PRCD: A full-chain parallel residual compensation debiasing framework

Hate speech detection models often suffer from systemic misclassification due to biases in training data, particularly cognitive distortions in semantic associations learned in deep networks, making it especially challenging to identify implicit biases. While existing neuron pruning methods can mitigate explicit biases to some extent, removing parameters weakens the model’s semantic representation capability and struggles to address deeply ingrained cognitively distorted associations in deep networks. To tackle this issue, this paper proposes a full-chain parallel residual compensation debiasing framework (PRCD). This framework introduces a residual compensation method for soft pruning (RCM), which enables soft pruning of bias signals across the entire model, from shallow to deep layers—without compromising the model’s semantic representation ability. Additionally, a toxicity constraint enhancement method based on sensitivity attribution prediction (ToxCon) is incorporated to generate contrastive samples that expose bias, effectively guiding RCM in correcting implicit biases stemming from cognitive distortions in semantic associations. Experimental results on three public datasets demonstrate that the PRCD significantly improves model performance and fairness in detecting hate speech, achieving state-of-the-art performance.