Chain of Stance: Stance Detection with Large Language Models

TL;DR

The paper tackles stance detection by leveraging large language models through a novel prompting scheme called Chain of Stance (CoS), which decomposes the task into intermediate, stance-related assertions to guide reasoning. By treating LLMs as expert detectors that step through context, main ideas, emotional cues, and consistency checks before finalizing a stance, CoS achieves state-of-the-art results on SemEval-2016 Task 6 in both zero-shot and few-shot settings across multiple SOTA LLMs. Key contributions include the first application of chained reasoning to stance detection, a detailed six-step prompting protocol, and comprehensive experiments with ablation and error analyses demonstrating improved accuracy and interpretability. The work suggests that chain-of-thought style prompts can significantly bolster stance detection performance while offering more transparent rationales, albeit with higher prompt-length and computation costs that motivate future efficiency optimizations.

Abstract

Stance detection is an active task in natural language processing (NLP) that aims to identify the author's stance towards a particular target within a text. Given the remarkable language understanding capabilities and encyclopedic prior knowledge of large language models (LLMs), how to explore the potential of LLMs in stance detection has received significant attention. Unlike existing LLM-based approaches that focus solely on fine-tuning with large-scale datasets, we propose a new prompting method, called \textit{Chain of Stance} (CoS). In particular, it positions LLMs as expert stance detectors by decomposing the stance detection process into a series of intermediate, stance-related assertions that culminate in the final judgment. This approach leads to significant improvements in classification performance. We conducted extensive experiments using four SOTA LLMs on the SemEval 2016 dataset, covering the zero-shot and few-shot learning setups. The results indicate that the proposed method achieves state-of-the-art results with an F1 score of 79.84 in the few-shot setting.

Figures (4)

• Figure 1: The overall architecture of CoS.
• Figure 2: The specific implementation process of CoS.
• Figure 3: Error Analysis.
• Figure 4: Performance Variations Across Different Models.