Research On Natural Language Question Answering Method Based On Knowledge Graph

In the question answering scenario,answering questions needs to rely on a large amount of background knowledge,and natural language question answering based on rules and text matching cannot make use of the existing large amount of human background knowledge.The traditional retrieval-based question answering based on knowledge graph is implemented in a pipelined manner,which leads to error propagation.In order to reduce accumulated errors,the method of sorting after multiple recalls is generally adopted.However,this method will lead to too many candidate paths to be recalled,which increases the difficulty of path sorting.To solve the above problems,this thesis proposes a method that combines generative path model with information retrieval.The main contributions of this thesis are as follows:（1）A traditional retrieval-based question answering method is constructed.Aiming at the problem of superposition caused by error propagation,the method proposed in this thesis used a combination of traditional methods and semantic generative methods to reduce the error deviation of the final result caused by error propagation.（2）A question answering method integrating semantic parsing and feature engineering is constructed.Aiming at the problem of generating too many candidate paths,this thesis proposed a method of first generating and then expanding the candidate query paths by using the generative model,which reduced the number of candidate paths and the unnecessary generation of candidate paths.（3）In view of the problem of insufficient training data for the generative model,this thesis used simple rules to expand the training data set,so that the generative model integrates the nodes in the knowledge graph and the connection relationship between the nodes,which was more efficient for the task of generating query paths.（4）For the problem that the candidate path sorting model is too complex,this thesis converts the candidate query paths of natural language questions into text descriptions according to certain rules,and compares the semantic information of the questions and the converted text descriptions to the candidate paths.Sorting reduced the interference of other unnecessary information and simplifies the complexity of the candidate path sorting model.The knowledge map and dataset of the open domain knowledge base question answering task in the 2021 National Conference on Computational Semantics are selected in this thesis.For this task,a new natural language question answering method based on knowledge graph is proposed.This approach complements a more general information retrieval model and a more accurate semantic parsing model.The method proposed in this thesis ranks third on the final leaderboard of the competition,achieving an F1 score of 78.52 %（78.86 % for the first place）.