Multi-transformer based on prototypical enhancement network for few-shot relation classification with domain adaptation

Abstract

Few-shot relation classification aims to predict the relation between entity pairs in unseen sentences with a few labeled sentences. In real-world scenarios, the target domains often have limited data and could differ significantly from the source domains. Existing studies focus on the representation of entity pairs, ignoring the relation between the sentences and the entity pairs. Some researchers use feature-adaptive methods that do not consider that features contribute differently to relation classification. To solve these challenges, a multi-transformer based on a prototypical enhancement network (MTPEnet) is presented for few-shot relation classification with domain adaptation in this study. Specifically, for MTPEnet, two main parts are conducted. First, in the calculation of the prototype, a local and global representations (LGR) module is established to improve the representation of sentences. LGR considers not only the head and tail entities in a sentence but also the relation between the entity pair and the entire sentence. Second, a multi-transformer (MT) feature transformation module is introduced to address the cross-domain problem. Different weights are assigned to the target domain features according to the degree of fusion between the features in the target and source domains. Finally, the experimental results on the datasets FewRel 2.0 and FewTAC demonstrate the superiority of MTPEnet over several baseline models.