Few-shot entity linking prediction based on Graph-Transformer network

In the context of small sample scenarios, entity linking prediction aims to infer missing entities in query triples using a few reference triples. Nevertheless, the graph structure information of nodes in the entity encoding step is disregarded by popular entity linking prediction techniques. To address this issue, a Graph-Transformer network (GTNet) is proposed. In order to improve entity representation, we first create a structure-aware graph pooling layer that learns and fuses node graph structure information. The entity pair embeddings are then created by concatenating the head and tail entities, and their prototype embeddings are obtained by projecting reference entity pairs into a semantic prototype space. Finally, we calculate the similarity between the entity pair embeddings of the query and the reference entity pair prototype embeddings, and use this similarity as the link prediction score. Experiments on the NELL-One and Wiki-One datasets show that our proposed model outperforms the best baseline models by 0.012, 0.015, 0.028, 0.023 and 0.012, 0.05, 0.033, 0.031 in terms of mean reciprocal ranking (MRR), Hits@10, Hits@5, and Hits@1 metrics, respectively. This shows that by mining the graph structure information of nodes, our model may improve the entity representation ability, effectively predicting missing entities in triples and demonstrating improved generalization.