Leveraging Explainable AI and Sarcasm Features for Improved Cyberbullying Detection in Multilingual Settings

Cyberbullying has become a major social issue with the increasing prevalence of digital communication; it is especially difficult to identify in multilingual online environments. With an emphasis on the Bangla and Chittagonian languages, this paper presents a method that combines sarcasm detection with explainable AI techniques to improve the accuracy of cyberbullying detection across proposed languages. We compared several machine learning (ML) methods for cyberbullying detection and conducted a failure analysis to identify where models inaccurately classify cyberbullying. A prevalent issue was the misclassification of sarcastic texts as non-cyberbullying. To address this, we propose a method that incorporates ML-based sarcasm detection to mitigate these false classifications. To underpin our approach, we constructed a specialized sarcasm dataset, which was meticulously validated using inter-rater reliability measures such as Cohen’s Kappa and Fleiss’ Kappa, ensuring high-quality annotations. Our proposed methodology leverages a combination of machine learning models and LIME to not only improve detection rates but also provide transparency in model decisions, facilitating better understanding and trust among users. The results demonstrate a significant improvement in detection accuracy with the inclusion of sarcasm features, where the proposed method consistently outperforms previous work in accurately identifying cyberbullying instances.