Multilingual Toxic Comment Detection

• 

  This is the screenshot of the ROC-AUC score achieved by the model which is = 99.2% and this shows the precision and accuracy of this model

• 

  This is the screenshot of the final validation of the model

Inspiration

Toxic content crosses language barriers — hate speech, threats, and abuse exist in every language. Building a scalable multilingual moderation system that works across Hindi and English was the core motivation for this challenge.

What I Built

A binary toxic comment classifier using XLM-RoBERTa-base — a cross-lingual transformer model supporting 100 languages — fine-tuned to classify comments as Toxic (1) or Non-Toxic (0).

How I Built It

• Used XLM-RoBERTa-base for native multilingual support across Hindi and English
• Fine-tuned for 3 epochs on 9,000 labeled comments
• 90/10 train-validation split with seed=42 for reproducibility
• ROC-AUC computed using softmax probabilities for reliable ranking
• max_length=128, batch_size=16, fp16=True
• Platform: Google Colab T4 GPU

Results

Non-Toxic (0) → Precision: 0.97 | Recall: 0.96 | F1: 0.96

Toxic (1) → Precision: 0.96 | Recall: 0.97 | F1: 0.96

✅ ROC-AUC Score: 0.9922

Challenges Faced

• Handling multilingual text (Hindi + English) without any language-specific preprocessing
• Perfectly balanced dataset (50-50) required careful evaluation using ROC-AUC over accuracy
• Excel file format (.xlsx) required special handling compared to standard CSV

What I Learned

XLM-RoBERTa handles multilingual text natively without any translation or language detection preprocessing. Cross-lingual transfer learning is extremely powerful for low-resource language tasks like Hindi toxic comment detection.

Built With

• google-colab
• huggingface-transformers
• numpy
• openpyxl
• pandas
• python
• pytorch
• scikit-learn
• t4-gpu
• xlm-roberta