Idiom Token Classification using Sentential Distributed Semantics

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 194–204, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Idiom Token Classification using Sentential Distributed Semantics Giancarlo D. Salton and Robert J. Ross and John D. Kelleher Applied Intelligence Research Centre School of Computing Dublin Institute of Technology Ireland giancarlo.salton@mydit.ie {robert.ross,john.d.kelleher}@dit.ie Abstract Idiom token classification is the task of deciding for a set of potentially idiomatic phrases whether each occurrence of a phrase is a literal or idiomatic usage of the phrase. In this work we explore the use of Skip-Thought Vectors to create dis- tributed representations that encode fea- tures that are predictive with respect to id- iom token classification. We show that classifiers using these representations have competitive performance compared with the state of the art in idiom token classifi- cation. Importantly, however, our models use only the sentence containing the tar- get phrase as input and are thus less de- pendent on a potentially inaccurate or in- complete model of discourse context. We further demonstrate the feasibility of using these representations to train a competitive general idiom token classifier. 1 Introduction Idioms are a class of multiword expressions (MWEs) whose meaning cannot be derived from their individual constituents (Sporleder et al., 2010). Idioms often present idiosyncratic be- haviour such as violating selection restrictions or changing the default semantic roles of syntac- tic categories (Sporleder and Li, 2009). Conse- quently, they present many challenges for Natu- ral Language Processing (NLP) systems. For ex- ample, in Statistical Machine Translation (SMT) it has been shown that translations of sentences containing idioms receive lower scores than trans- lations of sentences that do not contain idioms (Salton et al., 2014). Idioms are pervasive across almost all lan- guages and text genres and as a result broad cov- erage NLP systems must explicitly handle idioms (Villavicencio et al., 2005). A complicating factor, however, is that many idiomatic expressions can be used both literally or figuratively. In general, idiomatic usages are more frequent, but for some expressions the literal meaning may be more com- mon (Li and Sporleder, 2010a). As a result, there are two fundamental tasks in NLP idiom process- ing: idiom type classification is the task of identi- fying expressions that have possible idiomatic in- terpretations and idiom token classification is the task of distinguishing between idiomatic and lit- eral usages of potentially idiomatic phrases (Fazly et al., 2009). In this paper we focus on this second task, idiom token classification. Previous work on idiom token classification, such as (Sporleder and Li, 2009) and (Peng et al., 2014), often frame the problem in terms of modelling the global lexical context. For exam- ple, these models try to capture the fact that the id- iomatic expression break the ice is likely to have a literal meaning in a context containing words such as cold, frozen or water and an idiomatic meaning in a context containing words such as meet or dis- cuss (Li and Sporleder, 2010a). Frequently these …

Content truncated to 500 words