You need to hit that thumbs down with the explanation so the model is trained with the penalty applied. Otherwise your explanations are not in the training corpus