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

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Stance detection identifies user attitudes in social media text.
• Knowledge distillation methods leverage soft labels but can inherit data bias.
• Bias from high-frequency and sentiment words weakens semantic learning in teacher models.

Purpose of the Study:

• To propose a novel counterfactual debiasing calibration method (CDCM) for knowledge distillation in stance detection.
• To address the issue of unintended bias in training data that affects model performance.
• To enhance the semantic learning capabilities of stance detection models.

Main Methods:

• CDCM formulates stance detection using a structural causal model.
• It employs factual-counterfactual reasoning for bias extraction and removal via element-wise subtraction.
• An adaptive sentiment lexicon and high-frequency feature threshold mechanism are used to mitigate lexical interference.

Main Results:

• Experimental results demonstrate the effectiveness of the CDCM algorithm.
• The method successfully improves the performance of stance detection models.
• CDCM mitigates the negative impact of shallow lexical cues and enhances semantic understanding.

Conclusions:

• CDCM offers an effective approach to debias knowledge distillation for stance detection.
• The proposed method enhances model robustness and accuracy by addressing data bias.
• This work contributes to more reliable stance detection in social media analysis.