Approaching what and how people with mental disorders communicate in social media–Introducing a multi-channel representation

Over the last few years, studies related to the detection of mental disorders in social media have been increasing. The latter because the awareness created by health campaigns that emphasizes the commonness of these disorders among all of us has motivated the creation of new datasets, many of them extracted from social media platforms. In this study, we aim to contribute to the analysis of three major mental disorders that are hitting the world: Anorexia, Depression and Self-harm. To this end, we propose a novel model that, first, extracts three different views, or information channels, from the posts shared by users: thematic interests, writing style, and emotions. Then, it optimally fusions the information from each channel by using a gated multimodal unit. We evaluate the feasibility of our approach in the aforementioned tasks, first by comparing its output against traditional and modern strategies, and later against the best contestants in the eRisk evaluation forum. In both evaluations, our approach clearly outperforms all of its competitors. Through an exhaustive analysis section, we provide evidence of what is being captured by each information channel, then highlighting the importance and robustness of a more holistic view in critical classification tasks.