FrameBERT: Conceptual Metaphor Detection with Frame Embedding Learning

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1558–1563 May 2-6, 2023 ©2023 Association for Computational Linguistics FrameBERT: Conceptual Metaphor Detection with Frame Embedding Learning Yucheng Li1∗, Shun Wang2∗, Chenghua Lin2†, Frank Guerin1, Loïc Barrault2,3 1 Department of Computer Science, University of Surrey, UK 2 Department of Computer Science, University of Sheffield, UK 3 Meta AI {yucheng.li, f.guerin}@surrey.ac.uk {swang209, c.lin}@sheffield.ac.uk loicbarrault@meta.com Abstract In this paper, we propose FrameBERT, a RoBERTa-based model that can explicitly learn and incorporate FrameNet Embeddings for concept-level metaphor detection. FrameBERT not only achieves better or comparable perfor- mance to the state-of-the-art, but also is more explainable and interpretable compared to ex- isting models, attributing to its ability of ac- counting for external knowledge of FrameNet. 1 Introduction Metaphor is a pervasive linguistic device, which attracts attention from both 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; a problem”, different from its literal meaning, “a continuous pain in the head”. Metaphor Detection presents a significant chal- lenge as it necessitates comprehending the intricate associations between the abstract concepts embod- ied by the metaphoric expression and the related context. Recent studies in this field have demon- strated its potential to positively impact various Natural Language Processing (NLP) applications, including sentiment analysis (Cambria et al., 2017; Li et al., 2022a), metaphor generation (Tang et al., 2022; Li et al., 2022b,c), and mental health care (Abd Yusof et al., 2017; Gutiérrez et al., 2017). Different strategies have been proposed for model- ing relevant context, including employing limited ∗ The two authors contributed equally to this work. † Corresponding author linguistic context such as subject-verb and verb- direct object word pairs (Gutiérrez et al., 2016), in- corporating a wider context encompassing a fixed window surrounding the target word (Do Dinh and Gurevych, 2016; Mao et al., 2018), and considering the complete sentential context (Gao et al., 2018; Choi et al., 2021). Some recent efforts attempt to improve context modelling by explicitly leveraging the syntactic structure (e.g., dependency tree) of a sentence in order to capture important context words, where the parse trees are typically encoded with graph convolutional neural networks (Le et al., 2020; Song et al., 2021). Despite the progress, we also observe the inad- equacy of existing models in semantic modelling, which plays a crucial role in metaphor detection. For instance, it has been noted that BERT’s ten- dency to aggregate shallow semantics instead of precise meaning, as its objective, may limit the con- text modelling ability (Xu …

Content truncated to 500 words