Adversarial Attacks on Parts of Speech: An Empirical Study in Text-to-Image Generation