Keywords: Symptom Identification, Mental Health, Large Language Model, Prompt Engineering
Abstract: Identifying depressive symptoms using social media data has received great attention. Most existing studies have used pre-trained language models (PLMs) in predicting depressive symptoms. However, these methods have faced challenges in understanding the context of entire sentences or paragraphs, which requires to capture implicitly expressed symptoms. In addition, the existing class-imbalanced and insufficient training datasets often fail to recognize unseen symptom expressions. To tackle these problems, we propose to use Large Language Models (LLMs) with powerful language processing capabilities to identify depressive symptoms. Experimental results from two publicly available datasets on various prompting strategies indicate that LLMs are not capable of inferring depressive symptoms. With a detailed error analysis, we suggest a future direction to align the reasoning process of LLMs with that of a human expert in identifying depressive symptoms.