Graph embedding based query construction over knowledge graphs

Graph-structured queries provide an efficient way to retrieve the desired data from large-scale knowledge graphs. However, it is difficult for non-expert users to write such queries, and users prefer expressing their query intention through natural language questions. Therefore, automatically constructing graph-structured queries of given natural language questions has received wide attention in recent years. Most existing methods rely on natural language processing techniques to perform the query construction process, which is complicated and time-consuming. In this paper, we focus on the query construction process and propose a novel framework which stands on recent advances in knowledge graph embedding techniques. Our framework first encodes the underlying knowledge graph into a low-dimensional embedding space by leveraging the generalized local knowledge graphs. Then, given a natural language question, our framework computes the structure of the target query and determines the vertices/edges which form the target query based on the learned embedding vectors. Finally, the target graph-structured query is constructed according to the query structure and determined vertices/edges. Extensive experiments were conducted on the benchmark dataset. The results demonstrate that our framework outperforms several state-of-the-art baseline models regarding effectiveness and efficiency.