基于问题与关系嵌入空间对齐的知识图谱问答
KGQA based on question and relation embedding space alignment
张志远 1张静1
作者信息
- 1. 中国民航大学计算机科学与技术学院,天津 300300
- 折叠
摘要
为解决基于嵌入的知识图谱问答中,因采用不同的训练模型导致问题嵌入与知识图谱嵌入处于不同语义空间的问题,提出一种知识图谱问答模型.将关系转换为手工构造的自然语言问题,通过神经网络训练问题的嵌入表示与知识图谱的关系嵌入表示尽可能靠近;通过训练集中的问题对神经网络参数进行微调,使答案获得最高评分.在WebquestionSP数据集上的实验结果表明,相较于EmbedKGQA,所提方法的hits@1指标提高了 10.5个百分点;在缺失50％三元组的情况下hits@1指标提高了 9.9个百分点.
Abstract
To solve the problems that in embedding based knowledge graph question answering,different training models are used for the problem embedding and knowledge graph embedding,leading to the different semantic spaces,a knowledge graph ques-tion answering model was proposed.The relationships of KG were converted into handcrafted natural language questions,a neu-ral network was trained to make the handcrafted questions embedding close to the relation embedding of KG.Parameters of neu-ral network were fine-tuned by question and answer pairs in the training set.Experimental results on the WebquestionSP dataset show that,compared with EmbedKGQA,the hits@1 index of the proposed method is increased by 10.5 percentage,and in the case of missing 50％triples,the hits@1 index is increased by 9.9 percentage.
关键词
知识图谱/问答系统/知识表示学习/问题嵌入/关系嵌入/关系对齐/答案选择Key words
knowledge graph/question answering system/knowledge representation learning/question embedding/relation em-bedding/relational alignment/answer choice引用本文复制引用
出版年
2024
NETL
NSTL
万方数据