Metaphor Detection with Effective Context Denoising

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1404–1409 May 2-6, 2023 ©2023 Association for Computational Linguistics Metaphor Detection with Effective Context Denoising Shun Wang1∗, Yucheng Li2∗, Chenghua Lin1†, Loïc Barrault1,3 , Frank Guerin2 1 Department of Computer Science, University of Sheffield, UK 2 Department of Computer Science, University of Surrey, UK 3 Meta AI {swang209, c.lin}@sheffield.ac.uk {yucheng.li, f.guerin}@surrey.ac.uk loicbarrault@meta.com Abstract We propose a novel RoBERTa-based model, RoPPT, which introduces a target-oriented parse tree structure in metaphor detection. Compared to existing models, RoPPT fo- cuses on semantically relevant information and achieves the state-of-the-art on several main metaphor datasets. We also compare our ap- proach against several popular denoising and pruning methods, demonstrating the effective- ness of our approach in context denoising. Our code and dataset can be found at https: //github.com/MajiBear000/RoPPT. 1 Introduction Metaphor is a pervasive linguistic device, which attracts attention from both the fields of psycholin- guistics and computational linguistics due to the key role it plays in the cognitive and communica- tive functions of language (Wilks, 1978; Lakoff and Johnson, 1980; Lakoff, 1993). Linguistically, metaphor is defined as a figurative expression that uses one or several words to represent another con- cept given the context, rather than taking the lit- eral meaning of the expression (Fass, 1991). For instance, in the sentence “This project is such a headache!”, the contextual meaning of headache is “a thing or person that causes worry or trouble”, different from its literal meaning, “a continuous pain in the head”.1 Metaphor detection is challenging, as it requires understanding the nuanced relationships between abstract concepts embodied by the metaphoric ex- pression and its surrounding context. Recent stud- ies on this direction show its potential in benefit- ing a wide range of NLP applications, including sentiment analysis (Li et al., 2022a), metaphor gen- eration (Li et al., 2022b,c) and mental healthcare (Abd Yusof et al., 2017; Gutiérrez et al., 2017). ∗ The two authors contributed equally to this work. † Corresponding author 1https://www.oxfordlearnersdictionaries.com When modelling relevant context for metaphor detection, various strategies have been proposed. These range from using highly restricted forms of linguistic context such as subject-verb and verb- direct object word pairs (Gutiérrez et al., 2016), to a wider context accounting for a fixed window sur- rounding the target word (Do Dinh and Gurevych, 2016; Mao et al., 2018), and modelling the full sen- tential context (Gao et al., 2018; Choi et al., 2021). While it has been argued that modelling a wider context is beneficial (Cheng et al., 2021), it has also been noted that a wider context is likely to intro- duce noise into the representations, and hence hin- der model’s performance in metaphor detection (Le et al., 2020). Some recent efforts (Le et al., 2020; Song et al., 2021a) attempt to improve context modelling by explicitly leveraging the syntactic structure (e.g., dependency parse tree) of a sentence in order to cap- ture important context words, where the parse trees are typically encoded …

Content truncated to 500 words