Automatic detection of psychological distress indicators in online forum posts

The stigma associated with mental health issues makes face-to-face discussions with family members, friends, or medical professionals difficult for many people. In contrast, the Internet, due to its ubiquity and global outreach, is increasingly becoming a popular medium for distressed individuals to anonymously relate experiences. In this paper, we present a system for automatically detecting psychological distress indicators in informal text interactions on Internet discussion forums. We compare a suite of innovative features and classifiers on data downloaded from an online forum discussing psychological health issues. Psychologists annotated individual messages with a comprehensive set of distress labels derived from the Diagnostic and Statistical Manual of Mental Disorders (DSM) IV. The noisy nature of the forum posts and the large set of distress labels for multi-label text classification (many of which cannot be detected by a mere surface form analysis of the text), make the task extremely challenging. A late fusion technique combines outputs from different classifiers resulting in promising accuracy on this challenging multi-label classification problem.