Large scale image annotation: learning to rank with joint word-image embeddings