Graph-based Bayesian Meta Relation Extraction

Abstract

Meta-learning methods accomplish rapid adaptation to a new task using few samples by first learning an internal representation that matches with similar tasks. In this paper, we focus on few-shot relation extraction. Previous works in few-shot relation extraction aim at predicting the relation for a pair of entities in a sentence by training with a few labeled examples in each relation. However, these algorithms implicitly require that the meta-training tasks be mutually-exclusive, such that no single model can solve all of the tasks at once, which hampers the generalization ability of these methods. To more effectively generalize to new relations, in this paper we address this challenge by designing a meta-regularization objective. We propose a novel Bayesian meta-learning approach to effectively learn the prototype vectors of relations via regularization on weights, and a graph neural network (GNN) is used to parameterize the initial prior of the prototype vectors on the global relation graph. Our approach substantially outperforms standard algorithms, and experiments demonstrate the effectiveness of our proposed approach.