Prompt-Based Self-training Framework for Few-Shot Named Entity Recognition

Exploiting unlabeled data is one of the plausible methods to improve few-shot named entity recognition (few-shot NER), where only a small number of labeled examples are given for each entity type. Existing works focus on learning deep NER models with self-training for few-shot NER. Self-training may induce incomplete and noisy labels which do not necessarily improve or even deteriorate the model performance. To address this challenge, we propose a prompt-based self-training framework. In the first stage, we introduce a self-training approach with prompt tuning to improve the model performance. Specially, we explore several label selection strategies in self-training to mitigate error propagation from noisy pseudo-labels. In the second stage, we fine-tune the BERT model over the high confidence pseudo-labels and original labels. We conduct experiments on two benchmark datasets. The results show that our method outperforms existing few-shot NER models by significant margins, demonstrating its effectiveness for the few-shot setting.