Tackling Solitary Entities for Few-Shot Knowledge Graph Completion

Few-Shot Knowledge Graph Completion (FSKGC) aims to predict new facts for relations with only a few observed instances in Knowledge Graph. Existing FSKGC models mostly tackle this problem by devising an effective graph encoder to enhance entity representations with features from their directed neighbors. However, due to the sparsity and entity diversity of large-scale KG, these approaches fail to generate reliable embeddings for solitary entities, which only have an extremely limited number of neighbors in KG. In this paper, we attempt to mitigate this issue by modeling semantic correlations between entities within an FSKGC task and propose our model YANA (You Are Not Alone). Specifically, YANA introduces four novel abstract relations to represent inner- and cross- pair entity correlations and construct a Local Pattern Graph (LPG) from the entities. Based on LPG, YANA devises a Highway R-GCN to capture hidden dependencies of entities. Moreover, a query-aware gating mechanism is proposed to combine topology signals from LPG and semantic information learned from entity's directed neighbors with a heterogeneous graph attention network. Experiments show that YANA outperforms the prevailing FSKGC models on two datasets, and the ablation studies prove the effectiveness of Local Pattern Graph design.