Dealing with textual noise for robust and effective BERT re-ranking