ContextMentalQA: Modeling Cultural, Social, and Religious Context in Arabic Mental Health Question Answering

Lama Ayash, Ashwag Alasmari, Hassan Alhuzali

Abstract: Question answering (QA) for mental health requires models that attend not only to clinical intent but also to the cultural, social, and religious frames through which individuals articulate distress and seek help. This paper introduces ContextMentalQA, a socio-culturally informed annotation schema and corpus for Arabic mental health questions, designed for multi-label classification across three main categories (Cultural, Social, Religious) and their finer-grained sub-categories. The developed corpus comprises 2,677 patient questions, of which 557 received at least one socio-cultural label, highlighting the predominance of social framing in Arabic mental health discourse. ContextMentalQA is paired with a multi-label classification pipeline based on AraBERT, incorporating imbalance-aware optimization, semisupervised augmentation via high-confidence pseudo-labeling, and adaptive per-class threshold calibration. Empirical analyses demonstrate that incorporating pseudo-labeled data yields consistent improvements across standard metrics, with reduced label-wise error and stronger performance to underrepresented categories. The proposed schema, dataset, and baseline models provide a foundation for developing Arabic mental health QA systems that are linguistically accurate, culturally grounded, and socially responsive.

External IDs:doi:10.36227/techrxiv.176369738.87142789/v1