Data augmentation via context similarity: An application to biomedical Named Entity Recognition

In this paper, we present COntext SImilarity-based data augmentation for NER (COSINER), a new method for improving Named Entity Recognition (NER) tasks using data augmentation. Unlike current techniques, which may generate noisy and mislabeled samples through text manipulation, COSINER uses context similarity to replace entity mentions with more plausible ones on the basis of available training data and considering the context in which entities typically appear. Through experiments on popular benchmark datasets, we show that COSINER outperforms existing baselines in various few-shot scenarios where training data is limited. Additionally, our method's computing times are comparable to the simplest augmentation methods and are better than approaches that rely on pre-trained models in their architecture.