Unsupervised Metaphor Identification Using Hierarchical Graph Factorization Clustering

Proceedings of NAACL-HLT 2013, pages 978–988, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Unsupervised Metaphor Identification Using Hierarchical Graph Factorization Clustering Ekaterina Shutova International Computer Science Institute and Institute for Cognitive and Brain Sciences University of California, Berkeley katia@icsi.berkeley.edu Lin Sun Computer Laboratory University of Cambridge lin.sun@cl.cam.ac.uk Abstract We present a novel approach to automatic metaphor identification, that discovers both metaphorical associations and metaphorical expressions in unrestricted text. Our sys- tem first performs hierarchical graph factor- ization clustering (HGFC) of nouns and then searches the resulting graph for metaphorical connections between concepts. It then makes use of the salient features of the metaphori- cally connected clusters to identify the actual metaphorical expressions. In contrast to pre- vious work, our method is fully unsupervised. Despite this fact, it operates with an encour- aging precision (0.69) and recall (0.61). Our approach is also the first one in NLP to exploit the cognitive findings on the differences in or- ganisation of abstract and concrete concepts in the human brain. 1 Introduction Metaphor has traditionally been viewed as a form of linguistic creativity, that gives our expression more vividness, distinction and artistism. While this is true on the surface, the mechanisms of metaphor have a much deeper origin in our reasoning. To- day metaphor is widely understood as a cognitive phenomenon operating at the level of mental pro- cesses, whereby one concept or domain is system- atically viewed in terms of the properties of another (Lakoff and Johnson, 1980). Consider the examples (1) “He shot down all of my arguments” and (2) “He attacked every weak point in my argument”. They demonstrate a metaphorical mapping of the concept of argument to that of war. The argument, which is the target concept, is viewed in terms of a battle (or a war), the source concept. The existence of such a link allows us to systematically describe arguments using the war terminology, thus leading to a num- ber of metaphorical expressions. Lakoff and John- son call such generalisations a source–target domain mapping, or conceptual 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 metaphor an important research area for com- putational and cognitive linguistics, and its auto- matic processing indispensable for any semantics- oriented NLP application. The problem of metaphor modeling is gaining interest within NLP, with a growing number of approaches exploiting statisti- cal techniques (Mason, 2004; Gedigian et al., 2006; Shutova, 2010; Shutova et al., 2010; Turney et al., 2011; Shutova et al., 2012). Compared to more traditional approaches based on hand-coded knowl- edge (Fass, 1991; Martin, 1990; Narayanan, 1997; Narayanan, 1999; Feldman and Narayanan, 2004; Barnden and Lee, 2002; Agerri et al., 2007), these more recent methods tend to have a wider cover- age, as well as be more efficient, …

Content truncated to 500 words