Distantly Supervised Named Entity Recognition with Category-Oriented Confidence Calibration

Anonymous

Abstract: In this work, we study the noisy-labeled named entity recognition under distant supervision setting. Considering that most NER systems based on confidence estimation deal with noisy labels ignoring the fact that model has different levels of confidence towards different categories, we propose a category-oriented confidence calibration(Coca) strategy with an automatically confidence threshold calculation module. We integrate our method into a teacher-student framework to improve the model performance. Our proposed approach achieves promising performance among advanced baseline models, setting new state-of-the-art performance on three existing distantly supervised NER benchmarks.