Year: 2025
Type: conference-paper
Abstract: Mental health detection in online discourse is a growing area of NLP research, particularly for low-resource languages such as Arabic, where stigma and limited access to professional care make anonymous, technology-driven solutions valuable.In the context of the AraHealth shared task, we were provided with three subtasks: multi-label classification for questions, multilabel classification for answers, and a QA system leveraging models developed in the previous two tasks.Our approach employed data augmentation to address class imbalance, as certain categories in the dataset were significantly underrepresented.Since our method relied on prompting models to classify questions and answers as well as to generate answers for the QA system, we utilized Gradient-free Edit-Based Instruction Search (GrIPS) to optimize prompt selection.Our system achieved strong results across all three subtasks, ranking 1st in answer classification and 3rd in both question classification and QA system answer generation.
Authors: AbdulRahman A. Morsy, Saad Mankarious, Ayah Zirikly
Institution: George Washington University
Language: English
Topic: Mental Health via Writing
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W4416037788