Exploring Multi-Agent Debate for Zero-Shot Stance Detection: A Novel Approach

Zero-shot stance detection aims to identify the stance expressed in social media text aimed at specific targets without relying on annotated data. However, due to insufficient contextual information and the inherent ambiguity of language, this task faces numerous challenges in low-resource scenarios. This work proposes a novel zero-shot stance detection method based on multi-agent debate (ZSMD) to address the aforementioned challenges. Specifically, we construct two debater agents representing the supporting and opposing stances. A knowledge enhancement module supplements the original tweet and target with relevant background knowledge, providing richer contextual support for argument generation. Subsequently, the two agents engage in debate over a predetermined number of rounds, employing rebuttal strategies such as factual verification, logical analysis, and sentiment analysis. If no consensus is reached within the specified rounds, a referee agent synthesizes the debate process and original input information to deliver the final stance determination. We evaluate ZSMD on two benchmark datasets, SemEval-2016 Task 6 and P-Stance, and compare it against strong zero-shot baselines such as MB-Cal and COLA. The experimental results show that ZSMD not only achieves higher accuracy than these baselines, but also provides deeper insights into subtle differences in opinion expression, highlighting the potential of structured argumentation in low-resource settings.