Mental Health Disorder Identification From Motivational Conversations

Mental health disorder continues to be a grievous concern plaguing humans worldwide. The scarcity of mental health professionals (MHPs) has driven novel efforts lately to combat mental illness by developing automated systems capable of assisting MHPs. However, lack of high-quality conversational data due to privacy concerns remains a bottleneck toward its study and automation. Also, distinguishing and identifying various mental disorders is a challenging task as these are expressed using similar patterns in terms of language usage and overall sentiment polarity. The need for classifying different mental illnesses is rather intuitive as their true diagnosis is imperative for determining their path to recovery. As a first step toward creating automated virtual agent (VA), in this article, we curate a large-scale dataset for multiple mental illnesses comprising of dyadic conversations between a mentally ill user (typically a support seeker) and the VA (typically a peer supporter imparting hope and motivation), collected from a peer-to-peer support platform. We present a hierarchical attention-based deep neural network classifier for modeling conversations to detect different mental disorders as the dialog progresses. The proposed network is equipped with lexicon-based sentiment scores in order to prioritize certain contributing features. The proposed model attained an accuracy of 83.91% and outperformed several strong baselines.