Can ChatGPT’s performance be improved on metaphor detection tasks? Bootstrapping and Combining Tacit Knowledge

Anonymous

Abstract: Metaphor detection, as a key task in the field of natural language processing, has received sustained academic attention in recent years. Current research focuses on the development of supervised metaphor detection systems, which usually require large-scale, high-quality labeled data support. With the rapid development of large-scale generative language models, e.g., ChatGPT, they have been widely used in multiple domains, including automatic summarization, sentiment analysis, and question and answer systems. However, it is worth noting that the use of ChatGPT for downstream metaphor detection tasks is often challenged with less-than-expected performance. Therefore, we propose a new method that aims to fully utilize the implicit knowledge of ChatGPT to support the task of detecting zero-shot verb metaphors. The method first uses ChatGPT to generate literal meaning collocations of verbs. For the text to be detected, subject-object pair of the target verbs in the text are parsed. Subsequently, these literal collocations and subject-object pair are mapped to the same set of topics, and the metaphors are finally identified through the analysis of entailment relations. The results show that the performance of ChatGPT in the verb metaphor detection task can be significantly improved by bootstrapping and integrating the implicit knowledge of ChatGPT.

Paper Type: long

Research Area: Efficient/Low-Resource Methods for NLP

Contribution Types: Approaches to low-resource settings

Languages Studied: English

Consent To Share Submission Details: On behalf of all authors, we agree to the terms above to share our submission details.