Multilingual Hate Speech Detection with XLM-RoBERTa
International Conference of Innovative Computer Engineering (ICE 2025) · S. Doghmash et al.
Designed and evaluated a multilingual ML pipeline for hate speech detection in Arabic, Turkish, and English. Implemented baseline models (TF-IDF + LR, SVM, RF) using a soft-voting ensemble, then fine-tuned XLM-RoBERTa with GPU acceleration, balanced sampling, and early stopping. Achieved 72.3% accuracy and 0.7079 Macro F1. Models were exported to ONNX and evaluated for real-time inference via API and Android deployment.