AUC Maximization for Low-Resource Named Entity Recognition

Current work in named entity recognition (NER) uses either cross entropy (CE) or conditional random fields (CRF) as the objective/loss functions to optimize underlying NER model. Both of these traditional objective for problem generally produce adequate performance when data distribution is balanced and there are sufficient annotated training examples. But since inherently an imbalanced tagging problem, model under low-resource settings could suffer using standard functions. Based on recent advances area ROC curve (AUC) maximization, we propose by maximizing AUC score. We give evidence that simply combining two binary-classifiers maximize score, significant improvement over loss achieved settings. also conduct extensive experiments demonstrate advantages our method highly-imbalanced To best knowledge, this first brings maximization setting. Furthermore, show agnostic different types embeddings, models domains. The code available at https://github.com/dngu0061/NER-AUC-2T.