Deriving Verbal and Compositional Lexical Aspect for NLP Applications Bonnie J. Dorr and Marl Broman Olsen University of Maryland Institute for Ad.vanced Computer Studies A.V. Williams Building College Park, MD 20742, USA bonnie ,molsen©umiacs. umd. edu Abstract selecting tense and lexical items for natural language Verbal and compositional lexical aspect generation ((Dorr and Olsen. 1996: Klavans and provide the underlying temporal struc- Chodorow, 1992), cf. (Slobin and Bocaz, 1988)). In ture of events. Knowledge of lexical as- addition, preliminary pyscholinguistic experiments pect, e.g., (a)telicity, is therefore required (Antonisse, 1994) indicate that subjects are sensi- for interpreting event sequences in dis- tive to the presence or absence of aspectual features course (Dowty, 1986; Moens and Steed- when processing temporal modifiers. Resnik (1996) man, 1988; Passoneau, 1988), interfacing showed that the strength of distributionally derived to temporal databases (Androutsopoulos, selectional constraints helps predict whether verbs 1996), processing temporal modifiers (An- can participate in a class of diathesis alternations. tonisse, 1994), describing allowable alter- with aspectual properties of verbs clearly influenc- nations and their semantic effects (Resnik, ing the alternations of interest. He also points out that these properties are difficult to obtain directly 1996; Tenny, 1994), and selecting tense and lexical items for natural language gen- from corpora. eration ((Dorr and Olsen, 1996; Klavans The ability to determine lexical aspect, on a large and Chodorow, 1992), cf. (Slobin and Bo- scale and in the sentential context, therefore yields caz, 1988)). We show that it is possible an important source of constraints for corpus anal- to represent lexical aspect--both verbal ysis and psycholinguistic experimentation, as well and compositional--on a large scale, us- as for NLP applications such as machine transla- ing Lexical Conceptual Structure (LCS) tion (Dorr et al., 1995b) and foreign language tu- representations of verbs in the classes cat- toring (Dorr et al., 1995a; Sams. 1995; Weinberg et aloged by Levin (1993). We show how al., 1995). Other researchers have proposed corpus- proper consideration of these universal based approaches to acquiring lexical aspect infor- pieces of verb meaning may be used to mation with varying data coverage: Klavans and refine lexical representations and derive a Chodorow (1992) focus on the event-state distinc- range of meanings from combinations of tion in verbs and predicates; Light (1996) considers LCS representations. A single algorithm the aspectual properties of verbs and affixes; and may therefore be used to determine lexical McKeown and Siegel (1996) describe an algorithm aspect classes and features at both verbal for classifying sentences according to lexical aspect. and sentence levels. Finally, we illustrate properties. Conversely. a number of works in the how knowledge of lexical aspect facilitates linguistics literature have proposed lexical semantic the interpretation of events in NLP appli- templates for representing the aspectual properties cations. of verbs (Dowry, 1979: Hovav and Levin, 1995; Levin and Rappaport Hovav. To appear), although these have not been implemented and tested on a large 1 Introduction scale. Knowledge of lexical aspect--how verbs denote situ- We show that. it is possible to represent the lexical ations as developing or holding in time--is required aspect both of verbs alone and in sentential contexts for interpreting event sequences in discourse (Dowty, using Lexical Conceptual Structure (LCS) represen- 1986; Moens and Steedman, 1988; Passoneau, 1988), tations of verbs in the classes cataloged by Levin interfacing to temporal databases (Androutsopou- (1993). We show how proper consideration of these los, 1996), processing temporal modifiers (Antonisse, universal pieces of verb meaning may be used t.o 1994), describing allowable alternations and their se- refine lexical representations and derive a range of mantic effects (Resnik, 1996; Tenny, 1994), and for meanings from combinations of LCS representations. 151 A single algorithm may therefore be used to deter- therefore proposed that aspectual interpretation be mine lexical aspect classes and features at both ver- derived through monotonic composition of marked bal and sentential levels. Finally, we illustrate how privative features [+/1~ dynamic], [+/0 durative] and access to lexical aspect facilitates lexical selection [+/0 relic], as shown in Table 2 (Olsen, To appear and the interpretation of events in machine transla- in 1997, pp. 32-33). tion and foreign language tutoring applications, re- With privative features, other sentential con- spectively. stituents can add to features provided by the verb but not remove them. On this analysis, the activity 2 Lexical Aspect features of march ([+durative, +dynamic]) propa- gate to the sentences in (1). with [+telic] added by Following Olsen (To appear in 1997), we distinguish the NP or PP, yielding an accomplishment interpre- between lexical and grammatical aspect, roughly tation. The feature specification of this composition- the situation and viewpoint aspect of Smith (1991). ally derived accomplishment is therefore identical to Lexical aspect refers to the '0ype of situation denoted that of a sentence containing a relic accomplishment by the verb, alone or combined with other sentential verb, such as produce in (2). constituents. Grammatical aspect takes these situa- tion types and presents them as impeffective (John (2) The commander produced the campaign plan. was winning the race/loving his job) or perfective (John had won/loved his job). Verbs are assigned to Dowry (1979) explored the possibility that as- lexical aspect classes, as in Table i (cf. (Brinton, pectual features in fact constrained possible units 1988)[p. 57], (Smith, 1991)) based on their behavior of meaning and ways in which they combine. In in a variety of syntactic and semantic frames that this spirit, Levin and Rappaport Hovav (To appear) focus on their features. 1 demonstrate that limiting composition to aspectu- A major source of the difficulty in assigning lex- ally described structures is an important part of an ical aspect features to verbs is the ability of verbs account of how verbal meanings are built up, and to appear in sentences denoting situations of multi- what semantic and syntactic combinations are pos- ple aspectual types. Such cases arise, e.g., in the sible. context of foreign language tutoring (Dorr et al., We draw upon these insights in revising our LCS 1995b; Sams, 1995; Weinberg et al., 1995), where lexicon in order to encode the aspectual features of a a 'bounded' interpretation for an atelic verb, e.g., verbs. In the next section we describe the LCS rep- march, may be introduced by a path PP to the bridge resentation used in a database of 9000 verbs in 191 or across the field or by a NP the length of the field: major classes, We then describe the relationship of aspectual features to this representation and demon- (1) The soldier marched to the bridge. strata that it is possible to determine aspectual fea- The soldier marched across the field. tures from LCS structures, with minimal modifica- The soldier marched the length of the field. tion. We demonstrate composition of the LCS and Some have proposed, in fact, that aspec- corresponding aspectual structures, by using exam- tual classes are gradient categories (Klavans and pies from NLP applications that employ the LCS Chodorow, 1992), or that aspect should be evaluated database. only at the clausal or sentential level (asp. (Verkuyl, 1993); see (Klavans and Chodorow, 1992) for NLP 3 Lexical Conceptual Structures applications). Olsen (To appear in 1997) showed that, although We adopt the hypothesis explored in Dorr and Olsen sentential and pragmatic context influence aspectual (1996) (cf. (Tenny. t994)), that lexical aspect fea- interpretation, input to the context is constrained in tures are abstractions over other aspects of verb se- large part by verbs" aspectual information. In par- mantics, such as those reflected ill the verb classes in titular, she showed that the positively marked fea- Levin (1993). Specifically we show that a privative tures did not vary: [+telic] verbs such as win were model of aspect provides an appropriate diagnostic always bounded, for exainple, In contrast, the neg- for revising [exical representations: aspectual inter- atively marked features could be changed by other pretations that arise only in the presence of other sentence constituents or pragmatic context: [-telic] constituents may be removed from the lexicon and verbs like march could therefore be made [+telic]. derived compositionally. Our modified LCS lexicon Similarly, stative verbs appeared with event inter- theu allows aspect features to be determined algo- pretations, and punctiliar events as durative. Olsen rithmically both from the verbal lexicon and from composed structures built from verbs and other sen- 1Two additional categories are identified by Olsen (To appear in 1997): Semelfactives (cough, tap) and Stage- tence constituents, using uniform processes and rep- level states (be pregnant). Since they are not assigned resentations. templates by either Dowty (1979) or Levin and Rappa- This project on representing aspectual struc- port Hovav (To appear), we do not discuss them in this ture builds on previous work, in which verbs were paper. grouped automatically into Levin's semantic classes