Research On Cardinality Constraint-based Filtering Methodology In Knowledge Representation Learning

Knowledge graphs(KG)are semantic networks that express relationships between entities.Nodes represent entities,and directions represent relationships between entities.They are often expressed in the form of triples,denoted as(head entities,relationships,Tail entity).In recent years,knowledge graphs have been widely used in various fields of artificial intelligence(AI).However,since the knowledge in reality is expanding over time,the completion of the knowledge graphs is also a dynamic process.Therefore,the knowledge graphs lack a large number of triples and is incomplete.The knowledge representation learning method embeds the entities and relationships in the knowledge graphs into dense low-dimensional real-value vectors,which can effectively predict the missing triples in the knowledge graphs.Normal triples are usually only stored in the knowledge graphs.However,negative triples are equally important in knowledge representation learning.Usually,random negative sampling is used to generate negative triples,but most of the generated negative triples are simple triples.After several epochs,simple negative triples have smaller contributions for training,or even almost no contribution.In order to alleviate this problem,some researchers have proposed a negative triplet screening method based on enhanced confrontation learning,KBGAN screening method,which can automatically generate high quality negative triples.However,this method has a serious "false negative example" problem.The so-called "false negative example" means that the negative triplet generated by the screening method is actually a positive triplet already stored by the knowledge graphs.The over-training of “fake negative examples” will affect the knowledge representation learning model obtained by training.At the same time,the KBGAN screening method does not distinguish the relationship types of the triples,and the ratio of the "false negative cases" generated by the triples of different relationships is different.Aiming at the above problems and phenomena,this paper proposes a method for different screening strategies based on the negative triplet screening method of knowledge representation learning--the cardinal constraint screening method ECCS.The specific method is as follows:Firstly,this paper proposes a method for determining the type of triples,which determines and classifies the relationship types of the triples in the training dataset of the knowledge graphs,and marks the type information in the triplet data.Subsequently,this paper proposes two types,difficult methods of screening considering cardinal constraints,ECCS-T method and ECCS-S method.The basic idea of the ECCS-T method is to reduce the probability value of the most likely "false negative" by automatically adjusting the probability distribution,thereby reducing the possibility that the "false negative" is screened;the basic idea of the ECCS-S method is to add a filter to remove the most likely "false negatives" through the filter to alleviate the problem of filtering the negative triples as "false negatives".On the three datasets of FB15k237,WN18 and WN18 RR,the screening method considering the cardinality constraint in the knowledge representation learning proposed in this paper--the ECCS method and the KBGAN method are compared.The results show that the ECCS-S method is on FB15k237 and WN18.The experimental results are better than ECCS-T and are better than the KBGAN method.The ECCS-T results on WN18 RR are slightly better than ECCS-S.Therefore,the ECCS method effectively alleviates the problem of over-training of “false negative cases” compared with the KBGAN method.