Self-harm Detection from Texts: A Comparative Study Utilizing BERT, Machine Learning, and Deep Learning Approaches

In a rapidly evolving world, the demands of modern life contribute to rising global anxiety and depression. Mental health, often neglected due to stigma, leads to untreated issues, including self-harm and suicide. Social media has become a platform for expressing mental health concerns, and employing suitable algorithms enables automated suicide sentiment detection. This research compares various BERT models to identify an efficient approach for leveraging social media to facilitate professional help-seeking without stigma. The study compared outcomes from different BERT models against those from conventional methods including Logistic Regression and Random Forest. The study also extended to incorporate deep learning models, specifically CNN, LSTM, and BiLSTM, offering a comprehensive analysis across diverse approaches to assess their effectiveness in text classification tasks. Through these experiments, we attained exceptional F1 scores of 99% for RoBERTa, 98% for AlBERT, and 96% for BERT base. In contrast, traditional models like Logistic Regression achieved 93%, Random Forest 89%, and deep learning models such as LSTM, BiLSTM and CNN achieved 82%, 93% and 90%, respectively. The results of these models are compared with each other to study and draw inferences on their performances. These findings hold significant implications for the development of more robust and language-specific Suicide text detection systems, contributing to the overall effort to curb the increasing suicide rates all over the world.