pre-trained language models (PLMs) like BERT are being used for almost all
language-related tasks, but interpreting their behavior still remains a
significant challenge and many important questions remain largely unanswered.
In this work, we re-purpose a sentence editing dataset, where