Metaphor Detection with Topic Transition, Emotion and Cognition in Context

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 216–225, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Metaphor Detection with Topic Transition, Emotion and Cognition in Context Hyeju Jang, Yohan Jo, Qinlan Shen, Michael Miller, Seungwhan Moon, Carolyn P. Ros´e Language Technologies Institute Carnegie Mellon University 5000 Forbes Ave, Pittsburgh, PA 15213, USA {hyejuj,yohanj,qinlans,millerm,seungwhm,cprose}@cs.cmu.edu Abstract Metaphor is a common linguistic tool in communication, making its detection in discourse a crucial task for natural lan- guage understanding. One popular ap- proach to this challenge is to capture se- mantic incohesion between a metaphor and the dominant topic of the surrounding text. While these methods are effective, they tend to overclassify target words as metaphorical when they deviate in mean- ing from its context. We present a new approach that (1) distinguishes literal and non-literal use of target words by exam- ining sentence-level topic transitions and (2) captures the motivation of speakers to express emotions and abstract concepts metaphorically. Experiments on an on- line breast cancer discussion forum dataset demonstrate a significant improvement in metaphor detection over the state-of-the- art. These experimental results also re- veal a tendency toward metaphor usage in personal topics and certain emotional con- texts. 1 Introduction Figurative language is commonly used in human communication ranging from literature to every- day speech. One of the most common forms of non-literal language is metaphor, in which two dissimilar concepts are compared. In the ut- terance, “Time is money” (Lakoff and Johnson, 1980), for example, the concept of “time” is com- pared to “money” to emphasize that time is valu- able. Bringing in information from another do- main allows more effective ways of expressing thoughts, feelings, and ideas than only using lit- eral language. Previous approaches to modeling metaphor have used either the semantic and syntactic in- formation in just the sentence that contains a metaphor (Turney et al., 2011; Tsvetkov et al., 2014), or the context beyond a single sentence (Broadwell et al., 2013; Strzalkowski et al., 2013; Schulder and Hovy, 2014; Klebanov et al., 2015; Jang et al., 2015) to detect topical discrepancy between a candidate metaphor and the dominant theme (See Section 2 for more detailed literature review). Although previous approaches were effective at capturing some aspects of the governing context of a metaphor, the space of how to best use the contextual information is still wide open. Previous context-based models tend to overclassify literal words as metaphorical if they find semantic con- trast with the governing context. These cases man- ifested in the work by Schulder and Hovy (2014) and Jang et al. (2015) as high recall but low preci- sion for metaphorical instances. We present a new approach that uses lexical and topical context to resolve the problem of low pre- cision on metaphor detection. To better capture the relevant context surrounding a metaphor, we approach the problem in two directions. First, we hypothesize that topic transition patterns be- tween sentences containing metaphors and their contexts are different …

Content truncated to 500 words