Modeling Zero-Shot Relation Classification as a Multiple-Choice Problem

Zero-shot relation classification (ZeroRC) aims to infer the semantic relations between entity pairs in sentences, while the relation sets at the training and testing stages are disjoint. It is a crucial task in information extraction and is much more challenging than traditional relation classification. We propose a novel method named MC-BERT to model ZeroRC as a Multiple-Choice problem employing BERT as the backbone model. A semantic template is designed to infuse the information of entities and context. It serves as the question stem of the multiple-choice, followed by a relation label which serves as a choice. Moreover, we propose a grouping strategy to improve training efficiency. We perform comprehensive experiments on two datasets, Wiki-ZSL and FewRel. The results show that our proposed method significantly outperforms previous works. Specifically, it achieves performance gains ranging from 1.17% to 6.87% in the F1 score with only 40% of the model parameters against to state-of-the-art method, demonstrating the simplicity and effectiveness of our proposed method.