Zero-shot stance detection (ZSSD) focuses on effectively transferring stance recognition models from source to unseen target domains, especially when target domain data is scarce or absent. Existing research relies predominantly on contrastive learning strategies to enhance model generalization or incorporates external knowledge resources to compensate for semantic-information deficits. However, these approaches often suffer from two critical issues: (1) the inherent high sparsity and heterogeneity of social-media data and (2) insufficient attention to domain relevance and implicit semantic targets within user-generated content. To address these limitations, this paper proposes a data-augmentation framework based on internal knowledge expansion to improve model performance in ZSSD tasks. Within the framework, an implicit target-mining model is designed to discover relevant instances from the training data that exhibit explicit or implicit semantic tendencies similar to the target sample with an unsupervised scheme, providing contextual supplementation and enhancing target representation capabilities. Meanwhile, a stance-classification module integrated with a semantic filtering mechanism can effectively suppress the interference of irrelevant or noisy information on stance determination to ensure the augmented information can be used effectively. Extensive experiments on three mainstream benchmark datasets demonstrate that the proposed framework achieves superior accuracy and robustness compared to existing state-of-the-art methods, thereby validating its effectiveness and generalizability for ZSSD tasks.
Wang et al. (Thu,) studied this question.