Stage-Aware Multi-Task Learning for Mental Health Conversation Classification

Abstract

Mental health chatbots need to be able to identify the stages of conversations-Exploration, Insight, Action, Crisis-in order to provide contextually relevant support. Current models are commonly unstage-conscious, so they cannot be used in practice. Our model is a multi-task neural architecture that has a shared BiLSTM encoder, position embeddings, and multi-head attention that simultaneously predicts user intent (80 classes) and conversation stage (4 classes). Heuristic keyword-based stage labeling and Word2Vec embeddings were used in training. We performed ablation experiments on three random seeds and tested how accurately stage transitions are in multi-turn situations. The model readily reached an intent accuracy of 90.4% and a stage accuracy of 98.8% on the test set. Stage-wise F1-scores were strong: Exploration (0.99), Insight (0.93), Action (1.00), Crisis (0.86). Word2Vec embeddings and multi-head attention were identified as important elements by ablation. The stateful chatbot was able to correctly transition between stages with 100% accuracy in the case of Anxiety-to-Coping and General Sadness, and 67% in Crisis Intervention in simulated conversations. This multi-task framework is stage-conscious and facilitates more compassionate and context-sensitive mental health care. Through the combined modeling of intent and conversation dynamics, the system is a step forward towards the AI chatbots being safer and more clinically aligned in their interventions, as well as what could be improved in the high-stakes crisis detection.