Research On Cross-lingual Knowledge Graph Alignment And Fusion

The concept of knowledge graph was proposed in 2012.As knowledge graph can accurately reflect facts in real world,knowledge graph has been applied in many fields in recent years,and large amount of researches have been carried out around it.Knowledge graph embedding model was put forward in 2013,and has been developing rapidly in these years.Researches in this field are still at the initial stage at present.Existing models are built based on relatively intuitive thoughts,whose rationality still needs to be improved,and the characteristics of embedded vectors are not fully utilized.This research focuses on the definition,training and application of multi-lingual knowledge graph embedding and alignment model,and studies cross-lingual knowledge graph alignment and fusion.On the basis of previous work,we propose a cross-lingual knowledge graph embedding and alignment model,and detailedly state the idea,knowledge model,alignment model and training process of the model.This model embeds multi-lingual knowledge graphs into low-dimensional real vector spaces,and uses a semantic vector and a space vector to depict each entity or relation,and constructs alignment model across languages via these vectors.Compared with knowledge expressed in discrete form in knowledge graph,vectors have continuity,thus can be used to model the correlations between multi-lingual entities and relations by the spatial relationships of vectors.On the basis of the model,we put forward a method of cross-lingual knowledge graph alignment and fusion,so that the model can be applied to cross-lingual knowledge alignment,and can represent multi-lingual knowledge graphs reasonably.From the need of research,aiming at the problem of alignment data auto-labeling in data preprocessing,we put forward an engineering method which builds cross-lingual entity links based on Wikipedia first,and then eliminates unreliable links based on diagraph decomposing method.By comparing with the existing cross-lingual knowledge graph embedding models,and comparing the different scoring functions in the model,the model is verified and analyzed.Experimental results show that our model can achieve better results than existing cross-lingual knowledge graph embedding models.