It lets BERT use both preceding and following context to refine each word, matching its pretraining and improving fluency.