Detecting Label Errors using Pre-Trained Language Models

We show that large pre-trained language models are extremely capable of identifying label errors in datasets: simply verifying data points in descending order of out-of-distribution loss significantly outperforms more complex mechanisms for detecting label errors on natural language datasets. We contribute a novel method to produce highly realistic, human-originated label noise from crowdsourced data, and demonstrate the effectiveness of this method on TweetNLP, providing an otherwise difficult to obtain measure of realistic recall.