Multi-relation Question Answering Based On Knowledge Graph Subgraph Fusion

With the gradual popularization of intelligent question answering technology in life,people's requirements for question answering systems are also increasing.How-ever,the traditional question answering system extremely relies on handwriting rules and templates with the problems of poor generalization performance and poor practi-cability.To improve the performance of the question answering system,knowledge reserves and the ability of knowledge reasoning are required.Therefore,it is a feasible method to supplement knowledge graph as external knowledge to question answering system.Knowledge graph is an efficient form of knowledge representation.It has a wide range of applications in the fields of search and recommendation,which can bring con-siderable performance improvement to question answering systems.Aiming at the re-search direction of knowledge graph question answering,this paper has mainly com-pleted two aspects of work,namely,the Chinese knowledge graph question generation method based on the pre-training model and the multi-relation question answering based on knowledge graph subgraph fusion.Most of the existing Chinese knowledge graph question answering corpora are small and of poor quality.Therefore,this paper proposes a method for generating Chi-nese knowledge graph problems based on pre-trained models.The architectural idea of this method comes from the conditional variational autoencoder.Based on this idea,a core architecture is designed in this paper.The architecture uses the pre-trained model BERT for pre-encoding,and uses the Transformer model to build the source-encoder and decoder.In addition,the method also combines and improves the answer coding method.We uses the NLPCC2017 KBQA data set for experiments.The results shows that the model has a significant improvement on the BLEU,ROUGE and human evalu-ation indicators compared with the baseline model and can generate more diverse prob-lems,proving the effectiveness of the method.This paper uses this method to construct a Chinese knowledge graph question and answer data set as one of the experimental data sets of the multi-relation question answering method in this paper.Multi-relation question answering is a task in which there are two or more triples in a question.The complexity is high and there are few related studies.Most studies have not considered the subgraph information of the triples involved in the problem.Therefore,this paper proposes a multi-relation question answering method based on knowledge graph subgraph fusion.This method improves knowledge representation through knowledge graph point-edge relation and subgraph structure.Combined with interpretable inference networks,it has the knowledge inference ability.The experiment was conducted on the real datasets PathQuestion and WorldCup2014.The experimental results show that this method has a certain improvement over other baseline models.On the other hand,a part of the baseline model and this experimental model were selected for an experimental comparison on the Chinese knowledge graph question answering dataset generated by the above problem generation model.The experimental results show that the multi-relation question answering method based on the subgraph fusion also has higher accuracy in the Chinese dataset.