Early Mental Health Risk Assessment through Writing Styles, Topics and Neural Models

This paper describes the participation of the RELAI team in the eRisk 2020 tasks. The 2020 edition of eRisk proposed two tasks: (T1) Early assessment of risk of self-harm and (T2) Signs of depression in social media users. The second task focused on automatically filling a depression questionnaire given user writing history. The RELAI team participated in both tasks, and addressed them using topic modeling algorithms (LDA and Anchor), neural models with three different architectures (Deep Averaging Networks (DANs), Contextualizers, and Recurrent Neural Networks (RNNs)), and an approach based on writing styles. For the second task related to early detection of depression, the system based on LDA performed well according to all the evaluation metrics, and achieved the best performance among participants according to the Average Difference between Overall Depression Levels (ADODL) with a score of 83.15%. Overall, the submitted systems achieved promising results, and suggest that evidence extracted from social media could be useful for early mental health risk assessment.