Sentiment Analysis Based on Social Media - Early Stress and Depression Detection

Depression has recently gained significant attention as a condition marked by persistent and profound mood disturbances. Extensive research suggests that depression can influence individuals’ online speech behavior, manifested through the use of depressive language and a reduction in posting frequency. Our system seamlessly integrates various sources of information, including historical tweets and user profile data. Concerning historical tweets, we propose two methods to navigate the extensive and intricate user tweet history. Our findings indicate that these methods yield more pertinent user information. Subsequently, we input this information into our meticulously constructed deep learning classification model. This model is built upon a pre-trained BERT (Bidirectional Encoder Representations from Transformers) and a bidirectional LSTM (Long Short-Term Memory) model that incorporates attention mechanisms. In the context of user information, we extract relevant details and directly incorporate them into a deep learning model based on bidirectional GRU (Gated Recurrent Unit) and MLP (Multi-Layer Perceptron). Concurrently, to address the challenge of imbalanced depression datasets, we introduce Focal Loss and Dice Loss. Our experimental results underscore the effectiveness of these loss functions in our model. To validate the efficacy of our system, we reprocess the depression tweet dataset and conduct a series of experiments. Through these experiments, we conclusively demonstrate the robustness of our model, effectively mitigating the challenge of sample imbalance to a considerable extent.