MUST: An Explainable AI-Based Framework for Multilingual Hate Speech Detection

A significant challenge to the society harmony, and safety has surfaced due to the widespread hatred content on diverse social media platforms. This work introduces an explainable AI framework for multilingual hate speech detection system, capable of classifying Urdu, Roman Urdu, and English hate speeches. It helps to counter, and prevent the growing challenge of online toxicity in multilingual societies such as Pakistani society. It deals with the issue of abusive language detection in code-mixed, and transliterated text, a common phenomenon in South Asia digital traffic. Previous research was mainly focused to high-resource languages, and the work on low-resource languages such as Urdu, and Roman Urdu is often neglected. The proposed method “MUST: An explainable AI-based framework for MUltilingual hate Speech deTection” is built using finetuning of fundamental large language models. These models are fine tuned on novel multilingual dataset to classify text into multi class classification. A novel dataset is constructed from three representative social media plate forms. Then, labeled by human annotators, and is preprocessed. By incorporating explainable artificial intelligence principles, we enhance user trust, and facilitate better understanding of the AI’s reasoning process. MUST performed well by achieving 95.79% accuracy, precision, recall, and F1 score. It enhances automated moderation, providing a scalable solution to detect harmful online content, and foster safer, more inclusive digital environments across various languages on social media platforms.