Context-Enhanced Mental Health Message Triage Using A Dual- Encoder Deberta–Bilstm Architecture

Abstract

Mental health support systems are increasingly intertwined with automated systems which need clin-ically-interpreted comprehension of user input. However, the subtleties of emotionally-ladened, non-finely tuned everyday text, present limitations in comparison to generalized NLP models. This study introduces a dual-encoder DeBERTa contextual encoder and BiLSTM sequential extractor architec-ture for context-based mental health message triage. The proposed model is trained and validated based on MHMT-Bench, a publicly accessible benchmark collection of diversified data on intent classification and crisis risk with the nuanced complexity and emotional depth of the real world in mental health messaging. Findings show that the proposed architecture performs comparably for in-tent classification and significantly reliable for crisis-risk detection, facilitated through knowledge translation of crisis-risk and learning based on calibration-driven assessments, while additional ex-periments confirm the superiority of the dual nature of the two encoders over a singular approach. Ultimately, the findings support hybrid embeddings' transferability for more explainable and contex-tualized triage processes for a safer digital mental health landscape.