ToxGuard — Multilingual Toxic Comment Classifier XLM-RoBERTa

XLM-RoBERTa fine-tuned for multilingual toxicity detection across multiple languages. Achieves Mean ROC-AUC of 0.9918 with LIME-based explainability and an interactive Gradio demo.

Comment 3

ToxGuard — Multilingual Toxic Comment Detection

Inspiration

Online hate speech and toxic comments represent a growing challenge for digital platforms worldwide. The majority of existing content moderation systems are designed primarily for English, leaving significant gaps in protection for users communicating in other languages.

ToxGuard was developed to address this limitation by building a single unified model capable of detecting toxic content across multiple languages — eliminating the need for separate, language-specific moderation systems and enabling more inclusive content safety at scale.

What We Built

ToxGuard is a multilingual toxic comment classification system fine-tuned on XLM-RoBERTa Base. The system classifies user-generated text as Toxic or Non-Toxic and incorporates Explainable AI techniques — specifically LIME and transformer attention heatmaps — to provide interpretable, auditable predictions suitable for real-world moderation pipelines.

Results

Epoch Train Loss Val Loss ROC-AUC Accuracy
1 0.5420 0.4736 0.9805 91.48%
2 0.3216 0.2758 0.9901 95.19%
3 0.2124 0.3445 0.9918 95.26%
  • Best Validation Mean ROC-AUC: 0.9918
  • Best Accuracy: 95.26%
  • Training Samples: 7,646 | Test Samples: 999

Links

System Architecture and Pipeline

The end-to-end pipeline was structured as follows:

  1. Dataset Ingestion — Automatic detection and loading of .csv and .xlsx files from Kaggle input directories
  2. Preprocessing — Null value removal, whitespace normalization, and label column standardization
  3. Data Splitting — Stratified 85/15 train-validation split to preserve class distribution
  4. Tokenization — XLM-RoBERTa SentencePiece tokenizer with a maximum sequence length of 128 tokens
  5. Model Fine-Tuning — Hugging Face Trainer API with AdamW optimizer, learning rate of 2e-5, warmup ratio of 0.1, and weight decay of 0.01
  6. Checkpoint Selection — Best model selected based on highest validation Mean ROC-AUC score
  7. Evaluation — Confusion matrix, ROC curve, precision-recall analysis, and full classification report
  8. Explainability Integration — LIME token importance and last-layer attention heatmap visualization
  9. Deployment — Interactive Gradio application with real-time inference and live explanation output

Explainable AI

A core design goal of ToxGuard was transparency. The system incorporates two complementary explainability methods:

LIME (Local Interpretable Model-Agnostic Explanations)

Identifies the contribution of individual tokens toward the final classification decision, highlighting which words drive the model toward a Toxic or Non-Toxic prediction.

Attention Heatmaps

Visualizes token-level attention weights from the final transformer layer, providing a secondary view of where the model focuses during inference.

These features make ToxGuard suitable for deployment contexts where predictions must be explainable, auditable, and defensible — not just accurate.

Interactive Demo

The Gradio-based application provides the following capabilities:

  • Text input in any of the multiple supported languages
  • Real-time binary classification with probability scores
  • Visual toxicity probability indicator
  • Live LIME explanation chart generated per prediction

🔗 Try it live: https://huggingface.co/spaces/ayushtiwari18/toxgaurd

Challenges

Multilingual Text Handling

Processing text across different scripts, linguistic structures, and code-mixed inputs required careful tokenization strategy and preprocessing design.

Environment-Specific File Handling

Kaggle input paths reference directories rather than individual files. This required implementing glob-based auto-detection to correctly identify dataset files at runtime.

Library Compatibility

Recent updates to the Hugging Face Transformers and Hub libraries introduced parameter deprecations — specifically evaluation_strategy and use_auth_token — which required targeted fixes during development.

Subword Tokenization and Explainability

Generating word-level LIME explanations from SentencePiece subword tokens required additional post-processing to produce human-readable output.

Key Learnings

  • Cross-lingual transfer learning via XLM-RoBERTa generalizes effectively across languages, including code-mixed inputs such as Hindi-English
  • ROC-AUC is a significantly more informative metric than accuracy for imbalanced classification tasks
  • Explainability tools such as LIME serve both as a debugging mechanism and a validation layer for model behavior
  • Preprocessing quality and evaluation strategy have a measurable impact on final model performance

Future Work

  • Extend classification to a multi-label schema covering threat, insult, identity-based hate, and obscenity categories
  • Deploy as a production-grade REST API for real-time platform integration
  • Improve performance on low-resource languages through targeted language-specific fine-tuning
  • Integrate multilingual SHAP for deeper, gradient-based explainability

Built With

+ 3 more
Share this project:

Updates