Structure-Aware Reasoning for Knowledge Base Question Answering

Lu Ma, Peng Zhang, Xi Zhu, Dan Luo, Bin Wang

Abstract: Answering question according to knowledge base (i.e. KBQA) has attracted extensive attention recently. Information retrieval is one of the mainstream methods for the KBQA task that first finds the topic entity in the question via entity linking systems, and then selects the most related entities as answers from the subgraph (nodes in it are called candidate answers) of topic entity on the knowledge base (KB). However, existing methods generally separately perform reasoning over every candidate answer by considering the semantic relationships between question and the features extracted from KB, breaking away from the graphical structure of the KB and suffering from long-term dependency problem of entities. To address that, we propose a structure-aware reasoning method, which enables to exploit the graphical structure of entities on KB via Graph Convolutional Network and capture deep semantic relationships between question and candidate answers. Our method reasons about the correct answer by jointly considering information of all candidate answers, and focusing on important components in the question and on KB . We conduct experiments on the WebQuestions dataset, and the results demonstrate the effectiveness of our proposed method.