Research On Cross-Lingual Knowledge Graph Entity Alignment Based On Pseudo-Siamese Network

With the continuous development of Q&A system,decision support,semantic search and personalized recommendation,the demand for large-scale unified multilingual knowledge graph is becoming more and more urgent.Therefore,more and more studies begin to focus on the integration of cross-lingual knowledge graphs.Cross-language knowledge graph entity alignment aims to identify common entities between two knowledge graphs constructed by different languages,which is an effective method to realize cross-language knowledge graph fusion.The current research on entity alignment mainly focuses on the three views of the text,relationship and structure of the knowledge graph,and there are few studies on the attribute view.Because the feature extraction process of attribute view is easily interfered by noise attributes and entity common attributes,and the information of attribute view is more dispersed,and the composition of attribute and attribute value text is more complex,it will lead to insufficient feature extraction of attribute view.In addition,for the features obtained from multiple views,simple stitching cannot effectively combine the feature advantages of each view,and even reduce the entity alignment effect of the original features.In view of the above problems,this paper designs a cross-language entity alignment model based on attribute view and fusion of multiple views.The main research work is as follows:(1)Aiming at the problem of insufficient feature extraction of interference attributes and attributes in the feature extraction process of attribute view,the attention-based crosslanguage entity alignment model of deep bidirectional gated recurrent network(AB-GRU)is constructed.The model initializes the name text of attribute and attribute value by pretraining model,and improves the representation effect of attribute and attribute value text.The global and local attention mechanisms are designed to screen out noise and general attributes,and reduce the interference of noise attributes on subsequent feature extraction;we construct a deep bidirectional gated recurrent network to aggregate decentralized attribute view features,reduce the impact of the location information of the input sequence,and extract the deep features of the attribute view.(2)Aiming at the problem of low quality of multi-view feature fusion in cross-language knowledge graph,a multi-view cross-language entity alignment model based on pseudosiamese network(PSN-MEA)is constructed.The model introduces a multi-layer graph convolution network on the basis of AB-GRU to extract the features of the corresponding knowledge graph structure view of the entity,and builds a highway network between two adjacent graph convolution networks to refine the name features and structural features of the entity.The constructed pseudo-siamese network can train different deep neural networks by constructing a unified objective function,so that the multi-view features obtained by learning can be better distinguished.The design of benchmark multi-view fusion strategy further improves the effect of entity alignment by configuring different weights to fuse the characteristics of three views,text,structure and attribute.The experimental results on ZH＿EN,JA＿EN and FR＿EN of three subsets of DBP15 K in cross-language public dataset show that AB-GRU effectively improves the feature extraction effect of attribute view.On the basis of retaining the advantages of AB-GRU,PSNMEA effectively integrates multi-view features.Compared with the entity alignment results of 11 mainstream models,PSN-MEA is superior to other models in three evaluation indexes of Hits@1,Hits@10 and MRR.