Lexical Structures for Linguistic Inference 
Peter Anick and Sabine Bergler 
Computer Science Department 
Brandeis University 
Waltham, MA 02254-9110 
anick@aiag, enet. dec. com 
sabine @cs. brandeis, edu 
Abstract 
In order to resolve metonymy and other violations of selectional restrictions between lexical 
items, a language understander must be able to infer relationships that do not have explicit 
lexical analogs in tile sentence. Although such inferencing has typically been relegated to the 
world knowledge portion of a natural language processing system, there is also evidence, from 
both theoretical analysis in compositional semantics and distributional analysis of corpus data, 
that some cases of metonymy may best be processed with respect to more specific lexical and 
syntactic constructions. In this paper, we argue how the richer vocabulary for lexical semantics 
proposed in Pustejovsky's "Generative Lexicon" theory allows one to explore the role of lexical 
information in such cases, and therefore sheds more light on the distinction between lexical 
inferences, which follow from defaults associated with lexical items and rules of composition, and 
pragmatic inferences, which depend on reasoning with respect to the context of the utterance. 
1 Introduction 
Is a lexicon merely a set. of entry points into a conceptual, "real world" knowledge base, or does 
it have its own rich structure and rules of inference, related to, yet distinct from the body of 
world knowledge? Ill this paper, we explore this question from the point of view of the inferential 
machinery required to relate a verb and its arguments. We will argue that a rich lexical semantic 
structure makes it possible to express rules of composition with the degree of precision necessary to 
account for the many nuances of actual language use, capturing linguistic generalizations based on 
the semantic content of lexical items rather than relying solely on general purpose inferencing over 
an encyclopedic knowledge base. 
One of the fundamental properties of most computational lexicons is an accounting of the rela- 
tionship between a verb and its arguments. Typically, each argument position of a verb is annotated 
with "selectional restrictions" which are intended to constrain the set of nouns which may legally 
fill the argument position 1. Such constraints are useful for word sense disambiguation, since se- 
mantically anomalous colnbinations of senses can be readily identified and discarded during parsing. 
A number of researchers have noted that such selectionai restrictions may legitimately be violated 
under certain circumstances, as ill the sentences below: 
(1) "The car drank gasoline." (\[Wilks, 1978\]) 
(2) "Ted played Bach." (\[Fass, 1988\]) 
1 If nouns are organized into a conceptual "isa" hierarchy, then a selectlonal restriction may be no more than a 
reference to a position in the hierarchy. 
102 
In (1), the "car" violates the selectional restriction that the subject be animate. In (2), the 
object "Bach" is a person, violating the verb's selectional restrictions (i.e. some form of music). 
Following \[Wilks,1978\], Fass \[1988\] interprets selectional restrictions as preferences and goes on to 
account for violations of preferences via operations such as metonymy and metaphor. For example, 
Fass postulates a number of metonymic substitution rules, such as "Artist for Artform", "Container 
for Contents", and "Part for Whole." When a preference violation is encountered, his algorithm 
tries applying these metonymic inference rules to the offending noun in an attempt to find the 
"shortest" coherent path between the verb's preference and some metonymically related concept. 
Failing this, the algorithm goes on to explore metaphorical interpretations. In a similar vein, \[Parisi 
and Castelfranchi, 1988\] make use of a semantic network encoding "encyclopedic" relations among 
lexical concepts to derive paths between verbs and arguments which violate selectional restrictions. 
Using a shortest path heuristic, their algorithm interprets the senetence "The pot is boiling." as 
"The liquid which is contained in the pot is boiling." 
While these algorithms work well on a small set of interesting examples, we feel that distance met- 
rics operating on unconstrained encyclopedic knowledge insufficiently characterize the relationship 
between lexical entries and metonymic inference. In this paper, we present a richer representational 
framework for lexical knowledge, based on Generative Lexicon theory \[Pustejovsky, 1991\], and de- 
scribe a methodology for exploring the interplay between word senses and the metonymic processes 
they enter into. \[Pustejovsky 1991\] distinguishes between "logical metonymy", whose resolution 
falls naturally out of operations on information contained directly in lexical entries, and "pragmatic 
metonymy", whose resolution depends on interpretation with respect to the proper pragmatic con- 
text. We begin with an overview of the major components of Generative Lexicon theory and present 
data from the semantic domain of musical terms to illustrate how metonymic reasoning is facilitated 
in a generative lexicon. Then we indicate how in the case of reporting verbs the compositional 
semantics of the matrix verb and the subject produce pragmatic constrains on the complement. We 
conclude that it is advantageous to represent the conceptual knowledge that bears on the linguistic 
behavior of words in the lexicon. 
2 Generative Lexicon 
Pustejovsky's Generative Lexicon theory \[Pustejovsky, 1991\] rejects the characterization of a lexicon 
as a static listing of word senses and focuses instead on elucidating the set of generative devices 
which give rise to word senses and allow for the flexible and creative use of language. 
Generative Lexicon theory postulates a number of representational subsystems for the description 
of lexical items and the rules of syntactic and semantic composition. These include, in addition 
to argument and event structures, (1) a qualia structure, which partitions the aspects of a noun's 
meaning into formal, constitutive, agentive, and telic roles, or qualia, (2) a lexical inheritance network 
reflecting the same partition, (3) Lexical Conceptual Paradigms, which describe sets of syntactic 
behaviors which correspond to lexicai semantic categories, and (4) generative devices for extending 
the logical senses of lexical items dynamically, as for instance coercion rules, mentioned later in this 
paper. By attempting to tease apart various aspects of the lexical semantics, the theory provides a 
new vocabulary with which to approach traditional linguistic problems. 
Qualia 
• Constitutive Role: Relation t,o constituent parts. 
• Formal Role: Distinguishes the word within a larger domain. 
103 
• Telie Role: Purpose and function. 
• Agentive Role: Whatever brings it about. 
To illustrate the use of qualia structures for handling selectional restrictions, consider three 
schematic qualia structures for food-related nouns: 
food(x) 
\[const: phys_obj(x)\] 
\[formal: \] 
\[relic: eat(E,z,x)\] 
\[agentive: cook(w,x), prepare(y,x)\] 
meal(T,x) 
\[const: food(y)\] 
\[formal: amount(x,y)\] 
\[relic: eat(T,z,x)\] 
\[agentive: "food\] 
doughnut (x) 
\[eonst : dough(u)\] 
\[formal: cake(x) k ring-shaped(x)\] 
\[relic : "food\] 
\[agentive : bake(w,x)\] 
The empty formal role in the structure for "food" indicates that food is a mass noun and, as the 
constitutive role indicates, composed of some physical material. The telic role is doing most of the 
work of defining food as something which is eaten. The agentive role adds that food typically arises 
by cooking or preparing processes (although this is certainly defeasible). 
Doughnut also inherits the telic quale from food, but differs from food in that it specifies a value in 
its formal quale. That is, it "packages" the mass noun in a specific way (cf. Pelletier and Schubert 
\[1986\] , Krifka \[1987\] ). This in fact, is the structural semantic characterization of a packaging 
coercion: taking a lexical structure with no formal role, and specifying the formal role (Pustejovsky 
\[1991\] for details). 
Now consider tile representation for meal. A meal directly denotes both the event pertaining to 
thc activity of eating the food, as well as tile food itself. The fact that meal is defined with respect 
to food is made evident in the fact that the formal role is a certain packaging of an amount of food, 
while the telic role includes the event denoting the occasion of the eating. Now consider the following 
sentences making use of these nouns. 
1. John ate the doughnut. 
2. John ate the meal. 
3. John left after the meal. 
4. John left after tile doughnut. 
Tile first problem presented by such data is how to describe the seleetional restriction imposed 
by the verb eat on its direct object, since it clearly allows for both physical objects and events. The 
104 
qualia structure suggests a straightforward answer. Instead of stating the restriction in terms of a 
formal or constitutive property, we may use the telic role. That is, eat may take as direct object any 
noun whose telic role is eat. This may at first seem like a circular definition, but why should the 
telic role be any less accessible to the selecting verb than, say, the formal role? This approach also 
renders it unnecessary to create an explicit subclass of physical objects called edible-objects. The 
class is implicitly defined via the telic quale. 
The next problem is how to accomodate doughnut as the object of a preposition selecting for an 
event. This sentence requires an inference (i.e. that John left after eating the doughnut) that can 
only be verified in the discourse context. However, the lexical structures do provide a means for 
quickly identifying the potential interpretation. 
Pustejovsky \[1989, 1991\], has proposed a system of type coercion to explain how verbs like want 
and believe can accept arguments of type S, VP, and NP without requiring multiple verb entries in 
the lexicon. Ite defines type coercion as "a semantic operation that converts an argument to the 
type which is expected by a function, where it would otherwise lead to a type error." This principle 
of type coercion can be applied to coercing an object into an event, as required for our doughnut. 
The telic and agentive roles identify readily available, definitional events associated with a nominal. 
In most of the examples of this kind of selectional restriction violation that we have encountered, it 
is the telic role that fills the expected event. 
By examing what systematic semantic relationships exist in the grammatical expressions of lexical 
items, we hope to apply the principles of the generative lexicon to new data and illustrate new 
phenomena in support of the view of the "non-static" lexicon. Making use of data from human 
interest "notes" conferences, Wall Street Journal and Time magazine articles, our methodology is to 
analyze a set of semantically related verbs with respect to their argument types in order to ascertain 
how the lexical semnantic structure of the verb relates to the range of potential (e.g. metonymic) 
concepts that may appear in an argument position. 2 We consider first a number of verbs in the 
domain of music, such as "play", "perform", and "strum." Then we turn to verbs of reporting, as 
found in newspaper and magazine articles. 
3 Hierarchical classification and qualia structure 
One question is how to assign tile definitional information to the respective qualia. While it is beyond 
the scope of this paper to adequately address this question, let us suggest that the very structure of 
the definitions of words often suggests which qualia are of relevance. Consider for example a faceted 
classification scheme (cf. Vickery \[1975\] ), as one might do for a subject matter indexing task for 
the music domain (the data presented here is based on an electronic "notes" conference on music): 
music production 
creation "compose" 
performance 
production means 
voice "sing" 
lips "whistle" 
musical instrument "play" 
in a particular manner 
brushing across strings "strum." 
sounding strings with a pick "pick" 
2We have made extensive use of Levin's \[1990\] verb classification, and the methodology of verbal diathesis therein. 
105 
purpose 
entertaining an audience "perform" 
skill improvement "practice" 
The category "production means" describes, for example, verbs that incorporate different "tools" 
for making the music, i.e. sing incorporates in its lexical meaning the use of voice, whistle the use of 
lips and play the use of an instrument. Note that play does not specify the instrument, it implies an 
instrument but can take as an argument a specification of the particular instrument. Neither sing 
nor whistle can take an instrument as argument. The implied argument pertains to the constitutive 
role. 
Perform and practice, both purposeful activities by definition, define that purpose in their relic 
role. Both words select for an event in object position. An interesting difference between these two 
words is the selectionai restriction on the event type for the object position: perform selects for the 
result of an accomplishment, practice for an activity. This fact explains a difference in metonymic 
behavior that we will sketch below. 
3.1 Possible violations in object position 
We can arrive in a similar mnanner at a categorization of nouns in object position of verbs from our 
musical domain: 
Music "packagings" 
unit tune, song, melody, piece, chord, lick, note, repertoire, arrangement 
nmsical form jig, horT~pipe, symphony, sonata 
relation to other performers lead, backup 
location in a piece ending, intro 
number of perfomers solo, duet 
physical presentation line, measure, bar, page 
musical instruments guitar, piano 
musical instrument parts string, key, soundboard, bridge 
genres jazz, blues 
events containing music gig, concert, job, festival, wedding 
locations where music is played bar, club, concert hall 
composers Bach, Beethoven 
instrument makers Fender, Martin 
This faceted classification together with data analysis suggests that "music packagings" tune, 
song, ... are related to packagers \[Pelletier and Schubert 1989\], in fact "packaging" the mass term 
music into events. This category can be further refined into "accomplishments" (sonata, piece, ...) 
and "achievements" (chord, note, ... ). This corresponds with the behavior of perform and practice: 
to practice a chord 
to practice a high note 
to perform a sonata 
to perform a piece 
This discussion gives only a very rough first impression on the workings of the structures of 
the Generative Lexicon theory when it comes to violations of seleetional restrictions. We showed, 
however, that it is not impossible to characterize the common sense knowledge that pertains to 
106 
linguistic behavior. This is an important goal for a generative lexicon, as it aims to capture exactly 
that body of knowledge that bears on the linguistic behavior of words (in context) in a structured 
way, thus constraining common sense inferencing at an appropriate level of granularity. 
4 Lexical constraints on common sense reasoning 
In the previous section we have shown how to identify and categorize the conceptual or world 
knowledge that directly bears on the syntactic behavior of verbs towards their direct objects in 
the musical domain. Let us now consider another semantic field, namely reporting verbs in the 
newspaper context. These verbs frequently occur with a metonymy in subject position. We show how 
the violations of the selectional restrictions can be dealt with in the Generative Lexicon framework, 
producing at the same time a pragmatic coherence constraint for the complement clause. 
Most utterance verbs can be used in reported speech in newspaper articles, selecting for two 
arguments, namely a proposition as complement clause and a source as subject. The interpretation 
of these instances of reported speech is that the journalist quotes or reports the proposition in 
the complement clause giving as evidence for its credibility or reliability the original source of the 
reported material. (cf. Bergler, forthcoming) 
Two examples of reported speech in newspaper text.s are: 
A Lincoln spokesman said its management "never authorized or participated in any bugging of 
anyone." 
He said "I have not misled anyone" about US policy in Iraq. 
-Wall Street Journal, 10/27/89 
-Boston Globe, 4/5/91 
Reporting verbs are pervasive in newspaper articles \[Bergler, 1991\] . The sense of an utterance 
verb as a reporting verb is sometimes slightly different from its literal sense. We therefore introduce 
a semantic type REPORTING-VERB, which allows us to define a small part of the lexical hierarchy of 
reporting verbs as follows: 
utter(A,B) 
\[Form: produce(P1, A, B) 
& is (B, sound) 
& is(B, words)\] 
\[Constitutive: intentional(P 1)\] 
\[Agentive: human(A) & 
individual(A) & use(A, voice)\] 
REPORTING-VERB(A,B) 
\[Form: utter(T1, C, utter(P1, A, B)) 
& proposition(B)\] 
\[Telic: has-expertise(A, B) or has-insight(A, B) 
\[Constitutive: interpretation(T1, P1)\] 
\[Agentive: use(C, voice or writing)\] 
assert(A, B) 
\[Form: utter(A, B) & true(B)\] 
107 
say(A, B) 
\[Form: 
\[Literal: utter(A,B)\] 
\[REPORTING-VERB: 
assert(A,B)\] \] 
announce(A,B) 
\[Form: say(T1, A, B)\] 
\[Constitutive: new(B)\] 
\[Agentive: 
legitimation(A, T1) 
claim(A,B) 
\[Form: 
\[Literal: assert(A, P1) & own(P1, A, B); 
demand(T1, A, B)\] 
\[REPORTING-VERB: assert(A, B) 
& opposition(C, B)\] 
insist(A,B) 
\[Form: claim(A, B) 
\[Constitutive: MANNER: vehement\] 
This partial hierarchy is a very crude classification of the four reporting verbs say, announce, 
claim, and insist. We have indicated a literal word sense where it differs significantly, but have 
made no attempt to define the respective lexical entries completely. Note that the semantic type 
REPORTING--VERB specifies the basic "reporting situation", especially that the telic role (roughly 
describable as its "purpose") is that the source has expertise on the topic of the complement clause 
or at least some (relevant) insight. 
As reported in Bergler \[1991\] , we find that reporting verbs frequently occur with different kinds 
of metonymy. The violated selectional constraint is that the source of the utterance, i.e. the agent, 
is not a single individual human being, but a city or institution or even a building, as illustrated 
below: 
1. Marlin Fitzwater said ... 
2. Washington said ... 
3. The White Ilouse said ... 
4. IBM announced ... 
All reporting verbs accept metonymic extensions where the original human source is replaced by 
some significant superclass they fall into. In the case of reporting verbs of special importance is the 
employer (IBM) or the membership in a group (the French). But we find also the characterization 
of the source in terms of a role he or she played , as in the witness. This "knowledge" can best be 
stored in the concept of REPORTING-VERB -- all the verbs defined with a reporting verb sense will 
then inherit the "permission" to use metonymic extensions to a superclass or a specific role played 
in an event. 
But this is not a sufficient restriction on the possible metonymies. Consider the following scenario: 
a German makes a disparaging remark ¢ about tim German government in the United States. 
The types of expected descriptions of the source, that would be deemed appropriate, might be 
the following: 
Hans Glueck said that ¢ (Name) 
A German said that ¢ (Nationality) 
? A Mercedes Benz employee said that ¢ (Affiliation) 
however, given what ¢ denotes the following descriptions would be deemed inapproprite and in fact 
uninformative (of. Grice \[1967\], Gazdar \[1979\] ) 
A human said that ¢ 
A vegetarian said that ¢ 
A European said that ¢ 
108 
This example illustrates that not all true descriptions of a source can be used as subject in 
reporting contexts. The source has to be relevant to the topic of the complement clause. Bergler 
\[forthcoming\] describes that the lexical realization of the source in newspaper articles generally adds 
to the evaluation of the credibility or reliability the source has with respect to the topic. Thus 
a German in general has more competence and thus reliability on issues concerning the German 
government because of direct affectedness; this fact however is lost if we characterize him (correctly) 
as a European. 
This is little surprising; general pragmatic considerations such as Gricean maxims ("Be relevant", 
"Be informative") account for this fact. It has been noticed that it is extremely difficult to put these 
pragmatic considerations into computational form \[Wilks, 1986\] . We will show here that some 
pragmatic coherence can, however, be established on the basis of lexical semantics. 
When analyzing the subjects for all occurrences of seven reporting verbs in a corpus of 250,000 
words from the TIME Magazine (1963), we found that a small semantic grammar described the 
different lexicalizations of the subject, moreover, that grammar shows that the source is often lex- 
calized as some institution or as holding a particular position. It is beyond the scope of this paper to 
discuss why this should be so; suffice it to allude to the fact that newspapers and magazines report 
very frequently on government activity and business developments and that iusider information is 
considered more valuable. 
The dilemma here is that a basically pragmatic problem is part of the selectional restrictions 
of reporting verbs and that therefore the occurring systematicity has to have some reflection in a 
lexical semantics. What does the Generative Lexicon solution to this problem look like? 
Let us first state the problem with an example. Announce selects for a subject that has some 
"legitimization" for making the statement that is repeated or paraphrased in the complement. The 
relation between subject and complement that constitutes the "legitimization" is established by the 
projective conclusion space, a device that generates certain semantic relationships between lexical 
items dynamically, much in the way a semantic net records semantic relationships between words 
statically. 3 
The iexical definition of officiafis in the American lleritage Dictionary is: 
official adj 
I. Of, pertaining to, or authorized by a proper authority; authoritative. 
2. Formal or ceremonious: an official banquet. 
n 
I. One who holds an office or position. 
2. A referee in a sport. 
and in the Longman Dictionary of Contemporary English we find: 
offlc|al adj of or about a position of trust, power, and responsability: an official position/an 
official occasion/an official manner o\] speaking. -- opposite: unofficial; compare officious 
which results in a GL entry of the fornl 4 
otfieial(*x*, y) 
\[Form: human(*x*), organization(y), position(.') ~ in(z,y) 
has-authority(z) ~i hold(*x*, :)\] 
aFor detail, see \[Pustejovsky 1991\] 
4see \[Bergler, forthcoming\] for detail. 
109 
\[Telie: work-forC*x*, y), trust(y, *x*)\] 
\[Constitutive: individual(*x*)\] 
\[hgentive: ~ y\] 
The combination Officials announced sounds very appropriate (semantically "comfortable") re- 
gardless of the content of the complement clause, because the semantic selectional restriction imposed 
by announce is met at the lexicai semantic level by the definition of official as somebody of authority 
or trust, implicitly "legitimized". This is why Officials announced sounds more like a phrasal pattern 
than like an ad hoc generative combination. 
This is an obvious case where a simple feature matching procedure can establish the coherence 
relation given two lexical entries. Consider the slightly more complex case of Chrysler officials 
announced (b, which introduces constraints on the content of the complement, restricting it to matters 
related to the automobile industry in general. Chrysler officials is a particularization of officials in 
general and is thus more specific. A more specific source has a more specific domain of expertise 
(and is thus often more reliable). This means that the lexical coherence between Chrysler officials 
and announce is established just like above with the additional constraint for the complement to be 
coherent (i.e. to relate to something about) Chrysler. 
Lastly, the metonymic use of Chrysler announced ~ causes a violation of the selectional restriction 
of announce, which leads to a coercion of Chrysler to an equivalent of Chrysler officials s. 
The interpretation process of the three subject NP phrases Officials, Chrysler officials, and 
Chrysler in the context of announce is entirely due to the lcxical semantics of the words involved, 
and has thus far not involved random common sense inference, which is one of the main ideas behind 
GL (namely to minimize common sense reasoning to the really hard cases). 
Let us now make the leap to say, a reporting verb that has no explicit semantic constraints. We 
know that anybody can say anything without in fact being of much consequence. The use of say as 
a reporting verb in the newspaper context, however, restricts its use to statements where the source 
has some insight or experience that justifies repeating or rephrasing his/her words. This is a fact 
related to the concept of REPORTING-VERB, in fact an anchor for coercion. 
The case of Officials said ~ is analog to the case of Officials announced ~b, with officials being in a 
position of authority or trust filling the requirement associated with the concept REPORTING-VERB, 
an inherited and therefore weaker constraint than that for announce. Similarly, Chrysler officials 
said that fb derives the coherence relation from the concept REPORTING-VERB and constrains the 
topic of the complement clause to coher with Chrysler. 
Chrysler said (b is more marked than Chrysler announced qS, but indeed does occur frequently, 
for example in the Wall Street Journal corpus. Interestingly, we found that in the sentences that 
contained a company name as subject of say, the complement phrase was frequently referring back 
to the subject with the pronoun it as in: 
• AEG also said it expects group operating profit to remain at last year's level of 115 million 
marks. 
• Chantal said it is in advanced stages of testing one of the drugs, Cyoctol, as a topical treatment 
for mild to severe acne. 
• And Nissan Motor Co., reacting to foreign pressure on Japanese auto makers, said it plans to 
slash annual vehicle exports in half by the late 1990s. 
• San Miguel said in a report to its stockholders that higher wages, production costs, and interest 
rates threaten growth prospects for the second half. 
5 We expect the process at work here to be very close to that of ellipsis resolution. 
II0 
This enhances the markedness of the sentences enforcing the same coercion based on a stipulation 
of metonymy as for announce. But note that it is possible to use say in its reporting sense with a 
much weaker coherence relation between the source and the complement than the one required for 
announce. 
We have derived coherence relations between three different source NPs, Officials, Chrysler oJ~i- 
cials, and Chrysler to two reporting verbs, namely announce and say from the lexical semantics of 
the words involved and we have shown how the compositional behavior of the matrix clause in turn 
introduces topic constraints on the complement. This means that an issue as inherently pragmatic 
sounding as restricting tile metonymy in subject position of reporting verbs to extensions that bear 
relevance to the complement clause can partly be resolved on the basis of lexical semantics, when 
treating the lexicon not as a static, passive list of definitions but a dynamic structure that can 
incorporate procedures that were previously only found in common sense reasoning systems. The 
important advantage to pure common sense reasoning systems lies in the fact that the reasoning in 
GL is constrained by and limited to syntactic and lexical semantic knowledge, incorporating con- 
ceptual knowledge only where it bears on language behavior. This results in much cheaper inference 
mechanisms, but will not explain the truly creative use of language. 
5 Conclusion 
In this paper we considered violations of selectional restrictions for verbs of different semantic fields. 
First, we discussed musical terms as they occur in an electronic "notes" conference and found that 
violations of the selectional restriction for the direct object (restricted to "music") in fact describe 
a systematic set of "aspects" of music, indeed allowing us to formulate coercion rules that make it 
possible to speak of the required semantic inferences ill general terms. Coercion rules are a generative 
device, illustrating one of the dynamic aspects of the Generative Lexicon formalism, which aims to 
represent world knowledge that is reflected in the syntactic and semantic behavior of words. The 
inference rules used in a generative lexicon are reminiscent of common sense reasoning but nmch 
more tightly constrained (and therefore limited). 
We illustrated on the musical terms that interesting differences in the acceptance of violations of 
the selectional restrictions for the object position can be represented (and explained) by abstractions 
that often hold in other domains. For instance we confirmed that the notion of "packager", occurring 
with mass terms ill general, plays all important role in the musical domain as well. 
This approach is advantageous compared to purely syntactic/semantic or purely common sen- 
sical approaches because it identifies certain pragmatic considerations needed to describe syntac- 
tic/semantic behavior (notoriously missing or assumed in syntactic or formal semantic theories) yet 
constrains the immense space for pragmatic inferences according to the lexical definitions and a 
small set of generative devices, thus leaving to common sense reasoning proper only truly novel uses 
of language. 
The semantic domain of reporting verbs served to illustrate that not only can we resolve selec- 
tional violations within this paradigm, but that from the very compositional meaning of a reporting 
verb with its subject, we can automatically derive pragmatic constraints on the topic of the comple- 
ment clause, thus limiting even here the work of the pragmatic inferencer. 
Acknowledgements 
We gratefully acknowledge James Pustejovsky for his invaluable help with this paper. 
111 

References 
S. Bergler. The semantics of collocational patterns for reporting verbs. In Proceedings 
of the Fifth Conference of the European Chapter of the Association for Computational 
Linguistics, Berlin, April 9-11, 1991, 1991. 
S. Bergler. Evidential Analysis of Reported Speech. PhD thesis, Brandeis University, 
forthcoming. 
D. Fass. An account of coherence, semantic relations, metonymy, and lexical ambiguity 
resolution. Ill S. Small, G. Cottrell, and M. Tanenhaus, editors, Lexical Ambiguity Res- 
olution: Perspectives from Psycholinguistics, Neuropsychology, and Artificial Intelligence. 
Morgan Kaufmann Publishers, San Mateo, 1988. 
G. Gazdar. Pragmatics: lmplicature, Presupposition, and Logical Form. Academic Press, 
New York, 1979. 
H.P. Grice. Logic and conversation. Unpublished MS. of the William James Lectures, 
\[Iarvard University, 1967. 
M. Krifka. Nominal reference and temporal constitution: Towards a semantics of quan- 
tity. In Proceedings of the Sixth Amsterdam Colloquium, pages 153-173, University of 
Amsterdam, 1987. 
B. Levin. Towards a Lexical Organization of English Verbs. University of Chicago Press, 
Chicago, forthcoming. 
D. Parisi and C. Castelfranehi. Disambiguation in a lexically based sentence understanding 
system. In S. Small, G. Cottrell, and M. Tanenhaus, editors, Lezical Ambiguity Resolution: 
Perspectives from Psycholinguistics, Neuropsychology, and Artificial Intelligence. Morgan 
Kaufmann Publishers, San Mateo, 1988. 
F.J. Pelletier and L.K.Schubert. Mass expressions. In D. Gabbay and F. Guenthner, 
editors, Handbook of Philosophical Knowledge, Vol. 4. Reidel, Dordreeht, Holland, 1986. 
J. Pustejovsky. The Generative Lexicon: A Theory of Computational Le~:ical Semantics. 
MIT Press, Cambridge, 1991. 
J. Pustejovsky. Towards a generative lexicon. Computational Linguistics, 17(3), 1991. 
B.C. Vickery. Classification and Indexing in Science. Butterworth and Co., London, 
England, 1975. 
Y.A. Wilks. Making preferences more active. Artificial Intelligence, 10:1-11, 1978. 
Y.A. Wilks. Relevance and beliefs. In T. Myers, K. Brown, and B. MeGonigle, editors, 
Reasoning and Discourse Processes, pages 265-289. Academic Press, New York, 1986. 
