Current evaluation metrics for image description may be too coarse. We therefore propose a
series of binary forced-choice tasks that each focus on a different aspect of the captions. We …