Research And Implementation Of Entity Alignment In Heterogeneous Knowledge Fusion

In recent years,the number of knowledge graphs and their applications has increased sharply.The knowledge graph of single data source is difficult to meet the diverse requirements of knowledge graph applications.Knowledge fusion can integrate the overlapping and complementary knowledge in multi-source heterogeneous knowledge graphs,and build a large-scale knowledge graph from the top.Entity alignment aims to judge whether entities from different knowledge graphs point to the same object in the real world,which is an important task in heterogeneous knowledge fusion.The entity alignment method based on knowledge embedding is a more efficient alignment method at present.This thesis aims at the problem that the integration of relation information in the current embedding-based entity alignment method is not comprehensive,ignoring the semantic information and direction information of the relation,as well as the problem of the poor alignment effect caused by the difference of original graph structure in entity alignment of different knowledge graphs.This thesis makes the following research:(1)Aiming at the problem of incomplete fusion of relation information in the current embedding-based entity alignment methods,ignoring the semantic information and direction information of the relation,this thesis proposes a context entity alignment model based on alignment relation(RA-CEA).Firstly,the model uses the word vector dictionary provided by Glo Ve to construct the initial vector representation of entity name and relation name,and then it carries out embedded learning and alignment calculation for relation and entity respectively.In the stage of relation embedding and alignment,the improved graph convolution neural network is used to embed and learn the relation,and the distance measurement formula and loss function are used to train the model.Finally,manual participation is revolved in the relation alignment task to provide reliable data support for the follow-up work;In the stage of entity embedding,the entity is also embedded in learning;In the stage of entity alignment,the entity is classified by using the alignment relation,the context entity set of alignment relation is generated,and the model training for entity alignment is carried out in the set.Finally,experiments are carried out on the same data set with other models to verify the effectiveness of the RACEA model proposed in this thesis.(2)Aiming at the poor alignment effect caused by the difference of graph structure between the knowledge graphs to be aligned,this thesis proposes a method based on cross-knowledge graph completion to balance the structural differences between the two knowledge graphs.In the case of ensuring that the entity domain and relation domain of a single knowledge graph are closed,firstly,the aligned entity pairs in the training samples are matched and learned in the cross-knowledge graph to supplement some missing triples;Then a rule mining algorithm is used to summarize the rules in a single knowledge,and the rules are reasoned within the knowledge graph.At the same time,we design constraint rules to transfer and reason the summarized rules in the crossknowledge graph,so as to further complete the triples.In the alignment-oriented entity embedding stage,the attention mechanism is introduced to assign different weights to different neighbor entities,so that the model focuses on important neighbor entities during the embedding process.Through the experimental comparison and analysis of the model,it is verified that improving the difference of the knowledge graph structure and introducing the attention mechanism can effectively improve the effect of entity alignment.