关键词: 知识图谱问答, 实体链接, 答案推理, 图神经网络

Abstract: Knowledge graph question answering technologies are widely used in dialogue systems, recommendation systems, and other intelligent services, which explicitly represents and models knowledge through graph structures to achieve entity linking and answer reasoning with good controllability and interpretability. However, existing entity linking methods suffer from low accuracy and the inability to fully utilize neighbor information. Meanwhile, existing answer reasoning methods lack effective information encoding mechanisms that are specifically tailored to both the question and the knowledge graph. To tackle these challenges, an entity linking method based on graph convolutional networks and relation matching aggregation is proposed. The method can aggregate neighbor information without relying on external tools. Additionally, an answer reasoning method is designed based on shared encoding and collaborative attention to facilitate efficient matching between the query and knowledge graph information. Unlike traditional methods that focus solely on answer reasoning tasks and implement entity linking based on external tools, the proposed knowledge graph question answering method can handle both entity linking and answer reasoning tasks simultaneously. Experimental results demonstrate that the proposed method outperforms traditional models. Furthermore, importance analysis experiments reveal the significance of knowledge graph information for entity linking and answer reasoning tasks.

Key words: knowledge graph question answering, entity linking, answer reasoning, graph neural networks