A Few-shot Learning Method Based on Bidirectional Encoder Representation from Transformers for Relation Extraction

Relation extraction is one of the fundamental subtasks of the information extraction. The purpose is to determine the implicit relation between two entities in a sentence. Therefore, Convolutional Neural Networks and Feature Attention-based Prototypical Networks (CNN-Proto-FATT), a typical few-shot learning method, is proposed and achieve competitive performance. However, convolutional neural networks suffer from the insufficient instances of relation in real scenes, leading to undesirable results. To extract long-distance features more comprehensively, the pre-trained model Bidirectional Encoder Representation from Transformers (BERT) is incorporated into CNN-Proto-FATT. In this model, named Bidirectional Encoder Representation from Transformers and Feature Attention-based Prototypical Networks (BERT-Proto-FATT), the multi-head attention helps the network extract semantic features cross long- and short-distance to enhance the encoded representations. Experimental results indicate that BERT-Proto-FATT demonstrates significant improvements on the FewRel dataset.