Sakinah-AI at MentalQA: A Comparative Study of Few-Shot, Optimized, and Ensemble Methods for Arabic Mental Health Question Classification
Year: 2025
Type: conference-paper
Abstract: This paper details the system developed by team Sakinah-AI for the MentalQA 2025 shared task, focusing on Arabic mental health question classification.We compare few-shot learning with Large Language Models against fine-tuning of BERT-based models (CAMeL-BERT and AraBERTv2).Few-shot learning with Palmyra-Med-70B achieved the highest weighted F1-score of 0.605, followed by hyperparameter-optimized CAMeL-BERT at 0.597.Notably, 5-fold ensemble methods proved detrimental to performance.Our results demonstrate that for low-resource specialized domains, both few-shot learning and optimized fine-tuning of appropriate base models outperform ensemble strategies.To ensure reproducibility all experimental code and final fine-tuned models are made publicly available.
Institutions: University of South Wales, Cairo University
Language: English
Topic: Mental Health via Writing
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W4416037841