Toxic Comment Classification using Deep Learning

Overview

A multilingual toxic comment classification system that can identify toxic content across 7 languages (English, Russian, Turkish, Spanish, French, Italian, Portuguese). The system uses language-aware transformers with custom attention mechanisms and advanced deep learning techniques to accurately classify comments into different toxicity categories.

Key Features

• Language-aware transformer model with XLM-RoBERTa base
• Support for 7 languages 360k+ comments
• Classification across 6 toxicity categories
• High performance with AUC scores above 0.92 for main categories
• Efficient data processing with caching system

Challenges & Solutions

• Handling multilingual text processing efficiently
• Implementing language-aware attention mechanisms
• Optimizing model performance across different languages
• Balancing memory usage with model complexity