Research On Multi-modal Cross-lingual Entity Alignment

Although the development of knowledge graphs is becoming more and more mature,entity alignment has become one of the indispensable technologies in knowledge graph research with the emergence of a large number of heterogeneous knowledge graphs.The diversity of languages,uneven quality,and different data modalities are the main reasons for the heterogeneity of knowledge graphs.Therefore,cross-language entity alignment technology is needed to integrate heterogeneous knowledge graphs to form a large-scale,high-quality,multi-language,and multi-modal knowledge graph,which can provide strong support for related research based on knowledge graphs.Even if the existing methods have achieved good results in the cross-language entity alignment task,there is still an issue that the equivalent entities produce noise in the alignment process,which makes it difficult to improve the accuracy of entity alignment.Therefore,this paper proposes a neighbor entity screening rule to remove redundant equivalent neighbor entities,which can ensure the quality of aligned entities.At the same time,in order to reduce the interference of noise,the entity screening rules are further optimized,and the cross-language entity alignment is realized by combining the entity descriptions and the entity image.Compared with the existing cross-lingual entity alignment methods on the public data set,the experimental results show that the proposed method effectively improves the effect of entity alignment.The main research work in this paper is as follows.1)This paper designs a cross-language entity alignment method based on dual relation graphs and neighbor entity screening.Aiming at the problem of equivalent neighbor entities,this paper proposes a neighbor entity screening rule based on entity names and attributes to delete redundant equivalent neighbor entities.At the same time,in order to avoid the problem that entity attributes may be insufficient and bring trouble to entity screening,the dual relation graph is constructed as auxiliary evidence to complete entity screening,and the dual relation graph is used to strengthen the close connection between entities,so that the neighbor information of entities can be fully utilized.2)This paper designs an optimized neighbor entity screening method based on entity hypothesis replacement.In the real corpus,the attribute information of two entities with the same name may be similar or have inclusion relationships,usually,the two entities are considered equivalent,but they may not be equivalent in fact.Therefore,based on this situation,this paper proposes an optimized neighbor entity screening strategy.This strategy assumes that two entities are equivalent,and replaces their positions in knowledge graph,and then combines the information of neighbor entities to determine whether the structure of the knowledge graph after the replacement entity is valid.Finally,it compares the similarity of the neighborhood feature vectors of the entities before and after the replacement to determine whether the two entities are equivalent.3)This paper designs a cross-language entity alignment method based on entity descriptions and images.The Pulse Coupled Neural Network is used to extract the implicit relations in entity descriptions,and then the entity names and attributes are combined to interact with the pre-aligned entity pairs iteratively to achieve the preliminary alignment.This method uses the image feature extraction method to extract image feature knowledge,and realizes dimensionality reduction through image feature projection,and then combines it with entity descriptions feature vector.The relation aggregation network is used to aggregate the neighbor output feature vector of each layer of the entity.Then,the pre-aligned entity pairs are used to guide the alignment,and whether the two entities can be aligned is judged by calculating the similarity of the output feature vectors of the two entities.Finally,the two alignment results are compared,and the minimum value is taken as the final alignment result.