End-to-End Pre-trained Dialogue System for Automatic Diagnosis

Yuan Wang, Zekun Li, Leilei Zeng, Tingting Zhao

Abstract: With the development of medical technology, Chinese medical resources are extremely scarce. At this necessary time, the development of dialogue agents to interact with patients and provide clinical advice has attracted more and more attention. In the task of generative medical dialogue, the end-to-end method is often used to establish the model. However, traditional end-to-end models often generate deficient relevance to medical dialogue. Towards this end, we propose to integrate medical information into initial pre-trained model and use the division of sentence based on “words and expressions” to improve the accuracy of medical entity recall, which will make the model have a deeper understanding of medical field. Finally, we use the Chinese medical dialogue MedDG [1] to fine-tune the model, so that the model can give the reply to the doctor’s clinical inquiry for the disease content from the patient. The experimental results show that our framework achieves higher accuracy in disease diagnosis, which won the fourth place in the 2021 medical dialogue generation task containing Chinese.