WET -- Weighted Ensemble Transformer for Identifying Psychiatric Stressors Related to Suicide on X (formerly Twitter)

TL;DR

This paper introduces WET, a dual-branch deep learning model combining Transformer-based semantic analysis with emotional and behavioral features to improve detection of suicide-related stressors on social media.

Contribution

The paper presents a novel Weighted Ensemble Transformer architecture that integrates semantic and auxiliary signals for better suicide risk detection on X.

Findings

WET achieves 0.9901 accuracy in binary classification.

Hybrid approach outperforms traditional machine learning and baseline transformers.

Incorporating emotional and behavioral features significantly enhances detection performance.

Abstract

Suicide remains one of the leading causes of death worldwide, particularly among young people, and psychological stressors are consistently identified as proximal drivers of suicidal ideation and behavior. In recent years, social media platforms such as X have become critical environments where individuals openly disclose emotional distress and conditions associated with suicidality, creating new opportunities for early detection and intervention. Existing approaches, however, predominantly rely on raw textual content and often neglect auxiliary emotional and contextual signals embedded in user metadata. To address this limitation, we propose a Weighted Ensemble Transformer (WET), a dual branch deep learning architecture designed to identify psychiatric stressors associated with suicide in X posts. Our model integrates semantic representations extracted through Transformer encoders with an engineered feature vector capturing sentiment, subjectivity, polarity, and user engagement characteristics. We collected, filtered, and annotated 125,754 English tweets for suicide-related psychological stressors and evaluated the proposed model under two configurations. Extensive comparative experiments against traditional machine learning methods, advanced recurrent networks, and transformer baselines demonstrate that WET achieves state-of-the-art performance, reaching 0.9901 accuracy in binary classification. These findings show that hybridizing deep semantic signals with auxiliary emotional and behavioral features substantially improves suicidality detection accuracy.