Descriptive Image Captioning with Salient Retrieval Priors