Research On Entity Alignment Method Based On Joint Robust Knowledge Graph And Attribute Fusion

With the development of the times and the development of science and technology,the amount of information from all aspects is becoming more and more large,and how to store information and obtain the intricate relationship between information has become very important.The birth of the Knowledge Graph integrates knowledge from different sources through knowledge fusion to assist data for decision-making reasoning.Entity alignment(EA),as one of the most basic and critical means in the process of knowledge fusion,aims to find the corresponding entities existing in the real world in different artificially constructed knowledge graphs.Nowadays,a large number of research mostly use supervised or semi-supervised methods to carry out entity alignment research,such as using manually labeled seed sets for entity alignment,and then getting new entity matches,and then putting them into the original seed set for training to form a semi-supervised way to expand the seed set.On the one hand,due to the wide range of information sources and the differences in manually annotated entities,errors will accumulate in the task of entity alignment,which will affect the effect of entity alignment to some extent.On the other hand,the entity alignment research method considers a single structural embedding,ignoring the favorable external condition that the knowledge graph contains attribute information.This paper conducts research on the above two challenges,and the main work as follows:1.Aiming at the difference and misleading artificial labeling,a neighborhood matching entity alignment based on Robust knowledge graphs(NMR)of a robust knowledge graph is proposed.A robust knowledge graph is to solve the error through certain methods to ensure the correctness of the seed entity pair under the condition that there is an error in the seed entity pair,and the adversarial model is used to seed the data in the existing dataset DBP15 K,which is mainly composed of two parts,one is the noise perception optimization generator.The other is a noise-aware optimization discriminator.The two iteratively train each other to a steady state and set a confidence threshold to determine whether the seed entity pair matches the real alignment,and merge the pure entity pairs produced in each iteration to provide high-quality seed entity pairs for subsequent alignments.At present,the research work on entity alignment tasks often uses neighborhood information to enhance the effect of entity alignment,so this method is used in the alignment process to carry out entity alignment work,and the heterogeneity problem of knowledge graph is alleviated by neighborhood sampling,matching and aggregation after embedding in graph convolutional neural networks.2.Aiming at the unity of structural embedding,combined with the existing research NMR model,the attribute information of entities is further integrated,and an entity alignment for Jointly Robust knowledge graph and attribute fusion(JRAF)is proposed,due to the high complexity in the process of sampling,matching and aggregating neighbor information.Therefore,on the JRAF model,we remove the highly complex calculation of neighborhood matching in NMR,and fuse with the attributes to obtain good experimental results,which is improved again compared with the NMR model.3.To evaluate the model,three cross-lingual knowledge graph datasets from DBP15 K were used.Compared with the existing classical entity alignment research model,the efficiency and robustness of the two experimental model methods proposed in this paper are proved,and the effect of entity alignment is improved.