Effects of Human Adversarial and Affable Samples on BERT Generalization