Adversarial Learning with Domain-Adaptive Pretraining for Few-Shot Relation Classification across Domains

Abstract

The existing methods for domain-adaptive few-shot relation classification based on word embeddings or pretraining models trained on massive corpora, are not strong enough to cover the wide disparity of text and relation definitions to the specific target domain, leading to the inferior performance. To fill in this gap, here we propose an enhanced adversarial approach utilizing domain-adaptive pretraining model to obtain semantic features of relations, which continues unsupervised pretraining on corpus in target domain. We also construct a classification enhancer module to emphasize the class differentiation by making greater use of the supporting and query data, which not only helps to deal with few-shot problem, but also diminishes the negative effect of domain alignment caused by adversarial learning. Experimental results on FewRel2.0-DA dataset demonstrate that our proposed method achieves strong performance, which can improve the best reported result by up to 5.3 % on average accuracy for few-shot relation classification across domains.