具有关系敏感嵌入的知识库错误检测

Knowledge base error detection with relation sensitive embedding

准确性与质量对于知识库而言尤为重要,尽管已经有很多关于知识库不完整性的研究,但是很少有工作者考虑到对于知识库存在的错误进行检测,按照传统方法通常无法有效捕捉知识库中错误事实内在相关性。本文提出了一种知识库具有关系敏感嵌入式方法NSIL,以获取知识库各关系之间的相关性,从而检查出知识库中的错误,以此提高知识库的准确性与质量。该方法分为相关性处理和错误检测两阶段。在相关性处理阶段,使用NSIL的相关函数以分值形式获取各关系之间的相关度;在错误检测阶段,基于相关度分值进行错误检测,对于缺失主体或客体的三元组进行缺失成分预测。最后在知识库之一Freebase生成的基准数据集"FB15K"上进行了广泛验证,证明了该方法在知识库错误知识检测方面有着很高的性能。

Accuracy and quality are very important for the knowledge base.Although there have been many researches on the incompleteness of knowledge base,few workers consider the detection of errors in the knowledge base.According to the traditional methods,it is usually unable to effectively capture the internal correlation of errors in the knowledge base,so as to check the errors.In this paper,a relational sensitive embedded method NSIL for knowledge base is pro-posed to obtain the correlation among the relationships between them,so as to check out the errors in the knowledge base,so as to improve the accuracy and quality of the knowledge base.This method is divided into two stages:correla-tion processing and error detection.In the correlation processing stage,correlation function of NSIL is used to obtain the correlation degree of each relationship in the form of score;in the error detection stage,error detection is based on the score of correlation degree,and missing component prediction is carried out for the triplet of missing subject or object.At last,the method is verified on the benchmark data set"FB15 K"which is generated by Freebase,one of the largest knowledge bases.It is proved that the method has high performance in knowledge base error detection.