Unsupervised Chinese Verb Metaphor Recognition Based on Selectional Preferences

Unsupervised Chinese Verb Metaphor Recognition Based on Selectional Preferences* Yuxiang Jia, Shiwen Yu Institute of Computational Linguistics, Peking University, Beijing, China, 100871 {yxjia, yusw}@pku.edu.cn Abstract. Metaphors are pervasive in human language and developing methods to recognize and deal with metaphors is an indispensable task in Natural Language Processing (NLP). This paper proposes an unsupervised method to recognize metaphors from real texts. Firstly, source domain candidates are determined based on automatically acquired selectional preferences. And then metaphors are recognized with the source domain knowledge. Experiment results show that this unsupervised method outperforms the baseline by a great improvement. In addition, the source domain knowledge can also be used for metaphor comprehension. Keywords: Selectional Preference, Source Domain, Metaphor Recognition, Concept Concreteness 1. Introduction Conceptual Metaphor Theory (Lakoff and Johnson, 1980) considers metaphor as a mapping from the concrete source domain to the abstract target domain. Abstractions and enormously complex situations are routinely understood via metaphors. Metaphorical expressions are pervasive in human languages and must be treated for Natural Language Understanding (NLU) (Carbonell, 1982). As an important figure of speech, metaphor processing has interesting applications in many Natural Language Processing (NLP) tasks like machine translation, paraphrasing, information retrieval and question answering. Metaphor processing can be divided into three tasks, recognition, comprehension and generation, among which recognition is the basic step. Metaphor recognition is to decide whether a sentence contains metaphorical expressions (a word, phrase or the whole sentence). This paper focuses on verb metaphor, to decide whether a verb is in metaphorical usage or literal usage. In selectional preference violation view, a satisfied preference indicates a literal semantic relation, while a violated preference indicates a metaphorical one. Take the following two sentences as examples. (1) 农民 在 精心 培植 幼苗。 Nong2min2 zai4 jing1xin1 pei2zhi2 you4miao3 Farmer at carefully cultivate young plants “Farmers are cultivating young plants carefully.” * This research is funded by National Basic Research Program of China (No.2004CB318102). The authors are grateful to the two anonymous reviewers for their helpful comments and suggestions. Copyright 2008 by Yuxiang Jia and Shiwen Yu 207 22nd Pacific Asia Conference on Language, Information and Computation, pages 207–214 (2) 我们 要 大力 培植 人才。 Wo3men2 yao4 da4li4 pei2zhi2 ren2cai2 We should devote great effort train talents “We should devote great effort to train talents.” Sentence 1 is a literal usage while sentence 2 is a metaphorical one. The fact that literally 培植 ‘cultivate’ requires the object to denote some plants suggests that selectional preferences offer a cue to the presence of a metaphor. But the selectional preferences automatically induced by conventional computational models may not reflect semantics in the literal usage. On the other hand, concept concreteness or abstractness is an important indicator of literal usage, where concrete concepts usually indicate literal usage while abstract concepts correspond to non-literal usage. This paper makes use of concept concreteness based on automatically acquired selectional preferences for verb metaphor recognition. Though metaphorical usage could be considered as a different sense of the target word, but when performing inference, it …

Content truncated to 500 words