Multi-task metaphor detection based on linguistic theory

Metaphorical expressions are widely present in natural language, posing significant challenges to a variety of natural language processing tasks such as machine translation. How to obtain richer contextual representations is an urgent problem to be solved. To address this issue, this paper proposes a model that combines syntax-aware local attention (SLA), a simple contrastive sentence embedding framework (SimCSE), and linguistic theories, called a combination of syntax-aware and semantic methods (CSS). Specifically, we apply linguistic theory in metaphor detection. Additionally, we simultaneously conduct metaphor identification and contrastive learning tasks. The SimCSE contrastive learning framework effectively captured more information, and the concurrent execution of these two tasks helped increase the sensitivity of the semantic space to metaphors. The integration of SLA with the pre-trained language model BERT enhanced the attention weights between grammatically relevant words, assisting the encoder in focusing more on grammar-related words. Overall, CSS prioritizes the sentence itself, avoiding the introduction of excessive additional information. Experimental results on the VU Amsterdam-verb (VUA), TroFi, and MOH-X metaphorical corpora show that our method is superior to state-of-the-art models.