TRANSFORMER-BASED CYBERBULLYING DETECTION IN MOROCCAN DARIJA: A WILLARD TAXONOMY APPROACH

Cyberbullying remains a serious threat, yet detection tools for Arabic dialects are scarce. Moroccan Darija poses particular challenges for automated systems. This paper evaluates three transformer models on the Offensive Moroccan Comments Dataset. Results show MARBERT achieves 85.07% F1-score, outperforming AraBERT (83.84%) and the multilingual baseline (80.10%), confirming that dialectal pre-training matters for low-resource varieties. We also propose a severity framework based on Willard's (2007) cyberbullying taxonomy, which distinguishes eight behavioral types. Our system maps these into three levels: CRITICAL (cyberstalking, harassment, outing) for immediate action, MODERATE (flaming, denigration, exclusion, trickery, impersonation) for standard review, and NONE for safe content. Dataset analysis shows 9.6% of comments fall into the CRITICAL category. On this high-risk class, MARBERT reaches 86.05% F1-score with 84.80% recall, ensuring most dangerous content gets flagged. These results offer practical guidance for deploying content moderation systems for Arabic communities.