Combining resources for MWE-token classification

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 100–104, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Combining resources for MWE-token classification Richard Fothergill and Timothy Baldwin Department of Computing and Information Systems The University of Melbourne VIC 3010 Australia r.fothergill@student.unimelb.edu.au, tb@ldwin.net Abstract We study the task of automatically disam- biguating word combinations such as jump the gun which are ambiguous between a lit- eral and MWE interpretation, focusing on the utility of type-level features from an MWE lexicon for the disambiguation task. To this end we combine gold-standard idiomatic- ity of tokens in the OpenMWE corpus with MWE-type-level information drawn from the recently-published JDMWE lexicon. We find that constituent modifiability in an MWE-type is more predictive of the idiomaticity of its tokens than other constituent characteristics such as semantic class or part of speech. 1 Introduction A multiword expression (MWE) is a phrase or sequence of words which exhibits idiosyncratic be- haviour (Sag et al., 2002; Baldwin and Kim, 2009). The nature of this idiosyncracy may be purely dis- tributional — such as hot and cold being more com- mon than cold and hot — but in this paper we study MWEs with idiosyncratic semantics. Specifically we are concerned with expressions such as jump the gun which are ambiguous between a literal interpre- tation of “to leap over a firearm”, and an idiomatic interpretation of “to act prematurely”. While MWEs are increasingly entering the main- stream of NLP, the accurate identification of MWEs remains elusive for current methods, particularly in the absence of MWE type-specialised training data. This paper builds on the work of Hashimoto et al. (2006) and Fothergill and Baldwin (2011) in ex- ploring whether type-level MWE properties sourced from an idiom dictionary can boost the accuracy of crosstype MWE-token classification. That is, we attempt to determine whether token occurrences of ambiguous expressions such as Kim jumped the gun on this issue are idiomatic or literal, based on: (a) annotated instances for MWEs other than jump the gun (e.g. we may only have token-level annotations for kick the bucket and throw in the towel), and (b) dictionary-based information on the syntactic prop- erties of the idiom in question. We find that constituent modifiability judgments extracted from the idiom dictionary are more predic- tive of the idiomaticity of tokens than other features of the idiom’s constituents such as part of speech or lexeme. However, violations of the dictionary’s modifiability rules have variable utility for machine learning classification, being suggestive of the literal class but not definitive. Finally, we present novel re- sults illuminating the effectiveness of contextual se- mantic vectors at MWE-token classification. 2 Related Work The OpenMWE corpus (Hashimoto and Kawahara, 2009) is a gold-standard corpus of over 100, 000 Japanese MWE-tokens covering 146 types. It is the largest resource we are aware of which has hand- annotated instances of MWEs which are ambiguous between a literal and idiomatic interpretation, and has been used by Hashimoto and Kawahara (2009) and Fothergill and Baldwin …

Content truncated to 500 words