Metaphor Detection via Linguistics Enhanced Siamese Network

Proceedings of the 29th International Conference on Computational Linguistics, pages 4149–4159 October 12–17, 2022. 4149 Metaphor Detection via Linguistics Enhanced Siamese Network Shenglong Zhang Ying Liu Tsinghua University, Beijing, China, 100084 zsl18@mails.tsinghua.edu.cn yingliu@mail.tsinghua.edu.cn Abstract In this paper we present MisNet, a novel model for word level metaphor detection. MisNet con- verts two linguistic rules, i.e., Metaphor Identi- fication Procedure (MIP) and Selectional Pref- erence Violation (SPV) into semantic matching tasks. MIP module computes the similarity be- tween the contextual meaning and the basic meaning of a target word. SPV module per- ceives the incongruity between target words and their contexts. To better represent basic meanings, MisNet utilizes dictionary resources. Empirical results indicate that MisNet achieves competitive performance on several datasets. 1 Introduction Metaphor is an omnipresent figurative language in daily communication. Conceptual Metaphor The- ory proposes that metaphor is a mapping mech- anism between the source domain and the target domain (Lakoff and Johnson, 2008). e.g. 1 The scream pierced the night. In e.g. 1, the literal meaning of the verb pierce is "some sharp object goes into or on through some- thing" . However, the contextual meaning is "break silence". Here, the source domain is a highly ab- stract action, while the target domain can present the corresponding meaning in a more concrete way. In general, there exists two types of metaphors, i.e., novel metaphors and conventional metaphors. e.g. 2 He attacked the government’s defence policy. Semantic Shift Theory shows that new lexical senses can derive from metaphors (Blank, 2013). Once a metaphorical usage of a word is accepted by most people, the metaphorical lexical sense is fixed. Thus a polysemant may have metaphorical marginal meanings (Bloomfield, 1994). In e.g. 2, the metaphorical target word attack means criticize, which is also a sense of it. It is a conventional metaphor for the metaphorical meaning has been fixed. While e.g. 1 is a novel metaphor, for the metaphorical usage is temporary. Linguistic rules instruct us how to identify metaphors. According to Metaphor Identification Procedure (MIP) (Crisp et al., 2007; Steen, 2010), a metaphor is identified if the contextual meaning of the target word contrasts with one of its more basic meaning. More basic meanings are: 1) More concrete; what they evoke is easier to imagine, see, hear, feel, smell, and taste; 2) related to bodily ac- tion; 3) more precise (as opposed to vague); 4) his- torically older(Group, 2007; Do Dinh et al., 2018). The basic meaning of pierce in e.g. 1 contrasts with its contextual meaning so it is a metaphor. Researchers are divided on how to represent ba- sic meanings. Gao et al. (2018) and Mao et al. (2019) used dynamic ELMo embeddings and static GloVe embeddings to encode contextual target meanings and basic target meanings respectively. Choi et al. (2021) proposed that a target used alone is literal. Su et al. (2021) and Wan et al. (2021) used the gloss (brief definition) of a target to rep- resent its literal meaning. All the methods are not linguistically …

Content truncated to 500 words