A new preprint describes an explainable system for suicide-risk assessment on social media. Rather than only outputting a severity score, the model is designed to show its work: it identifies the supporting language in a post and distinguishes between risk factors and protective factors. The authors built the system for the IEEE BigData shared task, suggesting a focus on practical, competition-style evaluation.
The approach relies on multi-task QLoRA, a parameter-efficient fine-tuning method that adapts large language models for several related tasks at once. In this case, the tasks include estimating risk severity, extracting evidence spans, and classifying factor types. The abstract does not report quantitative results, so the system's accuracy and generalizability remain to be assessed.
If it performs well, the method could help clinicians and moderators understand why an automated system flags a post as high-risk, making the output more trustworthy and actionable. The preprint is available on arXiv under identifier 2610.00610.