VerbNet Representations: Subevent Semantics for Transfer Verbs Susan Windisch Brown Julia Bonn James Gung University of Colorado University of Colorado University of Colorado [email protected] [email protected] [email protected] Annie Zaenen James Pustejovsky Martha Palmer Stanford University Brandeis University University of Colorado [email protected]@[email protected] Abstract verb senses and introduce a means for automati- cally translating text to these representations. We This paper announces the release of a new ver- explore the format of these representations and the sion of the English lexical resource VerbNet types of information they track by thoroughly ex- with substantially revised semantic represen- amining the representations for transfer of posses- tations designed to facilitate computer plan- sions and information. These event types are ex- ning and reasoning based on human language. We use the transfer of possession and trans- cellent examples of complex events with multi- fer of information event representations to il- ple participants and relations between them that lustrate both the general framework of the change across the time frame of the event. By representations and the types of nuances the aligning our new representations more closely new representations can capture. These repre- with the dynamic event structure encapsulated by sentations use a Generative Lexicon-inspired the Generative Lexicon, we can provide a more subevent structure to track attributes of event precise subevent structure that makes the changes participants across time, highlighting opposi- tions and temporal and causal relations among over time explicit (Pustejovsky, 1995; Pustejovsky the subevents. et al., 2016). Who has what when and who knows what when are exactly the sorts of things that we 1 Introduction want to extract from text, but this extraction is dif- ficult without explicit, computationally-tractable Many natural language processing tasks have seen representations. These event types also make up a rapid advancement in recent years using deep substantial portion of VerbNet: 37 classes of verbs learning methods; however, those tasks that re- deal with change of possession and transfer of in- quire precise tracking of event sequences and par- formation out of VerbNet’s 300+ classes, covering ticipants across a discourse still perform better 810 verbs. using explicit representations of the meanings of each sentence or utterance. To be most useful 2 Background for automatic language understanding and genera- tion, such representations need to be both automat- The language resource VerbNet (Kipper et al., ically derivable from text and reasonably format- 2006) is a hierarchical, wide-coverage verb lexi- ted for computer analysis and planning systems. con that groups verbs into classes based on sim- For applications like robotics or interactions with ilarities in their syntactic and semantic behavior avatars, commonsense inferences needed to under- (Schuler, 2005). Each class in VerbNet includes stand human language directions or interactions a set of member verbs, the thematic roles used are often not derivable directly from the utterance. in the predicate-argument structure of these mem- Tracking intrinsic and extrinsic states of entities, bers (Bonial et al., 2011), and the class-specific such as their existence, location or functionality, selectional preferences for those roles. The class currently requires explicit statements with precise also provides a set of typical syntactic patterns and temporal sequencing. corresponding semantic representations. A verb In this paper, we describe new semantic rep- can be a member of multiple classes; for exam- resentations for the lexical resource VerbNet that ple, run is a member of 8 VerbNet classes, in- provide this sort of information for thousands of cluding the run-51.3.2 class (he ran to the store) 154 Proceedings of the First International Workshop on Designing Meaning Representations, pages 154–163 Florence, Italy, August 1st, 2019 c 2019 Association for Computational Linguistics and the function-105.2.1 class (the car isn’t run- these caused-motion frames. ning). These memberships usually correspond to Each frame also includes a semantic represen- coarse-grained senses of the verb. The resource tation that uses basic predicates to show the rela- was originally based on Levin’s (1993) analysis of tionships between the thematic role arguments and English verbs but has since been expanded to in- to track any changes over the time course of the clude dozens of additional classes and hundreds of event. Thematic roles that appear in the ”syntax” additional verbs and verb senses. should always appear somewhere in the semantic VerbNet representations previously formed the representation. Overall, this linking in each frame basis for Parameterized Action Representation of the syntactic pattern to a semantic representa- (PAR) providing a conceptual representation of tion is a unique feature of VerbNet that emphasizes different types of actions (Badler et al., 1999). the close interplay of syntax and semantics. These actions involve changes of state, changes of location, and exertion of force and can be used 2.2 Revision of the Semantic Representations to animate human avatars in a virtual 3D environ- VerbNet’s old representations included an event ment (R. Bindiganavale and Palmer, 2000). They variable E as an argument to the predicates. are particularly well suited for motion and con- Representations of states were indicated with tact verb classes, providing an abstract, language- either a bare E, as for the own-100 class: independent representation (Kipper and Palmer, has possession(E, Pivot, Theme), or During(E), 2000). The more precise temporal sequencing de- as for the contiguous location-47.8 class (Italy scribed here is even more suitable as a foundation borders France): contact(During(E), Theme, Co- for natural language instructions and human-robot Theme). Most classes having to do with change, or human-avatar interactions. such as changes in location, changes in state and changes in possession, used a path rel predicate 2.1 VerbNet in combination with Start(E), During(E), and End/Result(E) to show the transition from one VerbNet has long been used in NLP for seman- location or state to another (1). tic role labeling and other inference-enabling tasks (Shi and Mihalcea, 2005; Giuglea and Moschitti, 2006; Loper et al., 2007; Bos, 2008). In addi- (1) The rabbit hopped across the lawn. tion, automatic disambiguation of a verb’s Verb- Theme V Trajectory Net class has been used as a stand-in for verb sense motion(during(E), Theme) disambiguation (Abend et al., 2008; Brown et al., path rel(start(E), Theme, ?Initial location1, 2014; Croce et al., 2012; Kawahara and Palmer, CH OF LOC, prep) 2014). path rel(during(E), Theme, Trajectory, VerbNet’s semantic representations use a CH OF LOC, prep) Davidsonian first-order-logic formulation that path rel(end(E), Theme, ?Destination, incorporates the thematic roles of the class. Each CH OF LOC, prep) frame in a class is labeled with a flat syntactic pattern (e.g., NP V NP). The ”syntax” that follows Efforts to use VerbNet’s semantic representa- shows how the thematic roles for that class appear tions (Zaenen et al., 2008; Narayan-Chen et al., in that pattern (e.g., Agent V Patient), much 2017), however, indicated a need for greater con- like the argument role constructions of Goldberg sistency and expressiveness. We have addressed (2006). A previous revision of the VerbNet se- consistency on several fronts. First, all neces- mantic representations made the correspondence sary participants are accounted for in the repre- of these patterns to constructions more explicit sentations, whether they are instantiated in the by using a common predicate (i.e., path rel) for syntax, incorporated in the verb itself (e.g., to all caused-motion construction frames(Hwang, drill), or simply logically necessary (e.g., all en- 2014). At the request of some users, we are sub- tities that change location begin in an initial loca- stituting more specific predicates for the general tion, whether it is commonly mentioned or not). path rel predicate, such as has location, has state 1A question mark in front of a thematic role indicates a and has possession, although the subevent pat- role that appears in the syntax in some frames for the class terns continue to show the commonality across but not in this particular frame. 155 Second, similar event types are represented with troduced a new event variable, e¨, to distinguish a similar format; for example, all states are rep- a process from other types of subevents, such as resented with E, never with During(E). Finally, states. For example, see the motion predicate in predicates are given formal definitions that apply (2). across classes. (2) The rabbit hopped across the lawn. In order to clarify what is happening at each Theme V Trajectory stage of an event, we turned to the Generative has location(e , Theme, ?Initial Location) Lexicon (Pustejovsky, 1995) for an explicit the- 1 motion(e¨ , Theme, Trajectory) ory of subevent structure. Classic GL character- 2 :has location(e , Theme, ?Initial location) izes the different Aktionsarten in terms of struc- 2 has location(e , Theme, ?Destination) tured subevents, with states represented with a 3 simple e, processes as a sequence of states char- acterizing values of some attribute, e1...en, and Although people infer that an entity