Exploring Automatically Perturbed Natural Language Explanations in Relation Extraction