Bengali Fake-News & Hate-Speech Detection with Explainable AI
A controlled comparison of BanglaBERT and multilingual BERT on Bengali fake-news and hate-speech tasks, paired with statistical testing and quantitative explanation evaluation.
The project fine-tunes BanglaBERT and mBERT on two Bengali datasets and compares them with TF-IDF classical machine-learning baselines.
Beyond predictive metrics, the work evaluates explanation faithfulness using comprehensiveness and sufficiency, measures SHAP–LIME agreement, and tests model differences with McNemar's test and bootstrap confidence intervals.
What the project demonstrates.
~95k news articles and ~50k comments across two Bengali datasets
Transformer comparison against TF-IDF classical ML baselines
Accuracy, macro-F1, ROC-AUC, McNemar test, and bootstrap confidence intervals
Quantitative SHAP/LIME faithfulness and agreement analysis
How the work was approached.
The challenge
Compare Bengali language classifiers while testing whether the explanations are faithful rather than only visually plausible.
Technical approach
- Fine-tune BanglaBERT and multilingual BERT against classical TF-IDF models.
- Use statistical comparisons and quantitative explanation tests including comprehensiveness and sufficiency.
Evidence & outcomes
- Two Bengali tasks: fake-news classification and hate-speech detection.
- Evaluation includes macro-F1, ROC-AUC, McNemar and bootstrap confidence intervals.
Scope note: Dataset sizes and outcome metrics are described in the project repository and CV; no benchmark rank is claimed here.
Want to inspect the implementation? Explore the source repository ↗