Arabic Mental Health Question Answering: A Multi-Task Approach with Advanced Retrieval-Augmented Generation

Résumé

Arabic-speaking communities face persistent challenges in mental health support due to linguistic complexity, cultural nuances, and limited specialized resources.This study introduces AraHealthQA 2025, a multi-task framework for Arabic mental health question answering, tackling three subtasks: (i) question classification, (ii) answer strategy classification, and (iii) generative question answering using a Retrieval-Augmented Generation (RAG) pipeline.For classification,finetuned AraBERTv2, MARBERTv2, and Arabic RoBERTa on multi-label mental health data.For generation, developing a culturallyaware RAG system that integrates semantic chunking, query enhancement, and hybrid retrieval.Dense retrieval via akhooli/Arabic-SBERT-100K, sparse retrieval via rank_bm25, and generation using Sakalti/Saka-14B finetuned with culturally aligned mental health terminology (e.g., respecting religious sensitivities in advice).The approach achieves weighted F1-scores of 0.742 (question classification) and 0.718 (answer classification), and a BERTScore F1 of 0.821 representing up to 15% improvement over retrieval-only baselines.These findings demonstrate the potential of culturally sensitive, Arabic-focused NLP systems to advance accessible mental health support.

Citer ce document

Accès au document

Texte intégral en lecture en ligne, réservé aux abonnés SPHAERO et aux membres de l'institution. Se connecter

Auteur(s)

Statistiques

Consultations : 1

Téléchargements : 0