Abstract: We constructed a three-stage representation learning model by combining convolutional neural network and transformer encoder with meta-learning as the core idea. In order to express the interaction between  entities and task relations in the reference set,  we used convolutional neural network to obtain relation-meta, applied the transformer encoder to enhance the entity representation in query set, and  designed a processor for calculating matching score of incomplete triples to  solve the problem of few-shot knowledge graph completion, i.e., the phenomenon that the large-scale knowledge graph was sparse, and the number of entity pairs corresponding to the long-tail relations with low frequency was large. The experimental results on the NELL-One and Wiki-One datasets show that the proposed model performs well  in predicting head and tail entities corresponding to long-tail relations in large-scale knowledge graphs, and can achieve efficient feature representation generation and missing entity completion for entities and relations in knowledge graphs.

Key words: knowledge graph completion, meta learning, Transformer encoder, convolutional neural network, knowledge graph embedding