NSP-NER: A Prompt-based Learner for Few-shot NER Driven by Next Sentence Prediction

Abstract

Recently, prompt-based learning has achieved great success in few-shot learning. This paradigm does better integrate pre-training and downstream tasks and mine the knowledge inherent in the pre-train language model itself. Most research on prompt-learning has been conducted, but the application of prompt-based learning in NER has not been fully explored. The main obstacles of prompt-based learning's application to NER are that, each unit to be predicted is a token or span instead of the whole input text, and the existing prompt-based methods are not sensitive to the boundaries of tokens or spans. To address these issues, we propose a prompt-based learner for few-shot NER driven by Next Sentence Prediction (NSP), reformulating NER as a NSP task with span boundary information enhancement. We conduct experiments on three NER datasets in true few-shot learning setting, which supposes that only a small training set and a small validation set are available. Experimental results show that our method outperforms previous state-of-the-art models and yields promising performance even in such a challenging scenario.

Tasks

Reproductions