LLM-Empowered Stance Detection Based on the Expanded Stance Triangle Framework

ACL ARR 2025 February Submission6480 Authors

Abstract: Stance detection on social media refers to the task of predicting the attitudes (favor, against or neutral) of documents toward a specified target. Recently, there has been an increasing interest in employing Large Language Models (LLMs) to detect stance, demonstrating impressive performance without relying on labeled data. However, these models tend to be conservative and thus often classify documents as neutral, since users typically express their attitudes implicitly through other objects, rather than directly mentioning the target. In this paper, we present LLMTriStance, a novel LLM-empowered approach for stance detection in social media, integrating the expanded stance triangle framework from linguistics. Leveraging pseudo labels generated by LLMs and nouns extracted via syntactic tools, we apply pattern mining to actively discover the common objects associated with specific evaluations when expressing attitudes toward a target. These stance expression rules are then purified through conflict identification and resolving, enabling the generation of valuable prompts for LLMs across various cases. This process forms an iterative cycle, leading to progressive improvements in accuracy. Experimental results on multiple stance detection datasets show that our model outperforms state-of-the-art methods, providing interpretable object-attitude pairs as rationales for its predictions.

Paper Type: Long

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

Research Area Keywords: stance detection

Contribution Types: Model analysis & interpretability, Approaches to low-resource settings

Languages Studied: English

Submission Number: 6480