Causal-Aware Classification of Social Media Hate Speech: Enhancing Robustness and Fairness with BERT

Social media platforms face increasing challenges in moderating hate speech effectively. While deep learning models like BERT have advanced detection performance, they often rely on spurious correlations and may exhibit bias toward marginalized communities. This paper proposes a causal-aware classification framework integrating causal inference techniques with BERT fine-tuning to improve robustness and fairness in hate speech detection. Using the HateXplain dataset, which includes labeled social media posts and annotator rationales, we construct a causal graph identifying potential confounders. Our model incorporates backdoor adjustment and invariant risk minimization (IRM) during training. Experiments demonstrate enhanced accuracy under distribution shifts and reduced demographic bias compared to baseline models.