MetaR: Few-Shot Named Entity Recognition With Meta-Learning and Relation Network

While conventional named entity recognition (NER) has achieved performance close to human ability, few-shot NER research has seen little use. Few-shot NER aims to recognize novel classes in a sentence using only a few labeled samples. Previous related work has mostly implemented entity classification based on methods such as Matching Networks, k-Nearest Neighbors (KNN), and Prototypical Networks; however, these methods cannot discover and represent nonlinear relationships between samples and classes. In response to the above problems, this paper proposes a new few-shot NER meta-learning framework, MetaR, which contains two sub-models, an entity location recognizer, and an entity type classifier. Specifically, we first train the entity location recognizer and then use it to recognize the entity's location, thus accomplishing both entity localization and non-entity filtration. Then, we propose a MAML-Relation network as the entity type classifier to categorize the entity location pieces extracted in the previous stage. By designing a new Relation Network for entity classification and applying MAML for parameter updating, the model can find a good set of initialization parameters, quickly adapt to novel classes, and fully exploit the nonlinear relationship between samples and classes to achieve better classification performance. Extended experiments on several benchmark datasets demonstrate that our few-shot NER framework achieves state-of-the-art performance compared to other methods.