Few-Shot Knowledge Graph Completion with Data Fusion and Augmentation

Abstract

This paper addresses the few-shot knowledge graph completion problem, which aims to infer facts for long-tail distributed relations for completing knowledge graphs. The few-shot knowledge graph completion task confronts with the difficulties of insufficient structural evidence caused by structure sparsity and heterogeneity, and few training samples caused by the long-tail distribution property of few-shot relations. To overcome the above difficulties, we propose to adaptively fuse entity structure information with rich textual content features, and leverage a generative approach to augment entity samples to enrich the training samples for few-shot relations. We seamlessly integrate the adaptive feature fusion and generative sample augmentation components with the few-shot learning task into an end-to-end framework, with the feature fusion and sample augmentation able to be adjusted for the few-shot learning objective through error backpropagation. We conduct few-shot knowledge graph completion experiments on two real-world knowledge graphs, showing the significant advantage of the proposed algorithm over state-of-the-art baselines, and the effectiveness of the proposed feature fusion and sample augmentation components.

Tasks

Reproductions