Multilingual knowledge graph completion without aligned entity pairs

The goal of multilingual knowledge graph completion (MKGC) is to improve link prediction performance on the target knowledge graph by leveraging data from other language-specific knowledge graphs. Existing methods usually use pre-aligned entities between different knowledge graphs to accomplish knowledge transfer. However, there are usually no pre-aligned entities between different knowledge graphs in practical scenarios, making knowledge transfer difficult to achieve. Considering the MKGC without aligned entity pairs, a pseudo-aligned entity generation module that integrates a pre-trained language model is proposed to iteratively generate new aligned entities for knowledge transfer. It is suggested to use a graph neural network based on multi-graph attention (MGA-GNN) to encode the triples in order to differentiate the information in various language-specific wisdom graphs. Finally, the plausibility of the triples is calculated via the embeddings output by the network to conduct the link prediction task. Experimental results on the DBP-5L and E-PKG datasets show the effectiveness of the proposed method and its superior performance in more practical scenarios.