PEMRC : a positive enhanced machine reading comprehension method for few-shot named entity recognition in biomedical domain

Abstract

In this paper, we propose a simple and effective few-shot named entity recognition (NER) method for biomedical domain, called PEMRC (Positive Enhanced Machine Reading Comprehension). PEMRC is based on the idea of using machine reading comprehension reading comprehension (MRC) framework to perfome few-shot NER and fully exploit the prior knowledge implied in the label information. On one hand, we design three different query templates to better induce knowledge from pre-trained language models (PLMs). On the other hand, we design a positive enhanced loss function to improve the model’s accuracy in identifying the start and end positions of entities under low-resources scenarios. Extensive experimental results on eight benchmark datasets in biomedical domain show that PEMRC significantly improves the performance of few-shot NER.