In this article, we will introduce a method that use word or sentence similarity matrix for classification.
In NLP, we do not mask any input embeddings for Bert in text classifcation task. However, in paper: Spelling Error Correction with Soft-Masked BERT proposed a masked method.