Transfer Learning for Self-Report Questionnaire Completion and Measurement of the Severity of Signs of Depression based on Social Media Posts

This study pertains to the domain of Data Science, specifically focusing on the application of advanced computational methods to enhance societal well-being. We leverage the contribution of Social Media Networks (SMN) in detecting mental disorders. Specifically, we focus on measuring the severity of a SMN user’s depression symptoms. We exclusively use text data sourced from the SMN Reddit. Our research process is based on the automatic completion of a self-report mental state questionnaire, Beck’s Depression Inventory (BDI). This consists of 21 questions-statements and their respective multiple-choice scaled answers. The goal is to develop prediction systems for both individual answers and the overall depression state of users. Our approached is based on state-of-the-art Natural Language Processing. Specifically, we utilize models based on BERT, which achieve Transfer Learning from large text collections to tasks with significantly fewer training data. We implement both approaches recommended by literature for BERT-based models: feature-based approach and fine-tuning. We evaluate our methods using metrics recommended by the literature and developed specifically for the needs of this work, such as Average Hit Rate (AHR), Average Closeness Rate (ACR), Average Difference between Overall Depression Levels (ADODL), and Depression Category Hit Rate (DCHR). The first two evaluation metrics refer to the success of predicting the questionnaire’s answers per se, while the last two refer to the success of predicting the general depression state of the SMN user. The classification systems we develop achieve competitive results compared to similar implementations that have been reported.