Enhancing Sinhala Hate Speech and Offensive Language Detection through XAI

Abstract

The rapid growth of social media has sped up the spread of hatred, abusive and offensive content, creating an urgent need for automated detection systems especially for low-resource languages such as Sinhala. This study develops, and evaluates standard deep learning (DL) models, transformerbased architectures, and a newly proposed hybrid model using DL models with Sinhala Offensive Language Dataset (SOLD) for detecting hate speech in Sinhala. Among Conventional DL models, Bi-LSTM demonstrated the strongest performance with 82% accuracy and a ROC score of 0.88, despite the challenge of informal expressions, and rich morphology in Sinhala language. Among transformer-based models, XLM-R large achieved the best results with 84% accuracy and an ROC of 0.92, demonstrating their effectiveness in modeling nuanced semantic and syntactic structures in Sinhala online discourse. Moreover, a hybrid model integrating multiple DL models was developed and evaluated, achieving superior performance with an accuracy of 86% with a ROC of 0.93 on Sinhala hate speech detection task, outperforming all baseline deep learning and transformer-based models. Beyond predictive performance, this study also contributes important interpretability of model predictions with Explainable AI (XAI) techniques - SHAP and LIME - which detail the local and global contributions of tokens. These explanations provide clear pathways to make decisions about models. Overall, this study presents a comprehensive evaluation about model’s performance and interpretability analysis of model’s predictions to detect hate speech in Sinhala with exciting brevity for future multilingual and explainable NLP applications.