Towards Transparent Stance Detection: A Zero-Shot Approach Using Implicit and Explicit Interpretability

TL;DR

The paper tackles zero-shot stance detection by introducing IRIS, an interpretable framework that combines implicit (text subsequences) and explicit (linguistic) rationales within an information retrieval ranking paradigm. It delineates a three-stage pipeline—Rationale Generation, Relevance Ranking, and Grouping/Selection—followed by a rationale-encoded classification with a composite loss that ties rationale usefulness to stance accuracy. Empirical results on VAST, EZ-STANCE, P-Stance, and RFD show IRIS achieves strong generalization and interpretable predictions, outperforming baselines especially with limited training data. The work advances transparent stance analysis with practical implications for content moderation and bias detection, while acknowledging language and ethical constraints and proposing future multilingual extensions.

Abstract

Zero-Shot Stance Detection (ZSSD) identifies the attitude of the post toward unseen targets. Existing research using contrastive, meta-learning, or data augmentation suffers from generalizability issues or lack of coherence between text and target. Recent works leveraging large language models (LLMs) for ZSSD focus either on improving unseen target-specific knowledge or generating explanations for stance analysis. However, most of these works are limited by their over-reliance on explicit reasoning, provide coarse explanations that lack nuance, and do not explicitly model the reasoning process, making it difficult to interpret the model's predictions. To address these issues, in our study, we develop a novel interpretable ZSSD framework, IRIS. We provide an interpretable understanding of the attitude of the input towards the target implicitly based on sequences within the text (implicit rationales) and explicitly based on linguistic measures (explicit rationales). IRIS considers stance detection as an information retrieval ranking task, understanding the relevance of implicit rationales for different stances to guide the model towards correct predictions without requiring the ground-truth of rationales, thus providing inherent interpretability. In addition, explicit rationales based on communicative features help decode the emotional and cognitive dimensions of stance, offering an interpretable understanding of the author's attitude towards the given target. Extensive experiments on the benchmark datasets of VAST, EZ-STANCE, P-Stance, and RFD using 50%, 30%, and even 10% training data prove the generalizability of our model, benefiting from the proposed architecture and interpretable design.

Figures (7)

• Figure 1: Flow diagram of IRIS followed by stage-wise detailed architecture design. In stage-wise diagram: [Left]: Relevance Ranking; [Center]: Grouping and Selection; [Right]: Classification. Notations: $R_i$: LLM-generated $i^{th}$ implicit rationale; $\{S^{rf}_i, S^{ra}_i, S^{rn}_i\}$: Relevance scores of $i^{th}$ implicit rationale towards favor, against, and neutral stances; ${Re}_k$, ${Ie}_k$: k relevant and irrelevant implicit rationales; ${IR}_i$: Encoded $i^{th}$ relevant implicit rationale; ${ER}$: Encoded explicit linguistic rationale; $\{S^{pf}_i, S^{pa}_i, S^{pn}_i\}$: stance scores for each encoded relevant $i^{th}$ implicit rationale and explicit rationale; $S_{final}$: Final predicted stance. The details are clearly mentioned in Section \ref{['sec_methodo']}.
• Figure 2: Instruction for Ranking Algorithm
• Figure 3: LLM Prompt for Explicit Rationale
• Figure 4: LLM Prompt for Implicit Rationale
• Figure 5: Results of IRIS with few-shot targets VAST