A Fine-tuned Language Model for Enhanced Metaphor Analysis
Prof. Dahui Dong
Guangdong Province, China
Abstract: Metaphors are a fundamental aspect of human cognition and communication, involving the understanding of one concept (the target domain) in terms of another (the source domain). Lakoff and Johnson’s Conceptual Metaphor Theory (CMT) provides a foundational framework for understanding this cognitive process [1]. While computational metaphor processing has seen significant advancements, current metaphor detection tasks largely focus on identifying metaphorical language rather than exploring the underlying conceptual mappings or providing justifications for metaphor usage. Moreover, recent efforts have highlighted a critical gap in the availability of large-scale, structured datasets that detail conceptual metaphors and their components, which hampers the development of more nuanced metaphor analysis systems [2]. Existing manually annotated datasets, while valuable, are often relatively small and lack sufficient structure [3, 4], limiting the scope of metaphor detection research.
With the rise of Transformer-based pre-trained language models (PLMs) such as BERT [5] and RoBERTa [6], recent advances in natural language processing (NLP) have enabled improved metaphor detection over earlier approaches that relied on feature engineering or recurrent neural networks (RNNs). More recently, the potential of large language models (LLMs) such as GPT-3 [7] and ChatGPT [8] for metaphor detection in zero-shot or few-shot settings has been explored. For instance, Tian et al. [9] introduced a Theory Guided Scaffolding Instruction (TSI) framework, which aids LLMs in reasoning through metaphor detection using metaphor theories like CMT, MIP, and SPV, with results demonstrating significant improvements over earlier LLM-based approaches. Current efforts focus on transforming metaphor detection tasks into various NLP tasks, including sequence labeling, reading comprehension, and relation classification, while incorporating linguistic theories (e.g., MIP, SPV, CMT) and external knowledge resources (e.g., WordNet, FrameNet, word glosses) to improve model performance in zero-shot or few-shot learning contexts [10–17].
This study aims to address the lack of structured resources for conceptual metaphors and explore the potential of fine-tuning LLMs to enhance metaphor analysis. The central research question is: Can a systematically constructed dataset containing key types of conceptual metaphors, when used to fine-tune a pre-trained LLM, significantly improve performance on conceptual metaphor analysis tasks? This research is crucial for advancing computational metaphor processing and enabling more sophisticated analyses of texts across domains such as literary studies, discourse analysis, and philosophical or religious texts [18].
Brief Biography of the Speaker: Dahui Dong is an Associate Professor (Distinguished) at the School of Foreign Languages, Guangzhou Institute of Science and Technology. With a Ph.D. in Linguistics, his research focuses on corpus and computational linguistic studies, machine and AI translation studies, and the bibliometrics studies of scientific literature.
Dr. Dong has published widely in international journals and books published by Routledge and Springer. His recent work, “A Scientometrics Research Perspective in Applied Linguistics” (2023), reflects his ongoing contributions to epistemology and research methodology in humanity and social sciences.