A Knowledge Graph Question Answering Approach Based on Graph Attention Networks and Relational Path Encoding

Shuxin Cao, Xiaoxu Zhu, Peifeng Li

Abstract: Most of the previous knowledge graph question answering (KGQA) methods focus on question parsing and ignore the knowledge graph information. In order to improve the utilization of knowledge graph structure information, based on the good performance of graph networks in the field of knowledge graph embedding, this paper proposes to utilize relational graph attention networks to pre-train the knowledge graph Q&A database. Entity and relation representations are obtained by fusing multi-hop information and the results are used to initialize the vector representation of nodes and edges in the graph. Inference is then performed using a multi-hop Q&A network based on relational paths. Since most current inference models do not consider path-level information, LSTM is introduced to encode paths in the model fusion process. In order to improve the reliability of inference, a multilayer attention mechanism is introduced to calculate the similarity between the paths and the original problem, so as to obtain the weight coefficients of the paths, which are fused into the relation and entity scores. Experiments demonstrate that the optimized Q&A inference method outperforms the original and baseline models on the MetaQA, WebQSP and CWQ datasets.