Medical Dialogue Generation via Extracting Heterogenous Information.

The goal of medical dialogue generation is to produce precise doctor responses so that patients can receive trustworthy medical advice. Medical dialogue generation attracts more and more attention as a result of the strict requirements for response accuracy. The majority of current research, however, simply extracts the patient's status from the dialogue history, which is insufficient for extracting other medical information from the medical discourse and overlooks the key features of the dialogue history itself. Even if these techniques gather pertinent information, such as patient symptoms and diseases, they are still unable to generate accurate and instructive answers. To deal with this problem, we propose a dialogue generation model that can accomplish Heterogenous Medical Information Extraction (HMIE), including patient attributes, dialogue topics, and doctor decisions. Through the attention mechanism, we present a patient attribute classifier to comprehend the variety of patient-related information in the dialogue. Then, using the gating mechanism, we suggest a selector acquires more precise patient attributes according to various dialogue context factors. The dialogue topic locator initially deduces the dialogue topic and direction from the dialogue history to aid the generation of the doctor's decision before the doctor diagnosis network reasons the doctor's decision. We conduct experiments on two large medical dialogue datasets, and a large number of experimental results show that our model HMIE outperforms the existing baseline.