Abstract: In order to improve the efficiency of information retrieval in power engineering cost data, this paper designs a cost information question answering system based on knowledge graph, named entity recognition and question intention recognition techniques. This system takes the budget quota and historical power grid cost documents as the data source, and generates the knowledge graph through entity and relationship extraction and structured data transformation. The BERT-Bi-LSTM-CRF model is constructed for entity recognition of the user input question text, and the BERT＿TextCNN text classification algorithm is used to capture the semantic information and feature patterns in the question, thus to realize the question intention recognition. An answer retrieval module is designed to query the knowledge graph and generate the exact answer. Experiments show that the accuracy of the proposed entity and intention recognition methods can reach 90.2% and 88.2% respectively, and the overall performance is superior to the existing algorithms. The corpus test proves that the system can return accurate answers to user input questions, which can provide reference for the cost information management and cost estimation in power engineering.