DECEN: A deep learning model enhanced by depressive emotions for depression detection from social media content

Abstract

Depression is a serious and recurrent mental illness that significantly affects an individual's life and the society as a whole. Automatic detection of depression is crucial for early intervention and minimizing negative consequences. Existing studies on building deep learning models for automated depression detection have mainly used post-level emotion polarity (i.e., positive and negative emotions) and word embeddings as predictive features. Few have considered depressive emotions (e.g., anhedonia) expressed in those posts, despite that depressive emotions are essential to clinical depression diagnosis. Moreover, existing approaches for depression detection often ignore the relationship between emotions and their context. This study proposes a Depressive Emotion-Context Enhanced Network (DECEN) that consists of a pre-trained depressive emotion recognition module and an emotion-context enhanced representation module to address those limitations. DECEN first integrates semantic and syntactic structure representations of textual content of social media posts to identify depressive emotions conveyed through terms either explicitly or implicitly, rather than general emotion words. Furthermore, we propose an emotion-context enhanced representation method to enhance the role of the context of depressive emotions in depression detection. The evaluation using real social media data demonstrates that DECEN outperforms the state-of-the-art models in depression detection. The results of an ablation experiment also reveal that the proposed depressive emotion recognition and emotion-context enhanced representation modules, the two novel design artifacts, improve model performance. This study contributes to depression diagnostic decisions by introducing a novel method and providing new technical and practical insights for detecting depression from social media content.