Leveraging knowledge graph embeddings for natural language question answering

A promising pathway for natural language question answering over knowledge graphs (KG-QA) is to translate natural language questions into graph-structured queries. During the translation, a vital process is to map entity/relation phrases of natural language questions to the vertices/edges of underlying knowledge graphs which can be used to construct target graph-structured queries. However, due to linguistic flexibility and ambiguity of natural language, the mapping process is challenging and has been a bottleneck of KG-QA models. In this paper, we propose a novel framework, called KemQA, which stands on recent advances in relation phrase dictionaries and knowledge graph embedding techniques to address the mapping problem and construct graph-structured queries of natural language questions. Extensive experiments were conducted on question answering benchmark datasets. The results demonstrate that our framework outperforms state-of-the-art baseline models in terms of effectiveness and efficiency.