Degrees of Stativity: the Lexical Representation of Verb Aspect

Degrees of Stativity: The Lexical Representation of Verb Aspect

Judith L. Klavans and Martin Chodorow IBM T.J.Watson Research Yorktown Heights, NY and ~Hunter College of the City University of New York

Abstract 1. Les ptopri6t6s lexicales inh6rentes peuvent fitre dfitermin&s en appli- L'acqnisition automatique de connaissance qurmt aux textes une s&ie de tests lexicale h partir de larges corpus s'est llngnistiques bien &ablis. Cela donne essentiellement oceup& des phfinom~nes l'analyse de corpus une dimension de co-occurrence, aux dfipens des traits suppl~mentaire qui va au-delh des lexicaux inh~rents. Nous prfisentons ici ph~nom~nes de coocurrence (eomme une m&hodologie qui permet d'obtenir par example information mutuelle, l'information sfimantique sur l'aspect du substitutabilitfi). verbe en analysa~t automatiquement un corpus et en appliquant des tests lin- 2. Une propri~t~ lexieale n'a pas be- guistiques h l'aide d'une stifle d'outils soin d'etre discrete mats peut &re d'anrdyse structura]e. Lorsque ces deux repr~sent~e en tant que valeur ~ in- terprdter comme une tendance ou t~:hes sont accomplies, nous proposons probabilitY; de cette mani~re, ]e oh- une rfipresentation de l'aspect du verbe qni associe une valeur de mesure pour les .jet lexieal exprime la propri~t~ donn~e difffirents types d'fiv~nements. Les mesures darts un contexte non-marqu& Les valeurs der/v~es du corpus sont vari- zefl~tent l'usage typique du verbe, et par ables, e'est-h-dire des valeurs h degr~s. consfiquent une mesure de rfisistance ou de non-r~sistance h la coercion dans le Des tests linguistiques ont &~ automa- eontexte de la phrase. Les rfisultats que tiquement appliques aux corpus analys~s nous rapportons ici ont fitfi obtenus de de mani~re k d&erminer la valeur initiale deux mani~res: en extrayant l'information de l'aspect pour la stativit~, et ce, pour un n&essrSre h partir du corpus &iquetfi de ensenble de verbes frequents, repr~sentant Francis and Ku~era (1982), et en fais~nt plus de 90% des occurrences de verbes dans tourner un analyseur syntaxlque (MeCord un corpus d'un million de roots. 1980, 1990) sur le corpus du Reader's Di- gest afin d'extraire une information plus pr&ise sur l'usage du verbe dans le texte. Notre travail iUustre deux aspects:

ACTESDE COLING-92, NArTr~. 23-28 AOl~r 1992 1 1 2 6 PROC. OFCOLING-92. NANTES, AUG. 23-28, 1992 Degrees of Stativity: The Lexical Representation of Verb Aspect

Judith L. Klavans and Martin Chodorow t IBM T.J.Watson Research Yorktown Heights, NY amd ~Hunter College of the City University of New York

Abstract value for stativity for u set of frequent verbs, cov- eting over 90% of verb occurrences in a one million The automatic acquisition of lexiced knowledge word corpus. from large corpora has dealt primarily with coocurrence phenomena, at the expense of inherent lexic~l features. We present here u method- 2 Event types ology for obtaining semantic information on verb aspect by parsing a corpus and automatically up- Aspect can be informally defined as a property plying linguistic tests with a set of structural anal- which reflects the temporal organization of an ysis tools. Once applied, we propose a represen- event, such as duration (whether an event involves tation for verb aspect that associates a value with change over time), telicity (whether it has a defi- weights for event types. Weights reflect typical nite endpoint or is ongoing), iteratlvlty (whether verb use, and thus represent a measure of the resis- or not it is repetitive) and so on (see Comrie 1976.) tance or ease of coercion in sentential context. The We assume three event types, following Vendler results we report here have been obtained in two 1967, refined by many others, and recently re- ways: by extracting relevant information from the cast from the perspective of computational lexicon tagged Brown corpus (Francis and Ka~era 1982), building by Pastejovsky 1991, and Pustejovsky and by running ~ parser (McCord 1980, 1990) on and Boguraev (to appear) 1: the Reader's Digest corpus to extract more accurate information on verb usage in text. State(S): knon, resemble, b~, love Process(P): run, walk, eviM Transition(T): give, open, build, destroy 1 Overview A verb can enter into u construction which may Our work illustrates two points: change the overall phrasal or sentential event structure, as ttte result of event-coercion. Coer- 1. Inherent lexical properties can be deterufined cion can be defined as the process by which verbs by applying a battery of established hnguistic in context appear to demonstrate some regular tests to corpora~ This adds to the utility of ambiguities, i.e. they appear to change categories. corpus analysis a dimension beyond coocnr- Pustejovsky argues that u verb is inherently (lex- fence phenomena (e.g. mutual information, fealty) specified as being of a certain event type substitutability). (e-type).

2. A ]exical property need not be discrete bat Schema: e-type(Verb) - SIPIT can be represented as a value to be inter- e-type(reeenble) ~ S preted as a tendency or probability for the e-type(run) ~ P lexical item to exhibit the given property in e-type(give) - T an unmarked context. Corpus-derived values One are variable, i.e. DEGREE ValU~ Figure INolice that accomplishment verbs (e.g. ~palnt • pie- Linguistic tests have been automatically applied ture"), and Itchleveme, t veltbw ("reach a goM") are both to parsed corpora to determine an initial aspectual considered tr~neition verbi,

AcrEs DE COLING-92. NANIJ£S. 23-28 AO~I 1992 1 1 2 7 PnOC. OF COLING-92, NAhrrgs, AUG. 23-28, 1992 We call this the NON-DEGREE APPROACH, since mapping of resources into a common central lexi- there is no degree specified. Our claim is that, cat datu base, rather than being dictionary bound. rather than the representation in Figure One, the The view is further expanded in Byrd 1989. Puste- event structure is a vector of values, e-type (Verb) jovsky and Boguraev (to appear) argue that a the~ = (S,P,T). We deal here only with the statics/non- ory of lexlcal semantics making use of a knowl- statics distinction, since there are regular differ- edge representation framework offers an expressive ences between states versus activities and accom- vocabulary for the representation of such lexical plishments (Lakoff 1965). We propose a simplified information. They outline a concrete framework representation e-type(Verb)=Vo(x) with a single consisting of four levels of lexlcal meaning: (1) Ar- value for stativity, Vs, the non-statics value being gument Structure; (2) Event Structure; (3) Quails merely the complement, i.e. 1I, + V~s = 1. Our Structure; and (d) Lexical Inheritance Structure. position can be summarized as: Pustejovsky 1991 provides formal details of the event structure specification for verbs, with a for- Schema: e-type(Verb) ~ Vs(Z) real explanation of semantic type coercion. where 0 < Vs(X)< 1 and Specification of verb and phrasal aspect mat- nhore VH + V~s = 1 . ter to NLP systems in several ways. For example, when a stative verb is used in the present tense, Figure T~o it involves only one occasion of the event. In con- We call our position the DEGREE APPROACH trust, a non-stative usage involves a frequentative, since a numeric value or DEGREE iS specifted. We iterative, or repetitive interpretation. Thus: agree with the notion of assuming basic verb types, John knous the answers. (single) with coercions, rostra the position proposed by John runs. (repetitive) Dowry 1979 that there be different entries for dif- Sue builds houses. (repetitive) ferent usages. To allow the process interpretation in sentences It is necessary to understand the interpretation of like: statlvity for several reasons. In language genera- tion, adverbial adjuncts which have a durational (i) The child iv being a fool today. interpretation are, under most circumstances, dis- (2) She i~ resembling her mothez allowed: more and more each day. *John knew the answers for three days. the phrase "more (and more)" in (2), a temporal John ran for three hours. expression, acts to "coerce" or force a non-stative Sue built houses for three decades. event (in this case a process). ~ In fact, the verb "be", often touted as fundamentally the most sta- For text analysis, if a verb which is usually stative tics verb in the language is frequently coerced into is used with a non-statlve adverb, such as "de- process and transition. liberately", it can be inferred that the event is non-stative, and not the reverse. If the event is non-statics, then it can be inferred that it could 3 Computational Lexicons be a repetitive or iterative event. This can effect not only the semantic interpretation of the text it- The work reported here is part of an ongoing self, but also translation and the choice of adverb. project in the bexical Systems Group at IBM Re- 3 search to extract and represent lexical knowledge 3Many of these issues are discussed in the CL Special in CompLex, a computational lexicon which will Issue on Tense and Aspect (June, 1988) in articles by Hin- be able to provide information to natural language niche, Moens and Steedman, Nakhimovsky,Passoneau, and processing systems. The seeds of this project are Webber. For example, Passoneau demonstrates how, with- out an ~ccurate specification of the ~pectual tendencies presented in Klavans 1988, where it is argued that of the verb coupled with the effect of temporal and aspec- the building of a computational lexicon requires tual adjuncts, messages, which tend to be in the present tense, ttre not correctly understood nor generated in the ~Our position might be viewed in terms of ~signing PUNDIT system. For instance, "the pressure is low" must probabilities to the arc tranmltions in ~ system luck as that be interpreted at statlve, whereas "the pump operates" proposed by Moths and Steedman 1988, in which coercions must be interpreted as a process. In machine translation, ~re recursive but carefullyconstrained along several key pa- for example, the verb parecerse meaning 'to resemble one r~neters defining possible transitions within a finite-state another' is even leHs statics in Spanish than in English. trtmsition network. Thus, sentence (2) above should be translated into Spanish

ACRES DE COLING-92, NAntES, 23-28 AO'dr 1992 1 1 2 8 PRec. OF COLING-92, NANTES.AUG. 23-28, 1992 Aspect is complex. Stativity is not u simple over its frequency as a verb in the same corpus: feature such as animate(+/-). To ob ..... this, it suffices to look in any comprehensive grammar, F(Verb(X) for example, Quirk et al. 1972, in wtfich lexieal whore 0 < V, < 1 verbs are divided into the classes "dynamic" and "stative", with the caveat that "it would be more A value closer to 0 indicates that a verb prefers sta- accurate to speak of 'dynamic' and "stative' uses tivlty. ]'his basic value can be modified by other of verbs" (p. 94-95). Dowty 1979 observes that the tests, such the force/persuade test, and the delib- issue of interpretation and aspect involves " the erately/carefully test. thorny problems of polysemy and honrophony"(p. We are aware that this is an overMmplificatlon 62). Since some verbs are "more statlve" than of the property of stativity. Ilowever, our goal is others, meaning that the most common unmarked to search for the most robust and pragmatically use is as a marker of sentential or event stativity, possible tests to start with. Our technique has the lexicon must embody this lexieal fact. Our given results which concur with our intuitions, al- proposal provides the capability in the lexicon of though there are limitations discussed below, s In representing that variability, combined with auto- order to do this, we needed u text tagged for part matic means of inducing variability values. of speech. Otherwise, instances such as "hear- ing aid", '% knowing look" would be taken as instances of progressive verbs. 4 Procedure

Some standard tests for the stative/nou-statlve 5 Results distinction are given in Dowty 1979, such as the progressive, imperative, cmnplement of Tagged Ku~era and Francis Data force/persuade, intentionM adverbs, and pseudo- The Brown corpus s provided a convenient starting cleft, as in: 4 point, since words are tagged for part of speech. Progressive : non-statives ltowever, a closer look at the tag set itself reveals * John is kneeing the anuuer. a weakness which could bias our results. The John in running. tag VBG is used to tag "verb, present partici- John in building s house. ple, and gerund." Thus, there is no distinction Complement of force/persuadu: in the label for the different usages of the "-ing" * flail forced John to know the ansner. form, although some "-ing" forms are labelled NN flail pursuadud Amy to run. for noun or JJ for adjective. Despite this prob- John forced Bill to build a house, lem, we chose to use the Brown data, knowing Adverbs, e-g- deliberately, carufully: that the numbers aright be distorted. We started * Gull doliburatuly knou the anBeer. with a list of the 100 most frequent words la- Evelyn ran carefully. belled as verbs (i.e. the 100 most frequent verbs, Sue carefully built a house. which account for over 90verbs which have been discussed in the literature on stativity (such as Even though tests for stativity involve interactions "resemble", "matter", "intend"). Figure Three between semantic and syntactic facets of a sen- lists some results, ordered by degree of stativity. tence, we have chosen three tests for stativlty, the most robust being the tendency for a verb to occur e-type(try)= Vs(.3326) in the progressive. Unlike other tests, the progres- u-type(work)= V0(.3064) sive itself is a statement of duration and process. e-typu(nit)= V,(.2929) We hypothesized that degree of stativity, i.e. a u-type(run)= V, (.2853) value for stativity Vo, conld be inferred by deter- e-typu(play)- Vs(.2552) mining the ratio of total frequency of occurrence aPustejovsky, personM communication, hem pointed out of a verb in a progressive usage, past or present, some problem cases where the progressivefMsely indicates that root is non-Jtative, such as lie/sit verbs (The book is in the simple present, not the progre,mive,a~ i*t EngllJh. lying on the shelf, The c~p is sitting on the cownteO, and eada d~a .lla se parsee I~s a ¢~ ~dre. mental attitude verbs (John il thinking that he should 90 For activity verbs, the progressive in EngliJh translates to home now, Sdohn is knowing that he should go home note, the progressive in SpalfiJh. Marll is suspecting that John will propose tonight. ) Such 4Activltes and accompllshmentl are subject to other dim- problemt van be rewolvedby fine-tuning testJ. tinguilhing tests which are not the wubject of this paper. °We thank Slav~ Kate for pointin[ us to this resource.

ACTESDE COLING-92, NANTES,23-28 AOt~V 1992 1 1 2 9 PROC. OF COLING-92, NANTES,AUO. 23-28, 1992 e-type(move)= Vs(.2326) e-type(go)= 1I,(.2901) (n = 107t) e-type(go)= 11,(.2304) e-type(foiler)= 1I,(.0375) (n = t33) e-type(follow)= V,(.1234) e-type(give)= V,(.0209) (n = 430) e-type(give)= V,(.0783) e-type(become)= V,(.0507) (n = 493) e-type(become)= V,(.0718) e-type(hear)= V,(.OO00) (n = 242) e-type(hoar)= V,(,0669) o-type(feel)= V,(.0317) (n = 315) e-type(~eel)= V,(.0637) e-type(uppenm)= V,(.0272) (n = 147) e-type(uppettr)= V,(.0481) e-type(knov)= V,(.OOO0) (n = 630) e-type(know)= 1/',(.0332) e-type(want)= V,(.0000) (n = 466) e-type(vent)= V,(.0253) o-type(need) = V,(.O000) (n = 258) e-type(need)= Vw(.0197) e-type(be)= V,(.0124) (n = 12482) e-type(be)= V,(.0173) e-type(seem)= V,(.0000) (n = 260) e-type(eeem)= V,(.Ol20) Figure Four Figure Three Additional Syntactic Tests As can be seen, the ranking roughly reflects intuitions about stativity, so, for example, seem is The progressive test is only one of several tests, more stative than hear, which is in turn more st~ and in and of itself is certainly inadequate. Sev- tire than run. eral tests mast be run, and then event values must be computed for each linguistic test. Two param- eters are involved: the strength of each test as an Parsing with English Slot Grammar indicator of e-type, and the sparsity of data. The second more refined method utilizes a parser We have preliminary results on two tutditional to analyze text, and to record verb usages. For this tests: the force/persuade test and the deliber- purpose, we used the English Slot Grammar (Me- ately/carefully test. Synonyms and taxonyms Cord 1980, 19901 a broad-coverage parser writ- were collected for each (ad)verb, data were ex- ten in PKOLOG. 7 To obtain counts of verb us- tracted frmn the corpus and parsed. For example, ages from the representations produced by ESG, the following shows how a sentence with "force" we used a tool for querying trees (QT), built by was analyzed. However, more datais needed, from the second author, also in PROLOG. The cor- a larger corpus, for the results to be significant. pus is the Reader's Digest (RD) corpus, consist- The same applies for the adverb test. ing of just over one million words. We took the Difficultiue forced him to abandon ... same llst of the 115 most frequent and most fre- verb(force) inf_ camp_verb (abandon) quently discussed verbs that was used for obtaining values from the Brown corpus. We ex- Figure Five - Verb "Force" tracted all sentences under 30 words containing the inflectional variants of these verbs from the Results of running and computing the weights of RD corpus. We then ran the parser on this sub- different tests on larger corpora will be reported corpus, ran QT, and obtained values for the dif- in future publications. ferent verb usages. Unlike the Brown data, dis- tinctions are made between the gerundive and participial usages. Figure Four gives results for 6 Discussion some verbs, listed in the same order as in Figure Three, with n indicating the number of tokens: As expected, the results from each corpus differ considerably; we believe this is due primarily e-type(try)= V,(.2167) (n = 286) to surface tagging vs. full parsing. The results e-type(york)= V,(.1311) (n ffi 244) from the second method using ESG do not carry e-type(sit)= Vs(.1506) (n = 146) the noise from the ambiguous VBG tag from the e-type(run)= V,(.1565) (n = 230) Brown corpus. However, there are two important e-type(play)= V,(.2315) (n = 95) points to be made: (1) One million words is simply e-type(move)= V,(.0798) (n = 2131 not enough. More data need to be (and will be) run to get a more complete and accurate count. TAn exception to the quality it imperatives, **here there were some errors in the parsing; they were removed from These are to be viewed as preliminary data, us~ our cMculationt. able but not complete. (2) The value V,(.0000)

AcrEs DF.COLING-92, NANTES, 23-28 Aour 1992 l 1 3 0 PREC. OV COL1NG-92, NANTES, AUG. 23-28, 1992 cannot be considered categorical. Verbs are gen- 7 References erally adaptable in context, s It is a known fact that value* either for words with low frequencies 1. Byrd, Roy. 1989. "Discovering Relationships or words in low frequency constructions must be unsung Word Sense*", Dictionaries in the Elec- computed on very large corpora (Liberman 1989.) tronic Age: Proceedings of the Fifth Annual Con- The current limitations of this approach must Jerenec o] the University of Waterloo Centre ]or be clearly stated. First of all, this method con- the New Oxford English Dictionary rates the polysemons usages of certain verbs and, 2. Comrie, Bernard 1976. Aspect. Cambridge in English, of the verb-particle construction. It University Press: Cambridge, English. could be argued that with enough corpus data, 3. Dowry, David. 1979. Word Meaning and Mon. this would become unimportant, but we believe *ague Grammar. Dordrecht: tlolland. this position not correct. What is required is a 4. Francis, W. Nelson and Henry Ku6era. 1982. fuller analysis of adjuncts in order to know if n Frequency Analysis of Enlglish Usage: Lexicon verb has been coerced. For example, it could and Grammar. tlonghton Mifflin: Boston. be the case that a verb which is S in the an- 5. Klavaas, Judith L. (1988) "Building a Corn- marked case (i.e. in a neutral context) tends to puts*loaM Lexicon using Machine Readable Dic- appear as a T verb frequently, since that verb tionaries", Proceedings of the Third International might not occur frequently in a null context at Congress of the European Association for Lexicog- all. As another example, consider the case of a raphy. Budapest, Hungary. typical stative verb "know". With the object "an- 6. Lakoff, George. 1965. "On the Nature of swer", "know" becomes typically inchoative, e.g. Syntactic Irregularity", PhD Dissertation, Depart- "know the answer by tomorrow" meaning "be- meat of Linguistics. Indiana University: Bloom- come knowledgeable of the answer", or it could ingtoa, Indiana. be used in the transition sense, e.g. "he will know 7. Liberman, Mark. 1989. "How Many Words Do the answer by tomorrow" meaning "he does not People Know?" Invited Talk delivered at the 27th know now and will know then." Thus, it could Annus.l Meeting of the Association for Compata~ be underlying semantic structure, and not sur- *tonal Linguistics. Vancouver, B.C., Canada. face syntactic behavior, that determines coercion 8. McCord, M. C. 1980. "Slot Grammars" Com- possibilities. 9 In conclusion, The DEGREE AP- putational Linguistlcs, Vol 6:31-43. PROACH captures the fact that verbs have degrees 9. MeCord, M. C. 1990. "SLOT GRAMMAR: of e-type, i.e. that some verbs are more pliable A System for Simpler Construction of Practi- than others. Thus, rather than the non-degree cal Natural Language Grammar*," in R. Sender values in Figure One, we argue for entries like: (Ed.), International Symposium on Natural Lan- guage and Logic, Lecture Notes in Computer Sci- e-eype(reeeablo) = V,(.0740) ence, Springer Verlag. e-typo(go) = V#(.2304) 10. Moensp Marc and Marc Steedman. 1988 e-type(seem)= V,(.012O) "Temporal Ontology and Temporal Reference" A corpus-breed method can be used to antomatl- Computational Linguistics. Vol. 14:2:15-28. cally derive values for e-type, i.e. under n certain 11. Pustejovsky, James. 1991. "The Syntax of cut-off, the verb is stative, but alternbh in con- Event Structure" Cognition Vol. 41:103:47-82. text. More importantly, it gives a degree of likeli- 12. Pustejovsky, James and Bran Bognrnev. to hood that given any context, the verb will be used appear. "Lexical Knowledge l~.epreaentation and statively or non-statively. Natural Language Processing" IBM Journal of *This is a tact which any semanticist who is trying to Research and Development. argue a firm point can at*el* to. 13. Quirk, Randolph, Sidney Greenbaum, Geof- tAho, there appear to be some clashes on the resulting frey Leech, Jan Svartvik. 1972. A Comprehen- values and intuitions, thus leading to the suggeJtion that either our intuitions are not correct or that the method it sive Grammar of the English Language, Longman: unreliable. We have not, in fact, addressed the issue of London. underlyin~ semantic representation (e.g. in terms of prim. 14. Teany, Carol Lee. 1987. "Grnmmatlcalizing itivea) in this paper. It has been suggested that syntactic Aspect and Affectedness", PhD Dissertation, De~ tests [or aspect might be flawed, and that the only way to distinguish aspectual classes it via the semantic conse- par*men* of Linguistics. M.I.T., Cambridge: Mas- quences ot a restive vs. non0tative proposition. If correct, sachusetts. the approach of extracting values based on syntactic tests 15. Vendler, Zeno. 1967. Linguistics iu Philoso. will fail by definition, regardless of whether the value, are phy, Coruell University Press: Ithaca, New York. assigned manually or automatically.

ACTES DE COLING-92, NAgrES, 23-28 ho~rr 1992 1 1 3 1 PROC. oF COLING.92, NANTES. AUG. 23-28, 1992