Language Models for Image Captioning: The Quirks and What Works