Research On Co-Training Of Internal And External Semantics For Entity Alignment

With the development of Knowledge Graph(KG),various knowledge graphs have emerged at home and abroad.Scholars can enhance the data coverage of one KG by integrating multiple KGs from different data sources into it.As an important sub task of KG fusion,entity alignment can find entities referencing the same real-world object in two KGs,which plays an important role in automatically integrating multiple KGs.However,due to the complexity of entity information in different KGs,the traditional entity alignment methods are difficult to use;Secondly,due to the emergence of more and more specific language KGs,there are great multilingual differences between KGs;Finally,due to the huge volume of KG and the sparsity of training data which is marked artificially,it is difficult to achieve the expected training effect.Therefore,the crosslingual entity alignment is a necessary but challenging task.This paper mainly studies the cross-lingual entity alignment task,and analyzes the current researches at home and abroad,proposing a promising entity alignment algorithm by KG embedding based on co-training of internal and external semantics,and uses this algorithm to build a visual prototype system of cross-lingual entity alignment.The main contents of this paper are as follows:We propose an entity alignment algorithm,im-GCN-align,based on relation-aware GCN with simplified internal semantics,which simplifies the attribute characteristics of entities,and improves the connectivity matrix of graph,and mines the complete internal semantics of entities,and then improves the entity alignment effect from the perspective of internal semantics.We propose an entity alignment algorithm,BRCEA,by KG embedding based on co-training of internal and external semantics,which uses im-GCN-align and RDGCN to fully mine the internal and external semantics of entities respectively,and alleviates the problem of insufficient training data through co-training in order to improve the effect of entity alignment with limited training data.We construct a visual prototype system of entity alignment,which applies the entity alignment results of the above research,BRCEA algorithm,to DBP15k datasets to provide multilingual search services in Chinese,Japanese,English and French.