Text-based Decision Fusion Model for Detecting Depression

With about 300 million people in the world suffer from depression, depressive disorder has become a major health problem in the world. The 2017 Audio/Visual Emotion Challenge required Participants to build a model in order to detect depression based on audio, video, and text data. In this paper, we use single-modality, transcribed text data, for depression detection. We proposed a decision fusion model which combines Bert text embedding of interview transcript and key phrases recognition. Text embedding module is composed of Bert embedding model and LSTM network. Key phrases recognition module recognizes words such as “depression”, “cannot sleep” that are believed to be valuable in improving the recognition accuracy. We fuse the two identification methods at the decision level. Our proposed decision fusion model outperforms previous single-modality approaches in terms of classification accuracy. The F1 scores and precision is 0.81 and 0.82, respectively.