A BERT-based sentiment analysis model for depressive text

Depression is a big public health problem around the world. Like the traditional methods of screening, clinical assessments, and self-rating scales are among the limits. These methods are based on the subjectivity of patient reports and the experience of doctors and clearly can have bias. In addition, they do not permit real-time monitoring of changes in mood. For this reason, this field of mental health is challenged with accurately and efficiently detecting probable mood changes in individuals at risk for depression. To resolve these challenges, we then propose a BERT-based sentimental analysis model adapted for depression-related text. Current methods for screening for traditional depression and established sentiment analysis models are hindered by data scarcity, data imbalance, and the problems of deriving emotional changes in a dynamic and clean manner. To leverage the strengths of BERTs in our model, we train the last layer of BERTs and then construct a classification layer tailored to the depressive emotion, which is more sensitive to the semantic information of the mood. Moreover, we also present data augmentation in its simplest form, which is the rearrangement of positive samples to increase the diversity of samples while still preserving the semantic meaning. To tackle data imbalance, we introduce a dynamic weight mechanism to the BCEWithLogitsLoss function by adjusting the weight of the positive samples with data distribution. The model effectively captures the complex emotional changes by combining global and local semantic dynamic representations. In our empirical study, we analyze the open-source dataset of interview transcripts of 162 volunteers. Our model is evaluated using the experimental results, and the precision, the recall rate, and the F1 score obtained are 0.667, 0.667, and 0.667, respectively. Our model has achieved 2.62 higher in terms of precision and F1 score as compared to the existing models. The results indicate that our model provides a big benefit in identifying screened individuals at risk for depression.