Logit Adjustment with Normalization and Augmentation in Few-Shot Named Entity Recognition

We study the problem of few-shot learning in Name Entity Recognition(FS-NER). Specifically, unlike other sequence labeling-based models, that mainly focus on better representations, we leverage logit adjustment technology to alleviate the problem that the different distribution between training and test dataset. Furthermore, we propose a simple but effective method, called Logit Adjustment with Normalization and Augmentation (LANA), for FS-NER. In detail, LANA first combines moving average and logit adjustment to retain the information of pre-training to overcome the representation drop problem in FS-NER. We also involve logit normalization to deal with the overfitting problem in FS-NER, and further improve the generalization ability of LANA. Our method achieves competitive performance on seven widely used FS-NER datasets and significantly reduces the influence of overfitting and representation drop.