Learning to Identify Metaphors from a Corpus of Proverbs

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2060–2065, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Learning to Identify Metaphors from a Corpus of Proverbs G¨ozde ¨Ozbal† and Carlo Strapparava† and Serra Sinem Tekiro˘glu† and Daniele Pighin‡ †FBK-Irst, Trento, Italy, ‡Google - Z¨urich, Switzerland gozbalde@gmail.com, {strappa, tekiroglu}@fbk.eu, biondo@google.com Abstract In this paper, we experiment with a re- source consisting of metaphorically annotated proverbs on the task of word-level metaphor recognition. We observe that existing feature sets do not perform well on this data. We design a novel set of features to better cap- ture the peculiar nature of proverbs and we demonstrate that these new features are signif- icantly more effective on the metaphorically dense proverb data. 1 Introduction Recent years have seen a growing attention to- wards attempts to understand figurative language in text (Steen et al., 2010, Shutova and Teufel, 2010, Turney et al., 2011, Neuman et al., 2013, Klebanov et al., 2015). Recently, ¨Ozbal et al. (2016) published a resource consisting of 1,054 proverbs annotated with metaphors at the word and sentence level, mak- ing it possible for the first time to test existing mod- els for metaphor detection on such data. More than in other genres, such as news, fiction and essays, in proverbs metaphors can resolve a significant amount of the figurative meaning (Faycel, 2012). The rich- ness of proverbs in terms of metaphors is very fas- cinating from a linguistic and cultural point of view. Due to this richness, proverbs constitute a challeng- ing benchmark for existing computational models of metaphoricity. In this paper, we devise novel feature sets es- pecially tailored to cope with the peculiarities of proverbs, which are generally short and figuratively rich. To the best of our knowledge, this is the first attempt to design a word-level metaphor rec- ognizer specifically tailored to such metaphorically rich data. Even though some of the resources that we use (e.g., imageability and concreteness) have been used for this task before, we propose new ways of encoding this information, especially with respect to the density of the feature space and the way that the context of each word is modeled. On the proverb data, the novel features result in compact models that significantly outperform existing features de- signed for word-level metaphor detection in other genres (Klebanov et al., 2014), such as news and es- says. By also testing the new features on these other genres, we show that their generalization power is not limited to proverbs. 2 Background In this section we provide a brief overview of the efforts of the NLP community to build metaphor datasets and utilize them to develop computational techniques for metaphor processing. Steen et al. (2010) construct the Amsterdam Metaphor Cor- pus (VUAMC) by annotating a subset of BNC Baby1. Linguistic metaphors in VUAMC are an- notated by utilizing the Metaphor Annotation Pro- cedure (MIP) proposed by Group (2007). VUAMC contains 200,000 words in sentences sampled from various genres (news, fiction, …

Content truncated to 500 words