Research On Representation Learning Method Of Noisy Knowledge Graph

Knowledge graphs are used to store structured facts,which are represented in the form of triples,namely(head entity,relation,tail entity).The construction and updating of large-scale knowledge graphs usually adopt(semi)automatic methods for knowledge extraction,and noise will inevitably be introduced in the process.However,most traditional knowledge representation learning methods assume that the triples in the knowledge graph are correct,and represent the knowledge in the distributed manner.Therefore,the noise detection of knowledge map is very important.In addition,even large-scale knowledge graphs can not completely cover all the knowledge in reality,which leads to the problem of incompleteness.The noise and incompleteness of the knowledge graphs will adversely affect the downstream application tasks,so it has become an urgent problem to be solved.In order to solve the above problems,based on the translation model,this paper proposes a knowledge representation learning method based on the perception of difference and support by combining the triple structure information,entity hierarchical type information and relations path information.On this basis,in order to solve the problem that relation path information plays a negative role in the noisy knowledge graphs,this paper proposes a knowledge representation learning method based on the combination of logical rules and relation path information.In addition,based on the above two models,a knowledge graph noise detection prototype system is designed to assist users to judge the knowledge.The main contents and innovations of this paper are as follows:(1)To address the above-mentioned problems of knowledge graphs,this paper proposes a novel knowledge representation learning framework for noisy knowledge graphs which combines entity hierarchical type information with relation path information.It can generate noise-free knowledge representations while detecting the possible noise in knowledge graphs.Specifically,the model is divided into two parts: a triple dissimilarity estimator and a triple support estimator.The triple dissimilarity estimator,based on the translation model,generates the matching extent between entities and relations in triples by combining the triple structure information,entity hierarchical type information and relation path information of knowledge graphs.The triple support estimator,by further using these information,makes judgements on the matching extent of entities and relations.The two estimators are combined to form a nested judgement of the triple to measure whether the triple contains noise.This research has been verified by experiments on public datasets,and the results show the effectiveness of the proposed method.(2)Through the further research,it is found that the relationship path information will have a negative impact on the noise detection task on the noise knowledge map.To address this problem,this paper proposes a knowledge representation learning framework of noise knowledge graphs that fuses logic rule information and relation path information.It follows the method proposed in Chapter 3 and is improved on this basis.It can improve the accuracy of path inference by introducing logic rule information to guide the synthesis of relation paths and using the accuracy of logic rules.Meanwhile,the lack of interpretability of relation path inference is well compensated by the interpretability of logic rules.The experimental results show that both models proposed in this paper outperform the baseline methods in all metrics.Experimental results on public datasets show the effectiveness of the proposed method.(3)We design a prototype knowledge graph noise detection system based on the above two models,which can assist users in making judgements about knowledge.The input of the prototype system is a triple and the output is the score of the triple.The likelihood of a triple being noisy is indicated by the triple score.The prototype system can be used for knowledge-driven applications to reduce the impact of noisy knowledge.