Knowledge graph is one of the key technologies to realize question answering in open domain. Open domain question answering tasks often require enough knowledge information, and the incompleteness of knowledge graph becomes an important factor restricting the performance of question answering system. Existing related work utilizes external unstructured text combined with knowledge graph-based structured knowledge to fill in the missing information. At this time, the accuracy and efficiency of retrieving external texts are particularly critical, and selecting texts that are highly relevant to the problem can improve system performance. Conversely, selecting texts that are less relevant to the question will introduce knowledge noise, thereby reducing the accuracy of question answering tasks. Therefore, this paper designs a question answering system model that integrates text and knowledge graphs, in which the text retriever can fully mine the semantic information of questions and texts to improve the quality of retrieval and the accuracy of query subgraphs; Combine knowledge from text and knowledge bases to build fusion representations of knowledge. The experimental results show that the model has certain advantages in performance compared with the baseline method.