MSME: Zero-Shot Stance Detection via Multi-Stage Multi-Expert Reasoning

ACL ARR 2025 May Submission6481 Authors

Abstract: In recent years, zero-shot stance detection based on LLMs has garnered increasing attention and demonstrated promising results. However, it continues to face three significant challenges: a heavy reliance on accurate event background knowledge, poor performance in reasoning about complex targets, and difficulties in handling rhetorical expressions such as irony and metaphor. To address these challenges,we design a Multi-Stage Multi-Expert zero-shot stance detection framework(MSME). In the preparation stage, MSME automatically retrieves background knowledge related to the target and constructs explicit stance labels. In the analysis stage, the social media expert focuses on developing fine-grained stance labels, the knowledge reasoning expert emphasizes the logical connections between background information and the target, while the pragmatics expert analyzes the implicit influence of rhetorical devices on stance expression. In the decision-making stage, the decision-maker integrates the results of multi-dimensional analyses to produce a final, interpretable stance judgment. Experimental results show that the proposed MSME achieves higher F1 scores than current SOAT baselines across three public datasets, particularly for texts containing complex targets and rhetorical structures.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: stance detection

Contribution Types: Reproduction study, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 6481