Answering Complex Questions on Knowledge Graphs

Xin Wang, Min Luo, Chengliang Si, Huayi Zhan

Abstract: The topic of knowledge-based question answering ( $$\mathsf {KBQA}$$ ) has attracted wide attention for a long period. A series of techniques have been developed, especially for simple questions. To answer complex questions, most existing approaches apply a semantic parsing-based strategy that parses a question into a query graph for result identification. However, due to poor quality, query graphs often lead to incorrect answers. To tackle the issue, we propose a comprehensive approach for query graph generation, based on two novel models. One leverages attention mechanism with richer information from knowledge base, for core path generation and the other one employs a memory-based network for constraints selection. The experimental results show that our approach outperforms existing methods on typical benchmark datasets.