The Suitability of Generative Adversarial Training for BERT Natural Language Generation