Less is More: Understanding Word-level Textual Adversarial Attack via n-gram Frequency Descend