Modeling Co-Existing Mental Health Risks in Social Media via Multi-Label Learning and LLM

Abstract

Mental health risk detection from user-generated social media text has become increasingly important as psychiatric conditions such as depression, anxiety, and suicidal ideation continue to rise. However, most existing computational studies operationalize the problem using single-label datasets, implicitly assuming mutually exclusive conditions and thereby under-modeling clinically prevalent comorbidity. To better align modeling assumptions with real-world mental health phenomena, we formulate social-media-based risk identification as a multi-label text classification problem, enabling the simultaneous prediction of co-existing risks within a single post. We construct a hybrid corpus by integrating AIMH/SWMH and the Sentiment Analysis for Mental Health dataset, and by adding neutral/positive samples from Sentiment140 to strengthen healthy-content representation. Multi-label annotations are generated via Meta Llama-3-70B-Instruct using deterministic zero-shot prompting (temperature=0); a manual audit of 1,000 randomly sampled instances yields 94% agreement with the LLM-generated labels. We then benchmark multiple transformer-based encoders under a unified multi-label training protocol and report macro/micro F1 performance, highlighting the feasibility of transformer-based multi-label learning for modeling overlapping mental health risks at scale in social media.