Denoising Distantly Supervised Named Entity Recognition via a Hypergeometric Probabilistic Model

Denoising is the essential step for distant supervision based named entity recognition. Previous denoising methods are mostly on instance-level confidence statistics, which ignore variety of underlying noise distribution different datasets and types. This makes them difficult to be adapted high rate settings. In this paper, we propose Hypergeometric Learning (HGL), a algorithm distantly supervised NER that takes both into consideration. Specifically, during neural network training, naturally model samples in each batch following hypergeometric parameterized by noise-rate. Then instance regarded as either correct or noisy one according its label derived from previous training step, well sampled batch. Experiments show HGL can effectively denoise weakly-labeled data retrieved supervision, therefore results significant improvements trained models.