Enhancing metaphor identification in large language models: a step towards autonomous language understanding

Abstract

Metaphors are inherent in daily lives, influencing thought, perception, and interaction. It's ability to clarify, obscure, or identify likeness between different items, creates new meaning and develops richness in our communications. However, identifying and interpreting metaphors remains challenging for LLMs, which often struggle with analogical reasoning, sentence semantics and novel words. This study introduces two transformer-based architectures, DisBERT and WordBERT, designed to enhance metaphor identification capabilities of LLMs. DisBERT incorporates WSD mechanism based on Black’s interaction theory of metaphor, embettering semantic comparison. Conversely, WordBERT employs explicit definitions from WordNet to generate definition-based semantic embeddings. Empirical evaluations on four standard benchmarks—VUA18, VUA20, MOH-X, and TroFi—demonstrate that DisBERT effectively captures additional contextualized semantic features, consistently achieving results comparable to state-of-the-art models. Meanwhile, WordBERT exhibits notably high recall, highlighting the potential of explicit definition-based embeddings for metaphor identification. This project underscores the importance of integrating contextual and lexical semantics in metaphor detection, advancing toward human-like autonomous language understanding.