Phase: Learning emotional phase-aware representations for suicide ideation detection on social media

Mental health conditions often manifest through distinct emotional phases that can be observed in social media activity. We introduce PHASE, a novel approach for suicide ideation detection that learns emotional phase-aware representations from social media posts. Our method recognizes that users experiencing suicidal ideation may go through different emotional phases, each characterized by distinct linguistic patterns and emotional expressions. PHASE employs a multi-phase learning framework that captures these temporal emotional transitions and uses them to improve detection accuracy. The model learns to identify and represent different emotional states such as despair, hopelessness, anger, and isolation, and models how these phases evolve over time. Through extensive experiments on social media datasets, we demonstrate that incorporating emotional phase awareness significantly improves the performance of suicide ideation detection systems compared to methods that treat all posts uniformly.