LangSafe: A Multilingual Hate Speech Detection System Using Hybrid CNN-BiLSTM-Attention and Gradient Boosting Ensemble
Author : Greshma P Sebastian and Hameesha PS
Abstract :
Online hate speech poses serious threats to individuals and communities across linguistic boundaries. Existing automated detection systems predominantly target English and fail on multilingual platforms, especially in diverse nations like India. This paper presents LangSafe, a multilingual hate speech detection system supporting English, Hindi, and Malayalam. LangSafe employs a hybrid ensemble combining a CNN-BiLSTM-Attention deep learning model with XGBoost and LightGBM gradient boosting classifiers trained on dual TF-IDF feature representations. A weighted voting strategy fuses model outputs, refined by a language-specific lexicon heuristic layer. The ensemble achieves 88.7% accuracy and 87.3% macro F1-score on a combined multilingual test set, outperforming each individual component. The system is deployed as a web application featuring real-time scanning, bulk processing, dashboard analytics, and PDF audit report generation.
Keywords :
Hate speech detection, multilingual NLP, ensemble learning, CNN-BiLSTM, XGBoost, LightGBM, Malayalam, Hindi, code-mixed text.
