This project proposes a pre-training model that can integrate knowledge graph completion and question-answering tasks, and emphatically tests the application of this model in question-answering in the field of Chinese medical diagnosis. Without the use of other pre-training corpus and pre-training model parameters, the model can complete the parameterized embedding and derivation of triplet knowledge. The performance of MRR in the Wikipedia atlas completion task is close to that of TransE, but the number of parameters is only 20.8% of the latter. In the case of providing certain question-and-answer function, the number of parameters is limited to 50M, which greatly saves computing resources and is conducive to the expansion and update of knowledge graph.