Knowledge-driven Chinese Medical Dialogue Text Generation with Corpus Construction

This study proposes a multi-task multi-label doctor–patient dialogue dataset (TLMedDialog) that focuses on gastrointestinal medicine. First, the study describes the labelling specification of the doctor–patient dialogue dataset and then discusses the dataset’s labelling process and statistics. Based on this dataset, the study then develops a knowledge-based medical dialogue model. The current medical dialogue text generation process has many problems, such as insufficient medical entities in the reply, confusion regarding medical entity relationships, and a lack of role information etc. This study proposes using entity prediction, dialogue history, and medical knowledge graphs to assist in the generation of medical dialogue text in order to improve the credibility and fluency of the process. The study also performs corresponding performance comparisons for natural language understanding (including intent classification and medical entity recognition (MER)) and dialog policy learning.