Few-Shot Named Entity Recognition with Biaffine Span Representation

Anonymous

Abstract: While Named Entity Recognition (NER) is a widely studied task, making inferences of entities with only a few labeled data (i.e., few-shot NER) has been challenging. Correspondingly, the N-way K-shot NER task is proposed to recognize entities in the given N categories with only K labeled samples for each category. Existing methods treat this task as a sequence labeling problem, while this paper regards it as an entity span classification problem and designs a Biaffine Span Representation (BSR) method to learn contextual span dependency representation to fit into the classification algorithm. The BSR applies a biaffine pooling module to establish the dependencies of each word on the whole sentence and to reduce the dimension of word features, thus, the span representation could gain contextual dependency information to help improve recognition accuracies. Experimental study on four standard NER datasets shows that our proposed BSR method outperforms pre-trained language models and existing N-way K-shot NER algorithms in two types of adaptations (i.e., Intra-Domain Cross-Type Adaptation and Cross-Domain Cross-Type Adaptation). Notably, F_1 value has increased by an average of 13.77% and 18.30% on the 5-way 1-shot task and the 5-way 5-shot task, respectively.