Answers are grounded in this paper and cite their sources.
Racism and body shaming are among the most damaging forms of abuse on Bangla social media, yet they are rarely treated as distinct detection targets, and existing studies seldom report controlled comparisons, statistical validation or interpretability. This paper presents an optimized multi-transformer framework that separates racism from body shaming in Bangla text already identified as abusive. A dataset of 4,989 verified instances (2,573 body shaming, 2,416 racism; 1,214 naturally collected comments and 3,775 ChatGPT-assisted sentences) was constructed. BanglaBERT, mBERT, MuRIL and XLM-RoBERTa were fine-tuned under identical conditions with Optuna search, stratified five-fold cross-validation, exact McNemar tests and SHAP with a faithfulness evaluation. BanglaBERT achieved the highest numerical performance (96.39% accuracy at baseline, 96.79 ± 0.59% under cross-validation, 96.29% after tuning), although its differences from the other models were not statistically significant and a wider search favoured XLM-RoBERTa (97.19%). Racism was recovered less reliably than body shaming, and SHAP faithfulness ran inversely to accuracy. Additional evaluations showed stable results across seeds and under leakage-aware grouped cross-validation (96.07–96.63%), but accuracy fell to 86.60% to 89.95% on naturally collected test comments and to 55.49–63.81% when models were trained on generated sentences alone, because generation changed what each class expresses, notably racism as regional stereotyping rather than colourism. On the same data, the prior study’s support vector machine reached 89.98%, while a character n-gram support vector machine matched the transformers and exceeded them when trained on collected comments (93.16%). Validation on naturally collected comments is therefore needed before deployment.
No takes yet. Share an insight, caveat, or question.
Hena et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: