Research On Cross-lingual Entity Alignment Technology Based On Bootstrapping Learning And Multi-perspective Learning

The construction and application of multilingual knowledge graphs(KGs)have promoted the development of many artificial intelligence-related cross-lingual tasks.The cross-lingual entity alignment task refers to the task of finding the matched entity pairs in different languages among multilingual KGs.Through cross-lingual entity alignment,KGs of different languages can be connected and fused to form new knowledge and increase the information density.Existing methods mainly rely on a small number of cross-lingual links and triples(subject-predicate-object)to learn entity embeddings.With the development of the Internet and the maturity of crowdsourcing technique,a large number of knowledge graphs also contain rich entity descriptions,providing conditions for text information encoding.In order to address the problems of the structural embedding,the textual coding,and the scarcity of the alignment data,this paper proposes two corresponding cross-lingual entity alignment models.The main contributions of this thesis are as follows:(1)This thesis proposes a bootstrapping model based on TransD for cross-lingual entity alignment,which incorporates TransD to add interaction between entities and relations,and optimize the embeddings of them through the loss of triples.For the problem of insufficient seed entity alignments,it uses the bootstrapping method to iteratively add new entity alignments obtained in the training process to expand training data.The results on DBP15 K prove the advantages of this model in structural encoding and the effectiveness of the bootstrapping process.(2)This thesis proposes a multi-perspective cross-lingual entity alignment model via graph convolutional networks(GCNs).This model trains two-layer GCNs to encode structural embedding and textual embedding based on triples and entity descriptions.The two perspectives are then combined to calculate entity alignments.For further performance improvement,it uses machine translation and long short-term memory network to encode cross-lingual entity descriptions.The ablation study of textual embedding module and the experiments on DBP15 K prove the superiority of this model in structural encoding and textual encoding,and the effectiveness of text description encoding for the entity alignment task.