Incorporating Prior Type Information for Few-Shot Knowledge Graph Completion

Abstract

Few-shot knowledge graph completion aims to infer unknown triple facts with only a small number of reference triples. Existing methods have shown a strong capability on this problem by combining knowledge representation learning and meta learning. They ignore prior knowledge in the few-shot scenario, while prior knowledge can boost useful information to handle the challenges brought by limited referenced instances. To address the above issue, we propose a few-shot knowledge graph completion model PiTI-Fs, with entity type information as prior knowledge in a two-module learning framework. In the prior knowledge learning module, we propose to extract a metagraph for capturing prior type information by entity clustering where entities in the same cluster are considered to have the same attribute. We pre-train the metagraph to learn the prior knowledge features and fuse them into the embeddings of entities. In the meta learning module, we introduce a transformer-based relation learner to model the interactions within reference entity pairs and implement an optimization-based meta learning paradigm to train our model. Our method outperforms most of baseline models for the few-shot knowledge graph completion task. The experimental results demonstrate the effectiveness of the proposed modules.