Social bias in language models continues to create fairness risks in multilingual and multicultural environments. Existing datasets provide limited cultural diversity, insufficient support for overlapping bias categories, and minimal availability of human-interpretable reasoning, which reduces transparency and reliability in the bias detection. The ToxicBias-Reasoning dataset addresses these gaps by providing 7,562 annotated statements representing bias related to caste, religion, race, gender, political identity, and LGBTQ issues, including 247 caste-specific samples and 1,923 non-biased samples. The dataset includes manually validated labels and instance-level reasoning, with training and validation reasoning generated through a GPT-4o-assisted human-in-the-loop process and a fully manual test set for high-fidelity evaluation. This study evaluates transformer-based classification models trained using hierarchical and multi-task strategies. The results demonstrate strong performance for bias detection, with RoBERTa achieving a macro F1 score of 0.9099, and for multilabel category classification, where incorporating a logic-aware loss improves consistency and yields F1 scores of up to 0.95 for major categories such as race and religion, along with notable gains in minority categories, including caste and political bias. A text-to-text knowledge distillation framework additionally trains a compact generative model for reasoning, with BART-Large attaining a ROUGE-L score of 45.22 and a BLEU score of 16.90. These findings support the practical deployment of explainable and culturally grounded bias detection systems in fairness-critical natural language processing applications.
Kumar et al. (2026) studied this question.