CL&CD: Contrastive Learning and Cluster Description for Zero-Shot Relation Extraction

Abstract

Zero-shot Relation Extraction (ZRE) is designed to identify new relations when the model is adapted to a new environment in a new domain. The majority of existing ZRE methods employ distant supervision for data labeling, which inevitably leads to incomplete annotations and noise. To handle this problem, we propose a ZRE framework based on Contrastive Learning and Cluster Description (CL&CD), a two-stage contrastive learning method is used to train labeled data which consists pseudo-labeled data and labeled data. The framework can effectively promote the model mapping the instances of the same relations to the adjacent vector space better. The module of clustering description reasonably optimizes the time of manual intervention and enormously reduces the consumption of human resources. Experimental results on Wiki-ZSL and FewRel show the superior performance of the CL&CD, which outperforms all baseline models including the state-of-the-art RCL. Furthermore, CL&CD improves F1-Score by up to 8% in both Wiki-ZSL and FewRel with 15 unseen relations.