Improve embedded knowledge graph multi-hop question answering based on relational path

Abstract

Multi-hop Knowledge Graph Question Answering (KGQA) requires reasoning about multi-hop inference relations between topic entities and answers on the knowledge graph(KG) and returning correct answers. The difficulty of obtaining the implied inference relations in multi-hop questions described in natural language and the sparse knowledge graph bring challenges to the multi-hop KGQA. In this paper, we propose an embedded knowledge graph multi-hop Q&A model based on relational paths, which exploits the relational chains in the knowledge graph and the semantic similarity of multiple questions to improve the accuracy of multi-hop KGQA task. The comparative experiments prove that our method significantly outperforms the state-of-the-art counterparts. Comprehensive ablation experiments also validate the effectiveness of our approach on multi-hop KGQA tasks.