MentalMARBERT detects Arabic mental health disorders with 86% F1 score

Detecting mental health disorders from Arabic social media is notoriously difficult due to dialectal variation, informal language, scarce labeled data, and class imbalance. Most NLP advances have focused on English, leaving Arabic multi-class classification under-explored. To address this, Fatimah Almalki and colleagues at the University of Jeddah and King Saud University propose a two-phase framework. In Phase 1, they perform domain-adaptive and task-adaptive pretraining (DAPT and TAPT) on three Arabic models—AraBERT, CAMeLBERT, and MARBERT—using a large unlabeled corpus of Arabic mental health tweets. They then select the best backbone (MARBERT) based on a unified evaluation protocol.

In Phase 2, the authors compare four configurations: single-stage vs. hierarchical two-stage classification, combined with full fine-tuning vs. Low-Rank Adaptation (LoRA). They also constructed a novel annotated dataset of 50,670 tweets across six disorder categories (e.g., depression, anxiety), achieving high inter-annotator reliability (Krippendorff's Alpha = 0.733, pairwise agreement = 0.797). Results show that the domain-adapted MentalMARBERT with hierarchical two-stage full fine-tuning delivers the best performance: macro-F1 of 0.861 and accuracy of 0.877. These findings demonstrate that domain-specific adaptive pretraining and hierarchical classification significantly improve Arabic mental health detection, paving the way for real-world screening tools.