How can entities improve the quality of medical dialogue generation?

Studies on Medical Dialogue Generation (MDG) contribute to a highly reliable medical dialogue system. As medical dialogue involves professional entities, such as symptoms and diagnoses, recent MDG models follow the pipeline composed of entity prediction and entity-aware dialogue generation to enhance professionalism and reliability. However, existing entity-aware models have not taken full use of entities. In this paper, we propose a pre-training model of mutual reconstruction of entities and corresponding sentences to establish a strong association between entities and sentences. Our model can increase language fluency while strengthening the role of entities in professional reliability. Moreover, we found a certain pattern of entity transitions during dialogues, so we use bidirectional GRU to encode historical text information and historical entity information separately. Then we use an attention mechanism to fuse them. We performed several ablation experiments to demonstrate the validity of our two models. Finally, we propose Entity-enhanced Dialogue Generation (EDG) model and improve 63%-129% in BLEU and Distinct metrics compared to the baseline.