Wheres the Meeting that was Cancelled?

Existential Implications of Transitive Verbs

Pat r icia Amar al 1 Val er ia de Pai va2 C l eo C ond or avd i 3 Annie Z aenen3
(1)Department of Romance Languages, University of North Carolina at Chapel Hill, USA (2)School of Computer Science, University of Birmingham, UK (3)CSLI, Stanford University, USA

pamaral@[Link], [Link]@[Link], cleoc@[Link], azaenen@[Link]

ABSTRACT
This paper describes a preliminary classication of transitive verbs in terms of the implications of existence (or non-existence) associated with their direct object nominal arguments. The classication was built to underlie the lexical marking of verbs in the lexical resources that the automated system BRIDGE developed at Xerox PARC used for textual inference. Similar classications are required for other logic-based textual inference systems, but very little is written about the issue.

KEYWORDS: textual inference, lexical resources, transitive verbs. 1 Motivation

A computational system cannot be said to understand natural language if it cannot draw some rather direct inferences from a text. Central among them are inferences about the existence or non-existence of the entities and eventualities referred to. In this paper we look at two types of referentially opaque transitive verbs: verbs that are inherently negative and thus imply non-existence, and verbs with what we could call time-dependent opacity. In our discussion we make the simplifying assumption that the reader/the system considers the speaker trustworthy so that anything that the speaker is committed to as being true (or false) by virtue of the linguistic expression used, is treated as true (or false). Our notion of speaker commitment covers both entailments and presuppositions/implicatures (see (Karttunen and Zaenen, 2005) for a short discussion and (Potts, 2005) for an extensive motivation.) The detection of existential implications is an essential part of computing textual inferences, as conceived, for instance, in the RTE (Recognizing Textual Entailment) Pascal challenge (Dagan et al., 2006). A simplied example is given in (1). (1) Ed built a spacious hut. There was a spacious hut. YES

Our inquiry and classication extends (Nairn et al., 2006), which looks at implicative verbs with clausal complements. The relation between the two problems can be seen by comparing the implications of the verb prevent construed with a clausal complement or a nominal (event-denoting) complement:
Proceedings of the 3rd Workshop on Cognitive Aspects of the Lexicon (CogALex-III), pages 183194, COLING 2012, Mumbai, December 2012.

183

(2) (3)

Ed prevented Mary from leaving. Mary left. NO Ed prevented an accident. There was an accident. NO

The work reported here, as the one in (Nairn et al., 2006) and elsewhere, takes the view that inferential aspects are one of the main challenges that lexicographers interested in cognitive features of the lexicon need to address. The treatment proposed in (Nairn et al., 2006) aims at capturing the authors commitment to the truth or falsity of the complement clause of the verb. This classication is based both on the semantics of the complement-taking verb and on the syntactic type of the clause (e.g. factive forget that vs. implicative forget to). In the case of nominal complements, different factors need to be taken into consideration to determine the speakers commitment to the existence or non-existence of the denotation of the complement of the verb. These include: syntactic alternations of the verb,

the aspectual class of the verb phrase,

whether the nominal complement is event-denoting or not, the tense and aspect of the verb.

the aspectual properties of the nominal complement if it is event-denoting

A nal factor is the (in)deniteness of the direct object. Denite NPs tend to presuppose the existence of their referents. We will try to control for this by constructing examples with indenite NPs. Further complications will be discussed in the relevant sections.

Constraints on the classication

The classication was conceived to be used in conjunction with the representations produced by the automated system BRIDGE (Bobrow et al., 2007a),(Bobrow et al., 2007b). An important feature of these representations is the (un)instantiability of concepts, which corresponds to a claim of (non-)existence of an entity or occurrence of an event denoted by the concept. For instance, the sentence Negotiations prevented a strike involves events of the type negotiation, strike and preventing. Intuitively, the negotiations (whatever they may be) are presented as having occurred in the real world and so is the preventing event. In our representation, the terms corresponding to the words negotiations and prevented are instantiated in the top context, which corresponds to what the author of the sentence considers as true. But the term corresponding to strike should only be instantiable in the prevent-context; in the top context the term should be declared uninstantiable. See (Condoravdi et al., 2001), (Crouch et al., 2003) for motivation and details. The use of contexts, which correspond logically to partial possible worlds, allows us to represent further aspects of the situation prevented (for instance, how long that strike would have been or how bitter, etc.) without running into contradictions. The BRIDGE system, by default, treats the nominal arguments of verbs as carrying existential commitments at least in the context of the predication. This is obviously inadequate for the phenomena that we discuss in this paper and in section 5 we will propose an extension of the system that allows us to treat these phenomena more adequately.

184

Criteria for the classication

The main criterion adopted in the classication of the verb classes is whether the verb meaning indicates or does not indicate that the referent of the direct object existed before the eventuality denoted by the verb took place (pre-state) or not and whether the referent of the direct object exists after this eventuality (post-state). We focus on verbs that affect the existence of its object; for example, annul meets this description, but touch does not; only in the former case is the change with respect to the existence of the referent of the direct object part of the meaning of the verb. This difference in the implications that we intend to capture is exemplied in the contrast between (4) and (5):1 (4) Ed touched a teapot. Pre-state: There was a teapot. YES Post-state: There is a teapot. YES The judge has annulled Eds marriage. Pre-state: Ed was married. YES Post-state: Ed is married. NO

(5)

We will call verbs that indicate a change in the existence of the referent of their nominal complement, existential change verbs. In this paper we focus on this type of verbs and leave aside verbs that do not encode pre- and poststates (or with pre- and poststates that are the same). We present the different types of existential change verbs in the next section. In section 5, we discuss the representation of information about existence in our system.

Existential change verbs

In this class we identied eight sub-classes. They can be subdivided broadly into verbs of causation (the rst ve subclasses) and aspectual verbs (the last three subclasses).

4.1 Cause-type verbs

In this subsection we look at verbs whose nominal complement is headed by a noun that denotes an eventuality. Examples are cause, provoke, force, produce, bring about, induce. They share the following implications: (i) In positive environments: the predicate entails the occurrence of an eventuality/situation as a post-state, (ii) the caused eventuality/situation does not exist in the pre-state, (iii) in negative environments it is unknown whether the caused eventuality/situation has taken place. This indeterminacy is due to the indeterminacy of the scope of the negation. This pattern of inferences is exemplied in (6) and (7): (6) The decree caused trouble. pre-state: There was trouble. NO post-state: There was trouble. YES

1 This representation of the lexical meaning of verbs abstracts away from many factors that may intervene in a factual situation. For instance, (5) may be used in a case where Ed has gotten married again.

185

(7)

The decree didnt cause trouble. pre-state: There was trouble. UNKNOWN post-state: There was a trouble. UNKNOWN

Similarly, in (8), under normal circumstances, the speaker is committed to the non-occurrence of the revolution before the decree and to its occurrence after the decree. (8) The decree caused a revolution. pre-state: There was a revolution before the decree. NO post-state: There was a revolution after the decree. YES

When the progressive is used we nd some cases of the Imperfective Paradox((Dowty, 1979)), whereas in others there seems not to be such effect. Compare (9) and (10): (9) (10) The decree was causing trouble when it was revoked. The decree was causing a revolution when it was revoked.

Whereas in the rst example we conclude that there was trouble (i.e., the decree has caused some trouble), in the second we conclude that a revolution was avoided. We hypothesize that this is due to the nature of the eventuality that the nominal refers to. It is well-known that eventualities can be homogeneous (states and processes) or not (accomplishments and achievements). Trouble is homogeneous: a little bit of trouble is trouble but revolution is not: for instance, the beginning of a revolution is not a revolution. Causetype verbs can also be used to express a change of degree rather that a change from non-existence to existence as exemplied in (11). (11) The medicine induced an increase in blood pressure. pre-state: There was an increase of blood pressure. NO post-state: There was an increase in blood pressure. YES

Here the event that occurs is not the coming into existence of blood pressure but the increase in it. That is, of course, as expected: here the caused eventuality is the increase.

4.2

Verbs of creation

Closely related to the previous class are verbs of creation. They are different in that their complement refers to an object (physical or not) and not to an eventuality. This class includes verbs like build, bake (as in bake a cake), write, coin, compose, compute(as in compute a solution), concoct, construct (see Create verbs 24.4 and 24.1 in (Levin, 1993)) with the following cluster of implications: (i) in positive environments, there is a speakers commitment to the non-existence of the object before the event (entailment), and (ii) a commitment to the existence of the object after the event in the simple past tense, (iii) in negative environments, it is unknown whether the object exists, all we know is that the referent of the subject of the sentence did not bring it into existence.

186

(12)

John built a house. There is a [Link] John didnt build a house. There is a house. UNKNOWN

(13)

The effect of the Imperfective Paradox, however, is much stronger with these verbs than with the previous class: in positive environments, the use of the progressive form changes the speakers commitment as to the existence of the object. Therefore, we need a conditional marking in the rules, i.e. if the verb occurs in the simple past, the speaker is committed to the existence of the object, and if the verb occurs in the progressive, the speaker is committed to the non-existence of the object. (14) John is building a house. There is a house. NO

Verbs like draw, picture,sculpt etc. behave like verbs of creation when their nominal complement denotes the material or eventive result (draw a picture, sing a song). But they belong in the class of intensional verbs when their nominal complement denotes the content of the act: because what you draw may or may not exist in the real world (e.g. draw a unicorn).

4.3 Verbs of destruction

Verbs like destroy, extinguish, terminate, annul, invalidate, nullify, break off, annihilate, demolish, undo, wreck, resolve, share the following cluster of implications: In positive environments, (i) the speaker is committed to the existence of the object before the event (entailment) and (ii) the speaker is committed to the non-existence of the object after the event (entailment), and (iii) in negative environments, there is no commitment as to the existence of the object, but in common usage the speaker seems to be committed to the existence of the object (plausible, not strict inference). This is exemplied below: (15) The reghters extinguished a re. pre-state: There was a re. YES post-state: There is a re. NO The reghters didnt extinguish a re. pre-state: There was a re. UNKNOWN post-state: The re continues. UNKNOWN (The reghters didnt extinguish the re, but the rain did.)

(16)

The following two classes of verbs differ from the previous ones in that there is a modal component to their meaning; the nominal complement of the verb may denote either an eventuality that is true in the actual world or whose existence is restricted to a possible world other than the actual world.2
2 Speakers commitments of existence (or non-existence) allowed by the verbs presented in sections 4.4 and 4.5 may receive a morphological marking.

187

4.4 Avoid-type verbs

Verbs like avoid, elude, escape, whose meaning can roughly be paraphrased as manage not to experience something evaluated as bad, share the following cluster of implications when the nominal denotes an eventuality: In positive environments, (i) the speaker is committed to the potential occurrence of the eventuality denoted by the nominal complement before the event, and (ii) these verbs allow for both a wide and a narrow scope interpretation: in the wide scope interpretation, the speaker is committed to the occurrence of the eventuality after the event, and in the narrow scope interpretation, the speaker is committed to the non-occurrence of the eventuality after the event. (iii) In negative environments, the speaker is committed to the occurrence of the eventuality after the event. This is exemplied below for avoid: Narrow scope reading: (17) So heres some good news about how hundreds of workers avoided a layoff and didnt lose the jobs to downsizing . . . pre-state: There was a potential layoff. YES post-state: There was a layoff. NO Wide scope reading: (18) We landed in Lima only to nd that yet again we had narrowly avoided an earthquake (Tokyo all over again). This one was a massive quake of around 7-8 on the richter scale . . . pre-state: There was a potential earthquake. YES post-state: There was an earthquake. YES Note that we are concerned here with the inferences that are licensed by the lexical meaning of avoid. What has changed between the prestate and the poststate is precisely the speakers commitment as to the existence of the eventuality denoted by the nominal complement of the verb in the poststate: the non-occurrence of the eventuality in (17) and its occurrence in (18). When the nominal dependent denotes an object rather than an event, the object is assumed to exist in the pre- and in the poststate (wide scope reading): (19) We avoided a tree. pre-state: There was a tree. YES post-state: There was a tree. YES.

When the nominal complements direct denotation is an object, by semantic coercion the complement is interpreted as denoting an eventuality: (20) We avoided a tree. We avoided hitting a tree. We avoided the ball. We avoided being hit by the ball.

188

The direct denotation of these objects is assumed to exist before and after the act of avoidance but the eventuality described in the expansions is asserted not to take place.

4.5 Prevent-type verbs

Verbs like prevent, avert, block, inhibit, impede, hinder, deter, preclude, forbid, forestall, and cancel (in the sense of cause not to,prevent from happening), spare (in the meaning refrain from harming) share the following cluster of implications: In positive environments: (i) the speaker is committed to the potential existence of the object before the event, and (ii) the speaker is committed to the non-existence of the object after the event, (iii) there is a causal implication, and (iv) in negative environments, the speaker is committed to the existence of the object (plausible inference, seems to be the common usage). The nominal complement of this class of verbs is event-denoting. (21) The government prevented an oil spill in the bay. post-state: There was an oil spill in the bay. NO And nobody questions him because this mayor of a large American city who didnt prevent a major terrorist attack but seemed emotional in its aftermath has some special insight into the nature of terrorism . . . post-state: There was a major terrorist attack. YES

(22)

4.6 Begin-type verbs

This class includes aspectual verbs like begin, start, initiate that denote the beginning of an eventuality. When the referent of the nominal complement is an eventuality, these verbs share the following cluster of implications: (i) in positive environments, there is a speakers commitment as to the non-occurrence of the eventuality before the event (entailment), and (ii) there is no commitment as to the occurrence of the eventuality after the event; (iii) in negative environments, there is a commitment as to the non-occurrence of the eventuality after the event. We illustrate (i) in (24), 25, and (23)): (23) Ed and Mary didnt begin a relationship. pre-state: There was a relationship. NO Ed and Mary began a relationship. pre-state: There was a relationship. NO The queen began a visit to India. pre-state: There was a visit. NO

(24)

(25)

The status of (ii) depends on the properties of the eventuality referred to by the nominal complement. We hypothesize that the same distinction as discussed above in subsection 4.1 holds here too: when the eventuality is homogeneous, there is an existence commitment, when it is not, there is no commitment. Compare (26) and (27):

189

(26)

Ed began a trip to Paris. post-state: Ed made a trip to Paris. UNKNOWN The queen began a visit to India. post-state: The queen made a visit to India. UNKNOWN Ed and Mary began a relationship. post-state: There was a relationship. YES A boy playing with matches started a Southern California wildre. post-state: There was a wildre. YES

(27)

Adapting test for verbs we can illustrate the difference between the nouns in (26) and (27) as follows: (28) #They had a relationship in 2 months. #There was a wildre in two weeks. They made a trip to Paris in 2 weeks. They made a visit to India in two weeks.

As is the case with verbs, homogeneous events do not take temporal modiers that express the duration, whereas accomplishments do. When the verbs in this class take a nominal complement which is not eventdenoting, as is the case of book in (29), semantic coercion changes the denotation to an eventuality. As has been argued inter alia in (Pustejovsky, 1995), a sentence like (29) is ambiguous (at least) between starting to write a book and starting to read a book. As the combination of the verb and the nominal complement does not tell us which reading we have to choose, and this choice bears on the existential commitment about the object (see (30) and (31)), we mark the implications of (29) as UNKNOWN. (29) John started a book. pre-state: There is a [Link] post-state: There is a book. UNKNOWN John started to write a book. pre-state: There is a book. NO post-state: There is a book. UNKNOWN John started to read a book. pre-state: There is a book. YES post-state: There is a [Link]

(30)

(31)

In negative environments, the entailments are the same regardless of whether the denotation of the complement is an object or an eventuality.

190

4.7 Continue-type verbs

Verbs like continue and pursue, which we dont illustrate here, share the following cluster of implications: (i) the speaker is committed to the occurrence of the eventuality in the prestate (presupposition), (ii) in positive environments, the speaker is committed to the occurrence of the eventuality in the poststate (entailment), (iii) in negative environments, there is a speakers commitment as to the non-occurrence of the eventuality in the poststate. As with the previous class the implications depend on the aspectual class of the noun.

4.8 End-type verbs

Examples of end-type verbs are end, stop, cease, nish, discontinue, suspend, interupt. When the direct denotation of their nominal complement isnt an eventuality, its interpretation is coerced to an eventuality reading as is the case with begin and continue-type verbs. The end-type verbs share the following cluster of implications: (i) In positive environments, the speaker is committed to the non-occurrence of the eventuality after the end-event (entailment), and (ii) in negative environments, there is no commitment as to the occurrence of the eventuality, but in common usage the speaker seems to be committed to the occurrence of the eventuality after the end-event (plausible, not strict inference). Again, these verbs, as well as interrupt and discontinue, have different entailments depending on the aspectual properties displayed by the nouns that they take as complement. With nouns that denote activities, the speaker is committed to the existence of the activity, whereas this is not the case for nouns denoting accomplishments (or achievements): (32) Ed interrupted a discussion between the students. pre-state: The students had been discussing. YES post-state: There was a discussion between the students. NO Ed stopped the bleeding. pre-state: There was a bleeding. YES post-state: There is bleeding. NO John stopped the evaluation of the system. pre-state: There was an evaluation of the system. NO post-state: The system was evaluated. NO

(33)

(34)

However, we must further distinguish between two sub-classes within this class of verbs. The verbs end and nish behave differently from stop with accomplishment predicates: (35) Ed stopped a repair. post-state: There was a repair. NO Ed ended/nished a repair. post-state: There was a repair. YES

(36)

But this is not the case for nouns that denote activities or states, where both end and stop display the same pattern of implications:

191

(37)

The president ended/stopped a war. post-state: There is a war. NO

With respect to nominals whose primary denotation are objects, the interpretation depends again on the eventuality to which the interpretation of the nominal is plausibly coerced. For example, in (38) what is understood to have stopped is the ticking of the clock. Again nish and end behave differently. (38) John stopped a clock. post-state: There is a clock. YES Ed didnt nish a dissertation. post-state: There is a dissertation. NO

(39)

It is clear, then, that the entailments of sentences containing aspectual verbs like start, continue, end and stop, among others, depend on the aspectual properties of the nouns that they take as complements. For event-denoting nouns that are not deverbal (e.g. crime, accident, earthquake, ceremony, game, violence) little is known about these properties.

Representing existence information

Representation of the kind of information within the system BRIDGE is mediated by the relevant lexical information being imported into the Unied Lexicon (UL) (Crouch and King, 2005). Similarly to complement taking implicative verbs (Nairn et al., 2006), we expect to mark by hand the new implication signatures discussed, using some frequency criteria. We envisage using the British National Corpus (BNC) frequency list to uncover transitive verbs with these new kinds of implicative behavior. The appropriate lexical markings would then trigger rules constructing representations that encode the corresponding implications. We also envisage leveraging some of the Verbnet semantics information to check our proposed pre and post conditions. Notions of pre- and post conditions are widely used in logics for verication of programs, in the so-called Hoare logics. These kinds of conditions are also used in AI planning and in formal models of concurrency. However, they have found little use in semantics of natural languages. We propose to use these conditions as a rst approximation for the inferential meaning of verbs.

Conclusion

The present investigation of existential implications of transitive verbs shows that any implementation of logic-based textual inference needs to take into consideration different types of factors: the implicative behavior of a set of transitive verbs as a function of their lexical meaning, tense and aspect of the verbs, aspectual properties of the nominal complements, and deniteness. The combination of these factors as clues for the identication of the commitment of the speaker with respect to the existence of the entity or event denoted by the nominal complement of the verb is a challenge for any Entailment and Contradiction Detection system. Our attempt to spell out the existential inferences leads to theoretical problems: it shows that we need an ontological classication of the nominal complements in eventuality-denoting and object-denoting, that we need a coercion mechanism for the object-denoting nouns and

192

a distinction between the existential implications for the denoted object and for the denoted coerced eventuality and it forces us to look at the ill-understood aspectual properties of eventuality-denoting nouns whether they are morphologically deverbal or not.

References
Bobrow, D., Cheslow, B., Condoravdi, C., Karttunen, L., King, T. H., Price, L., Nairn, R., de Paiva, V ., [Link], and Zaenen, A. (2007a). Precision-focused textual inference. Proceedings of ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, pages 1621. Bobrow, D., Condoravdi, C., Karttunen, L., King, T. H., Price, L., Nairn, R., de Paiva, V ., [Link], and Zaenen, A. (2007b). Parcs bridge and question answering system. Proceedings of Grammar Engineering Across Frameworks, pages 2645. Condoravdi, C., Crouch, R., van den Berg, M., Everett, J., Stolle, R., Paiva, V ., and Bobrow, D. (2001). Preventing existence. In Proceedings of the Conference on Formal Ontologies in Information Systems (FOIS), Ogunquit, Maine. Crouch, D., Condoravdi, C., de Paiva, V ., and Stolle, R. (2003). Entailment, intensionality and text understanding. In Proceedings of the HLT-NAACL Workshop on Text Meaning, Edmonton, Canada. Crouch, D. and King, T. H. (2005). Unifying lexical resources. In Proceedings of the Workshop on the Identication and Representation of Verb Features and Verb Classes, Saarbruecken, Germany. Dagan, I., Glickman, O., and Magnin, B. (2006). The pascal recognising textual entailment challenge. In Lecture Notes in Computer Science, 3944, pages 177 190. Dowty, D. (1979). Word Meaning and Montague Grammar. Reidel, Dordrecht. Karttunen, L. and Zaenen, A. (2005). Veridicity. In Katz, G., Pustejovsky, J., and Schilder, F., editors, Annotating, Extracting and Reasoning about Time and Events, volume Dagstuhl Seminar Proceedings 05151. Dagstuhl, Germany. Levin, B. (1993). English Verb Classes and Alternations: A Preliminary Investigation. The University of Chicago Press, Chicago. Nairn, R., Condoravdi, C., and Karttunen, L. (2006). Computing relative polarity for textual inference. In Proceedings of ICoS-5 (Inference in Computational Semantics), Buxton, UK. Potts, C. (2005). The Logic of Conventional Implicatures. Oxford Studies in Theoretical Linguistics. Oxford University Press, Oxford. Pustejovsky, J. (1995). The generative Lexicon. MIT Press.

193