Research Paper
A Novel Multi-Task Approach to Analyzing Bipolar Disorder on Social Media Abstract
Mental health conditions affect millions of people today. While existing work on predicting mental health conditions from social media text focuses largely on depression and similar conditions, other less prominent disorders like bipolar tend not to receive in-depth analysis. Furthermore, these works tend not to analyze or model the correlated nature of these different disorders and conditions. To account for the coexistence and correlation of multiple mental health conditions, this paper introduces DeMHeM, a novel multitask framework designed for the descriptive classification of bipolar and related mental health topics on online platforms like Reddit. By treating each mental health category as a separate task, DeMHeM leverages both the shared latent and task-specific semantic feature space by integrating sentence-level and topic-level embeddings. It further incorporates Focal Loss for joint learning, inter-task parameter sharing, and regularization decay to optimize the prediction for the naturally skewed imbalanced dataset. Hence, the model distinguishes between different mental health categories and also models the correlation among them by categorizing each post into potentially multiple mental health categories. Our results show that DeMHeM surpassed the baseline models and can be used to understand the multi-faceted discussion of mental health topics for a given community.
Link to paper