A Dual-Perspective Metaphor Detection Framework Using Large Language Models

TL;DR

The paper tackles metaphor detection with a transparency-focused approach, addressing the opaque decision processes of traditional models. It introduces DMD, a dual-perspective framework that guides LLMs through implicit theory-driven guidance (from MIP and SPV representations) and explicit theory-driven guidance (dictionary-backed, theory-informed prompts), complemented by a self-judgment module to reconcile outputs. By constructing a MelBERT-based datastore and performing k-nearest-neighbor retrieval, DMD enables in-context learning that leverages theory-aligned similarities; simultaneously, it augments this with retrieval of lemma definitions and multi-step reasoning. The self-judgment stage then fuses the two perspectives into a final, substantiated answer. Experiments on MOH-X and TroFi with GPT-3.5 turbo and GPT-4o show state-of-the-art performance and robust ablations, highlighting the value of dual guidance for metaphor understanding and the reliability of LLM-driven reasoning in NLP tasks.

Abstract

Metaphor detection, a critical task in natural language processing, involves identifying whether a particular word in a sentence is used metaphorically. Traditional approaches often rely on supervised learning models that implicitly encode semantic relationships based on metaphor theories. However, these methods often suffer from a lack of transparency in their decision-making processes, which undermines the reliability of their predictions. Recent research indicates that LLMs (large language models) exhibit significant potential in metaphor detection. Nevertheless, their reasoning capabilities are constrained by predefined knowledge graphs. To overcome these limitations, we propose DMD, a novel dual-perspective framework that harnesses both implicit and explicit applications of metaphor theories to guide LLMs in metaphor detection and adopts a self-judgment mechanism to validate the responses from the aforementioned forms of guidance. In comparison to previous methods, our framework offers more transparent reasoning processes and delivers more reliable predictions. Experimental results prove the effectiveness of DMD, demonstrating state-of-the-art performance across widely-used datasets.

Figures (1)

• Figure 1: Overview of our DMD framework. It consists of three parts: (a) Implicit Theory-Driven Guidance: For each sample, theoretical representations are computed using a pre-trained MelBERT model to find the $k$ nearest neighbors for in-context learning; (b) Explicit Theory-Driven Guidance: For each sample, the multi-step thoughts are generated by the LLM, and the information related to target word's lemma is retrieved from the Oxford Dictionary, both of which are used to guide the LLM explicitly; (c) Self-Judgment: Further evaluate the responses produced by Implicitly Theory-Driven Guidance and Explicitly Theory-Driven Guidance, and the final answer is extracted from the judgment.