Dynamic BERT-Reinforcement Learning Model for Intent Recognition in Medical Dialogue Systems

Abstract

This study proposes a dynamic reasoning model of medical dialogue intentions that integrates BERT and reinforcement learning, aiming to solve the recognition difficulties caused by complex multiple rounds of interaction contexts and changeable user intentions in medical scenarios. Although the traditional BERT model is excellent in semantic modeling, it has limitations such as poor adaptability and static response strategy in the face of dynamic changes in intention expression in medical dialogue. Therefore, this paper introduces reinforcement learning mechanism, and realizes dynamic intention reasoning and policy optimization through state modeling, reward function and policy network. The experimental results highlight the robustness of our model in complex and dynamic medical dialogue scenarios. In high-complexity intent recognition tasks, our model achieved an accuracy improvement of 8.5%. Moreover, in extended multi-round dialogues, the BRL model demonstrated a significant increase in recognition accuracy—from 32% in the 70th round to 60% in the 140th round. This performance was notably better than that of the BLN model, which achieved about 40% accuracy. These improvements underscore the effectiveness of integrating reinforcement learning to adapt to evolving user intents and provide more accurate and contextually relevant responses in long-duration medical dialogues. In the sensitivity analysis of reward function, different reward functions have a significant impact on the model performance. Among them, RWA and RWF perform best when the weight numbers are 2 and 4, with an accuracy rate of more than 70%, while RWN and RWS are often below 40%. To sum up, the model combining BERT and reinforcement learning not only improves semantic understanding capabilities, but also realizes dynamic strategy adaptation, providing an efficient and intelligent intentional reasoning solution for medical dialogue systems.