Research On Knowledge Gragh Reasoning Method For Question Answering

Knowledge question answering is a classic research content in natural language processing.With the development of knowledge graph technology,multi-hop knowledge question answering for knowledge graph has become a hot topic.Due to the lack of links in the knowledge graph,in the multi-hop question answering,it is necessary to reason on the relation query path of the knowledge graph to infer the hidden relation.Complete the completion of the knowledge graph while answering multi-hop questions.Knowledge graph embedding is an effective method to solve knowledge reasoning,but traditional embedding methods do not incorporate the entity semantic information in the reasoning path,and the existing neural network model does not fully consider the context characteristics of the path sequence in the process of extracting reasoning path features.Therefore,it is difficult to cover the complicated and hidden information of the reasoning path.In response to the above problems,this paper proposes an reinforce path reasoning model,and improve the query path reasoning effect by learning the complex context information and sufficient path local information in the query path.The main work of the thesis is as follows:(1)This paper proposes a weight distribution knowledge graph embedding model based on the neighbor graph network.According to the query path,the entity may have potential feature information for auxiliary path reasoning.The entity uses an aggregator to assign weights to its neighbor node information in the knowledge graph,use the attention mechanism to improve the accuracy of feature extraction in both coarse-grained and fine-grained.At the same time,this model can make full use of the entity information in the query path,improve the feature representation of the query path,and improve the accuracy of reasoning.(2)This paper proposes a Transformer-based Multi-hop Question Answering Reinforcement Path Reasoning Model.By analyzing the structural information of the path sequence,considering that paths of different lengths have different effects on the contextual semantic feature extraction at different distances,combined BI-GRU and Transformer to embed the query path,and fully extract the path sequence each element local and overall contextual characteristics through the multi-head attention mechanism.While combining the path entity information,the improved semantic representation of the stride path is used to enhance the representation of the weight distribution of entities and relationships in the path,and reduce the risk of reducing the inference effect due to the inclusion of too much feature information in the reasoning model.(3)By experimenting the proposed method on the query path data sets FB15K-237 and WNRR18,and comparing it with other path query reasoning methods,the validity of the Transformer reinforcement path reasoning model based on the neighbor graph network proposed in this paper is verified.