BERT for lexical substitution

Question

I\'ve been trying to let the BERT generate lexical substitution candidates that are semantically close to the target word. One option that I was using is to concatenate the