Detecting Label Errors by using Pre-Trained Language Models

Derek Chong, Jenny Hong, Christopher D. Manning

We show that large pre-trained language models are inherently highly capable of identifying label errors in natural language datasets: simply examining out-of-sample data points in…