A Relation Representation Method Based on Global Relation Graph for Few-Shot Relation Extraction

Few-Shot relation extraction (FSRE) aims at recognizing novel relations by training with merely a handful of labelled examples in each relation. And the mainstream approach for FSRE currently is to use Prototype Network. Recent applications of prototype networks in FSRE have primarily focused on introducing external relation information for assisting relation classification. However, these approaches may introduce unnecessary noise and ignore the relationships information between different relations. In this paper, we propose a novel relation representation approach based on global relation graph for FSRE. Specifically, the relation representation for each relation class is obtained by concatenating label information and global relation graph information, and then directly added to the original prototype during training and prediction. Then, we evaluate our relation representation approach on the dataset FewRel1.0 and NYT-25 which are most used few-shot relation extraction. The Experimental results indicate that introducing a global relation graph to construct relation representations enables better consideration of the relationships information between different relations and reduces noise.