Detection of Severity of Depression from Social Media Data using Attention-Based Approach

Abstract

In clinical psychology, diagnostic of mental illness is mostly done by patient’s self reported experiences, behaviors reported by the patients themselves, their relatives and a mental status examination. This method can lead to a variety of biases, such as cognitive bias, in which patients hide their illness for fear of judgement. Popular social networks can serve as a tool for dealing with this problem. But mental health research in this domain has been hindered by a lack of standard typology, scarcity of adequate data and lack of a robust classification network. In this thesis, the clinical articulation of depression is leveraged to build a typology for social media texts for detecting the severity of depression. It emulates the standard clinical assessment procedure Diagnostic and Statistical Manual of Mental Disorders (DSM-5) and Patient Health Questionnaire (PHQ-9) to encompass subtle indications of depressive disorders from social media texts. The typology has been developed with the association of two expert psychologists. To examine the typology, a dataset is constructed by scraping posts from Twitter, followed by a standard annotation method to label each tweet as ‘non-depressed’ or ‘depressed’, while three severity levels are considered for ‘depressed’ tweets: (1) mild, (2) moderate, and (3) severe. To classify severity of depression in this dataset, two attention-based models, namely BERT and DistilBERT are pre-trained and fine-tuned and a strong baseline result is provided. The findings of this study ought to provide strong directions for further research in this domain.