基于对比训练图嵌入的知识图谱多跳问答  

Multi-hop knowledge graph question answering based on contrastive learning graph embeddings

作　　者：陈秋元 李梁[1] CHEN Qiu-yuan;LI Liang(College of Computer Sciences and Engineering,Chongqing University of Technology,Chongqing 400054,China)

机构地区：[1]重庆理工大学计算机科学与工程学院,重庆400054

出　　处：《计算机工程与设计》2024年第11期3434-3440,共7页Computer Engineering and Design

摘　　要：在不完全知识图谱上进行问答的研究中,现有方法通过知识图谱嵌入与问题嵌入的联合训练对问题进行推理时,多跳问题上的性能比单跳问题上的性能有显著下降。为缓解以上问题,提出一种使用对比训练的知识图谱嵌入提高模型节点预测能力,结合关系链筛选答案的方法。使用对比训练图嵌入和问题嵌入联合训练,得出候选答案并筛选出正确答案。在Meta-QA数据集上验证该方法的有效性,在不完全知识图谱上对比基准模型Embed-KGQA有6.4%的3-hop准确率提升。In the research of question answering on incomplete knowledge graphs,existing methods use the joint training of knowledge graph embedding and problem embedding to infer problems,and the performance on multi hop problems is significantly reduced compared to that on single hop problems.To alleviate the above problems,a method of embedding knowledge graphs through comparative training was proposed to improve the prediction ability of model nodes,and to filter answers through relationship chains.Comparative training graph embedding and question embedding were used to jointly train,identify candidate answers,and filter out correct answers.The effectiveness of this method is verified on the Meta-QA dataset.Compared with the benchmark model Embed-KGQA,the accuracy of 3-hop is improved by 6.4%on the incomplete knowledge graph.

关 键 词：多跳问答 不完全知识图谱 对比训练 知识图谱嵌入 关系链 链路预测 语言模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]