Continuous Entity Reasoning for Multi-Turn Medical Dialog Generation

Ye Wang, Xinyang Li, Hong Yu, Feng Hu, Guoyin Wang, Dajiang Lei

Abstract: Medical Dialogue Systems (MDS) are indispensable tools for medical assistance, serving as a vital component of consumer electronics. Consumer technology transitions from simple manual devices to advanced intelligent systems that integrate seamlessly into daily life. In this process, MDS contributes significantly to the improvement of health management. Current MDS research mainly focuses on extracting medical entities from the dialogue history as key information to improve response quality. However, these approaches often neglect the diverse relationships between entities, especially in multi-turn medical dialogues. To tackle this challenge, we introduce the Continuous Entity Reasoning Medical Dialogue Generation model (CRMED), by leveraging large language models and centering on entity reasoning in multi-turn dialogues to significantly improve the accuracy and informativeness of generated responses, for supporting the health management needs of consumer health technologies. Specifically, we first propose a novel strategy for encoding historical dialogues, which draws on structured background knowledge to investigate the interaction between contextual information and external knowledge, fully integrating text features and knowledge features. Then, based on contextual information and background knowledge, we construct a tree-like reasoning structure to simulate the relationships between entities. Finally, to enhance the quality of the dialogue, we incorporate the entities and paths involved in each step of the reasoning process as part of the dialogue generation. The effectiveness of the proposed model is validated through extensive experiments on two large medical dialogue datasets. The model demonstrates significant potential in consumer health technology applications such as anomaly detection, fault diagnosis, and predictive maintenance. Moreover, it provides a robust, adaptive, and scalable solution for medical assistance, driving consumer health technology towards a more efficient and secure future. Further details are available in the implementation at https://github.com/xinyang183/CRMED.

External IDs:dblp:journals/tce/WangLYHWL25