Do Sentiment and Emotion Affect Mental Health? A Multi-task Classification Framework for Comprehensive Understanding of Mental Health, Emotion, and Sentiment from Motivational Conversations

In light of the escalating challenges posed by our modern lifestyle, cultivating a nuanced approach for addressing mental health issues becomes imperative. Navigating the complexities of daily life demands a thoughtful and comprehensive strategy to identify and address the diverse array of mental health issues that may manifest. The challenges in accurately identifying mental health expressions stem from their complex character of communication, which frequently shares linguistic patterns and similar expressive nuances as communicated by humans. However, we hypothesize that mental health conditions are closely associated with affective factors in particular feelings, moods, and emotions. These states define how we think, feel, and behave. Thus, in this article, we aim to explore and analyze the association of the affective states such as sentiment and emotion with mental health in the view of identifying mental health conditions accurately once the feelings and emotions of humans are understood. In this regard, this article investigates multi-task classification encompassing mental health disorder identification (MHDI), emotion recognition (ER), and sentiment analysis (SA) in non-clinical conversations where MHDI forms the primary task and ER-SA the auxiliary tasks boosting the identification of the primary one. To demonstrate our hypothesis, we propose Core Fusion Network (CFN), a variation of multi-tasking in light of the significance that sentiment and emotion plays in understanding mental health. This method adeptly considers private and shared features across tasks, significantly enhancing classification precision. For our study, we extend the recently released MotiVAte dataset containing dyadic conversations between support seekers and a virtual assistant imparting hope and motivation to enclose emotion and sentiment tags for each conversation in a semi-supervised manner. Our hypothesis is reinforced by an extensive ablation study with state-of-the-art multi-task models and the proposed Core Fusion Network (CFN), which exhibits increased accuracy of 89.12% for MHDI, 64.24% for ER, and 79.04% for SA in the tri-task variant as opposed to its corresponding uni-task and bi-task variants. These outcomes underscore the potential of multi-task learning in streamlining mental health classification by integrating emotional and sentiment dimensions.