Complex Knowledge Base Question Answering via Structure and Content Dual-Driven Method

Jinhao Zhang, Lizong Zhang, Jinchuan Zhang, Yichen Xin, Xu Zheng

Abstract: Benefiting from the extensive availability of large-scale knowledge graphs and the improvements in structural methods of question-answering systems, knowledge-based question answering (KBQA) has attracted much attention in the application field of knowledge base. Semantic-parsing-based KBQA methods perform question-answering tasks by executing constructed query graphs on the graph database. However, current semantic-parsing-based KBQA methods still face two non-negligible drawbacks: (1) The semantic-parsing-based methods are usually composed of independent structure prediction and content filling phases, where the structure prediction step lacks effective analysis and utilization of the filled contents. (2) It is still manpower-intensive to convert complex question query labels into equivalent schema structures. To address the challenges above, we propose a structure-and-content dual-driven KBQA method, Mutual Optimization of comBining detaIled and strUctural graphS (MOBIUS), which effectively couples structure prediction with content filling and supports each other to predict the SPARQL trees deconstructed by the official library. The newly generated structure fills in the new data, and the newly added data assists in the generation of the new structure. Then SPARQL trees are used to predict SPARQL queries. Extensive experiments on two datasets demonstrate that our model has achieved good results totally.