A HYBRID DIALOGUE SYSTEM FOR STRUCTURED HEALTH CONVERSATIONS: INTEGRATING BILSTM-BASED INTENT CLASSIFICATION WITH FINITE-STATE CONTROL

Abstract

This thesis presents a hybrid dialogue system designed to support structured decision-making inhealth-related conversations, specifically focused on dietary salt intake. In contrast to end-toend generative models like GPT-4, which excel at producing fluent responses but suffer from unpredictability, opacity, and high computational demands, the proposed system emphasizes interpretability, low-latency operation, and offline capability. The architecture integrates a trainable intent classification module—implemented using Bidirectional Long Short-Term Memory (BiLSTM) networks, with and without attention mechanisms—with a transparent finite-state machine (FSM) controller developed using the Pytransitions library. A domain-specific dataset was synthetically generated using large language models (LLaMA-2 and Mistral) and manually labeled across six intent categories relevant to sodium counseling. A separate evaluation benchmark was established using GPT-4 via API. Empirical results show that the BiLSTM+Attention model achieves a balanced F1 score of 0.45 and an average inference time under 0.003 seconds, outperforming the baseline BiLSTM in all metrics while offering significant speed and transparency advantages over GPT-4. Furthermore, integration with the FSM ensures predictable, auditable dialogue flow—an essential feature in high-stakes domains like healthcare. The system demonstrates that hybrid AI approaches can bridge the gap between linguistic flexibility and operational reliability. Future extensions include memory-aware dialogue tracking, multilingual support, GUI deployment, and adaptive dialogue policies through reinforcement learning.