Use Perturbations when Learning from Explanations