Learning Outside the Box: Discourse-level Features Improve Metaphor Identification

Proceedings of NAACL-HLT 2019, pages 596–601 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 596 Learning Outside the Box: Discourse-level Features Improve Metaphor Identification Jesse Mu1,3, Helen Yannakoudakis2, Ekaterina Shutova4 1Computer Science Department, Stanford University, USA 2The ALTA Institute, 3Dept. of CS & Technology, University of Cambridge, UK 4ILLC, University of Amsterdam, The Netherlands muj@stanford.edu, hy260@cl.cam.ac.uk, e.shutova@uva.nl Abstract Most current approaches to metaphor identi- fication use restricted linguistic contexts, e.g. by considering only a verb’s arguments or the sentence containing a phrase. Inspired by pragmatic accounts of metaphor, we argue that broader discourse features are crucial for bet- ter metaphor identification. We train simple gradient boosting classifiers on representations of an utterance and its surrounding discourse learned with a variety of document embedding methods, obtaining near state-of-the-art results on the 2018 VU Amsterdam metaphor iden- tification task without the complex metaphor- specific features or deep neural architectures employed by other systems. A qualitative analysis further confirms the need for broader context in metaphor processing. 1 Introduction From bottled up anger to the world is your oys- ter, metaphor is a defining component of lan- guage, adding poetry and humor to communica- tion (Glucksberg and McGlone, 2001) and serv- ing as a tool for reasoning about relations between concepts (Lakoff and Johnson, 1980). Designing metaphor processing systems has thus seen signif- icant interest in the NLP community, with appli- cations from information retrieval (Korkontzelos et al., 2013) to machine translation (Saygin, 2001). An important first step in any metaphor process- ing pipeline is metaphor identification. To date, most approaches to its identification operate in re- stricted contexts, for instance, by only considering isolated verb–argument pairs (e.g. deflate econ- omy) (Rei et al., 2017) or the sentence containing an utterance (Gao et al., 2018). However, wider context is crucial for understanding metaphor: for instance, the phrase drowning students can be in- terpreted as literal (in the context of water) or metaphorical (in the context of homework). Of- “You can’t steal their ideas.” “No, idiot—not so I can steal them.” Britain still can’t decide when to play the mandarinate game of of silence [...] interests and concern of the Chinese government. Table 1: Metaphorical examples from the VUA dataset in context. Verb is bolded, arguments underlined. Im- mediate sentence in black, with further context in gray. ten the context required extends beyond the im- mediate sentence; in Table 1, coreferences (them) must be resolved to understand the arguments of a verb, and a game is metaphorical in a political context. Indeed, a rich linguistic tradition (Grice, 1975; Searle, 1979; Sperber and Wilson, 1986) ex- plains metaphor as arising from violations of ex- pectations in a conversational context. Following these theories, in this paper we argue that metaphor processing models should expand beyond restricted contexts to use representations of wider discourse. We support this claim with two contributions: (1) we develop metaphor identifica- tion models which take as input an utterance, its immediate lexico–syntactic context, and broader discourse representations, …

Content truncated to 500 words