Depression Detection Using BERT on Social Media Platforms

Depression detection from social media has attracted significant attention for its potential to offer early intervention and support to individuals facing mental health issues. In this study, we present a comprehensive evaluation of deep learning techniques for depression detection, with a specific focus on leveraging BERT, a powerful Natural Language Processing (NLP) Transformer model. Our exploration encompasses tailored preprocessing techniques for social media text, diverse feature extraction methods, and optimized model architectures tailored for depression detection tasks using BERT. Through rigorous experimentation and evaluation, we compare the performance of different BERT-based strategies, considering metrics such as accuracy, efficiency, and scalability. Additionally, we conduct a comparative analysis of labeled and unlabeled data from the same dataset. For labeled data, we employ BERT directly, while for unlabeled data, an autoencoder is utilized following label removal. The findings indicate that BERT outperformed other methods, achieving a high F1-score of 93% on the Reddit dataset. BERT achieved an impressive test accuracy of 91.92%, surpassing the Autoencoder model, which attained 84.84%.