Learning to Imagine: Visually-Augmented Natural Language Generation