Depression Detection in Social Media using XLNet with Topic Distributions

Depression Detection in Social Media using XLNet with Topic Distributions

Due to the complexity of depressive diseases, detecting depressed users on social media platforms is a challenging task. In recent years, with an increasing number of users of social media sites, this field of re-search has begun to develop rapidly. To improve the detection performance of traditional methods, two challenges need to be overcome. The first challenge is that textual content posted on social media plat-forms suffers from serious data sparseness. The second one is how to effectively use emotions, user in-formation, and behavior characteristics to predict potentially depressed users. In this paper, we propose a novel model called the Topic-enriched Depression Detection Model (TDDM), which combines topic in-formation and user behavior to predict depressed users on social media platforms. TDDM first employs a Conditional Random Field Regularized Topic Model (CRFTM) to extract the topic knowledge of user posts. XLNet is used to encode posts to further expand the semantic features of short texts. Finally, we integrate user behavior features into TDDM to improve the detection performance of the model. The ex-perimental results on a real-world Twitter dataset demonstrate that the proposed model performs better than baseline models in detecting depressed users at both pseudo-document level and user level.