Heterogeneous graph reasoning for knowledge-grounded medical dialogue system

Beyond the common difficulties faced in task-oriented dialogue system, medical dialogue has recently attracted increasing attention due to its huge application potential while posing more challenges in reasoning over medical domain knowledge and logic. Existing works resort to neural language models for dialogue embedding and neglect the explicit logical reasoning, leading to poor explainable and generalization ability. In this work, we propose an explainable Heterogeneous Graph Reasoning (HGR) model to unify the relational dialogue context understanding and entity-correlation reasoning into a heterogeneous graph structure. HGR encodes entity context according to the corresponding utterance and deduces next response after fusing the underlying medical knowledge with entity context by attentional graph propagation. To push forward the future research on expert-sensitive task-oriented dialogue system, we first release a large-scale Medical Dialogue Consultant benchmark (MDG-C) with 16 Gastrointestinal diseases for evaluating consultant capability and a Medical Dialogue Diagnosis benchmark (MDG-D) with 6 diseases for measuring diagnosis capability of models, respectively. Extensive experiments on both MDG-C and MDG-D benchmarks demonstrate the superiority of our HGR over state-of-the-art knowledge grounded approaches in general fields of medical dialogue system.