Hybrid Text Representation for Explainable Suicide Risk Identification on Social Media

Abstract

Social media data that characterize users can provide mental health signals, including suicide risks. Existing methods for suicide risk identification on social media have demonstrated promising results; however, the limitation of existing methods is that they are unable to capture low-and high-level features with complex structured data on social media and are incapable of explaining the predicted labels. Explainable models are more useful when translated, so we aimed to evaluate a novel method that would produce explainable models. This article presents a hybrid text representation method that integrates word and document-level text representations to explain suicide risk identification on social media. The proposed method is then fed to a transformer-based encoder with ordinal classification to determine suicide risk. Our results show that our method outperforms state-of-the-art baselines with an FScore of 0.79 (an absolute increase of 15%) on a public suicide dataset. Our method shows that an explainable model can perform at a comparable level to the best nonexplainable models but has advantages if translated for use in clinical and public health practice.