Comprehensive and Consistent PropBank Light Verb Annotation

Comprehensive and Consistent PropBank Light Verb Annotation Claire Bonial,1 Martha Palmer2 1U.S. Army Research Laboratory, 2University of Colorado, Boulder 12800 Powder Mill Rd., Adelphi, MD, USA E-mail: Claire.N.Bonial.civ@mail.mil, Martha.Palmer@colorado.edu Abstract Recent efforts have focused on expanding the annotation coverage of PropBank from verb relations to adjective and noun relations, as well as light verb constructions (e.g., make an offer, take a bath). While each new relation type has presented unique annotation challenges, ensuring consistent and comprehensive annotation of light verb constructions has proved particularly challenging, given that light verb constructions are semi-productive, difficult to define, and there are often borderline cases. This research describes the iterative process of developing PropBank annotation guidelines for light verb constructions, the current guidelines, and a comparison to related resources. Keywords: light verb constructions, semantic role labeling, NLP resources 1. Introduction The goal of PropBank (Palmer et al., 2005) is to supply consistent, general-purpose labeling of semantic roles across different syntactic realizations. With over two million words from diverse genres, the benchmark annotated corpus supports the training of automatic semantic role labelers, which in turn support other Natural Language Processing (NLP) areas, such as machine translation. PropBank annotation consists of two tasks: sense and role annotation. Previously, PropBank annotation was focused on verbs (a separate, but related project, NomBank (Meyers et al., 2004), focused on noun relations), but recent efforts have shifted to expanding the annotation coverage of PropBank from verb relations to adjective and noun relations, as well as light verb constructions (LVCs; e.g., have a nap, do an investigation, give a kiss) (Bonial et al., 2014). This shift has allowed PropBank to capture more comprehensively all mentions of events of states, which can be realized alternately as verbs (fear), nouns, (fear), adjectives (afraid) or LVCs (have fear). Although each new relation type has presented unique annotation challenges, ensuring consistent and comprehensive annotation of LVCs has proved particularly challenging because LVCs are semi-productive, difficult to define objectively, and there are often borderline cases. This research describes the iterative process of developing PropBank annotation guidelines for LVCs (for a description of the early development stages of LVC annotation, see Hwang et al., 2010), the current guidelines,1 and a comparison to the annotation practices of two related resources: the TectoGrammatical Treebank (Cinková et al., 2004) and Tu and Roth’s (2011) LVC dataset that was developed specifically for enabling automatic LVC detection. 1 http://propbank.github.io/ 2 https://github.com/propbank/propbank-frames 2. PropBank Background PropBank annotation consists of two tasks: sense annotation and role annotation. The PropBank lexicon provides a listing of the coarse-grained senses of a verb, noun, or adjective relation, and the set of roles associated with each sense (thus, called a “roleset”). The roles are listed as argument numbers (Arg0 – Arg6) and correspond to verb-specific roles. For example: Offer-01 (transaction, proposal): Arg0: entity offering Arg1: commodity Arg2: price Arg3: benefactive, or entity offered to Previous versions of PropBank maintained separate rolesets for the verb offer, and the related nouns offer/offering, but these have been combined (Bonial et al., …

Content truncated to 500 words