plane or planes and whose main verb is fly.We
would also expect fly to be modified by the adverb fast or quickly. In the general case,
we would like to gather from the corpus sentences indicating what planes do fast.
Similarly, for the adjective-noun combination fast scientist, we would like to find in
the corpus information indicating what the activities that scientists perform fast are,
whereas for easy problem we need information about what one can do with problems
easily (e.g., one can solve problems easily) or about what problems are (e.g., easy to
solve or easy to set).
(39) a. fast plane
b. fast scientist
c. fast programmer
d. easy problem
In sum, in order to come up with a paraphrase of the meaning of an adjective-
noun combination, we need to know which verbs take the head noun as their subject or
object and are modified by an adverb corresponding to the modifying adjective. This
can be expressed as the joint probability P(a, e, n, rel), where e is the verbal predicate
modified by the adverb a (directly derived from the adjective present in the adjective-
noun combination) bearing the argument relation rel (i.e., subject or object) to the head
noun n. By choosing the ordering 〈e, n, a, rel〉 for the variables a, e, n, and rel,wecan
rewrite P(a, e, n, rel), using the chain rule, as follows:
P(a, e, n, rel)=P(e)· P(n | e)· P(a | e, n)· P(rel | e, n, a)(40)
Although the terms P(e) and P(n | e) can be straightforwardly estimated from the BNC
(see (12) for P(e); P(n | e) can be obtained by counting the number of times a noun n
co-occurs with a verb e either as its subject or object), the estimation of P(a | e, n) and
P(rel | e, n, a) faces problems similar to those for metonymic verbs. Let us consider
more closely the term P(rel | e, n, a), which can be estimated as shown in (41).
ˆ
P(rel | e, n, a)=
f(rel, e, n, a)
f(e, n, a)
(41)
286
Computational Linguistics Volume 29, Number 2
One way to obtain f(rel, e, n, a) would be to parse fully the corpus so as to identify
the verbs that take the head noun n as their subject or object and are modified by the
adverb a, assuming it is equally likely to find in a corpus the metonymic expression
(e.g., fast plane) and its paraphrase interpretation (i.e., plane that flies quickly). As in the
case of verb-noun metonymies, this assumption is unjustified: For the adjective-noun
combination fast plane, there are only six sentences in the entire BNC that correspond
to f(rel, e, n, a). According to the sentences in (42), the most likely interpretation for
fast plane is plane that goes fast (see examples (42a)–(42c)). The interpretations plane that
swoops in fast, plane that drops down fast, and plane that flies fast are all equally likely, since
they are attested in the corpus only once (see examples (42d)–(42f)). This is rather
unintuitive, since fast planes are more likely to fly than swoop in fast. Similar problems
affect the frequency f(e, n, a).
(42) a. The plane went so fast it left its sound behind.
b. And the plane’s going slightly faster than the Hercules or
Andover.
c. He is driven by his ambition to build a plane that goes faster
than the speed of sound.
d. Three planes swooped in, fast and low.
e. The plane was dropping down fast towards Bangkok.
f. The unarmed plane flew very fast and very high.
In default of a corpus explicitly annotated with interpretations for metonymic
adjectives, we will make the following independence assumptions:
P(a | e, n) ≈ P(a | e)(43)
P(rel | e, n, a) ≈ P(rel | e, n 44)
The rationale behind the approximation in (43) is that the likelihood of seeing an
adverbial a modifying a verb e bearing an argument relation to a noun n is largely
independent of that specific noun. For example, flying can be carried out fast or slowly
or beautifully irrespective of whether it is a pilot or a bird who is doing the flying.
Similarly, the adverb peacefully is more related to dying than killing or injuring irrespec-
tive of who the agent of these actions is. Accordingly, we assume that the argument
relation rel is independent of whether the verb e (standing in relation rel with noun n)
is modified by an adverb a (see (44)). In other words, it is the verb e and its argument
n that determine the relation rel rather than the adjective or adverb a. Knowing that
flying is conducted slowly will not affect the likelihood of inferring a subject relation
for plane and fly. Yet we are likely to infer an object relation for plane and construct
irrespective of whether the constructing is done slowly, quickly, or automatically. We
estimate the probabilities P(e), P(n | e), P(a | e), and P(rel | e, n) as follows:
ˆ
P(e)=
f(e)
N
(45)
ˆ
P(n | e)=
f(n, e)
f(e)
(46)
ˆ
P(a | e)=
f(a, e)
f(e)
(47)
287
Lapata and Lascarides Logical Metonymy
Table 12
Most frequent verbs modified by the adverb fast.
f(fast,e) f(fast,e)
go 29 work 6
grow 28 grow in 6
beat 27 learn 5
run 16 happen 5
rise 14 walk 4
travel 13 think 4
move 12 keep up 4
come 11 fly 4
drive 8 fall 4
get 7 disappear 4
Table 13
Most frequent verbs taking as an argument the noun plane.
f(SUBJ,e,plane) f(OBJ,e,plane)
fly 20 catch 24
come 17 board 15
go 15 take 14
take 14 fly 13
land 9 get 12
touch 8 have 11
make 6 buy 10
arrive 6 use 8
leave 5 shoot 8
begin 5 see 7
ˆ
P(rel | e, n)=
f(rel, e, n)
f(e, n)
(48)
By substituting equations (45)–(48) into (41) and simplifying the relevant terms, (41)
can be rewritten as follows:
P(a, e, n, rel)=
f(rel, e, n)· f(a, e)
f(e)· N
(49)
Assume we want to discover a meaning paraphrase for the adjective-noun combination
fast plane. We need to find the verb e and the relation rel (i.e., subject or object) that
maximize the term P(fast, e, plane, rel). Table 12 gives a list of the most frequent verbs
modified by the adverb fast in the BNC (see the term f(a, e) in equation (49)), and
Table 13 lists the verbs for which the noun plane is the most likely object or subject
(see the term f(rel, e, n) in equation (49)). In the following section, we describe how
the frequencies f(rel, e, n), f(a, e), and f(e) were estimated from a lemmatized version
of the BNC.
Table 12 can be thought of as a list of the activities that can be fast (i.e., going,
growing, flying), whereas Table 13 specifies the events associated with the noun plane.
Despite our simplifying assumptions, the model given in (49) will come up with plau-
sible meanings for adjective-noun combinations like fast plane. Note that the verbs fly,
come, and go are most likely to take the noun plane as their subject (see Table 13). These
288
Computational Linguistics Volume 29, Number 2
verbs also denote activities that are fast (see Table 12, in which the underlined verbs
are events that are associated both with the adverb fast and the noun plane). Further
note that a subject interpretation is more likely than an object interpretation for fast
plane, since none of the verbs likely to have plane as their object are modified by the
adverb fast (compare Tables 12 and 13).
As in the case of metonymic verbs, the probabilistic model outlined above ac-
quires meanings for polysemous adjective-noun combinations in an unsupervised
manner without presupposing annotated corpora or taxonomic information. The ob-
tained meanings are not discourse-sensitive; they can be thought of as default semantic
information associated with a particular adjective-noun combination. This means that
our model is unable to predict that programmer that runs fast is a likely interpretation
for fast programmer when the latter is in a context like the one given in (5) (repeated
here as (50)).
(50) a. All the office personnel took part in the company sports day last
week.
b. One of the programmers was a good athlete, but the other was
struggling to finish the courses.
c. The fast programmer came first in the 100m.
3.2 Parameter Estimation
As in the case of verbs, the parameters of the model were estimated using a part-of-
speech-tagged and lemmatized version of the BNC. The counts f(e) and N (see (49))
reduce to the number of times a given verb is attested in the corpus. The frequency
f(rel, e, n) was obtained using Abney’s (1996) chunk parser Cass (see Section 2.2 for
details).
Generally speaking, the frequency f(a, e) represents not only a verb modified by
an adverb derived from the adjective in question (see example (51a)), but also con-
structions like the ones shown in (51b) and (51c), in which the adjective takes an
infinitival VP complement whose logical subject can be realized as a for PP (see ex-
ample (51c)). In cases of verb-adverb modification we assume access to morphological
information that specifies what counts as a valid adverb for a given adjective. In most
cases adverbs are formed by adding the suffix -ly to the base of the adjective (e.g.,
slow-ly, easy-ly). Some adjectives have identical adverbs (e.g., fast, right). Others have
idiosyncratic adverbs (e.g., the adverb of good is well). It is relatively straightforward
to develop an automatic process that maps an adjective to its corresponding adverb,
modulo exceptions and idiosyncracies; however, in the experiments described in the
following sections, this mapping was manually specified.
(51) a. comfortable chair → a chair on which one sits comfortably
b. comfortable chair → a chair that is comfortable to sit on
c. comfortable chair → a chair that is comfortable for me to sit on
In cases in which the adverb does not immediately succeed the verb, the parser
is not guaranteed to produce a correct analysis, since it does not resolve structural
ambiguities. So we adopted a conservative strategy, in which to obtain the frequency
f(a, e), we looked only at instances in which the verb and the adverbial phrase mod-
ifying it were adjacent. More specifically, in cases in which the parser identified an
AdvP following a VP, we extracted the verb and the head of the AdvP (see the exam-
ples in (52)). In cases where the AdvP was not explicitly identified, we extracted the
verb and the adverb immediately following or preceding it (see the examples in (53)),
289
Lapata and Lascarides Logical Metonymy
assuming that the verb and the adverb stand in a modification relation. The examples
below illustrate the parser’s output and the information that was extracted for the
frequency f(a, e).
(52) a. [
NP
Oriental art] [
VP
came] [
AdvP
more slowly.]
come slowly
b. [
NP
The issues] [
VP
will not be resolved] [
AdvP
easily.]
resolve easily
c. [
NP
Arsenal] [
VP
had been pushed] [
AdvP
too hard.]
push hard
(53) a. [
NP
Some art historians] [
VP
write well] [
PP
about the present.]
write well
b. [
NP
The accidents] [
VP
could have been easily avoided.]
avoid easily
c. [
NP
A system of molecules] [
VP
is easily shown] [
VP
to stay
constant.]
show easily
d. [
NP
Their economy] [
VP
was so well run.]
run well
Adjectives with infinitival complements (see (51b) and (51c)) were extracted from
the parser’s output. We concentrated solely on adjectives immediately followed by
infinitival complements with an optionally intervening for PP (see (51c)). The adjective
and the main verb of the infinitival complement were counted as instances of the
quantity f(a, e). The examples in (54) illustrate the process.
(54) a. [
NP
These early experiments] [
VP
were easy] [
VP
to interpret.]
easy interpret
b. [
NP
It] [
VP
is easy] [
PP
for an artist] [
VP
to show work
independently.]
easy show
c. [
NP
It] [
VP
is easy] [
VP
to show] [
VP
how the components interact.]
easy show
Finally, the frequency f(a, e) collapsed the counts from cases in which the adjective
was followed by an infinitival complement (see the examples in (54)) and cases in
which the verb was modified by the adverb corresponding to the related adjective
(see the examples in (52)–(53)). For example, assume that we are interested in the
frequency f(easy, show). In this case, we will take into account not only sentences
(54b) and (54c), but also sentence (53b). Assuming this was the only evidence in the
corpus, the frequency f(easy, show) would be three.
Once we have obtained the frequencies f(a, e) and f(rel, e, n), we can determine
what the most likely interpretations for a given adjective-noun combination are. If
we know the interpretation preference of a given adjective (i.e., subject or object), we
may vary only the term e in P(a, n, rel, e), keeping the terms n, a, and rel constant.
Alternatively, we could acquire the interpretation preferences automatically by varying
both the terms rel and e. In Experiment 4 (see Section 3.3) we acquire both meaning
paraphrases and argument preferences for polysemous adjective-noun combinations.
In what follows we illustrate the properties of the model by applying it to a small
number of adjective-noun combinations (displayed in Table 14). The adjective-noun
290
Computational Linguistics Volume 29, Number 2
Table 14
Paraphrases for adjective-noun combinations taken from the literature.
easy problem → a problem that is easy to solve (Vendler 1968, page 97)
easy text → text that reads easily (Vendler 1968, page 99)
difficult language → a language that is difficult to speak, learn, write, understand (Vendler 1968, page 99)
careful scientist → a scientist who observes, performs, runs experiments carefully (Vendler 1968, page 92)
comfortable chair → a chair on which one sits comfortably (Vendler 1968, page 98)
good umbrella → an umbrella that functions well (Pustejovsky 1995, page 43)
Table 15
Object-related interpretations for adjective-noun combinations, ranked in order of likelihood.
easy problem easy text difficult language comfortable chair good umbrella
solve −15.14 read −17.42 understand −17.15 sink into −18.66 keep −21.59
deal with −16.12 handle −18.79 interpret −17.59 sit on −19.13 wave −21.61
identify −16.83 use −18.83 learn −17.67 lounge in −19.15 hold −21.73
tackle −16.92 interpret −19.05 use −17.79 relax in −19.33 run for −21.73
handle −16.97 understand −19.15 speak −18.21 nestle in −20.51 leave −22.28
Table 16
Subject-related interpretations for adjective-noun combinations,
ranked in order of likelihood.
easy text good umbrella careful scientist
see −19.22 cover −23.05 calculate −22.31
read −19.50 proceed −22.67
understand −19.66 investigate −22.78
achieve −19.71 study −22.90
explain −20.40 analyze −22.92
combinations and their respective interpretations are taken from the lexical semantics
literature (i.e., Pustejovsky (1995) and Vendler (1968)). The five most likely model-
derived paraphrases for these combinations are shown in Tables 15 and 16.
The model comes up with plausible meanings, some of which overlap with those
suggested in the lexical semantics literature (underlined interpretations indicate agree-
ment between the model and the literature). Observe that the model predicts different
meanings when the same adjective modifies different nouns and derives a cluster of
meanings for a single adjective-noun combination. An easy problem is not only a prob-
lem that is easy to solve (see Vendler’s (1968) identical interpretation in Table 14) but
also a problem that is easy to deal with, identify, tackle, and handle. The meaning of
easy problem is different from the meaning of easy text, which in turn is easy to read,
handle, interpret, and understand. The interpretations the model arrives at for difficult
language are a superset of the interpretations suggested by Vendler (1968). The model
comes up with the additional meanings language that is difficult to interpret and language
that is difficult to use. Although the meanings acquired by the model for careful scientist
do not overlap with the ones suggested by Vendler (1968), they seem intuitively plau-
sible: a careful scientist is a scientist who calculates, proceeds, investigates, studies, and
analyzes carefully. These are all possible actions associated with scientists.
The model derives subject- and object-related interpretations for good umbrella,
which is an umbrella that covers well and is good to keep, good for waving, good to
hold, good to run for, and good to leave. A subject interpretation can be also derived for
291
Lapata and Lascarides Logical Metonymy
easy text. Our parser treats inchoative and noninchoative uses of the same verb as dis-
tinct surface structures (e.g., text that one reads easily vs. text that reads easily); as a result,
read is generated as a subject and object paraphrase for easy text (compare Tables 15
and 16). The object-related interpretation is nevertheless given a higher probability
than the subject-related one, which seems intuitively correct (there is an understood but
unexpressed agent in structures like text that reads easily). In general, subject and object
interpretations are derived on the basis of verb-subject and verb-object constructions
that have been extracted from the corpus heuristically without taking into account in-
formation about theta roles, syntactic transformations (with the exception of passiviza-
tion), or diathesis alternations such as the middle or causative/inchoative alternation.
Although the model can be used to provide several interpretations for a given
adjective-noun combination, not all of these interpretations are useful or plausible
(see the subject interpretations for easy text). Also, the meanings acquired by our model
are a simplified version of the ones provided in the lexical semantics literature. An
adjective-noun combination may be paraphrased with another adjective-noun combi-
nation (e.g., a good meal is a tasty meal) or with an NP instead of an adverb (e.g., a fast
decision is a decision that takes a short amount of time). We are making the simpli-
fying assumption that a polysemous adjective-noun combination can be paraphrased
by a sentence consisting of a verb whose argument is the noun with which the adjec-
tive is in construction (cf. earlier discussion concerning nonmetonymic uses of verbs
like enjoy).
In the following section we evaluate against human judgments the meaning para-
phrases generated by the model. As in the case of verbs, the model is tested on
examples randomly sampled from the BNC, and the linear relationship between the
subjects’ rankings and the model-derived probabilities is explored using correlation
analysis. In Section 3.5.1 we assess whether our model outperforms a naive baseline
in deriving interpretations for metonymic adjectives.
3.3 Experiment 4: Comparison against Human Judgments
The experimental method in Experiment 4 was the same as that in Experiment 1. Mean-
ing paraphrases for adjective-noun combinations were obtained using the model intro-
duced in Section 3.1. The model’s rankings were compared against paraphrase judg-
ments elicited experimentally from human subjects. The comparison between model
probabilities and their perceived likelihood enabled us to explore (1) whether there
is a linear relationship between the likelihood of a given meaning as derived by the
model and its perceived plausibility and (2) whether the model can be used to derive
the argument preferences for a given adjective, that is, whether the adjective is biased
toward a subject or object interpretation or whether it is equibiased.
3.3.1 Method.
3.3.1.1 Materials and Design. We chose nine adjectives according to a set of minimal
criteria and paired each adjective with 10 nouns randomly selected from the BNC. We
chose the adjectives as follows: We first compiled a list of all the polysemous adjectives
mentioned in the lexical semantics literature (Vendler 1968; Pustejovsky 1995). From
these we randomly sampled nine adjectives (difficult, easy, fast, good, hard, right, safe,
slow, and wrong). These adjectives had to be relatively unambiguous syntactically: In
fact, these nine adjectives were unambiguously tagged as “adjectives” 98.6% of the
time, measured as the number of different part-of-speech tags assigned to the word in
the BNC. The nine selected adjectives ranged in BNC frequency from 57.6 per million
to 1,245 per million.
292
Computational Linguistics Volume 29, Number 2
Adjective-noun pairs were extracted from the parser’s output. Recall that the BNC
was parsed using Abney’s (1996) chunker Cass (see Sections 2.2 and 3.2 for details).
From the syntactic analysis provided by the parser, we extracted a table containing the
adjective and the head of the noun phrase following it. In the case of compound nouns,
we included only sequences of two nouns and considered the rightmost-occurring
noun as the head. From the retrieved adjective-noun pairs, we removed all pairs with
BNC frequency of one, as we wanted to reduce the risk of paraphrase ratings being
influenced by adjective-noun combinations unfamiliar to the subjects. Furthermore,
we excluded pairs with deverbal nouns (i.e., nouns derived from a verb) such as fast
programmer, since an interpretation can be easily arrived at for these pairs by mapping
the deverbal noun to its corresponding verb. A list of deverbal nouns was obtained
from two dictionaries, CELEX (Burnage 1990) and NOMLEX (Macleod et al. 1998).
We used the model outlined in Section 3.1 to derive meaning paraphrases for the
90 adjective-noun combinations. We imposed no threshold on the frequencies f(e, a)
and f(rel, e, n). The frequency f(e, a) was obtained by mapping the adjective to its
corresponding adverb: the adjective good was mapped to the adverbs good and well,
the adjective fast was mapped to the adverb fast, easy was mapped to easily, hard was
mapped to hard, right to rightly and right, safe to safely and safe, slow to slowly and slow,
and wrong to wrongly and wrong. The adverbial function of the adjective difficult is
expressed only periphrastically (i.e., in a difficult manner, with difficulty). As a result we
obtained the frequency f(diﬃcult, e) only on the basis of infinitival constructions (see
the examples in (54)). We estimated the probability P(a, n, rel, e) for each adjective-
noun pair by varying both the terms e and rel. We thus derived both subject-related
and object-related paraphrases for each adjective-noun pair.
For each adjective-noun combination, the set of the derived meanings was again
divided into three “probability bands” (high, medium, and low) of equal size, and one
interpretation was selected from each band. The division into bands ensured that the
experimental stimuli represented the model’s behavior for likely and unlikely para-
phrases. We performed separate divisions for object-related and subject-related para-
phrases, resulting in a total of six interpretations for each adjective-noun combination,
as we wanted to determine whether there were differences in the model’s predictions
with respect to the argument function (i.e., object or subject) and also because we
wanted to compare experimentally derived adjective biases against model-derived bi-
ases. Example stimuli (with object-related interpretations only) are shown in Table 17
for each of the nine adjectives.
Our experimental design consisted of the factors adjective-noun pair (Pair), gram-
matical function (Func) and probability band (Band). The factor Pair included 90
adjective-noun combinations. The factor Func had two levels (subject and object),
whereas the factor Band had three levels (high, medium, and low). This yielded a
total of Pair × Func × Band =90×2×3 = 540 stimuli. The number of the stimuli was
too large for subjects to judge in one experimental session. We limited the size of the
design by selecting a total of 270 stimuli according to the following criteria: Our initial
design created two sets of stimuli, 270 subject-related stimuli and 270 object-related
stimuli. For each set of stimuli (i.e., object- and subject-related) we randomly selected
five nouns for each of the nine adjectives, together with their corresponding interpre-
tations in the three probability bands (high, medium, low). This yielded a total of Pair
× Func × Band =45× 2 × 3 = 270 stimuli. In this way, stimuli were created for each
adjective in both subject-related and object-related interpretations.
As in Experiment 1, the stimuli were administered to two separate subject groups
in order to limit the size of the experiment. Each group saw 135 stimuli consisting
of interpretations for all adjective-noun pairs. For the first group five adjectives were
293
Lapata and Lascarides Logical Metonymy
Table 17
Randomly selected example stimuli with log-transformed probabilities derived by the
model.
Probability Band
Adjective-noun
High Medium Low
difficult customer satisfy −20.27 help −22.20 drive −22.64
easy food cook −18.94 introduce −21.95 finish −23.15
fast pig catch −23.98 stop −24.30 use −25.66
good postcard send −20.17 draw −22.71 look at −23.34
hard number remember −20.30 use −21.15 create −22.69
right school apply to −19.92 complain to −21.48 reach −22.90
safe drug release −22.24 try −23.38 start −25.56
slow child adopt −19.90 find −22.50 forget −22.79
wrong color use −21.78 look for −22.78 look at −24.89
represented by object-related meanings only (difficult, easy, good, hard, slow); these adjec-
tives were presented to the second group with subject-related interpretations only. Cor-
respondingly, for the first group, four adjectives were represented by subject-related
meanings only (safe, right, wrong, fast); the second group saw these adjectives with
object-related interpretations.
Each experimental item consisted of an adjective-noun pair and a sentence para-
phrasing its meaning. Paraphrases were created by the experimenters by converting
the model’s output to a simple phrase, usually a noun modified by a relative clause. A
native speaker of English other than the authors was asked to confirm that the para-
phrases were syntactically well-formed (items found syntactically odd were modified
and retested). Example stimuli are shown in (55). A complete list of the experimental
items is given in Appendix B.
(55) a. high: difficult customer
a customer who is difficult to satisfy
b. medium: difficult customer
a customer who is difficult to help
c. low: difficult customer
a customer who is difficult to drive
(56) a. high: fast horse a horse that runs fast
b. medium: fast horse a horse that works fast
c. low: fast horse a horse that sees quickly
3.3.1.2 Procedure. The method used was magnitude estimation, with the same exper-
imental protocol as in Experiment 1.
3.3.1.3 Instructions, Demographic Questionnaire, and Training Phase. The instructions were
the same as in Experiment 1, with the exception that this time the subjects were asked
to judge how well a sentence paraphrased a particular adjective-noun combination.
The demographic questionnaire and the training phase were the same as in Experi-
ment 1.
3.3.1.4 Experimental Phase. Each subject group saw 135 metonymic sentences and their
paraphrases. A modulus item from the medium probability band was provided (see
294
Computational Linguistics Volume 29, Number 2
-26 -24 -22 -20 -18 -16
model probabilities
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
mean subject ratings
Figure 2
Correlation of elicited judgments and model-derived probabilities for metonymic
adjective-noun pairs.
Appendix B). The modulus was the same for all subjects and remained on the screen
the entire time the subject was completing the task. Subjects were assigned to groups
at random, and a random stimulus order was generated for each subject.
3.3.1.5 Subjects. Sixty-five native speakers of English participated in the experiment.
4
The subjects were recruited over the Internet by postings to relevant newsgroups and
mailing lists. Participation was voluntary and unpaid.
The data of one subject were eliminated after inspection of his response times
showed that he had not completed the experiment in a realistic time frame (average
response time < 1000ms). The data of four subjects were excluded because they were
non-native speakers of English.
This left 60 subjects for analysis. Of these, 54 subjects were right-handed, six left-
handed; 22 subjects were female, 38 male. The age of the subjects ranged from 18 to
54 years; the mean was 27.4 years.
3.4 Results
The data were normalized as in Experiment 1. We tested the hypothesis that para-
phrases with high probabilities are perceived as better paraphrases than paraphrases
assigned low probabilities by examining the degree to which the elicited judgments
correlate with the probabilities derived by the model. As in Experiment 1, the data
used for the judgments were the average of log-transformed and normalized subject
ratings per experimental item. The comparison between model probabilities and the
human judgments yielded a Pearson correlation coefficient of .40 (p <.01, N = 270).
Figure 2 plots the relationship between judgments and model probabilities. Descrip-
tive statistics for model probabilities and subject judgments are given in Appendix C.
In order to evaluate whether grammatical function has any effect on the relation-
ship between model-derived meaning paraphrases and human judgments, we split
the items into those that received a subject interpretation and those that received an
4 None of the participants of Experiment 1 took part in Experiment 4.
295
Lapata and Lascarides Logical Metonymy
object interpretation. A comparison between our model and human judgments yielded
a correlation of r = .53 (p <.01, N = 135) for object-related items and a correlation of
r = .21 (p <.05, N = 135) for subject-related items. Note that a weaker correlation is
obtained for subject-related interpretations. One explanation for this weaker correla-
tion could be the parser’s performance; that is, the parser may be better at extracting
verb-object tuples than verb-subject tuples. Another hypothesis (which we test below)
is that most adjectives included in the experimental stimuli have an object bias, and
therefore subject-related interpretations are generally less preferred than object-related
ones.
Using leave-one-out resampling (see Section 2.3.2 for details), we calculated how
well subjects agreed in their judgments concerning metonymic adjective-noun combi-
nations. For the first group, the average intersubject agreement was .67 (Min = .03,
Max = .82, StdDev = .14), and for the second group it was .65 (Min = .05, Max = .82,
StdDev = .14).
The elicited judgments can be further used to derive the grammatical function
preferences (i.e., subject or object) for a given adjective. In particular, we can deter-
mine the preferred interpretation for individual adjectives on the basis of the human
data and then compare these preferences against the ones produced by our model.
Argument preferences can be easily derived from the model’s output by compar-
ing subject-related and object-related paraphrases. For each adjective we gathered
the subject- and object-related interpretations derived by the model and performed
a one-way analysis of variance (ANOVA) in order to determine the significance of the
grammatical function effect.
We interpret a significant effect as bias toward a particular grammatical function.
We classify a particular adjective as object-biased if the mean of the model-derived
probabilities for the object interpretation of that adjective is significantly larger than the
mean for the subject interpretation; subject-biased adjectives are classified through a
comparable procedure, whereas adjectives for which no effect of grammatical function
is found are classified as equibiased. The effect of grammatical function was significant
for the adjectives difficult (F(1, 1806)=8.06, p <.01), easy (F(1, 1511)=41.16, p <.01),
hard (F(1, 1310)=57.67, p <.01), safe (F(1, 382)=5.42, p <.05), right (F(1, 2114)=9.85,
p <.01), and fast (F(1, 92)=4.38, p <.05). The effect of grammatical function was
not significant for the adjectives good (F(1, 741)=3.95, p = .10), slow (F(1, 759)=
5.30, p = .13), and wrong (F(1, 593)=1.66, p = .19). Table 18 shows the biases for
the nine adjectives as derived by our model. A check mark next to a grammatical
function indicates that its effect was significant in that particular instance, as well as
the direction of the bias.
Ideally, we would like to elicit argument preferences from human subjects in a
similar fashion. However, since it is impractical to elicit judgments experimentally for
all paraphrases derived by the model, we will obtain argument preferences from the
judgments based on the restricted set of experimental stimuli, under the assumption
that they correspond to a wide range of model paraphrases (i.e., they correspond
to a wide range of probabilities) and therefore are representative of the entire set of
model-derived paraphrases. This assumption is justified by the fact that items were
randomly chosen from the three probability bands (i.e., high, medium, low). For each
adjective we gathered the elicited responses pertaining to subject- and object-related
interpretations and performed an ANOVA.
The ANOVA indicated that the grammatical function effect was significant for the
adjective difficult in both by-subjects (subscript 1) and by-items (subscript 2) analyses
(F
1
(1, 58)=17.98, p <.01; F
2
(1, 4)=53.72, p <.01), and for the adjective easy in the
by-subjects analysis only (F
1
(1, 58)=10, p <.01; F
2
(1, 4)=8.48, p = .44). No effect of
296
Computational Linguistics Volume 29, Number 2
Table 18
Log-transformed model-derived and subject-based argument preferences for polysemous
adjectives.
Adjective Model Mean StdDev StdEr Subjects Mean StdDev StdEr
difficult
√
OBJ −21.62 1.36 .04
√
OBJ .0745 .3753 .0685
SUBJ −21.80 1.34 .05 SUBJ −.2870 .2777 .0507
easy
√
OBJ −21.60 1.51 .05
√
OBJ .1033 .3364 .0614
SUBJ −22.11 1.36 .06 SUBJ −.1437 .2308 .0421
fast OBJ −24.20 1.27 .13 OBJ −.3544 .2914 .0532
√
SUBJ −23.80 1.40 .14
√
SUBJ −.1543 .4459 .0814
good OBJ −22.12 1.28 .06 OBJ −.0136 .3898 .0712
SUBJ −22.27 1.10 .07 SUBJ −.1563 .2965 .0541
hard
√
OBJ −21.69 1.53 .06
√
OBJ .0030 .3381 .0617
SUBJ −22.12 1.35 .06 SUBJ −.2543 .2436 .0445
right
√
OBJ −21.65 1.36 .04
√
OBJ −.0054 .2462 .0450
SUBJ −21.84 1.24 .04 SUBJ −.2413 .4424 .0808
safe OBJ −22.75 1.48 .10
√
OBJ .0037 .2524 .0461
√
SUBJ −22.39 1.59 .12 SUBJ −.3399 .4269 .0779
slow OBJ −22.49 1.53 .08 OBJ −.3030 .4797 .0876
SUBJ −22.32 1.50 .07
√
SUBJ −.0946 .2357 .0430
wrong OBJ −23.15 1.33 .08
√
OBJ −.0358 .2477 .0452
SUBJ −23.29 1.30 .08 SUBJ −.2356 .3721 .0679
grammatical function was found for good (F
1
(1, 58)=2.55, p = .12; F
2
(1, 4)=1.01,
p = .37). The effect of grammatical function was significant for the adjective hard
in the by-subjects analysis only (F
1
(1, 58)=11.436, p <.01; F
2
(1, 4)=2.84, p = .17),
whereas for the adjective slow the effect was significant both by subjects and by items
(F
1
(1, 58)=4.56, p <.05; F
2
(1, 4)=6.94, p = .058). For safe and right the main effect
was significant in both by-subjects and by-items analyses (F
1
(1, 58)=14.4, p <.0005;
F
2
(1, 4)=17.76, p <.05, and F
1
(1, 58)=6.51, p <.05; F
2
(1, 4)=15.22, p = .018, respec-
tively). The effect of grammatical function was significant for wrong and fast only by
subjects (F
1
(1, 58)=5.99, p = .05; F
2
(1, 4)=4.54, p = .10, and F
1
(1, 58)=4.23, p = .05;
F
2
(1, 4)=4.43, p = .10). The biases for these adjectives are shown in Table 18. Check
marks again indicate instances in which the grammatical function effect was significant
(as determined from the by-subjects analyses), as well as the direction of the bias.
We expect a valid model to assign on average higher probabilities to object-related
interpretations and lower probabilities to subject-related interpretations for an object-
biased adjective; accordingly, we expect the model to assign on average higher prob-
abilities to subject-related interpretations for subject-biased adjectives. Comparison of
the biases derived from the model with ones derived from the elicited judgments
shows that the model and the humans are in agreement for all adjectives but slow,
wrong, and safe. On the basis of human judgments, slow has a subject bias and wrong
has an object bias (see Table 18). Although the model could not reproduce this result,
there was a tendency in the right direction.
Note that in our correlation analysis reported above, the elicited judgments were
compared against model-derived paraphrases without taking argument preferences
297
Lapata and Lascarides Logical Metonymy
into account. We would expect a valid model to produce intuitive meanings at least
for the interpretation that a given adjective favors. We further examined the model’s
behavior by performing separate correlation analyses for preferred and dispreferred
biases as determined previously by the ANOVAs conducted for each adjective (see
Table 18). Since the adjective good was equibiased, we included both biases (i.e., object-
related and subject-related) in both correlation analyses for that adjective. The com-
parison between our model and the human judgments yielded a Pearson correlation
coefficient of .52 (p <.01, N = 150) for the preferred interpretations and a correlation
of .23 (p <.01, N = 150) for the dispreferred interpretations. The result indicates that
our model is particularly good at deriving meanings corresponding to the argument
bias for a given adjective. However, the dispreferred interpretations also correlate sig-
nificantly with human judgments, which suggests that the model derives plausible
interpretations even in cases in which the argument bias is overridden.
In sum, the correlation analysis supports the claim that adjective-noun paraphrases
with high probability are judged more plausible than those with low probability. It also
suggests that the meaning preference ordering produced by the model is intuitively
correct, since subjects’ perception of likely and unlikely meanings correlates with the
probabilities assigned by the model.
The probabilistic model evaluated here explicitly takes adjective/adverb and verb
co-occurrences into account. However, one could derive meanings for polysemous
adjective-noun combinations by concentrating solely on verb-noun relations, ignoring
thus the adjective/adverb and verb dependencies. For example, in order to interpret
the combination easy problem, we could simply take into account the types of activities
that are related to problems (i.e., solving them, giving them, etc.). This simplification
is consistent with Pustejovsky’s (1995) claim that polysemous adjectives like easy are
predicates, modifying some aspect of the head noun and more specifically the events
associated with the noun. A naive baseline model would be one that simply takes into
account the number of times the noun in the adjective-noun pair acts as the subject
or object of a given verb, ignoring the adjective completely. This raises the question
of how well such a naive model would perform at deriving meaning paraphrases for
polysemous adjective-noun combinations.
In the following section we present such a naive model of adjective-noun pol-
ysemy. We compare the model’s predictions against the elicited judgments. Using
correlation analysis we attempt to determine whether the naive model can provide an
intuitively plausible ranking of meanings (i.e., whether perceived likely and unlikely
meanings are given high and low probabilities, respectively). We further compare the
naive model to our initial model (see Section 3.1) and discuss the differences between
them.
3.5 Experiment 5: Comparison against Naive Baseline
3.5.1 Naive Baseline Model. Given an adjective-noun combination, we are interested
in finding the events most closely associated with the noun modified by the adjec-
tive. In other words we are interested in the verbs whose object or subject is the
noun appearing in the adjective-noun combination. This can be simply expressed as
P(e | rel, n), the conditional probability of a verb e given an argument-noun relation
rel, n:
P(e | rel, n)=
f(e, rel, n)
f(rel, n)
(57)
The model in (57) assumes that the meaning of an adjective-noun combination
is independent of the adjective in question. Consider, for example, the adjective-noun
298
Computational Linguistics Volume 29, Number 2
pair fast plane. We need to find the verbs e and the argument relation rel that maximize
the probability P(e | rel, plane). In the case of fast plane, the verb that is most frequently
associated with planes is fly (see Table 13). Note that this model will come up with
the same probabilities for fast plane and wrong plane, since it does not take the identity
of the modifying adjective into account. We estimated the frequencies f(e, rel, n) and
f(rel, n) from verb-object and verb-subject tuples extracted from the BNC using Cass
(Abney 1996) (see Section 2.2 for details on the extraction and filtering of the argument
tuples).
3.6 Results
Using the naive model we calculated the meaning probability for each of the 270
stimuli included in Experiment 4 and explored the linear relationship between the
elicited judgments and the naive baseline model through correlation analysis. The
comparison yielded a Pearson correlation coefficient of .25 (p <.01, N = 270). Recall
that we obtained a correlation of .40 (p <.01, N = 270) when comparing our original
model to the human judgments. Not surprisingly the two models are intercorrelated
(r = .38, p <.01, N = 270). An important question is whether the difference between
the two correlation coefficients (r = .40 and r = .25) is due to chance. A one-tailed t-
test revealed that the difference between them was significant (t(267)=2.42, p <.01).
This means that our original model (see Section 3.1) performs reliably better than the
naive baseline at deriving interpretations for metonymic adjective-noun combinations.
We further compared the naive baseline model and the human judgments sepa-
rately for subject-related and object-related items. The comparison yielded a correlation
of r = .29 (p <.01, N = 135) for object interpretations. Recall that our original model
yielded a correlation coefficient of .53 (see Section 3.4). A one-tailed t-test revealed that
the two correlation coefficients were significantly different (t(132)=3.03, p <.01). No
correlation was found for the naive model when compared against elicited subject
interpretations (r = .09, p = .28, N = 135).
3.7 Discussion
We have demonstrated that the meanings acquired by our probabilistic model corre-
late reliably with human intuitions. Our model not only acquires clusters of meanings
(following Vendler’s (1968) insight) but furthermore can be used to obtain a tripartite
distinction of adjectives depending on the type of paraphrase they prefer: subject-
biased adjectives tend to modify nouns that act as subjects of the paraphrasing verb,
and object-biased adjectives tend to modify nouns that act as objects of the paraphras-
ing verb, whereas equibiased adjectives display no preference for either argument role.
A comparison between the argument preferences produced by the model and human
intuitions revealed that most of the adjectives we examined (six out of nine) display a
preference for an object interpretation (see Table 18), two adjectives are subject-biased
(i.e., fast, slow) and one adjective is equibiased (i.e., good).
We rigorously evaluated
5
the results of our model by eliciting paraphrase judg-
ments from subjects naive to linguistic theory. Comparison between our model and
human judgments yielded a reliable correlation of .40 when the upper bound for the
task (i.e., intersubject agreement) is approximately .65. We have demonstrated that a
naive baseline model that interprets adjective-noun combinations by focusing solely
on the events associated with the noun is outperformed by a more detailed model that
5 We have not compared the model’s predictions against norming data for adjectival metonymies
primarily because such data were not available to us. To our knowledge McElree et al.’s (2001) study is
the only example of a norming study on logical metonymy; however, it concentrates only on verbs.
299
Lapata and Lascarides Logical Metonymy
considers not only verb-argument relations but also adjective-verb and adverb-verb
dependencies.
Although the events associated with the different nouns are crucially important for
the meaning of polysemous adjective-noun combinations, it seems that more detailed
linguistic knowledge is needed in order to produce intuitively plausible interpreta-
tions. This is by no means surprising. As a simple example, consider the adjective-
noun pair fast horse. A variety of events are associated with the noun horse, yet only
a subset of those are likely to occur fast. The three most likely interpretations for fast
horse according to the naive model are a horse that needs something fast, a horse that gets
something fast, and a horse that does something fast. A model that uses information about
verb-adjective or verb-adverb dependencies provides a more plausible ranking: a fast
horse is a horse that runs, learns, or goes fast. A similar situation arises when one con-
siders the pair careful scientist. According to the naive model, a careful scientist is more
likely to believe, say, or make something carefully. However, none of these events is
particularly associated with the adjective careful.
Although our experiments on adjectival logical metonymy revealed a reliable cor-
relation between the model’s ranking and human intuitions, the fit between model
probabilities and elicited judgments was lower for metonymic adjectives (r = .40)
than for metonymic verbs (r = .64). One explanation for this lower degree of fit for
adjectives is that the semantic restrictions that adjectives impose on the nouns with
which they combine appear to be less strict than the ones imposed by verbs (con-
sider, for example, the adjective good, which can combine with nearly any noun). A
consequence of this is that metonymic adjectives seem to allow a wider range of in-
terpretations than verbs. This means that there will be a larger variation in subjects’
responses to the generated model paraphrases when it comes to adjectives than when
it comes to verbs, thus affecting the linear relationship between our model and the
elicited judgments. Our hypothesis is further supported by the intersubject agreement,
which is lower for metonymic adjectives than for verbs (.65 versus .74). As explained
in the previous sections, our model does not take into account the wider context
within which an adjective-noun combination is found. Precisely because of the ease
with which some adjectives combine with practically any noun, it may be the case
that more information (i.e., context) is needed in order to obtain intuitively plausible
interpretations for metonymic adjectives. The model presented here can be extended
to incorporate contextual information (e.g., intra- and extrasentential information), but
we leave this to future work.
4. General Discussion
In this article we have focused on the automatic interpretation of logical metonymy.
We have shown how meaning paraphrases for metonymic expressions can be acquired
from a large corpus and have provided a probabilistic model that derives a preference
ordering on the set of possible meanings in an unsupervised manner, without relying
on the availability of a disambiguated corpus. The proposed approach utilizes surface
syntactic analysis and distributional information that can be gleaned from a corpus
while exploiting correspondences between surface cues and meaning.
Our probabilistic model reflects linguistic observations about the nature of metony-
mic constructions: It predicts variation in interpretation for different verbs and adjec-
tives with respect to the noun for which they select and is faithful to Vendler’s (1968)
claim that metonymic expressions are usually interpreted by a cluster of meanings
instead of a single meaning. This contrasts with Pustejovsky’s (1995) approach, which
typically assigns a single reading to the metonymic construction, and with the account
300
Computational Linguistics Volume 29, Number 2
put forward by Copestake and Briscoe (1995), which assigns one interpretation con-
ventionally, albeit a default interpretation—in fact this interpretation might also be of
quite a general type (e.g., the event argument to enjoy in enjoy the pebble can be assigned
the general type act-on). Our model provides plausible alternatives to the default, and
it augments the general semantic type in the interpretation that’s assigned by these the-
ories with a plausible range of more specific values, although, in contrast to the hybrid
model of interpretation described in Copestake and Lascarides (1997), it does not pre-
dict the contexts in which a statistically dispreferred interpretation is the correct one.
Our approach can be viewed as complementary to linguistic theory: Although our
model does not identify odd metonymies in the way that a rule-based model might
(we return to this in Section 4.1), it does automatically derive a ranking of meanings,
thus distinguishing likely from unlikely interpretations. Even if linguistic theory is
able to enumerate all possible interpretations for a given adjective (note that in the
case of polysemous adjectives, we would have to take into account all nouns or noun
classes that the adjective could possibly modify), in most cases it does not indicate
which ones are likely and which ones are not. Our model fares well on both tasks.
It recasts the problem of logical metonymy in a probabilistic framework and derives
a large number of interpretations not readily available from linguistic introspection.
The information acquired from the corpus can be also used to quantify the argument
preferences of metonymic adjectives. These are only implicit in the lexical semantics
literature, in which certain adjectives are exclusively given a verb-subject or verb-object
interpretation.
4.1 Limitations and Extensions
We chose to model metonymic constructions and their meanings as joint distribu-
tions of interdependent linguistic events (e.g., verb-argument relations, verb-adverb
modification). Although linguistically informed, our approach relies on approxima-
tions and simplifying assumptions, partly motivated by the absence of corpora explic-
itly annotated with metonymic interpretations. We generate meaning paraphrases for
metonymic constructions solely from co-occurrence data without taking advantage of
taxonomic information. Despite the simplicity of this approach and its portability to
languages for which lexical semantic resources may not be available, there are certain
regularities about the derived interpretations that our model fails to detect.
Consider the interpretations produced for begin song, repeated here from Table 7:
sing, rehearse, write, hum, and play. Our model fails to capture the close correspondence
for some of these meanings. For example, hum and sing are sound emission verbs; they
additionally entail the performance or execution of the song. The verbs rehearse and
play capture only the performance aspect and can be thus considered supertypes of
hum and sing (one can play or rehearse a song by humming it, singing it, drumming
it, whistling it, etc.). The verb write, on the other hand, is neither a performance nor a
sound emission verb; it has to do with communication and creation. Another example
is comfortable chair, for which the model generates sink into, sit on, lounge in, relax in, and
nestle in (see Table 15). The verbs sink into, sit on, lounge in, and nestle in describe the
position one assumes when sitting in the chair, whereas relax in refers to the state of
the person in the chair. There is no notion of semantic proximity built into the model,
and correspondences among semantically related interpretations are not automatically
recognized.
An alternative to the knowledge-free approach advocated here is to use taxonomic
information to obtain some degree of generalization over the acquired interpretations.
Semantic classifications such as WordNet (Miller et al., 1990) or that in Levin (1993)
can be used to group the obtained verbs into semantically coherent classes. Further-
301
Lapata and Lascarides Logical Metonymy
more, the WordNet semantic hierarchy can be used to estimate directly probabilities
involving either nouns (e.g., P(rel | e, n), P(o | e)) or verbs (e.g., P(v | e), P(a | e)). Prob-
abilities can be defined in terms of senses from a semantic hierarchy by exploiting
the fact that the senses can be grouped into classes consisting of semantically similar
senses (Resnik 1993; Clark and Weir 2001; McCarthy 2000; Li and Abe 1998). So the
probability P(book | read) can be estimated by taking into account nouns that belong
to the same semantic class as book and can be read (e.g., journals, novels, scripts) or
by focusing on verbs that are semantically related to read (e.g., interpret, communi-
cate, understand). Note that estimation of probabilities over classes rather than words
can effectively overcome data sparseness and potentially lead to better probability
estimates. Currently our models cannot estimate probabilities for word combinations
unseen in the corpus, and WordNet could be used for re-creating the frequencies of
these combinations (Lapata, Keller, and McDonald 2001; Clark and Weir 2001). How-
ever, part of our aim here was to investigate whether it is at all possible to generate
interpretations for metonymic constructions without the use of prior knowledge bases
that might bias the acquisition process in uncontrolled and idiosyncratic ways.
A related issue is the fact that our models are ignorant about the potentially differ-
ent senses of the noun in the metonymic construction. For example, the combination
fast plane may be a fast aircraft, or a fast tool, or a fast geometrical plane. Our model
derives meanings related to all three senses of the noun plane. For example, a fast plane
is not only a plane (i.e., an aircraft) that flies, lands, or travels quickly, but also a plane
(i.e., a surface) that transposes or rotates quickly and a plane (i.e., a tool) that smoothes
something quickly. However, more paraphrases are derived for the “aircraft” sense of
plane; these paraphrases also receive a higher ranking. This is not surprising, since
the number of verbs related to the “aircraft” sense of plane are more frequent than the
verbs related to the other two senses. In contrast to fast plane, however, efficient plane
should probably bias toward the “tool” sense of plane, even though the “aircraft”
sense is more frequent in the corpus. One could make the model sensitive to this by
investigating the synonyms for the various senses of plane; moreover the “tool” sense
bias of efficient plane could also be inferred on the basis that efficient plane co-occurs
with different verbs from fast plane. There are also cases in which a model-derived
paraphrase does not provide disambiguation clues with respect to the meaning of the
noun. Consider the adjective-noun combination fast game. The model comes up with
the paraphrases game that runs fast and game that goes fast. Both paraphrases may well
refer to either the “contest,” “activity,” or “prey” sense of game. Note finally that our
model can be made sensitive to word sense distinctions by taking into account noun
classes rather than word forms; this modification would allow us to apply the model
to word sense–disambiguated metonymic expressions.
In this article we have focused on the automatic interpretation of logical metonymy
without explicitly dealing with the recognition of verbs or adjectives undergoing logical
metonymy. In default of the latter study, which we plan for the future, we sketch here
briefly how our proposal can be extended to recognizing logical metonymies. A very
simple approach would be to use the proposed model to generate interpretations for
metonymic and nonmetonymic constructions. The derived paraphrases and the range
of their probabilities could be then used to quantify the degree of “metonymic-ness”
of a given verb or adjective. One would expect that a larger number of paraphrases
would be generated for verbs or adjectives for which logical metonymy is possible. We
tested this hypothesis using a metonymic verb (i.e., enjoy) and a nonmetonymic one
(i.e., play). Enjoy is attested 5,344 times in the BNC in a verb-object relation, whereas
play is attested 12,597 times (again these numbers are based on information extracted
using Cass (Abney 1996)). Using the model presented in Section 2.1 we generated
302
Computational Linguistics Volume 29, Number 2
Table 19
Model-derived paraphrases for odd metonymies, ranked in order of likelihood.
begin dictionary begin rock begin keyboard begin highway
compile −19.32 crunch across −18.09 use −20.11 obstruct −20.40
flick through −19.59 climb −18.39 play −20.44 regain −20.79
use −19.80 run towards −18.70 operate −20.56 build −20.80
publish −20.34 percolate through −18.78 assemble −20.78 use −20.81
advance −20.39 dissolve −19.37 tune −20.80 detach −20.82
meaning paraphrases for all verb-object tuples found for enjoy and play. The model
generated 44,701 paraphrases for enjoy and 9,741 for play. Comparison between the
probabilities assigned to the interpretations for enjoy and play revealed that the para-
phrases obtained for enjoy were on average more likely (mean = −23.53, min = −27.02,
max = −15.32) than those discovered for play (mean = −24.67, min = −28.25, max
= −17.24). The difference was statistically significant (using an independent-samples
t-test; t(54, 440)=2.505, p <.01). This result indicates that enjoy is more likely to un-
dergo logical metonymy than play. Another potential indicator of metonymic use is
the likelihood that a given verb or adjective will be found in a certain syntactic con-
struction. Consider the adjective blue, for which our model (see Section 3.1) does not
generate any meaning paraphrases, presumably because blue is not attested as a verb
modifier (in contrast to adjectives like easy or fast).
Such an approach could potentially predict differences in productivity among
metonymic verbs. One would expect the metonymic uses of attempt, for example, to
be much less productive than the metonymic uses of enjoy and begin. One could con-
ceivably predict this on the basis of the frequency and diversity of NPs in attempt NP
constructions that are attested in the corpus, compared with those for enjoy NP and
begin NP. Furthermore, the model generates paraphrases for attempt that are on aver-
age less likely in comparison to those generated for enjoy or begin. However, we leave
this for future work.
As argued in Section 2.1, our model cannot distinguish between well-formed and
odd metonymic constructions; in fact, it will generally provide meaning paraphrases
even for combinations that are deemed by native speakers to be odd. Consider the
examples in (58) and their interpretations in Table 19. In general the paraphrases gen-
erated for problematic data are of worse quality than those produced for well-formed
metonymies. In most cases the model will generate unavailable interpretations (see be-
gin highway, begin keyboard). Consider, however, the pair begin rock. Pustejovsky (1995)
observes that although there is no generally available interpretation for a sentence like
Mary began the rock, because of what we understand begin to require of its argument
and our knowledge of what rocks are and what you can do to them, as speakers
and hearers we tend to accommodate information into the context so as to interpret
otherwise ill-formed expressions. Our model generates meaning paraphrases that are
relatively plausible assuming different pragmatic contexts for begin rock. One can begin
climbing or running towards a rock. Someone’s footsteps can crunch across a frozen rock,
a material can percolate through a rock, and rain water can dissolve a rock.
(58) a. ?John began the dictionary.
b. ?Mary began the rock.
c. *John began a keyboard.
d. *John began the highway.
303
Lapata and Lascarides Logical Metonymy
Table 20
Descriptives for odd and well-formed metonymies.
Well-formed N Mean StdDev StdErr
begin book 534 −21.54 1.515 .066
begin cigarette 104 −21.83 1.613 .158
begin coffee 104 −22.03 1.626 .159
begin story 381 −21.73 1.493 .076
easy problem 358 −21.14 1.606 .085
Odd N Mean StdDev StdErr
begin dictionary 76 −22.48 1.440 .165
begin keyboard 50 −22.40 1.337 .189
begin rock 50 −23.55 1.376 .193
begin highway 37 −22.52 1.337 .219
easy programmer 49 −23.23 1.289 .184
Despite the fact that the model does not recognize odd metonymies, one would
expect low probabilities to be assigned to ungrammatical constructions. Table 20 re-
ports some descriptive statistics on well-formed (top half) and odd (bottom half)
metonymies taken from the lexical semantics literature (Verspoor 1997; Pustejovsky
1995). A higher number of interpretations is generated for well-formed metonymies.
Using an independent-samples t-test, we can compare the differences in the probabili-
ties assigned to the two types of metonymies. Take, for example, begin dictionary: the av-
erage probability of its interpretations is lower than those for begin book (t(608)=5.07,
p <.01), begin cigarette (t(178)=2.77, p <.01), begin coffee (t(178)=2.1, p <.05), and
begin story (t(455)=4.1, p <.01). Similar results are obtained when comparing be-
gin keyboard against the well-formed metonymies in Table 19: the probability of its
interpretations is on average lower than those assigned to begin book,(t(582)=3.87,
p <.01), begin cigarette (t(152)=2.15, p <.01), and begin story (t(429)=3.02, p <.01).
The difference between begin coffee and begin keyboard is not statistically significant
(t(152)=1.39, p = 0.167). However, the mean for begin coffee is slightly higher than
that for begin keyboard. Although here we focus on verbs, similar comparisons can be
applied to adjectives. As shown in Table 20, the probabilities for easy programmer are
on average lower than those for easy problem (t(405)=8.74, p <.01).
Finally, recall from Section 2.1 that on the basis of Gricean reasoning, one would
expect to find in a corpus well-formed metonymies more often than their paraphrases
(see Tables 1–3). Following this line of reasoning, one might expect for conventionally
odd metonymies the opposite situation, that is, to find the paraphrases more often
than the metonymies proper. We tested this hypothesis for some examples cited in
the literature (Verspoor 1997; Pustejovsky 1995) by examining whether paraphrases
corresponding to odd metonymies are attested in the BNC as VP complements. We
found plausible paraphrases in the BNC for almost all verb-noun pairs illustrated
in Table 21. This suggests that the corpus data relating to the odd and well-formed
examples are largely compliant with the Gricean predictions. Corpus co-occurrences
of verb-noun combinations and their paraphrases could be exploited in creating a
system aimed at quantifying the grammaticality of metonymic expressions. However,
it is beyond the scope of the present study to develop such a system.
304
Computational Linguistics Volume 29, Number 2
Table 21
BNC frequencies for odd metonymic expressions.
Odd begin NP begin V-ing NP
begin chair 0 9
begin tunnel 0 4
begin keyboard 0 0
begin tree 1 13
begin highway 0 2
begin film 0 7
begin nail 0 4
begin door 0 18
begin dictionary 0 3
begin rock 0 17
Table 22
Five most likely interpretations for good author and good language.
good author good language
write SUBJ −21.81 use OBJ −17.39
work SUBJ −21.97 verse in OBJ −17.89
describe SUBJ −22.03 speak OBJ −18.32
know OBJ −22.05 know OBJ −19.18
engage with OBJ −22.23 learn OBJ −19.36
4.2 Relevance for NLP Applications
The meaning paraphrases discovered by our model could be potentially useful for
a variety of NLP tasks. One obvious application is natural language generation. For
example, a generator that has knowledge of the fact that fast plane corresponds to a
plane that flies fast can exploit this information either to render the text shorter (in cases
in which the input representation is a sentence) or longer (in cases in which the in-
put representation is an adjective-noun pair). Information retrieval is another relevant
application. Consider a search engine faced with the query fast plane. Presumably one
would not like to obtain information about planes in general or about planes that go
down or burn fast, but rather about planes that fly or travel fast. So knowledge about
the most likely interpretations of fast plane could help rank relevant documents before
nonrelevant ones or restrict the number of documents retrieved.
Note that in the case of adjectives, it is not just the paraphrase, but also the
grammatical function, that needs to be determined. How to render an adjective-noun
combination with an object- or subject-related paraphrase can be worked out by com-
puting the biases discussed in Section 3.4. So if we know that fast has a subject bias,
we can concentrate only on the subject-related interpretations. The choice of grammat-
ical function is less straightforward in the case of equibiased adjectives. In fact, it is
possible that the interpretation for a particular adjective varies depending on the noun
it modifies. For example, a good author writes well, whereas a good language is good to
learn, hear, or study. A simple way to address this is to select the interpretations with
the highest probability. For good author and good language, the five most likely inter-
pretations (and their grammatical functions) according to the model (see Section 3.3)
are given in Table 22. As can be seen from the table, subject-related interpretations are
ranked higher for good author; the opposite is true for good language. Another possibil-
305
Lapata and Lascarides Logical Metonymy
ity for determining the grammatical function for equibiased adjectives is to compare
verb-object and verb-subject interpretations directly for a particular adjective-noun
combination. As an example, consider the following. For good author, the model pro-
duces 107 object-related paraphrases and 199 subject-related ones. Furthermore, the
subject-related probabilities are on average higher than the object-related ones, and the
difference is statistically significant (using a one-tailed t-test, t(304)=3.26, p <.01).
For good language there are 253 object-related paraphrases and 180 subject-related ones.
The former are assigned higher probabilities than the latter, and the difference is sta-
tistically significant (t(431)=3.80, p <.01).
Machine translation is another related application. A logical metonymy may be
acceptable in a source language but unacceptable in the target language. Consider the
example in (59): Its direct translation into German produces a semantically unaccept-
able sentence (see (60a)). In this case we need to spell out the metonymy in order to
obtain an acceptable translation for German, and our model can be used to provide
the missing information by generating meaning paraphrases. Under such an approach,
we would not translate (59) directly, but one of its paraphrases (see (60b) and (60c)).
(59) Peter attempted the peak.
(60) a. Peter hat den Gipfel versucht.
Peter has the peak attempted
‘Peter attempted the peak.’
b. Peter hat den Gipfel zu besteigen versucht.
Peter has the peak to climb attempted
‘Peter attempted to climb the peak.’
c. Peter hat den Gipfel zu erreichen versucht.
Peter has the peak to reach attempted
‘Peter attempted to reach the peak.’
5. Related Work
In contrast to the extensive theoretical literature on the topic of logical metonymy, little
attention has been paid to the phenomenon from an empirical perspective. Briscoe,
Copestake, and Boguraev (1990) and Verspoor (1997) undertake a manual analysis of
logical metonymies found in naturally occurring text. Their results show that logical
metonymy is a relatively widespread phenomenon and that most metonymic exam-
ples can be interpreted on the basis of the head noun’s qualia structure, assuming a
theoretical framework similar to Pustejovsky’s (1991). Verspoor’s (1997) analysis of the
metonymic verbs begin and finish demonstrates that context plays a relatively small
role in the interpretation of these verbs: 95.0% of the logical metonymies for begin and
95.6% of the logical metonymies for finish can be resolved on the basis of information
provided by the noun for which the verb selects. Briscoe, Copestake, and Boguraev’s
(1990) work further suggests ways of acquiring qualia structures for nouns by com-
bining information extracted from machine-readable dictionaries and corpora. Our
probabilistic formulation of logical metonymy allows us to discover interpretations
for metonymic constructions without presupposing the existence of qualia structures.
In fact, we show that a simple statistical learner in combination with a shallow syn-
tactic analyzer yields relatively intuitive results, considering the simplifications and
approximations in the system.
306
Computational Linguistics Volume 29, Number 2
Perhaps more relevant to the work presented here are previous approaches to the
automatic interpretation of general metonymy (Lakoff and Johnson 1980; Nunberg
1995). This is slightly different from logical metonymy, in that the examples aren’t
usually analyzed in terms of semantic type coercion. But the phenomena are closely
related. Generally speaking an expression A is considered a metonymy if A deviates
from its literal denotation in that it stands for an entity B that is not expressed explicitly
but is conceptually related to A via a contiguity relation r (Markert and Hahn 1997).
A typical example of general metonymy is given in (61): in (61a) the bottle stands for
its content (i.e., the liquid in the bottle), and in (61b) Shakespeare stands for his works.
The contiguity relation r between the bottle and its liquid is Container for Contents; for
Shakespeare and his works the contiguity relation is Producer for Product.
(61) a. Denise drank the bottle.
b. Peter read Shakespeare.
Previous approaches to processing metonymy typically rely heavily on the avail-
ability of manually constructed knowledge bases or semantic networks (Fass 1991;
Inverson and Helmreich 1992; Bouaud, Bachimont, and Zweigenbaum 1996; Hobbs et
al. 1993). Furthermore, most implementations either contain no evaluation (Fass 1991;
Inverson and Helmreich 1992; Hobbs et al. 1993) or report results on the development
data (Bouaud, Bachimont, and Zweigenbaum 1996).
The approach put forward by Utiyama, Murata, and Isahara (2000) is perhaps the
most comparable to our own work. Utiyama et al. describe a statistical approach to
the interpretation of general metonymies for Japanese. Utiyama et al.’s algorithm in-
terprets verb-object metonymies by generating the entities for which the object stands.
These entities are ranked using a statistical measure. Given an expression like (62),
nouns related to Shakespeare are extracted from the corpus (e.g., si ‘poem’, tyosyo ‘writ-
ings’, sakuhin ‘works’) and ranked according to their likelihood. Two types of syntactic
relations are used as cues for the interpretation of metonymic expressions: (1) the noun
phrase AnoB, roughly corresponding to the English BofA, where A is the noun figur-
ing in the metonymic expression (e.g., Shakespeare in (62)) and B is the noun it stands
for (e.g., sakuhin ‘works’), and (2) nouns co-occurring with the target noun (e.g., Shake-
speare) within the target sentence.
(62) Shakespeare wo yomu
Shakespeare ACC read
‘read Shakespeare’
Given a metonymy of the form ARV, the appropriateness of a noun B as an
interpretation of A is defined as
L
Q
(B | A, R, V)=
P(B | A, Q)P(R, V | B)
P(R, V)
(63)
where V is the verb in the metonymic expression, A is its object, R is A’s case marker
(e.g., wo (accusative)), B is the noun A stands for, and Q is the relation Q bears to
A (e.g., no). The probabilities in (63) are estimated from a large, morphologically
analyzed Japanese corpus of newspaper texts (approximately 153 million words). A
Japanese thesaurus is used for the estimation of the term P(R, V | B) when the fre-
quency f(R, V, B) is zero (see Utiyama, Murata, and Isahara (2000) for the derivation
307
Lapata and Lascarides Logical Metonymy
and estimation of (63)). Utiyama et al.’s approach is tested on 75 metonymies taken
from the literature. It achieves a precision of 70.6% as measured by one of the au-
thors according to the following criterion: A metonymic interpretation was considered
correct if it made sense in some context.
Our approach is conceptually similar to that of Utiyama, Murata, and Isahara
(2000). Metonymies are interpreted using corpora as the inventory of the missing
information. In contrast to Utiyama et al., we use no information external to the cor-
pus (e.g., a thesaurus); sparse-data problems are tackled via independence assump-
tions, and syntactic information is obtained through shallow text analysis (Utiyama,
Murata, and Isahara (2000) rely on morphological analysis to provide cues for syn-
tactic information). The most striking difference, however, between our work and
Utiyama et al.’s is methodological. Their evaluation is subjective and limited to ex-
amples taken from the literature. The appropriateness of their statistical measure
(see (63)) is not explored, and it is not clear whether it can derive an intuitively
plausible ranking of interpretations or whether it can extend to examples found in
naturally occurring text. We test our probabilistic formulation of logical metonymy
against a variety of examples taken from the corpus, and the derived interpretations
are evaluated objectively using standard experimental methodology. Furthermore, the
appropriateness of the proposed model is evaluated via comparisons to a naive base-
line.
6. Conclusions
In this article we proposed a statistical approach to logical metonymy. We acquired the
meanings of metonymic constructions from a large corpus and introduced a probabilis-
tic model that provides a ranking on the set of possible interpretations. We identified
semantic information automatically by exploiting the consistent correspondences be-
tween surface syntactic cues and meaning.
We evaluated our results against paraphrase judgments elicited experimentally
from subjects naive to linguistic theory and showed that the model’s ranking of
meanings correlates reliably with human intuitions. Comparison between our model
and human judgments yields a reliable correlation of .64 for verb-noun combinations
and .40 for adjective-noun pairs. Furthermore, our model performs reliably better than
a naive baseline model, which achieves only a correlation of .42 in the case of verbs
and .25 in the case of adjectives.
Our approach combined insights from linguistic theory (i.e., Pustejovsky’s (1995)
theory of qualia structure and Vendler’s (1968) observations) with corpus-based ac-
quisition techniques, probabilistic modeling, and experimental evaluation. Our results
empirically tested the validity of linguistic generalizations and extended their cov-
erage. Furthermore, in agreement with other lexical acquisition studies (Merlo and
Stevenson 2001; Barzilay and McKeown 2001; Siegel and McKeown 2000; Light 1996;
McCarthy 2000; Rooth et al. 1999), we showed that it is possible to extract semantic
information from corpora even if they are not semantically annotated in any way.
Appendix A. The WebExp Software Package
As part of the evaluation of the probabilistic models presented in this article, we
conducted two psycholinguistic experiments. These experiments were administered
308
Computational Linguistics Volume 29, Number 2
using WebExp (Keller, Corley, and Scheepers 2001), a software package designed for
conducting psycholinguistic studies over the Web.
6
WebExp is a set of Java classes for conducting psycholinguistic experiments over
the World Wide Web. The software consists of two modules: the WebExp server, which
is a stand-alone Java application, and the WebExp client, which is implemented as a
Java applet. The server application runs on the Web server that hosts the experiment
and waits for client applets to connect to it. It issues experimental materials to clients
and records participants’ responses. The client applet is typically embedded into a
Web page that contains the instructions for the experiment. When a participant starts
the experiment, the WebExp client will download the experimental materials from
the WebExp server and administer them to the participant. After the experiment is
completed, it will send the participants’ responses to the server, along with other
participant-specific data.
As Java is a full-fledged programming language, it gives the Web designer max-
imal control over the interactive features of a Web site. WebExp makes use of this
flexibility to keep the experimental procedure as constant as possible across partici-
pants. An important aspect is that the sequence in which the experimental items are
administered is fixed for each participant: The participant does not have the ability to
go back to previous stimuli and to inspect or change previous responses. (If the partic-
ipant hits the Back button on the browser, the experiment will terminate.) WebExp also
provides timings of participant responses by measuring the response onset time and
the completion time for each answer. The studies reported in this article make no direct
use of these timings. Nevertheless, the timings were useful for screening the responses
for anomalies, that is, to eliminate the data for subjects who responded too quickly
(and thus probably did not complete the experiment in a serious fashion) or those who
responded too slowly (and thus were probably distracted while doing the experiment).
WebExp automatically tests the response timings against upper and lower limits pro-
vided by the experimenter and excludes participants whose timings are anomalous.
Further manual checks can be carried out on the response timings later on.
Apart from providing response timing, WebExp also offers a set of safeguards
that are meant to ensure the authenticity of the participants taking part and exclude
participants from participating more than once:
E-mail Address. Each participant has to provide his or her e-mail address. An automatic
plausibility check is conducted on the address to ensure that it is syntactically valid.
If the address is valid, then WebExp sends an e-mail to the address at the end of
the experiment (the e-mail typically contains a message thanking the participant for
taking part). If the e-mail is returned as undeliverable, the experimenter is effectively
informed that a participant is likely to have used a fake identity and has the option
of excluding that participant’s responses from further analysis.
Personal Data. Before the experiment proper commences, each participant has to fill
in a short questionnaire supplying name, age, sex, handedness, and language back-
ground. These data allow manual plausibility checks so that participants who provide
implausible answers can be eliminated from the data set.
6 The WebExp software package is distributed free of charge for noncommercial purposes. Information
on how to obtain the latest version is available at http://www.hcrc.ed.ac.uk/web exp/. A central entry
page for all experiments using WebExp can be found at http://www.language-experiments.org/.
309
Lapata and Lascarides Logical Metonymy
Responses. A manual inspection of the responses allows the experimenter to detect
participants who have misunderstood the instructions or who have responded in an
anomalous fashion (e.g., by giving the same response to every item).
Connection Data. The software also logs data related to the participant’s Web connec-
tion. This includes the Internet address of his machine and the operating system and
browser he uses. This information (in addition to the e-mail address) is valuable in
detecting participants who take part more than once.
In addition to making it possible to administer experiments over the Web, WebExp
can also be used in a conventional laboratory setting. In such a setting, WebExp has
the advantage of being platform independent (as it is implemented in Java); that
is, it will run on any computer that is connected to the Internet and runs a Web
browser. Comparisons of experimental data obtained from Internet experiments (using
WebExp) and their laboratory-based counterparts (Keller and Asudeh 2002; Keller and
Alexopoulou 2001; Corley and Scheepers 2002) revealed high correlations between the
two types of data sets; the comparisons also demonstrated that the same main effects
are obtained from Web-based and laboratory-based experiments.
Appendix B. Materials
B.1. Experiment 1
The following is a list of the materials used in Experiment 1. The modulus is shown
in (64). The verb-noun pairs and their selected interpretations (Interpr) are illustrated
in Table 23. In addition, the table shows the mean ratings (Rtg) for each paraphrase ac-
cording to the subjects’ responses and their probability (Prob) according to the model.
(64) David finished a course David finished writing a course
B.2. Experiment 4
The experimental item in (65) was presented as the modulus in Experiment 4. The
experimental materials are shown in Tables 24 and 25.
(65) hard substance a substance that is hard to alter
310
Computational Linguistics Volume 29, Number 2
Table 23
Materials for Experiment 1, with mean ratings and model probabilities.
High Medium Low
Verb-noun
Interpr Rtg Prob Interpr Rtg Prob Interpr Rtg Prob
attempt definition analyze −0.087 −21.44 recall −0.0338 −22.84 support −0.1571 −23.87
attempt peak climb 0.2646 −20.22 claim −0.0900 −23.53 include −0.4450 −24.85
attempt question reply to 0.1416 −18.96 set −0.2666 −21.63 stick to −0.3168 −22.55
attempt smile give 0.2490 −19.43 rehearse −0.1976 −22.21 look at −0.4722 −23.74
attempt walk take 0.1807 −19.81 schedule −0.1649 −22.71 lead −0.1863 −23.85
begin game play 0.2798 −15.11 modify −0.3403 −21.52 command −0.2415 −23.46
begin photograph develop 0.0816 −21.11 test −0.2147 −22.49 spot −0.4054 −23.69
begin production organize 0.0502 −19.09 influence −0.2329 −21.98 tax −0.4367 −22.78
begin test take 0.2699 −17.97 examine −0.1424 −21.78 assist −0.3440 −24.11
begin theory formulate 0.2142 −18.28 present 0.1314 −21.54 assess −0.1356 −22.40
enjoy book read 0.2891 −16.48 discuss −0.1515 −23.33 build −0.5404 −25.52
enjoy city live in 0.2028 −20.77 come to 0.1842 −23.50 cut −0.6957 −24.67
enjoy concert listen to 0.2779 −20.91 throw −0.2442 −23.61 make −0.2571 −24.97
enjoy dish cook −0.1223 −20.21 choose −0.2373 −24.61 bring −0.3385 −25.33
enjoy story write −0.0731 −19.08 learn −0.0887 −23.50 choose −0.2607 −24.61
expect order hear 0.2087 −20.29 read 0.0628 −22.92 prepare −0.3634 −23.25
expect poetry see 0.1100 −20.43 learn −0.0601 −22.81 prove −0.4985 −25.00
expect reply get 0.2696 −20.23 listen to 0.1178 −23.48 share −0.2725 −23.77
expect reward collect 0.2721 −21.91 claim 0.1950 −23.13 extend −0.3743 −23.52
expect supper eat 0.2487 −21.27 start 0.0285 −23.20 seek −0.3526 −23.91
finish gig play 0.2628 −20.34 plan −0.1780 −24.47 use −0.5341 −25.69
finish novel translate 0.0474 −21.81 examine −0.1323 −24.01 take −0.4375 −25.53
finish project work on 0.3113 −18.79 study 0.0679 −24.31 sell −0.1692 −25.05
finish room wallpaper 0.1497 −19.07 construct 0.0444 −22.48 show −0.6305 −24.59
finish video watch 0.3165 −22.37 analyze −0.0482 −24.33 describe −0.1718 −25.26
postpone bill debate −0.0401 −22.38 give −0.2028 −25.36 think −0.6238 −27.92
postpone decision make 0.3297 −20.38 publish −0.1087 −24.03 live −0.4984 −25.56
postpone payment make 0.2745 −21.85 arrange 0.0166 −23.21 read −0.4675 −25.91
postpone question hear −0.1974 −23.45 assess −0.0795 −24.70 go to −0.5383 −24.94
postpone trial go to −0.1110 −23.49 make −0.1425 −25.50 hear 0.0336 −25.75
prefer bike ride 0.2956 −20.64 mount −0.1617 −22.95 go for −0.2119 −24.49
prefer film go to 0.2456 −21.63 develop −0.1861 −23.07 identify −0.4657 −24.69
prefer gas use 0.3078 −20.28 measure −0.2903 −23.88 encourage −0.5074 −25.33
prefer people talk with 0.1235 −20.52 sit with 0.0283 −22.75 discover −0.2687 −25.26
prefer river swim in 0.0936 −19.53 sail 0.0433 −22.93 marry −0.7269 −24.13
resist argument contest −0.0126 −22.66 continue 0.0224 −24.43 draw −0.4551 −25.51
resist invitation accept 0.2497 −21.56 leave −0.3620 −25.10 offer −0.5301 −26.29
resist pressure take −0.1481 −22.67 make −0.3602 −24.98 see −0.4382 −25.22
resist proposal work on −0.0597 −23.56 take on 0.1286 −24.50 call −0.2906 −25.99
resist song whistle −0.0013 −22.11 start −0.1551 −24.47 hold −0.5691 −26.50
start experiment implement 0.1744 −21.57 study 0.0184 −22.70 need −0.5299 −24.09
start letter write 0.3142 −15.59 study −0.0877 −22.70 hear −0.4526 −24.50
start treatment receive 0.2888 −19.53 follow 0.1933 −22.45 assess −0.2536 −24.01
survive course give 0.0164 −22.87 make −0.1426 −24.48 write −0.1458 −26.27
survive journey make 0.2719 −22.31 take 0.2324 −24.43 claim −0.5388 −25.84
survive problem create −0.2697 −21.12 indicate −0.3267 −23.01 confirm −0.3625 −25.08
survive scandal experience 0.2200 −24.61 create −0.2115 −26.02 take −0.3068 −27.33
survive wound receive 0.2012 −24.49 produce −0.4361 −26.32 see −0.2668 −26.97
try drug take 0.1077 −17.81 grow −0.2516 −22.09 hate −0.4777 −23.88
try light turn 0.2397 −18.10 reach for −0.1163 −21.23 come with −0.5310 −23.93
try shampoo use 0.1404 −20.09 pack −0.3496 −21.56 like −0.2581 −24.56
try sport get into 0.1379 −19.65 encourage −0.2378 −21.09 consider −0.2223 −22.82
try vegetable eat 0.2068 −19.64 chop −0.0780 −21.38 compare −0.3248 −22.38
want bed lay on 0.1154 −19.17 reserve 0.0369 −21.24 settle in 0.0015 −22.26
want hat buy 0.2560 −17.84 examine −0.2127 −21.56 land on −0.6379 −22.38
want man marry 0.0826 −15.50 torment −0.2764 −20.99 assess −0.2510 −22.22
want money make 0.1578 −15.16 handle −0.0929 −20.91 represent −0.4031 −22.92
want program produce −0.1980 −18.86 teach −0.1743 −20.94 hate −0.6788 −22.63
311
Lapata and Lascarides Logical Metonymy
Table 24
Materials for Experiment 4, with mean ratings (object interpretations).
High Medium Low
Adjective-noun
Interpr Rtg Prob Interpr Rtg Prob Interpr Rtg Prob
difficult consequence cope with 0.3834 −18.61 analyze 0.0270 −20.43 refer to −0.3444 −24.68
difficult customer satisfy 0.4854 −20.27 help 0.3228 −22.20 drive −0.4932 −22.64
difficult friend live with 0.2291 −19.11 approach 0.0798 −21.75 miss −0.5572 −23.04
difficult group work with 0.3066 −18.64 teach 0.2081 −21.00 respond to 0.1097 −22.66
difficult hour endure 0.3387 −21.17 complete −0.1386 −21.83 enjoy 0.1600 −23.18
easy comparison make 0.4041 −17.73 discuss −0.0901 −22.09 come to 0.3670 −23.03
easy food cook 0.2375 −18.93 introduce −0.3673 −21.94 finish −0.1052 −23.15
easy habit get into 0.2592 −17.48 explain −0.2877 −21.79 support −0.0523 −23.70
easy point score 0.3255 −18.77 answer 0.1198 −20.65 know −0.0307 −23.43
easy task perform 0.4154 −17.56 manage 0.3094 −21.30 begin 0.0455 −22.48
fast device drive −0.0908 −22.35 make 0.2948 −23.65 see −0.4817 −25.00
fast launch stop −0.5438 −23.79 make 0.0075 −25.19 see −0.3963 −25.84
fast pig catch −0.5596 −23.98 stop 0.6285 −24.30 use −0.5350 −25.66
fast rhythm beat 0.0736 −20.46 feel −0.1911 −25.24 make −0.1296 −25.60
fast town protect −0.5896 −23.66 make −0.4564 −23.90 use −0.4996 −25.66
good climate grow up in 0.2343 −19.65 play in −0.6498 −22.18 experience −0.4842 −23.20
good documentation use 0.2549 −21.89 produce −0.2374 −22.38 include 0.1110 −23.73
good garment wear 0.2343 −20.23 draw −0.6498 −22.71 measure −0.4842 −23.16
good language know 0.2188 −19.18 reinforce −0.1383 −22.48 encourage −0.0418 −22.81
good postcard send 0.1540 −20.17 draw −0.3248 −22.71 look at 0.2677 −23.34
hard logic understand 0.2508 −18.96 express 0.0980 −22.76 impose −0.2398 −23.11
hard number remember 0.0326 −20.30 use −0.3428 −21.14 create −0.4122 −22.69
hard path walk 0.2414 −21.08 maintain 0.0343 −21.64 explore 0.1830 −23.01
hard problem solve 0.4683 −15.92 express 0.0257 −21.26 admit −0.2913 −23.39
hard war fight 0.2380 −17.31 get through 0.2968 −21.32 enjoy −0.5381 −23.18
slow child adopt −0.5028 −19.72 find −0.7045 −22.52 forget −0.6153 −23.93
slow hand grasp −0.5082 −18.03 win 0.2524 −22.07 produce −0.3360 −22.52
slow meal provide −0.0540 −19.55 begin 0.2546 −21.14 bring −0.3965 −23.29
slow minute take −0.1396 −19.48 fill 0.1131 −22.06 meet −0.6083 −23.30
slow progress make 0.3617 −18.50 bring −0.1519 −22.64 give −0.2700 −24.89
safe building use 0.1436 −20.83 arrive at −0.1640 −23.55 come in 0.0306 −23.96
safe drug release 0.1503 −23.23 try 0.1930 −23.62 start 0.1614 −24.31
safe house go to 0.2139 −20.87 get −0.3438 −22.41 make −0.3490 −23.19
safe speed arrive at −0.0242 −23.55 keep 0.1498 −23.59 allow 0.2093 −25.04
safe system operate 0.2431 −19.78 move −0.2363 −22.85 start 0.0013 −24.60
right accent speak in 0.1732 −19.90 know −0.1223 −22.50 hear 0.0946 −22.79
right book read 0.1938 −18.89 lend −0.0188 −22.60 suggest 0.0946 −24.90
right school apply to 0.2189 −21.76 complain to −0.3736 −22.82 reach −0.2756 −23.69
right structure build −0.1084 −19.88 teach −0.1084 −22.76 support −0.0505 −24.92
right uniform wear 0.1990 −19.79 provide −0.1084 −24.09 look at −0.0505 −25.24
wrong author accuse −0.1925 −21.90 read 0.0450 −24.09 consider 0.0653 −24.50
wrong color use 0.2366 −21.78 look for 0.0587 −22.78 look at −0.1907 −24.89
wrong note give 0.0222 −22.29 keep 0.2014 −22.64 accept −0.1462 −24.16
wrong post assume −0.3000 −20.10 make 0.2579 −23.81 consider −0.0466 −24.50
wrong strategy adopt 0.2804 −19.86 encourage 0.1937 −23.51 look for 0.0135 −24.39
312
Computational Linguistics Volume 29, Number 2
Table 25
Materials for Experiment 4, with mean ratings (subject interpretations).
High Medium Low
Adjective-noun
Interpr Rtg Prob Interpr Rtg Prob Interpr Rtg Prob
difficult customer buy −0.2682 −20.19 pick −0.2050 −22.58 begin −0.3560 −24.83
difficult friend explain −0.4658 −20.16 neglect −0.5274 −21.75 enjoy −0.4711 −23.59
difficult passage read 0.1668 −20.46 speak −0.3600 −22.62 appear −0.4030 −23.78
difficult piece read 0.1052 −20.30 survive −0.5080 −22.28 continue −0.1006 −23.89
difficult spell break −0.3047 −19.80 create −0.2412 −22.81 start −0.3661 −23.50
easy car start −0.1652 −21.01 move −0.2401 −21.42 close −0.5750 −23.94
easy change occur 0.1999 −20.32 prove 0.2332 −21.57 sit −0.0932 −23.27
easy food cook 0.0443 −20.28 change −0.6046 −22.25 form −0.3918 −22.96
easy habit develop 0.1099 −21.43 start 0.1156 −23.21 appear −0.3490 −24.70
easy task fit −0.3882 −20.77 end −0.0982 −22.90 continue 0.0474 −24.11
fast device go 0.2638 −22.80 come −0.3652 −23.91 add −0.2219 −24.68
fast horse run 0.4594 −20.78 work 0.0025 −22.98 add −0.5901 −24.64
fast lady walk −0.0261 −22.31 work −0.0716 −23.90 see −0.4816 −25.15
fast pig run 0.2081 −22.57 come −0.1807 −23.91 get −0.2764 −24.75
fast town grow −0.3601 −18.66 spread −0.3289 −22.84 sell −0.3462 −24.33
good ad read 0.1248 −22.39 sell 0.2154 −22.72 run −0.0832 −22.78
good climate change −0.3748 −21.30 improve −0.3312 −22.57 begin −0.4093 −23.36
good egg look −0.0581 −21.79 develop −0.2457 −22.02 appear 0.1149 −24.01
good light work −0.0022 −19.42 spread −0.2023 −21.75 increase −0.4349 −23.82
good show run 0.0787 −21.07 continue −0.0798 −22.79 die −0.6569 −23.79
hard fish bite −0.3583 −20.39 pull −0.2579 −22.53 appear −0.2568 −22.79
hard logic get −0.1211 −21.90 sell −0.4533 −22.18 go −0.4388 −24.59
hard substance keep 0.0227 −21.28 remain 0.0978 −22.93 seem 0.1971 −23.03
hard toilet flush −0.1796 −20.46 look −0.3465 −23.77 start −0.6835 −24.60
hard war break out −0.4969 −16.94 grow −0.2792 −22.21 increase −0.2602 −23.60
safe building approach −0.6815 −22.87 stay 0.0852 −23.09 start −0.5152 −23.53
safe drug come −0.3802 −19.41 play −0.5562 −22.24 try −0.3126 −22.45
safe man eat −0.5434 −21.23 ignore −0.6673 −21.89 agree −0.6509 −23.16
safe speed go −0.3116 −16.26 leave −0.6136 −20.24 remain 0.1267 −22.87
safe system operate 0.3697 −19.92 continue 0.0374 −21.48 think −0.4845 −22.90
slow child react 0.1485 −22.66 adapt 0.1556 −23.21 express −0.0256 −24.60
slow hand move 0.0738 −22.24 draw 0.0039 −23.38 work −0.0346 −25.56
slow meal go 0.1237 −20.57 run −0.1474 −21.47 become −0.2802 −23.17
slow minute pass 0.2717 −23.17 start −0.4423 −23.49 win −0.6709 −24.64
slow sleep come −0.0671 −19.56 follow −0.3108 −22.56 seem −0.2169 −24.85
right accent go −0.1727 −18.54 sound 0.1928 −21.14 fall −0.3926 −22.76
right book read −0.2429 −19.50 feel −0.1027 −21.74 discuss −0.2195 −23.61
right character live −0.2505 −22.65 set −0.4063 −23.14 feel −0.1651 −23.92
right people vote −0.4541 −21.70 eat −0.5921 −22.81 answer −0.0992 −24.68
right school teach 0.0159 −20.91 start −0.3466 −24.03 stand −0.3839 −24.88
wrong author go −0.4348 −20.59 think −0.4128 −22.96 read −0.5542 −24.50
wrong business think −0.4018 −23.15 spend −0.4416 −23.85 hope −0.5608 −24.18
wrong color go 0.1846 −20.93 show −0.0819 −23.54 seem 0.2869 −24.18
wrong note conclude −0.2575 −21.24 show −0.3480 −23.87 tell −0.2732 −24.45
wrong policy encourage −0.0401 −21.71 identify −0.2167 −23.56 accept 0.0183 −23.76
313
Lapata and Lascarides Logical Metonymy
Appendix C. Descriptive Statistics
Table 26 displays the descriptive statistics for the model probabilities and the subject
ratings for Experiments 1 and 4.
Table 26
Descriptives for model probabilities and subject ratings.
Rank Mean StdDev StdErr Min Max
Model probabilities, Experiment 1
High −21.62 2.59 0.26 −24.62 −15.11
Medium −22.65 2.19 0.22 −26.32 −20.90
Low −23.23 2.28 0.22 −27.92 −22.22
Subject ratings, Experiment 1
High 0.1449 0.2355 0.0309 −0.2697 0.3297
Medium −0.1055 0.2447 0.0321 −0.4361 0.2324
Low −0.3848 0.2728 0.0358 −0.7269 0.0336
Model probabilities, Experiment 4
High −20.49 1.71 0.18 −23.99 −15.93
Medium −22.62 0.99 0.10 −25.24 −20.24
Low −23.91 0.86 0.18 −25.85 −22.46
Subject ratings, Experiment 4
High −0.0005 0.2974 0.0384 −0.68 0.49
Medium −0.1754 0.3284 0.0424 −0.70 0.31
Low −0.2298 0.3279 0.0423 −0.68 0.37
Acknowledgments
This work was supported by ESRC grant
number R000237772 (Data Intensive
Semantics and Pragmatics) and the DFG
(Gottfried Wilhelm Leibniz Award to
Manfred Pinkal). Alex Lascarides is
supported by an ESRC research fellowship.
Thanks to Brian McElree and Matt Traxler
for making available to us the results of
their norming study and to Ann Copestake,
Frank Keller, Scott McDonald, Manfred
Pinkal, Owen Rambow, and three
anonymous reviewers for valuable
comments.

References
Abney, Steve. 1996. Partial parsing via
finite-state cascades. In John Carroll,
editor, Workshop on Robust Parsing, Prague.
European Summer School in Logic,
Language and Information, pages 8–15.
Apresjan, J. D. 1973. Regular polysemy.
Linguistics, 142:5–32.
Bard, Ellen Gurman, Dan Robertson, and
Antonella Sorace. 1996. Magnitude
estimation of linguistic acceptability.
Language, 72(1):32–68.
Barzilay, Regina and Kathy McKeown. 2001.
Extracting paraphrases from a parallel
corpus. In Proceedings of the 39th Annual
Meeting of the Association for Computational
Linguistics, Toulouse, France, pages 50–57.
Bouaud, Jacques, Bruno Bachimont, and
Pierre Zweigenbaum. 1996. Processing
metonymy: A domain-model heuristic
graph traversal approach. In Proceedings of
the 16th International Conference on
Computational Linguistics, Copenhagen,
Denmark, pages 137–142.
Briscoe, Ted, Ann Copestake, and Bran
Boguraev. 1990. Enjoy the paper: Lexical
semantics via lexicology. In Proceedings of
the 13th International Conference on
Computational Linguistics, Helsinki,
Finland, pages 42–47.
Burnage, Gavin. 1990. Celex–A guide for
users. Technical report, Centre for Lexical
Information, University of Nijmegen,
Nijmegen, the Netherlands.
Burnard, Lou. 1995. The Users Reference Guide
for the British National Corpus. British
National Corpus Consortium, Oxford
University Computing Service, Oxford,
U.K.
Clark, Stephen and David Weir. 2001.
Class-based probability estimation using a
semantic hierarchy. In Proceedings of the
Second Conference of the North American
Chapter of the Association for Computational
Linguistics, Pittsburgh, Pennsylvania.
Collins, Michael and James Brooks. 1995.
Prepositional phrase attachment through
a backed-off model. In David Yarowsky
and Kenneth W. Church, editors,
Proceedings of the Third Workshop on Very
Large Corpora, Cambridge, Massachusetts,
pages 27–38.
Copestake, A. 2001. The semi-generative
lexicon: Limits on lexical productivity. In
Proceedings of the First International
Workshop on Generative Approaches to the
Lexicon, Geneva.
Copestake, A. and E. J. Briscoe. 1995.
Semi-productive polysemy and sense
extension. Journal of Semantics, 12(1):15–67.
Copestake,A. and A. Lascarides. 1997. Inte-
grating symbolic and statistical represen-
tations: The lexicon pragmatics interface.
In Proceedings of the 35th Annual Meeting of
the Association for Computational Linguistics
and the Eighth Meeting of the European
Chapter for the Association for Computational
Linguistics, Madrid, pages 136–143.
Corley, Martin and Christoph Scheepers.
2002. Syntactic priming in English:
Evidence from response latencies.
Psychonomic Bulletin and Review,
9(1):126–131.
Cowart, Wayne. 1997. Experimental Syntax:
Applying Objective Methods to Sentence
Judgments. Sage, Thousand Oaks,
California.
Daelemans, Walter, Antal van den Bosch,
and Jakub Zavrel. 1999. Forgetting
exceptions is harmful in language
learning. Machine Learning, 34(1–3):11–43.
Fass, Dan. 1991. met*: A method for
discriminating metonymy and metaphor
by computer. Computational Linguistics,
17(1):49–90.
Godard, Dani `ele and Jacques Jayez. 1993.
Towards a proper treatment of coercion
phenomena. In Proceedings of the 31st
Annual Meeting of the Association for
Computational Linguistics, Columbus,
Ohio, pages 168–177.
Hobbs, Jerry R., Martin Stickel, Douglas
Appelt, and Paul Martin. 1993.
Interpretation as abduction. Artificial
Intelligence, 63(1–2):69–142.
Inverson, Eric and Stephen Helmreich. 1992.
An integrated approach to non-literal
phrase interpretation. Computational
Intelligence, 8(3):477–493.
Keller, Frank and Theodora Alexopoulou.
2001. Phonology competes with syntax:
Experimental evidence for the interaction
of word order and accent placement in
the realization of information structure.
Cognition, 79(3):301–371.
Keller, Frank and Ash Asudeh. 2002.
Probabilistic learning algorithms and
optimality theory. Linguistic Inquiry,
33(2):225–244.
Keller, Frank, Martin Corley, and Christoph
Scheepers. 2001. Conducting
psycholinguistic experiments over the
World Wide Web. Unpublished
manuscript, University of Edinburgh,
Edinburgh, U.K., and Saarland University,
Saarbruecken, Germany.
Lakoff, George and Mark Johnson. 1980.
Metaphors We Live By. University of
Chicago Press, Chicago.
Lapata, Maria, Frank Keller, and Scott
McDonald. 2001. Evaluating smoothing
algorithms against plausibility judgments.
In Proceedings of the 39th Annual Meeting of
the Association for Computational Linguistics,
Toulouse, France, pages 346–353.
Lascarides, Alex and Ann Copestake. 1998.
Pragmatics and word meaning. Journal of
Linguistics, 34(2):387–414.
Leech, Geoffrey, Roger Garside, and Michael
Bryant. 1994. The tagging of the British
national corpus. In Proceedings of the 15th
International Conference on Computational
Linguistics, Kyoto, Japan, pages 622–628.
Levin, Beth. 1993. English Verb Classes and
Alternations: A Preliminary Investigation.
University of Chicago Press, Chicago.
Li, Hang and Naoki Abe. 1998. Generalizing
case frames using a thesaurus and the
MDL principle. Computational Linguistics,
24(2):217–244.
Light, Marc. 1996. Morphological Cues for
Lexical Semantics. Ph.D. thesis, University
of Rochester.
Lodge, Milton. 1981. Magnitude Scaling:
Quantitative Measurement of Opinions. Sage,
Beverly Hills, California.
Macleod, Catherine, Ralph Grishman,
Adam Meyers, Leslie Barrett, and Ruth
Reeves. 1998. Comlex: A lexicon of
nominalizations. In Proceedings of the
Eighth International Congress of the European
Association for Lexicography,Li`ege,
Belgium, pages 187–193.
Markert, Katja and Udo Hahn. 1997. On the
interaction of metonymies and anaphora.
In Proceedings of the 15th International Joint
Conference on Artificial Intelligence, Nagoya,
Japan, pages 1010–1015.
McCarthy, Diana. 2000. Using semantic
preferences to identify verbal
participation in role switching
alternations. In Proceedings of the First
Conference of the North American Chapter of
the Association for Computational Linguistics,
Seattle, Washington, pages 256–263.
McElree, Brian, Matthew J. Traxler, Martin J.
Pickering, Rachel E. Seely, and Ray
Jackendoff. 2001. Reading time evidence
for enriched composition. Cognition,
(78)1:17–25.
Merlo, Paola and Susanne Stevenson. 2001.
Automatic verb classification based on
statistical distribution of argument
structure. Computational Linguistics,
27(3):373–408.
Miller, George A., Richard Beckwith,
Christiane Fellbaum, Derek Gross, and
Katherine J. Miller. 1990. Introduction to
WordNet: An on-line lexical database.
International Journal of Lexicography,
3(4):235–244.
Nunberg, Geoffrey. 1995. Transfers of
meaning. Journal of Semantics,
12(1):109–132.
Pustejovsky, James. 1991. The generative
lexicon. Computational Linguistics,
17(4):409–441.
Pustejovsky, James. 1995. The Generative
Lexicon. MIT Press, Cambridge.
Pustejovsky, James and Pierrette Bouillon.
1995. Logical polysemy and aspectual
coercion. Journal of Semantics, 12:133–162.
Resnik, Philip. 1993. Selection and
Information: A Class-Based Approach to
Lexical Relationships. Ph.D. thesis,
University of Pennsylvania, Philadelphia.
Resnik, Philip. 1999. Semantic similarity in a
taxonomy: An information-based measure
and its application to problems of
ambiguity in natural language. Journal of
Artificial Intelligence Research, 95–130.
Resnik, Philip and Mona Diab. 2000.
Measuring verb similarity. In Lila R.
Gleitman and Aravid K. Joshi, editors,
Proceedings of the 22nd Annual Conference of
the Cognitive Science Society. Erlbaum,
Mahwah, New Jersey, pages 399–404.
Rooth, Mats, Stefan Riezler, Detlef Prescher,
Glenn Carroll, and Franz Beil. 1999.
Inducing a semantically annotated lexicon
via em-based clustering. In Proceedings of
the 37th Annual Meeting of the Association for
Computational Linguistics, College Park,
Maryland, pages 104–111.
Sch ¨utze, Carson T. 1996. The Empirical Base of
Linguistics: Grammaticality Judgments and
Linguistic Methodology. University of
Chicago Press, Chicago.
Siegel, Eric and Kathleen McKeown. 2000.
Learning methods to combine linguistic
indicators: Improving aspectual
classification and revealing linguistic
insights. Computational Linguistics,
26(4):595–629.
Stevens, S. Smith. 1975. Psychophysics:
Introduction to Its Perceptual, Neural, and
Social Prospects. Wiley, New York.
Utiyama, Masao, Masaki Murata, and
Hitoshi Isahara. 2000. A statistical
approach to the processing of metonymy.
In Proceedings of the 18th International
Conference on Computational Linguistics,
Saarbr ¨ucken, Germany, pages 885–891.
Vendler, Zeno. 1968. Adjectives and
Nominalizations. Mouton, The Hague, the
Netherlands.
Verspoor, Cornelia Maria. 1997. Contextually-
Dependent Lexical Semantics. Ph.D. thesis,
University of Edinburgh, Edinburgh, U.K.
Weiss, Sholom M. and Casimir A.
Kulikowski. 1991. Computer Systems that
Learn: Classification and Prediction Methods
from Statistics, Neural Nets, Machine
Learning, and Expert Systems. Morgan
Kaufmann, San Mateo, California.
