Integrating Verb-Particle Constructions into CCG Parsing

Integrating Verb-Particle Constructions into CCG Parsing James W. D. Constable and James R. Curran School of Information Technologies University of Sydney NSW 2006, Australia {jcon6353,james}@it.usyd.edu.au Abstract Despite their prevalence in the English language, multiword expressions like verb-particle constructions (VPCs) are of- ten poorly handled by NLP systems. This problem is partly due to inadequacies in existing corpora; the primary corpus for CCG-oriented work, CCGbank, does not account for VPCs at all, and is inconsis- tent in its handling of them. In this paper, we apply some corrective transformations to CCGbank, and then use it to retrain an augmented version of the Clark and Cur- ran CCG parser. Using our technique, we observe no significant change in F-score, while the resulting parse is semantically more sound. 1 Introduction Multiword expressions (MWEs), compound lex- emes made up of two or more words that to- gether form a complete semantic unit, are one of the problems facing natural language processing systems. Verb-particle constructions (VPCs) are a common type of MWE, comprising a verb and a particle, most often a preposition. The meaning of some VPCs can be logically attributed to the com- ponent parts (e.g., picked out), but many are id- iomatic and semantically opaque (e.g., make out). Previous research into VPCs has focussed much attention on their automatic extraction and classi- fication (Baldwin and Villavicencio, 2002; Villav- icencio, 2003). However, research into how they should be handled by parsers is noticeably lacking. Their unusual ability to manifest in both a ‘joined’ and ‘split’ configuration (‘gunned down the man’ versus ‘gunned the man down’) prevents parsers from treating them as a single unit, and demands a system that is able to maintain the semantic bond between the components, even when they are non- adjacent. To compound the problem, existing corpora are not consistent in their handling of these construc- tions. The Penn Treebank (Marcus et al., 1993, 1994) has an RP tag for particles, but sometimes labels them as adverbs. The CCGbank (Hocken- maier and Steedman, 2007) analysis of particles varies, but leans towards treating all particles as adverbial modifiers. This is in itself problematic, since it fails to take into account the fact that par- ticles are a core part of the construct, whereas ad- verbs are optional. This lack of quality corpora for VPC-related work limits the power of corpus- trained parsers. In this paper we draw on the Penn Treebank and PropBank (Kingsbury and Palmer, 2003) to repair CCGbank’s representation of VPCs, and demon- strate how our approach is able to satisfactorily ac- count for most VPC-related phenomena. Retrain- ing the Clark and Curran parser (Clark and Cur- ran, 2007) on our modified corpus, we observe a very slight decrease in parser F-score, although this is balanced by the fact that the parses now make structural sense. 2 Combinatory Categorial Grammar Combinatory Categorial Grammar (CCG, Steed- man (2000)) is a lexicalised grammar formalism based on combinatory logic. One of the features that makes CCG so appealing to NLP researchers is …

Content truncated to 500 words