GISD: A Generative and Interpretable Framework for Explainable Zero-Shot Stance Detection

Zero-shot stance detection (ZSSD) seeks to identify the stance of a text in relation to an unseen target. Despite its importance, existing approaches rarely focus on explicitly explaining the reasoning process underlying their predictions or decisions. This paper represents the first attempt to address this gap by generating rationales for predictions in ZSSD, demonstrating that such explanations can effectively uncover the reasoning process. To this end, we propose a novel framework for stance detection called Generative and Interpretable Stance Detection (GISD). A key component of GISD is a three-hop logic elicitation (TH-LE) method, which leverages the chain-of-thought approach to extract if-then reasoning expressions from large language models (LLMs). These if-then expressions are then reformulated alongside the original inputs to create a new training dataset, which is subsequently used to train a generative model. To further enhance the training process, we introduce an iterative-instruction data augmentation (I2DA) method to expand the dataset. Experimental results reveal that the GISD framework significantly surpasses state-of-the-art methods in ZSSD performance. Moreover, manual evaluations confirm that the generated if-then expressions effectively explain the rationale behind the model's predictions.