Explainable Bangla Toxic Comment Detection using BanglaBERT with SHAP and LIME
A fine-tuned BanglaBERT detects toxic Bengali comments (86% accuracy, 87% F1, 0.91 ROC-AUC) and explains every prediction with SHAP and LIME.
Data-mining research (IEEE-format paper) that fine-tunes BanglaBERT on Bengali toxicity data using 5-fold stratified cross-validation and adds an explainability layer — SHAP and LIME reveal which words drove each toxic / non-toxic decision, tackling the black-box problem in low-resource Bangla NLP.
[ 01 ]
Research Overview
Toxic and hateful comments are rampant on Bengali social media, and the mental-health toll on adolescents is real. Most existing detection systems optimize accuracy alone and offer no explanation for their verdicts.
This study fine-tunes BanglaBERT for toxic-comment classification and wraps it in an explainability mechanism — LIME and SHAP — so moderators can see exactly which tokens triggered a flag.
[ 02 ]
Problem Statement
- Bengali toxicity detection is underserved: the language is low-resource, and informal writing, acronyms, and context-dependent insults defeat keyword rules and classical ML.
- Transformer models detect toxicity well but act as black boxes, which undermines trust in automated moderation.
- Applying both SHAP and LIME to a transformer-based Bangla toxicity framework had rarely been explored.
[ 03 ]
Objective
- Fine-tune BanglaBERT for binary toxic / non-toxic Bengali comment classification.
- Evaluate with 5-fold stratified cross-validation for reliable estimates.
- Add SHAP and LIME to surface the words driving each prediction.
- Contribute a transparent, trustworthy framework to low-resource Bangla NLP.
[ 04 ]
Methodology
- Data: multiple Bengali toxicity corpora compiled into a labeled benchmark for fine-tuning and evaluation.
- Model: BanglaBERT, the Bengali member of the BERT family, fine-tuned for sequence classification.
- Training: 5-fold stratified cross-validation with averaged accuracy, F1, and ROC-AUC reported.
- Explainability: LIME and SHAP both applied to individual predictions to identify important toxic words and validate model behaviour.
[ 05 ]
Models Used
BanglaBERT
Fine-tuned transformer with a sequence-classification head for toxic / non-toxic prediction.
SHAP + LIME
Post-hoc local explainers that attribute each prediction to the words that drove it.
[ 06 ]
Dataset
- Multiple Bengali toxicity datasets combined — a BanglaMedia corpus, a Bengali comments dataset, a contextual toxicity dataset, plus auxiliary and synthetic augmentation sets.
[ 07 ]
Implementation
- Fine-tuning loop over five stratified folds with the BanglaBERT transformer, aggregating per-fold metrics.
- Post-hoc analysis on held-out toxic and non-toxic examples using SHAP (additive attribution) and LIME (local surrogate) to highlight decisive tokens.
[ 08 ]
Key Features
- 5-fold stratified cross-validation for reliable generalization estimates.
- Dual explainability — both SHAP and LIME on the same framework.
- Word-level attribution that names the toxic terms behind each flag.
- Focus on low-resource Bangla NLP rather than high-resource English.
[ 09 ]
Results
- Averaged over the five folds, the fine-tuned BanglaBERT reaches 86% accuracy, 87% F1-score, and 0.91 ROC-AUC.
- SHAP and LIME attributions align with toxic keywords, showing the model reasons over the expected linguistic signals rather than spurious cues.
[ 10 ]
Outcome
- A validated, explainable toxicity detector that can underpin intelligent online moderation for Bengali content.
- A reproducible template for pairing transformer classifiers with SHAP + LIME in low-resource languages.
[ 11 ]
Tools & Technologies
[ 12 ]
Challenges
- Bengali linguistic nuance — context-dependent insults, slang, and transliterated forms.
- Limited high-quality labeled Bengali toxicity data; augmentation was needed.
- Explainability tools are compute-heavy over transformer embeddings.
[ 13 ]
Future Improvements
- Extend to multi-class severity (sarcasm, misogyny, religious hate) and cross-lingual transfer.
- Scale the labeled corpus and test other Bangla transformer variants (m-BERT, XLM-R, MuRIL).
- Ship the explainable scorer as a live moderation API.
[ 14 ]
Related work
Neuro-Screen: A Hybrid Ensemble Framework for Detection of Cognitive Impairment in Insomniac University Students
Undergraduate Thesis · Healthcare ML / DL
Thesis that fuses a gradient-boosting classifier with a three-layer neural network by averaging their probability outputs. Trained on 2,237 survey responses from students aged 20–35, the ensemble beats every standalone model across all metrics and pinpoints the lifestyle factors that most strongly predict cognitive decline.
Twitter Sentiment Analysis with TF-IDF and Multinomial Naïve Bayes
Natural Language Processing
Course research on the Sentiment140 corpus. 1.6M tweets are balanced down to 200k (100k positive + 100k negative), run through a deep preprocessing chain — lowercasing, URL/@/# removal, punctuation stripping, stopword removal, WordNet synonym substitution, Porter stemming, and lemmatization — then vectorized with TF-IDF and classified by Multinomial Naïve Bayes on a stratified 80/20 split.