Improving Zero-Shot Stance Detection by Infusing Knowledge from Large Language Models

The Zero-shot Stance Detection (ZSSD) task is designed to predict someone’s attitude towards unseen targets with limited training data. However, existing methods often need complex modifications to network structures or high training costs. Utilizing large language models directly for zero-shot tasks has so far yet yielded suboptimal results and is difficult to be deployed in offline environments. In this paper, we propose GPT-TiDA, a novel framework that leverages knowledge from LLMs to improve data augmentation for zero-shot stance detection. We employ LLM agents to iteratively generate background knowledge and pseudo-samples for new topics through role-playing conversations as a domain expert and an experienced reader. Experimental results on some public zero-shot datasets demonstrate that GPT-TiDA significantly outperforms baselines in limited training data scenarios for unseen targets. To the best of our knowledge, GPT-TiDA is the first work to infuse LLM knowledge into zero-shot stance detection.