CDCM: A counterfactual debiased calibration method based on knowledge distillation for stance detection

The stance detection task aims to accurately identify users' attitudes toward specific targets in social media texts. Recent knowledge distillation methods have shown promise by leveraging inter-class relational information from soft labels. However, they often overlook unintended bias-high-frequency and sentiment words-in training data, causing teacher models to over-rely on shallow lexical cues and weakening semantic learning. To address this, we propose CDCM, a counterfactual debiasing calibration method based on knowledge distillation. CDCM formulates stance detection within a structural causal model, identifies confounders, and uses factual-counterfactual reasoning to extract and remove bias via element-wise subtraction. Moreover, we design an adaptive sentiment lexicon construction strategy and a high-frequency feature threshold mechanism to mitigate lexical interference. Experimental results show that the CDCM algorithm effectively improves the performance.