Integrating Generative Lexicon Event Structures into VerbNet

Integrating Generative Lexicon Event Structures into VerbNet Susan Windisch Brown1, James Pustejovsky2, Annie Zaenen3, Martha Palmer1 1University of Colorado, 2Brandeis University, 3Stanford University {susan.brown, martha.palmer}@colorado.edu, jamesp@cs.brandeis.edu, azaenen@stanford.edu Abstract This paper focuses on specific changes to the semantic representations associated with classes of verbs in the English lexical resource VerbNet (Schuler, 2005). The new form has been restricted to first-order representations to simplify use by and integration with planners. More significantly, the modifications incorporate the Generative Lexicon’s event structure, with temporal ordering of subevents associated with explicit predications over the verb’s arguments. These changes allow for greater flexibility in representing complex events, for a more consistent treatment of the oppositions inherent in change-of-state classes, and for a more nuanced portrayal of the Agent’s role. Keywords: semantics, event representation, lexicon 1. Introduction Natural language processing has been moving from shallow semantic parsing to deeper semantic analysis of text in or- der to better support knowledge representation and reason- ing systems. Understanding sentences requires more than identifying events and participants and giving them the- matic role labels. In particular, it is essential to recognize any temporal sequencing within the event and any changes in state that might have occurred (Pustejovsky, 2005; Mani and Pustejovsky, 2012). The language resource VerbNet (Kipper et al., 2006) is a promising source of such information. It is a hierarchical, domain-independent verb lexicon that groups verbs into classes based on similarities in their syntactic and semantic behavior (Schuler, 2005). Each class in VerbNet defines a set of members, thematic roles for the predicate-argument structure of these members, selectional restrictions on the arguments, and frames consisting of a syntactic description and a corresponding semantic representation. It has long been used for semantic role labeling and other inference- enabling tasks (Shi and Mihalcea, 2005; Giuglea and Mos- chitti, 2006; Loper et al., 2007). Automatic disambiguation of a verb’s VerbNet class has also improved (Abend et al., 2008; Brown et al., 2014; Kawahara and Palmer, 2014). Efforts to use its semantic representations (Zaenen et al., 2008; Narayan-Chen et al., 2017) have revealed a need to revise them for consistency and greater expressiveness, in particular, a clearer representation of subevents. Recent work in Generative Lexicon (GL) has focused on further articulating the semantics of subevent structure in language (Pustejovsky and Moszkowicz, 2011; Puste- jovsky, 2013). Hence a reasonable undertaking is to revise VerbNet to take advantage of GL’s progress in representing subevent structure while preserving VerbNet’s strengths in linking predicate argument structure, thematic roles and se- mantic representations. The remainder of this paper will describe the changes being made to VerbNet’s semantic representations and the reasons behind those changes. Section 2 briefly describes the role semantic representations play in VerbNet and breaks down the structure of the ”old” semantic representations in ver- sion 3.3. Section 3 goes into more detail on the drawbacks of these representations and the requests from users for new functionality. An overview of GL event structure is given in section 4, highlighting how it can fulfill the needs identified …

Content truncated to 500 words