Dep-capsule: capsule network for depression detection of Chinese microblog users

Abstract

PurposeDepression has become one of the most serious and prevalent mental health problems worldwide. The rise and popularity of social networks such as microblogs provides a wealth of psychological data for early depression detection. Language use patterns reflect emotional states and psychological traits. Differences in language use between depressed and general users may help predict and diagnose early depression. Existing work focuses on depression detection using users' social textual emotion expressions, with less psychology-related knowledge.Design/methodology/approachIn this paper, we propose an RNN-capsule-based depression detection method for microblog users that improves depression detection accuracy in social texts by combining textual emotional information with knowledge related to depression pathology. Specifically, we design a multi-classification RNN capsule that enhances emotion expression features in utterances and improves classification performance of depression-related emotional features. Based on user emotion annotations over time, we use integrated learning to detect depression in a user’s social text by combining the analysis results with components such as emotion change vector, emotion causality analysis, depression lexicon and the presence of surprising emotions.FindingsIn our experiments, we test the accuracy of RNN capsules for emotion classification tasks and then validate the effectiveness of different depression detection components. Finally, we achieved 83% depression detection accuracy on real datasets.Originality/valueThe paper overcomes the limitations of social text-based depression detection by incorporating more psychological background knowledge to enhance the early detection success rate of depression.