Research on suicide risk identification technology based on social media

Suicide has become one of the severe global social issues. This study focuses on users' social media text data and adopts natural language processing techniques to conduct research on suicide risk identification, proposing a suicide risk recognition model based on dynamic segmentation and Transformer architecture. Users' social media posts over a period can often reach chapter length, but traditional Transformer-based models have input limitations concerning text length. To address this, the paper introduces a dynamic segmentation algorithm that minimizes cost to split all user posts within a month, sequentially inputs them into the RoBERTa model for feature extraction, and then utilizes a bidirectional Long Short-Term Memory network to extract features between segments that include contextual associations. The model also embeds its own attention mechanism to capture important local feature information. Eventually, it achieves an overall representation of the user's posting content, thus enabling effective identification of users' suicide risks. Compared to baseline models, there is a significant improvement in both F1 score and accuracy, reaching 73.1% and 74.67%, respectively.