XAI-Driven BERT Variants for Transparent Moroccan Darija Toxic Comments Detection in Social Media

In today’s globalized world, social media has become an integral part of our daily lives, where we interact, communicate, and express ourselves. Despite the multiple benefits, cyberbullying has given rise to the dissemination of negative and toxic content on social media. Therefore, it is essential to create a safe and positive environment, which requires clearly identifying and categorizing toxic content. The goal of this study is to identify malicious utterances in Moroccan dialect Darija on social media, particularly on the Instagram platform. Accordingly, we collected an Instagram dataset consisting of 91.559 instances, evenly balanced among positive, neutral, and toxic, with approximately 30.000 comments per class. We then fine-tuned the Arabic and Moroccan dialect pre-trained models to enhance their classification accuracy and performance. Leveraging BERT's strong capabilities in sentiment classification, we adopt a multi-class classification framework using several BERT variants such as DarijaBERT, DarijaBERT-mix, ARBERT, MarBERT, and CAMeLBERT-DA, with the implementations of Explainable AI (XAI) methods, particularly SHapley Additive exPlanations (SHAP) plots to describe the model behavior and the features' contribution to final predictions. The proposed method is highly effective, achieving an accuracy of 93.91% and F1-scores of 93.9%.