Metaphor Identification as Interpretation

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 1: Proceedings of the Main Conference and the Shared Task, pages 276–285, Atlanta, Georgia, June 13-14, 2013. c⃝2013 Association for Computational Linguistics Metaphor Identification as Interpretation Ekaterina Shutova International Computer Science Institute and Institute for Cognitive and Brain Sciences University of California, Berkeley katia@berkeley.edu Abstract Automatic metaphor identification and inter- pretation in text have been traditionally con- sidered as two separate tasks in natural lan- guage processing (NLP) and addressed in- dividually within computational frameworks. However, cognitive evidence suggests that hu- mans are likely to perform these two tasks si- multaneously, as part of a holistic metaphor comprehension process. We present a novel method that performs metaphor identification through its interpretation, being the first one in NLP to combine the two tasks in one step. It outperforms the previous approaches to metaphor identification both in terms of ac- curacy and coverage, as well as providing an interpretation for each identified expression. 1 Introduction Metaphor undoubtedly gives our expression more vividness, distinction and artistry, however, it is also an important linguistic tool that has long become part of our every-day language. Metaphors arise when one concept or domain is viewed in terms of the properties of another (Lakoff and Johnson, 1980). Consider the examples in (1) and (2). (1) My car drinks gasoline. (Wilks, 1978) (2) This policy is strangling business. The car in (1) and business in (2) are viewed as living beings and thus they can drink or be stran- gled respectively. The mapping between the car (the target concept) and living being (the source concept) is systematic and results in a number of metaphorical expressions (e.g. “This oil gives your car a second life”, “this car has is very temperamen- tal” etc.) Lakoff and Johnson call such generalisa- tions a source–target domain mapping, or concep- tual metaphor. The ubiquity of metaphor in language has been established in a number of corpus studies (Cameron, 2003; Martin, 2006; Steen et al., 2010; Shutova and Teufel, 2010) and the role it plays in human reasoning has been confirmed in psychological ex- periments (Thibodeau and Boroditsky, 2011). This makes its automatic processing an important prob- lem for NLP and its numerous applications (such as machine translation, information extraction, opin- ion mining and many others). For example, the use of the metaphorical verb strangle in (2) reflects the speaker’s negative opinion regarding the gov- ernment’s tight business regulations, which would be an important fact for an opinion mining system to discover (Narayanan, 1999). Other experiments (Agerri, 2008) have investigated and confirmed the role of metaphor interpretation for textual entailment resolution (RTE). The problem of metaphor modeling is rapidly gaining interest within NLP, with a growing number of approaches exploiting statistical techniques (Ma- son, 2004; Gedigian et al., 2006; Shutova, 2010; Shutova et al., 2010; Turney et al., 2011; Shutova et al., 2012a). Compared to more traditional ap- proaches based on hand-coded knowledge (Fass, 1991; Martin, 1990; Narayanan, 1997; Narayanan, 1999; Feldman and Narayanan, …

Content truncated to 500 words