Pre- Trained Language Models for Mental Health: An Empirical Study on Arabic Q&A Classification.

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Digital Mental Health

Background:

• Pre-trained language models (PLMs) offer promise for accessible mental health resources.
• Limited research exists on PLM efficacy in Arabic mental health applications.
• This study addresses the gap in evaluating PLMs for Arabic mental health tasks.

Purpose of the Study:

• To evaluate the performance of various pre-trained language models (PLMs) in classifying Arabic mental health questions and answers.
• To compare traditional machine learning approaches with PLM-based strategies.
• To investigate the impact of fine-tuning and prompt-based techniques on model performance.

Main Methods:

• Utilized the Arabic MentalQA dataset for question and answer classification.
• Experimented with traditional feature extraction, PLMs as feature extractors, PLM fine-tuning, and prompt-based methods (GPT-3.5, GPT-4).
• Evaluated Arabic-specific PLMs: AraBERT, CAMelBERT, and MARBERT.

Main Results:

• PLMs outperformed traditional methods, capturing semantic nuances effectively.
• MARBERT achieved top performance with Jaccard scores of 0.80 (question) and 0.86 (answer).
• Fine-tuning and prompt-based few-shot learning (GPT-3.5) significantly boosted classification accuracy.

Conclusions:

• PLMs and prompt-based approaches show significant potential for supporting Arabic-speaking individuals in mental health.
• This research highlights the improved accessibility and effectiveness of PLMs in Arabic mental health contexts.
• Findings contribute to understanding PLM applications in culturally sensitive digital mental health solutions.