WellDunn: On the Robustness and Explainability of Language Models and Large Language Models in Identifying Wellness Dimensions

This study investigates the use of language models in identifying and explaining Wellness Dimensions (WD) for mental health applications, focusing on models like GPT-3.5/4, MEDALPACA, and RoBERTa. It finds that while general-purpose models like RoBERTa perform better, domain-specific models and LLMs do not significantly improve performance or provide satisfactory explanations. The research highlights the need for more robust and explainable models, especially in mental health, as they often lack domain knowledge and generate inadequate justifications. The study also introduces datasets like MULTIWD and WELLXPLAIN, which emphasize the importance of evaluating models' alignment with expert annotations and their ability to handle complex, real-world mental health contexts.