Using Gaussian Mixture Models to Detect Figurative Language in Context

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 297–300, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Using Gaussian Mixture Models to Detect Figurative Language in Context Linlin Li and Caroline Sporleder Saarland University, Postfach 15 11 50 66041 Saarbr¨ucken, Germany {linlin, csporled}@coli.uni-saarland.de Abstract We present a Gaussian Mixture model for de- tecting different types of figurative language in context. We show that this model performs well when the parameters are estimated in an unsupervised fashion using EM. Performance can be improved further by estimating the pa- rameters from a small annotated data set. 1 Introduction Figurative language employs words in a way that de- viates from their normal meaning. It includes id- iomatic usage, metaphor, metonymy or other types of creative language. Being able to detect figurative language is important for a number of NLP applica- tions, e.g., machine translation. Simply checking the input against an idiom dic- tionary does not solve the problem. While some expressions (e.g., trip the light fantastic) are al- ways used idiomatically, many expressions (e.g., spill the beans), can take on a literal meaning as well. Whether such expression is used idiomatically or not has to be inferred from the discourse context. Likewise, simple dictionary look-up would not work for truly creative, one-off usages; these can neither be found in a dictionary nor can they be detected by standard idiom extraction methods, which apply statistical measures to accumulated corpus evidence for an expression to assess its ’idiomaticity’. An ex- ample of a fairly creative usage can be found in (1), which is a variation of the idiom put a sock in. (1) Take the sock out of your mouth and create a brand-new relationship with your mom. We propose a method for detecting figurative lan- guage in context. Because we use context informa- tion rather than corpus statistics, our approach works also for truly creative usages. 2 Related Work Most studies on the detection of idioms and other types of figurative language focus on one of three aspects: type-based extraction (detect idioms on the type level), token-based classification (given a po- tentially idiomatic phrase in context, decide whether it is used idiomatically), token-based detection (de- tect figurative expressions in running text). Type-based extractions exploit the fact that idioms have many properties which differentiate them from other expressions, e.g., they often exhibit a degree of syntactic and lexical fixedness. These properties can be used to identify potential idioms, for instance, by employing measures of association strength between the elements of an expression (Lin, 1999). Type-based approaches are unsuitable for expres- sions which can be used both figuratively and lit- erally. These have to be disambiguated in context. Token-based classification aims to do this. A num- ber of token-based approaches have been proposed: supervised (Katz and Giesbrecht, 2006), weakly su- pervised (Birke and Sarkar, 2006), and unsupervised (Fazly et al., 2009; Sporleder and Li, 2009). Finally, token-based detection can be viewed as a two stage task …

Content truncated to 500 words