Sensorimotor Enhanced Neural Network for Metaphor Detection

Sensorimotor Enhanced Neural Network for Metaphor Detection Mingyu Wan1,2, Baixi Xing3, Qi Su1, Pengyuan Liu3 and Chu-Ren Huang2 1School of Foreign Languages, Peking University 2Department of Chinese Bilingual Studies, The Hong Kong Polytechnic University 3School of Information Science Language, Beijing Language and Culture University {wanmy,Liupengyuan,sukia}@pku.edu.cn {xingbaixi}@gmail.com,{churen.huang}@polyu.edu.hk Abstract Detecting metaphors is challenging due to the subtle ontological differences between metaphorical and non-metaphorical expres- sions. Neural networks have been widely adopted in metaphor detection and become the main stream technology. However, linguistic insights have been less utilized. This work proposes a linguistically enhanced model for metaphor detection extending one published work (WAN et al., 2020) by incorporating the modality norms into attention-based Bi- LSTM. Results show that the current model outperforms most recent works by 0.5%- 11% F1, indicating the effectiveness of us- ing modality norms for metaphor detection. This work provides a new perspective to de- tect token-level metaphoricity by leveraging the modality mismatch between words. 1 Introduction Metaphors are prevalent in our everyday language even without our consciousness of its presence as we speak and write. It induces the unknown using the known, explains the complex using the simple, and helps us to emphasize the relevant aspects of mean- ing resulting in effective communication. In general, metaphor involves certain concept transfer from one domain (Source) to another (Tar- get), as in ‘sweet voice’ (using taste to describe sound). Lakoff (1980) describes metaphor as a cog- nitive mechanism (a property of language) reflected by our conceptual system for structuring our under- standing of the world. It is a fundamental way to relate our physical and familiar social experiences to a multitude of other subjects and contexts (Lakoff and Johnson, 2008). As a popular linguistic device, metaphors encode versatile ontological information, which usually in- volve e.g. domain transfer (Ahrens et al., 2003; Ahrens and Jiang, 2020), sentiment reverse (Steen et al., 2010) or modality shift (Winter, 2019) etc. Therefore, detecting the metaphors in texts is essen- tial for capturing the authentic meaning of the texts, which can benefit many natural language process- ing applications, such as machine translation, dia- logue systems and sentiment analysis (Tsvetkov et al., 2014). To better understand the intrinsic properties of metaphors and to provide an in-depth analysis to this phenomenon, we propose a linguistically- enriched deep learning model extending one pub- lished work (WAN et al., 2020) at ACL Figlang 2020 workshop by incorporating the modality norms into attention-based BiLSTM. As a continuation of their work, we conduct the current research to further tes- tify the effectiveness of leveraging conceptual norms for metaphor detection. For standard reference, we adopt the dataset of the first and second shared tasks of metaphor detection on verbs of the VUA cor- pus (Klebanov et al., 2018)1. Details about the ex- periment are given in Sections 3-5. 2 Related Work Research on metaphors have been mainly explored in the context of political communication, men- tal health, teaching, discourse analysis, assessment 1http://www.vismet.org/metcor/ documentation/home.html of English proficiency, among others (Ahrens and Jiang, 2020; Thibodeau …

Content truncated to 500 words