Suicide Ideation and Risk Detection from Social Media Using GPT Models

Abstract

As a reason for the sensitiveness of suicide ideation and its considerable impact on people's lives, the demand to treat and prevent the suicide ideation issue has become an obligation. Suicide ideation is a result of a combination of psychological pain and hopelessness. According to the World Health Organization, the task of reducing the global suicide mortality rate is a target, that should be attained. The entire population uses social media platforms to express their feelings, emotions, sentiments, and opinions. Social media platforms are among the most popular sources of datasets related to mental health issues. The process of detecting suicide ideation from social media platforms is based on recent methods of artificial intelligence such as machine learning and deep learning. In this study, we propose fine-tuning large language models to evaluate and find the level of suicide risk in posts published on Reddit. We fine-tuned four GPT-3 models using the UMD Reddit suicidality dataset, which is related to the subreddit of suicidal ideation. Our experimental results illustrate the efficiency of the LLMs in addressing our task. The model attains a high F1-score of 92.3%, an accuracy of 94.8%, and a training loss of 0.050.